test: verify mixed-scope doc priority survives reload and reopen

Co-authored-by: Mervin Praison <[email protected]>
test: document priority regression contracts
2026-11-07 17:33:33 -10:00 · 2026-10-07 09:52:42 +00:00 · 2026-10-03 22:48:43 +08:00 · 2026-10-03 22:17:51 +08:00 · 2026-10-03 20:44:05 +08:00 · 2026-10-03 08:40:19 +01:00
587 changed files with 54451 additions and 5987 deletions
@@ -0,0 +1,36 @@
+name: Pick Claude OAuth token
+description: Probe primary/secondary Claude OAuth tokens and select a working one (dual OAuth Option B).
+inputs:
+  token_primary:
+    description: Primary Claude OAuth token (CLAUDE_CODE_OAUTH_TOKEN)
+    required: false
+  token_secondary:
+    description: Secondary Claude OAuth token (CLAUDE_CODE_OAUTH_TOKEN_B)
+    required: false
+  active_slot:
+    description: Preferred slot — primary or secondary (repo variable CLAUDE_OAUTH_ACTIVE)
+    required: false
+    default: primary
+outputs:
+  token:
+    description: Selected OAuth token
+    value: ${{ steps.pick.outputs.token }}
+  slot:
+    description: Slot used (primary, secondary, or *-unprobed fallback)
+    value: ${{ steps.pick.outputs.slot }}
+runs:
+  using: composite
+  steps:
+    - id: pick
+      shell: bash
+      env:
+        CLAUDE_CODE_OAUTH_TOKEN: ${{ inputs.token_primary }}
+        CLAUDE_CODE_OAUTH_TOKEN_B: ${{ inputs.token_secondary }}
+        CLAUDE_OAUTH_ACTIVE: ${{ inputs.active_slot }}
+      run: |
+        if [[ -n "${HUB_SCRIPTS:-}" && -f "${HUB_SCRIPTS}/oauth-pick-token.sh" ]]; then
+          SCRIPT="${HUB_SCRIPTS}/oauth-pick-token.sh"
+        else
+          SCRIPT="${GITHUB_ACTION_PATH}/../../scripts/oauth-pick-token.sh"
+        fi
+        bash "$SCRIPT"
@@ -16,6 +16,9 @@ const PR_MAX_AUTO_ADDITIONS = 800;
 const PR_MAX_AUTO_FILES = 30;
 const MANUAL_ONLY_LABELS = new Set(['security', 'breaking-change', 'needs-manual-review', 'release']);
 const WORKFLOW_ONLY_LABEL = 'merge-gate-ci-only';
+const EXTERNAL_EXAMPLE_PATH_PREFIX = 'examples/tools/external/';
+const EXTERNAL_EXAMPLE_LABEL = 'pipeline/external-example';
+const MAINTAINER_ACCEPT_EXAMPLE_LABEL = 'maintainer-accept-example';
 const CI_ONLY_PATH_PREFIXES = ['.github/workflows/', '.github/actions/', '.github/scripts/merge-gate'];
 const SDK_PATH_PREFIXES = ['src/praisonai-agents/', 'src/praisonai/', 'src/praisonai-ts/'];
 const TS_SDK_SOURCE_PREFIX = 'src/praisonai-ts/src/';
@@ -328,7 +331,7 @@ function finalClaudeCompletedOnSha(comments, headPushedAt) {
 }

 const FINAL_CLAUDE_REVIEW_BODY =
-  '@claude You are the FINAL architecture reviewer. If the branch is under MervinPraison/PraisonAI (not a fork), you are able to make modifications to this branch and push directly. SCOPE: Review changes in this PR. Python SDK: praisonaiagents, praisonai. TypeScript SDK: src/praisonai-ts/. Do NOT modify src/praisonai-rust. Read ALL comments above from Gemini, Qodo, CodeRabbit, and Copilot carefully before responding.\n\n**MANDATORY READ (before reviewing):**\n- Always read src/praisonai-agents/AGENTS.md\n- If this PR touches src/praisonai-ts/, also read src/praisonai-ts/AGENTS.md §2.1.2 (TS triage + PR review checklist)\n\n**Phase 1: Review per AGENTS.md**\n1. Protocol-driven: check heavy implementations vs core SDK\n2. Backward compatible: ensure zero feature regressions\n3. Performance: no hot-path regressions\n4. SDK value: review in depth whether the change genuinely adds value to the SDK — never add features for the sake of adding them. It must strengthen the SDK (simpler, more user-friendly, robust, world-class, secure). If it does not clearly add value, request changes or recommend rejecting/closing rather than merging scope creep\n5. Do not bloat the Agent class with additional params — only if absolutely required; we already support many params.\n6. Repo routing: agent-callable tools → PraisonAI-Tools; lifecycle plugins → PraisonAI-Plugins; optional sandbox backends → PraisonAI-Plugins (`praisonai.sandbox` entry point) — request changes if wrongly added to praisonaiagents/\n\n**MANDATORY COMMENT FORMAT — include this Phase 1 table in your review comment:**\n#### Phase 1 — AGENTS.md review\n| Check | Result |\n|---|---|\n| Protocol-driven / no heavy impl in core | ✅ or ❌ + one-line rationale |\n| Backward compatible | ✅ or ❌ + one-line rationale |\n| Performance (hot path) | ✅ or ❌ + one-line rationale |\n| **SDK value** | ✅ or ❌ + one-line rationale (explicitly judge whether the change strengthens the SDK) |\n| No Agent param bloat | ✅ or ❌ + one-line rationale |\n| Repo routing | ✅ or ❌ + one-line rationale |\n\nFor TypeScript PRs (src/praisonai-ts/), also add:\n| TS types / parity / tests | ✅ or ❌ + one-line rationale (npm run build && npm test) |\n\n**Phase 2: FIX Valid Issues**\n7. For any VALID bugs or architectural flaws found by Gemini, CodeRabbit, Qodo, Copilot, or any other reviewer: implement the fix\n8. Also independently identify and fix any gaps or issues you find in the changed code — do not rely only on prior reviewer feedback\n9. Push all code fixes directly to THIS branch (do NOT create a new PR)\n10. Comment a summary of exact files modified and what you skipped\n\n**Phase 3: Final Verdict**\n11. If all issues are resolved, approve the PR / close the Issue\n12. If blocking issues remain, request changes / leave clear action items';
+  '@claude You are the FINAL architecture reviewer. If the branch is under MervinPraison/PraisonAI (not a fork), you are able to make modifications to this branch and push directly. SCOPE: Review changes in this PR. Python SDK: praisonaiagents, praisonai. TypeScript SDK: src/praisonai-ts/. Do NOT modify src/praisonai-rust. Read ALL comments above from Gemini, Qodo, CodeRabbit, and Copilot carefully before responding.\n\n**MANDATORY READ (before reviewing):**\n- Always read src/praisonai-agents/AGENTS.md\n- If this PR touches src/praisonai-ts/, also read src/praisonai-ts/AGENTS.md §2.1.2 (TS triage + PR review checklist)\n\n**Phase 1: Review per AGENTS.md**\n1. Protocol-driven: check heavy implementations vs core SDK\n2. Backward compatible: ensure zero feature regressions\n3. Performance: no hot-path regressions\n4. SDK value: review in depth whether the change genuinely adds value to the SDK — never add features for the sake of adding them. It must strengthen the SDK (simpler, more user-friendly, robust, world-class, secure). If it does not clearly add value, request changes or recommend rejecting/closing rather than merging scope creep\n5. Do not bloat the Agent class with additional params — only if absolutely required; we already support many params.\n6. Repo routing: agent-callable tools → PraisonAI-Tools; lifecycle plugins → PraisonAI-Plugins; optional sandbox backends → PraisonAI-Plugins (`praisonai.sandbox` entry point) — request changes if wrongly added to praisonaiagents/\n6b. **External tool examples (`examples/tools/external/**`):** If the diff is confined to that path, mark **SDK value** ❌ for monorepo merge — redirect the contributor to PraisonAI-Tools; do **not** Approve for merge unless the PR has the `maintainer-accept-example` label.\n\n**MANDATORY COMMENT FORMAT — include this Phase 1 table in your review comment:**\n#### Phase 1 — AGENTS.md review\n| Check | Result |\n|---|---|\n| Protocol-driven / no heavy impl in core | ✅ or ❌ + one-line rationale |\n| Backward compatible | ✅ or ❌ + one-line rationale |\n| Performance (hot path) | ✅ or ❌ + one-line rationale |\n| **SDK value** | ✅ or ❌ + one-line rationale (explicitly judge whether the change strengthens the SDK) |\n| No Agent param bloat | ✅ or ❌ + one-line rationale |\n| Repo routing | ✅ or ❌ + one-line rationale |\n\nFor TypeScript PRs (src/praisonai-ts/), also add:\n| TS types / parity / tests | ✅ or ❌ + one-line rationale (npm run build && npm test) |\n\n**Phase 2: FIX Valid Issues**\n7. For any VALID bugs or architectural flaws found by Gemini, CodeRabbit, Qodo, Copilot, or any other reviewer: implement the fix\n8. Also independently identify and fix any gaps or issues you find in the changed code — do not rely only on prior reviewer feedback\n9. Push all code fixes directly to THIS branch (do NOT create a new PR)\n10. Comment a summary of exact files modified and what you skipped\n\n**Phase 3: Final Verdict**\n11. If all issues are resolved, approve the PR / close the Issue\n12. If blocking issues remain, request changes / leave clear action items';

 async function getMergeState(github, owner, repo, prNumber) {
  const query = `
@@ -583,6 +586,11 @@ function isCiOnlyChange(files) {
  return files.every((f) => CI_ONLY_PATH_PREFIXES.some((p) => f.filename.startsWith(p)));
 }

+function isExternalExampleOnlyChange(files) {
+  if (!files.length) return false;
+  return files.every((f) => f.filename.startsWith(EXTERNAL_EXAMPLE_PATH_PREFIX));
+}
+
 function isInternalPullRequestLink(link, owner, repo) {
  if (link?.base?.repo?.full_name === `${owner}/${repo}`) return true;
  const baseUrl = link?.base?.repo?.url || '';
@@ -973,6 +981,15 @@ async function evaluatePipelineQuiescent(github, owner, repo, prNumber, core, op

  const pullFiles = await listPullFiles(github, owner, repo, prNumber);

+  if (
+    isExternalExampleOnlyChange(pullFiles) &&
+    !ctx.labels.includes(MAINTAINER_ACCEPT_EXAMPLE_LABEL)
+  ) {
+    reasons.push(
+      'external example only — belongs in PraisonAI-Tools (add maintainer-accept-example to opt in)'
+    );
+  }
+
  const agentChange = getAgentPyChangeFromFiles(pullFiles);
  const agentManual = manualReviewReasonForAgentPy(agentChange);
  if (agentManual) reasons.push(agentManual);
@@ -1019,6 +1036,13 @@ async function listPrNumbersForMergeGateScan(github, owner, repo, core) {
  }
  const mergeReady = mergeReadyIssues
    .filter((issue) => issue.pull_request)
+    .filter((issue) => {
+      const labels = (issue.labels || []).map((l) => (typeof l === 'string' ? l : l.name));
+      if (labels.includes(EXTERNAL_EXAMPLE_LABEL) && !labels.includes(MAINTAINER_ACCEPT_EXAMPLE_LABEL)) {
+        return false;
+      }
+      return true;
+    })
    .map((issue) => issue.number)
    .sort((a, b) => a - b);

@@ -1146,6 +1170,10 @@ module.exports = {
  claudeRunBlocksPr,
  hasBlockingClaudeRunForPr,
  isCiOnlyChange,
+  isExternalExampleOnlyChange,
+  EXTERNAL_EXAMPLE_PATH_PREFIX,
+  EXTERNAL_EXAMPLE_LABEL,
+  MAINTAINER_ACCEPT_EXAMPLE_LABEL,
  isInternalPullRequestLink,
  resolvePrNumberFromLinkedPullRequests,
  resolvePrNumberFromHeadBranch,
@@ -0,0 +1,101 @@
+#!/usr/bin/env bash
+# Pick a working Claude Code OAuth token (primary/secondary with failover).
+# Env: CLAUDE_CODE_OAUTH_TOKEN, CLAUDE_CODE_OAUTH_TOKEN_B, CLAUDE_OAUTH_ACTIVE
+# Out: GITHUB_OUTPUT token, slot
+set -euo pipefail
+
+PRIMARY="${CLAUDE_CODE_OAUTH_TOKEN:-}"
+SECONDARY="${CLAUDE_CODE_OAUTH_TOKEN_B:-}"
+ACTIVE="${CLAUDE_OAUTH_ACTIVE:-primary}"
+
+ACTIVE_LC=$(echo "$ACTIVE" | tr '[:upper:]' '[:lower:]')
+case "$ACTIVE_LC" in
+  secondary|b|2) PREFERRED=secondary ;;
+  *) PREFERRED=primary ;;
+esac
+
+probe_token() {
+  local token=$1
+  [[ -z "$token" ]] && return 1
+
+  local body='{"model":"claude-3-5-haiku-latest","max_tokens":1,"messages":[{"role":"user","content":"."}]}'
+  local tmp http_code body_resp
+
+  tmp=$(mktemp)
+  http_code=$(curl -sS -o "$tmp" -w "%{http_code}" \
+    -X POST "https://api.anthropic.com/v1/messages" \
+    -H "Authorization: Bearer ${token}" \
+    -H "anthropic-version: 2023-06-01" \
+    -H "content-type: application/json" \
+    -d "$body" 2>/dev/null || echo "000")
+
+  body_resp=$(cat "$tmp")
+  rm -f "$tmp"
+
+  if [[ "$http_code" == "200" ]] || [[ "$http_code" == "400" ]]; then
+    return 0
+  fi
+
+  if [[ "$http_code" == "401" ]] || [[ "$http_code" == "429" ]]; then
+    echo "::warning::OAuth probe failed (HTTP ${http_code})"
+    return 1
+  fi
+
+  if echo "$body_resp" | grep -qiE 'revoked|invalid.*token|authentication|weekly limit|rate_limit|overloaded|billing'; then
+    echo "::warning::OAuth probe rejected (HTTP ${http_code})"
+    return 1
+  fi
+
+  return 0
+}
+
+try_token() {
+  local slot=$1 token=$2
+  [[ -z "$token" ]] && return 1
+  if probe_token "$token"; then
+    SELECTED_SLOT=$slot
+    SELECTED_TOKEN=$token
+    return 0
+  fi
+  return 1
+}
+
+SELECTED_SLOT=""
+SELECTED_TOKEN=""
+
+if [[ "$PREFERRED" == "primary" ]]; then
+  try_token primary "$PRIMARY" || try_token secondary "$SECONDARY" || true
+else
+  try_token secondary "$SECONDARY" || try_token primary "$PRIMARY" || true
+fi
+
+if [[ -z "$SELECTED_TOKEN" ]]; then
+  if [[ "$PREFERRED" == "primary" && -n "$PRIMARY" ]]; then
+    SELECTED_TOKEN="$PRIMARY"
+    SELECTED_SLOT="primary-unprobed"
+  elif [[ -n "$SECONDARY" ]]; then
+    SELECTED_TOKEN="$SECONDARY"
+    SELECTED_SLOT="secondary-unprobed"
+  elif [[ -n "$PRIMARY" ]]; then
+    SELECTED_TOKEN="$PRIMARY"
+    SELECTED_SLOT="primary-unprobed"
+  else
+    echo "::error::No Claude OAuth token configured"
+    exit 1
+  fi
+  echo "::warning::OAuth probe failed for all tokens; using ${SELECTED_SLOT}"
+fi
+
+echo "::notice::Claude OAuth slot: ${SELECTED_SLOT}"
+
+if [[ -z "${GITHUB_OUTPUT:-}" ]]; then
+  echo "slot=${SELECTED_SLOT}"
+  exit 0
+fi
+
+{
+  echo "token<<EOF"
+  echo "$SELECTED_TOKEN"
+  echo "EOF"
+  echo "slot=${SELECTED_SLOT}"
+} >> "$GITHUB_OUTPUT"
@@ -73,4 +73,33 @@ assert(
  )
 );

+assert(
+  'external example only path',
+  mg.isExternalExampleOnlyChange([{ filename: 'examples/tools/external/arxiv/tool.py' }])
+);
+assert(
+  'mixed paths not external-example only',
+  !mg.isExternalExampleOnlyChange([
+    { filename: 'examples/tools/external/foo.py' },
+    { filename: 'src/praisonai-agents/foo.py' },
+  ])
+);
+
+assert(
+  'upstream head repo',
+  ps.isUpstreamHeadRepo(
+    { head: { repo: { full_name: 'MervinPraison/PraisonAI' } } },
+    'MervinPraison',
+    'PraisonAI'
+  )
+);
+assert(
+  'fork head repo still synced (not upstream)',
+  !ps.isUpstreamHeadRepo(
+    { head: { repo: { full_name: 'ai-mrscraper/PraisonAI' } } },
+    'MervinPraison',
+    'PraisonAI'
+  )
+);
+
 process.exit(failed ? 1 : 0);
@@ -37,8 +37,17 @@ const LABEL_SPECS = [
  { name: 'pipeline/blocked:cooldown', color: 'fef2c0', description: 'Blocked: post-push or @claude cooldown' },
  { name: 'pipeline/blocked:stale-final', color: 'f9d0c4', description: 'Blocked: FINAL stale after new commits' },
  { name: 'pipeline/blocked:no-final', color: 'ededed', description: 'Blocked: no FINAL @claude trigger yet' },
+  {
+    name: 'pipeline/external-example',
+    color: 'bfd4f2',
+    description: 'Diff only under examples/tools/external/ — not merge-ready without maintainer-accept-example',
+  },
 ];

+// PRs with pipeline/external-example cannot reach pipeline/merge-ready unless
+// maintainers add maintainer-accept-example (enforced in merge-gate evaluatePipelineQuiescent,
+// same fail-closed pattern as fork PRs).
+
 function deriveStage(comments, evalResult) {
  if (evalResult.ready) return 'pipeline/merge-ready';
  if (!chain.chainKickPosted(comments)) return 'pipeline/reviews-pending';
@@ -63,7 +72,8 @@ function reasonToBlockerLabel(reason) {
    r.includes('without test') ||
    r.includes('files changed') ||
    r.includes('possible secret') ||
-    r.includes('agent.py')
+    r.includes('agent.py') ||
+    r.includes('external example only')
  ) return 'pipeline/blocked:manual-review';
  if (r.includes('recent @claude') || r.includes('post-push buffer')) return 'pipeline/blocked:cooldown';
  if (r.includes('stale final')) return 'pipeline/blocked:stale-final';
@@ -133,11 +143,16 @@ async function syncPipelineLabels(github, owner, repo, prNumber, core) {
    github, owner, repo, prNumber, core
  );
  const { stage, blockers, all } = computePipelineLabels(ctx.comments, evalResult);
+  const pullFiles = await mergeGate.listPullFiles(github, owner, repo, prNumber);
+  const auxiliary = mergeGate.isExternalExampleOnlyChange(pullFiles)
+    ? [mergeGate.EXTERNAL_EXAMPLE_LABEL]
+    : [];
+  const labelSet = [...all, ...auxiliary];

  const current = ctx.labels.filter((l) => l.startsWith(PIPELINE_PREFIX));
-  const desired = new Set(all);
+  const desired = new Set(labelSet);
  const toRemove = current.filter((l) => !desired.has(l) && l !== 'pipeline/merged');
-  const toAdd = all.filter((l) => !current.includes(l));
+  const toAdd = labelSet.filter((l) => !current.includes(l));

  for (const name of toRemove) {
    try {
@@ -199,6 +214,11 @@ async function dispatchMergeGateForOldestReady(github, owner, repo, readyCandida
  return 0;
 }

+function isUpstreamHeadRepo(pr, owner, repo) {
+  const headFull = pr.head?.repo?.full_name;
+  return !headFull || headFull === `${owner}/${repo}`;
+}
+
 async function syncOpenPullRequests(github, owner, repo, options, core) {
  const { maxPrs = 20, dispatchMergeGate = true } = options || {};
  await ensurePipelineLabels(github, owner, repo, core);
@@ -223,9 +243,8 @@ async function syncOpenPullRequests(github, owner, repo, options, core) {
  for (const pr of prs) {
    if (synced >= maxPrs) break;
    if (pr.draft) continue;
-    if (pr.head?.repo?.full_name && pr.head.repo.full_name !== `${owner}/${repo}`) continue;
    const result = await syncPipelineLabels(github, owner, repo, pr.number, core);
-    if (result.ready) {
+    if (result.ready && isUpstreamHeadRepo(pr, owner, repo)) {
      readyCandidates.push({
        prNumber: pr.number,
        createdAt: result.createdAt || new Date(pr.created_at).getTime(),
@@ -248,6 +267,7 @@ module.exports = {
  STAGE_LABELS,
  BLOCKER_LABELS,
  ALL_PIPELINE_LABELS,
+  isUpstreamHeadRepo,
  deriveStage,
  deriveBlockerLabels,
  computePipelineLabels,
@@ -1,59 +0,0 @@
-# Why this exists
-#
-# praisonai-agents has ~11,900 tests. Until this gate, CI executed 15 of them:
-# one smoke step in test-optimized.yml naming two files. Everything else was
-# collected by nobody.
-#
-# The cost of that showed up as two test modules that could not even be
-# IMPORTED -- they referenced symbols moved or replaced months earlier, so 19
-# tests silently contributed nothing while still appearing to exist. A failing
-# test is loud; a test that never collects is not.
-#
-# This job does not run the suite. Running all 11,900 on every PR is a real
-# decision about Actions compute, and it would go red on arrival because there
-# are pre-existing failures nothing has been enforcing. A job that is red the
-# day it lands teaches people to ignore it.
-#
-# Collection alone takes ~40s, costs almost nothing, and would have caught both
-# dead modules the day they broke. It is the honest ratchet: it cannot pass
-# while a test file is unimportable.
-name: agents-collect-gate
-
-on:
-  push:
-    branches: [ main ]
-    paths:
-      - 'src/praisonai-agents/**'
-      - '.github/workflows/agents-collect-gate.yml'
-  pull_request:
-    paths:
-      - 'src/praisonai-agents/**'
-      - '.github/workflows/agents-collect-gate.yml'
-
-jobs:
-  collect:
-    name: every praisonai-agents test file imports
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-      - name: Install
-        run: |
-          set -euo pipefail
-          python -m pip install --upgrade pip
-          pip install -e src/praisonai-agents
-          pip install pytest pytest-asyncio pytest-timeout
-      - name: Collect the whole suite
-        working-directory: src/praisonai-agents
-        env:
-          # A key must be present or some modules bail at import; it is never
-          # used, because collection does not execute test bodies.
-          OPENAI_API_KEY: 'sk-collect-only-not-a-real-key'
-        run: |
-          set -euo pipefail
-          # No pipe here: piping pytest into tail discards its exit code, which
-          # is how a gate silently stops gating.
-          python -m pytest tests/ --collect-only -q --disable-warnings
@@ -0,0 +1,142 @@
+# Why this exists
+#
+# praisonai-agents has ~11,900 tests. Until this gate, CI executed 15 of them:
+# one smoke step in test-optimized.yml naming two files. Everything else was
+# collected by nobody.
+#
+# The cost of that showed up as two test modules that could not even be
+# IMPORTED -- they referenced symbols moved or replaced months earlier, so 19
+# tests silently contributed nothing while still appearing to exist. A failing
+# test is loud; a test that never collects is not.
+#
+# This job started as collection-only, because running the suite would have gone
+# red on arrival: 297 pre-existing failures nothing had been enforcing, and a job
+# that is red the day it lands teaches people to ignore it. Then a curated
+# green-set, ratcheted upward.
+#
+# Both are gone. The unit suite now runs in full and passes: 9373 tests, 0
+# failures. The `collect` job stays because it is the cheaper, stricter signal --
+# it cannot pass while a test file is unimportable, and it says so in 40s.
+#
+name: agents-suite
+
+on:
+  push:
+    branches: [ main ]
+    paths:
+      - 'src/praisonai-agents/**'
+      - '.github/workflows/agents-suite.yml'
+  pull_request:
+    paths:
+      - 'src/praisonai-agents/**'
+      - '.github/workflows/agents-suite.yml'
+
+jobs:
+  collect:
+    name: every praisonai-agents test file imports
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+      - name: Install
+        run: |
+          set -euo pipefail
+          python -m pip install --upgrade pip
+          pip install -e src/praisonai-agents
+          pip install pytest pytest-asyncio pytest-timeout
+      - name: Collect the whole suite
+        working-directory: src/praisonai-agents
+        env:
+          # A key must be present or some modules bail at import; it is never
+          # used, because collection does not execute test bodies.
+          OPENAI_API_KEY: 'sk-collect-only-not-a-real-key'
+        run: |
+          set -euo pipefail
+          # No pipe here: piping pytest into tail discards its exit code, which
+          # is how a gate silently stops gating.
+          python -m pytest tests/ --collect-only -q --disable-warnings
+
+  unit:
+    name: praisonai-agents unit tests
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+      - name: Install
+        run: |
+          set -euo pipefail
+          python -m pip install --upgrade pip
+          # Extras, not decoration. Tests across workflows/embedding/eval/context
+          # patch litellm and construct LLM(...), and the knowledge suite needs
+          # chonkie for chunking -- without [knowledge] 22 tests fail on
+          # "chonkie package not found", an install gap that reads as 22 code
+          # failures.
+          pip install -e "src/praisonai-agents[llm,knowledge]"
+          # praisonai-sandbox is a MONOREPO sibling, not a third-party optional.
+          # Without it 47 tests fail with "No module named
+          # praisonai_sandbox.compute" -- an install gap reported as 47 code
+          # failures. Installing it took the suite from 297 failures to 250.
+          pip install -e src/praisonai-sandbox
+          # Same story: a stale praisonai-bot in the environment made 22 gateway
+          # tests fail on a malformed pairing file. The monorepo version already
+          # handles it; the INSTALLED one did not.
+          pip install -e src/praisonai-bot
+          # The wrapper. tests/unit/test_managed_backend.py imports
+          # praisonai.integrations.managed_agents; without it 47 tests fail with
+          # "No module named 'praisonai'". Found by running the suite on 3.11 in
+          # a CI-shaped environment rather than trusting a developer machine
+          # where every sibling is already installed.
+          pip install -e src/praisonai
+          pip install pytest pytest-asyncio pytest-timeout
+      - name: Run the unit suite
+        working-directory: src/praisonai-agents
+        env:
+          # Never used: PRAISONAI_ALLOW_NETWORK=0 and the live marker keep these
+          # offline. Present because some modules bail at import without a key.
+          OPENAI_API_KEY: 'sk-not-a-real-key'
+          PRAISONAI_ALLOW_NETWORK: '0'
+        run: |
+          set -euo pipefail
+          # The whole suite, not a curated subset. tests/green-set.txt and
+          # tests/known-failures.txt are gone: 9373 pass, 0 fail.
+          #
+          # The 297 failures this job was built around were, in the end:
+          #   83  environment  -- monorepo siblings not installed, dirty HOME
+          #   19  ONE line     -- a module-level os.environ.setdefault in
+          #                       tests/unit/tools/test_delegation_tools.py that
+          #                       turned auto-approval on for the whole process
+          #                       at COLLECTION time
+          #   ~5  deprecation warnings that fire once per process
+          #  ~85  tests asserting contracts the product had deliberately changed
+          #    5  real product bugs
+          #  the rest: tests built on Mocks whose auto-created attributes
+          #            defeated the very code path under test
+          #
+          # Three traps worth knowing before adding a test here, each of which
+          # cost a full-suite bisect:
+          #   1. Module-level env mutation persists for the entire process, and
+          #      runs at collection. Use monkeypatch.
+          #   2. The approval registry is a process-global singleton. Anything
+          #      you add_requirement() stays for every later test.
+          #   3. `'litellm' in module_name` also matches praisonai's own
+          #      _litellm_loader; evicting it from sys.modules creates a second
+          #      module object with separate globals.
+          #
+          # No pipe: piping pytest into anything discards its exit code, which is
+          # how a gate silently stops gating.
+          #
+          # Nothing is excluded. test_prefetch.py used to be, recorded as "passes
+          # on macOS, fails on Linux" -- wrong variable. It was Python 3.12 vs the
+          # 3.11 pinned above: isinstance() against a runtime_checkable Protocol
+          # used hasattr on 3.11 (which any MagicMock satisfies) and
+          # inspect.getattr_static from 3.12 (which does not see Mock's
+          # auto-created attributes). The mock now names its methods, so both
+          # versions take the same branch.
+          python -m pytest tests/unit/ -q -p no:randomly --timeout=120 \
+            --disable-warnings
@@ -53,7 +53,7 @@ jobs:
      - name: Post Claude FINAL after prior reviewers
        uses: actions/github-script@v7
        with:
-          github-token: ${{ secrets.GH_TOKEN }}
+          github-token: ${{ secrets.GH_TOKEN != '' && secrets.GH_TOKEN || github.token }}
          script: |
            const path = require('path');
            const chain = require(path.join(process.env.GITHUB_WORKSPACE, '.github/scripts/pr-review-chain.js'));
@@ -97,7 +97,7 @@ jobs:
      - name: Post Claude FINAL (skip Copilot)
        uses: actions/github-script@v7
        with:
-          github-token: ${{ secrets.GH_TOKEN }}
+          github-token: ${{ secrets.GH_TOKEN != '' && secrets.GH_TOKEN || github.token }}
          script: |
            const path = require('path');
            const chain = require(path.join(process.env.GITHUB_WORKSPACE, '.github/scripts/pr-review-chain.js'));
@@ -126,7 +126,7 @@ jobs:
      - name: Re-post FINAL Claude review when stale or missing
        uses: actions/github-script@v7
        with:
-          github-token: ${{ secrets.GH_TOKEN }}
+          github-token: ${{ secrets.GH_TOKEN != '' && secrets.GH_TOKEN || github.token }}
          script: |
            const path = require('path');
            const mergeGate = require(path.join(process.env.GITHUB_WORKSPACE, '.github/scripts/merge-gate.js'));
@@ -182,7 +182,7 @@ jobs:
      - name: Sync pipeline status labels
        uses: actions/github-script@v7
        with:
-          github-token: ${{ secrets.GH_TOKEN }}
+          github-token: ${{ secrets.GH_TOKEN != '' && secrets.GH_TOKEN || github.token }}
          script: |
            const path = require('path');
            const ps = require(path.join(process.env.GITHUB_WORKSPACE, '.github/scripts/pipeline-status.js'));
@@ -203,7 +203,7 @@ jobs:
      - name: Post missing/stale FINAL @claude on open PRs
        uses: actions/github-script@v7
        with:
-          github-token: ${{ secrets.GH_TOKEN }}
+          github-token: ${{ secrets.GH_TOKEN != '' && secrets.GH_TOKEN || github.token }}
          script: |
            const path = require('path');
            const mergeGate = require(path.join(process.env.GITHUB_WORKSPACE, '.github/scripts/merge-gate.js'));
@@ -275,7 +275,7 @@ jobs:
      - name: Trigger CodeRabbit and Qodo reviews
        uses: actions/github-script@v7
        with:
-          github-token: ${{ secrets.GH_TOKEN }}
+          github-token: ${{ secrets.GH_TOKEN != '' && secrets.GH_TOKEN || github.token }}
          script: |
            const path = require('path');
            const chain = require(path.join(process.env.GITHUB_WORKSPACE, '.github/scripts/bot-pr-review-chain.js'));
@@ -285,7 +285,7 @@ jobs:
      - name: Sync pipeline status labels
        uses: actions/github-script@v7
        with:
-          github-token: ${{ secrets.GH_TOKEN }}
+          github-token: ${{ secrets.GH_TOKEN != '' && secrets.GH_TOKEN || github.token }}
          script: |
            const path = require('path');
            const ps = require(path.join(process.env.GITHUB_WORKSPACE, '.github/scripts/pipeline-status.js'));
@@ -28,8 +28,16 @@ jobs:
        with:
          token: ${{ steps.app-token.outputs.token }}

+      - name: Pick Claude OAuth token
+        id: oauth
+        uses: MervinPraison/github-automation-template/.github/actions/[email protected]
+        with:
+          token_primary: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+          token_secondary: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN_B }}
+          active_slot: ${{ vars.CLAUDE_OAUTH_ACTIVE != '' && vars.CLAUDE_OAUTH_ACTIVE || 'primary' }}
+
      - name: Run Claude Issue Triage
        uses: ./.github/actions/claude-issue-triage-action
        with:
-          claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+          claude_code_oauth_token: ${{ steps.oauth.outputs.token }}
          github_token: ${{ steps.app-token.outputs.token }} 
@@ -267,6 +267,14 @@ jobs:
      - name: Stash merge gate helpers
        run: cp .github/scripts/merge-gate.js /tmp/merge-gate.js

+      - name: Pick Claude OAuth token
+        id: oauth
+        uses: MervinPraison/github-automation-template/.github/actions/[email protected]
+        with:
+          token_primary: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+          token_secondary: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN_B }}
+          active_slot: ${{ vars.CLAUDE_OAUTH_ACTIVE != '' && vars.CLAUDE_OAUTH_ACTIVE || 'primary' }}
+
      - name: Fetch PR branch
        env:
          PR_NUMBER: ${{ matrix.pr_number }}
@@ -283,7 +291,7 @@ jobs:
          CLAUDE_CODE_SUBAGENT_MODEL: inherit
          ANTHROPIC_DEFAULT_SONNET_MODEL: claude-sonnet-4-6
        with:
-          claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+          claude_code_oauth_token: ${{ steps.oauth.outputs.token }}
          github_token: ${{ steps.app-token.outputs.token }}
          model: claude-opus-4-8
          trigger_phrase: '@claude-merge-gate-internal'
@@ -249,6 +249,14 @@ jobs:
          fetch-depth: 0
          token: ${{ steps.app-token.outputs.token }}

+      - name: Pick Claude OAuth token
+        id: oauth
+        uses: MervinPraison/github-automation-template/.github/actions/[email protected]
+        with:
+          token_primary: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+          token_secondary: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN_B }}
+          active_slot: ${{ vars.CLAUDE_OAUTH_ACTIVE != '' && vars.CLAUDE_OAUTH_ACTIVE || 'primary' }}
+
      - name: Fetch PR branch and Setup Remote
        if: steps.check_fork.outputs.is_pr == 'true'
        env:
@@ -275,7 +283,7 @@ jobs:
          GH_TOKEN: ${{ steps.check_fork.outputs.can_push_fork == 'true' && secrets.GH_TOKEN || steps.app-token.outputs.token }}
        with:
          allowed_bots: 'praisonai-triage-agent[bot]'
-          claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+          claude_code_oauth_token: ${{ steps.oauth.outputs.token }}
          github_token: ${{ steps.app-token.outputs.token }}
          model: claude-opus-4-8
          trigger_phrase: "@claude"
@@ -21,6 +21,9 @@ on:
      - '.github/actions/install-monorepo-packages/**'
  workflow_dispatch:

+permissions:
+  contents: read
+
 jobs:
  cross-repo-integration:
    runs-on: ubuntu-latest
@@ -41,13 +44,13 @@ jobs:
        ref: main
        
    - name: Set up Python
-      uses: actions/setup-python@v4
+      uses: actions/setup-python@v5
      with:
        python-version: '3.11'
        
    - name: Install dependencies
      run: python -m pip install --upgrade pip
-    - uses: PraisonAI/.github/actions/install-monorepo-packages
+    - uses: ./PraisonAI/.github/actions/install-monorepo-packages
      with:
        repo-root: PraisonAI/
        wrapper-extras: ".[ui,dev]"
@@ -138,13 +141,13 @@ jobs:
        ref: main
        
    - name: Set up Python
-      uses: actions/setup-python@v4
+      uses: actions/setup-python@v5
      with:
        python-version: '3.11'
        
    - name: Install dependencies
      run: python -m pip install --upgrade pip
-    - uses: PraisonAI/.github/actions/install-monorepo-packages
+    - uses: ./PraisonAI/.github/actions/install-monorepo-packages
      with:
        repo-root: PraisonAI/
        wrapper-extras: ".[ui,dev]"
@@ -162,4 +165,4 @@ jobs:
          python -m pytest tests/integration/test_aiui_host_agentic.py -v
        else
          echo "OPENAI_API_KEY not available, skipping agentic tests"
-        fi
+        fi
@@ -117,13 +117,25 @@ jobs:
              const labels = issue.labels.map((l) => l.name);

              if (status !== 'DIRTY' && labels.includes(LABEL)) {
-                await github.rest.issues.removeLabel({
-                  owner,
-                  repo,
-                  issue_number: prNumber,
-                  name: LABEL,
-                });
-                core.info(`Removed ${LABEL} from PR #${prNumber}`);
+                try {
+                  await github.rest.issues.removeLabel({
+                    owner,
+                    repo,
+                    issue_number: prNumber,
+                    name: LABEL,
+                  });
+                  core.info(`Removed ${LABEL} from PR #${prNumber}`);
+                } catch (err) {
+                  // Fork PRs may return 403/404 when the token cannot mutate labels on the head repo's issue.
+                  // For upstream PRs a 403/404 is a genuine failure that must surface, otherwise the
+                  // merge-blocking label stays in place while the check reports success.
+                  const isFork = headRepo && headRepo !== baseRepo;
+                  if (isFork && (err.status === 403 || err.status === 404)) {
+                    core.info(`Skip remove ${LABEL} on fork PR #${prNumber}: ${err.status}`);
+                  } else {
+                    throw err;
+                  }
+                }
              }

              if (status !== 'DIRTY') return { skipped: true, reason: 'not dirty' };
@@ -1,6 +1,6 @@
 name: Pipeline Status Sync

-# Sync pipeline/* stage and blocker labels on open internal PRs.
+# Sync pipeline/* stage and blocker labels on open PRs (fork heads included; merge gate only for upstream branches).

 on:
  schedule:
@@ -16,8 +16,6 @@ jobs:
  review:
    runs-on: ubuntu-latest
    if: >
-      secrets.PRAISONAI_APP_ID != '' &&
-      secrets.PRAISONAI_APP_PRIVATE_KEY != '' &&
      (
        (github.event_name == 'pull_request' && github.event.pull_request.draft == false) ||
        github.event_name == 'workflow_dispatch' ||
@@ -32,7 +30,28 @@ jobs:
      )
    
    steps:
+      - name: Check reviewer configuration
+        id: config
+        env:
+          HAS_APP_CREDENTIALS: ${{ secrets.PRAISONAI_APP_ID != '' && secrets.PRAISONAI_APP_PRIVATE_KEY != '' }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+        run: |
+          echo "configured=$HAS_APP_CREDENTIALS" >> "$GITHUB_OUTPUT"
+          openai_ready=true
+          if [ -z "$OPENAI_API_KEY" ] || [ "$OPENAI_API_KEY" = "not-needed" ]; then
+            openai_ready=false
+            if [ "$HAS_APP_CREDENTIALS" = "true" ]; then
+              echo "::notice::Skipping PraisonAI PR review: OPENAI_API_KEY is missing or a placeholder."
+            fi
+          fi
+          if [ "$HAS_APP_CREDENTIALS" = "true" ] && [ "$openai_ready" = "true" ]; then
+            echo "ready=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "ready=false" >> "$GITHUB_OUTPUT"
+          fi
+
      - name: Determine checkout ref
+        if: steps.config.outputs.ready == 'true'
        id: dest
        env:
          ISSUE_NUMBER: ${{ github.event.issue.number }}
@@ -48,6 +67,7 @@ jobs:
          fi

      - name: Checkout Repository
+        if: steps.config.outputs.ready == 'true'
        uses: actions/checkout@v4
        with:
          persist-credentials: false
@@ -55,6 +75,7 @@ jobs:
          fetch-depth: 0
          
      - name: Generate GitHub App Token
+        if: steps.config.outputs.ready == 'true'
        id: generate_token
        uses: tibdex/github-app-token@v2
        with:
@@ -62,14 +83,17 @@ jobs:
          private_key: ${{ secrets.PRAISONAI_APP_PRIVATE_KEY }}
          
      - name: Set up Python
+        if: steps.config.outputs.ready == 'true'
        uses: actions/setup-python@v5
        with:
          python-version: '3.11'
          
      - name: Install PraisonAI
+        if: steps.config.outputs.ready == 'true'
        run: pip install "praisonai[all]"
        
      - name: Run PraisonAI PR Review
+        if: steps.config.outputs.ready == 'true'
        env:
          GITHUB_TOKEN: ${{ steps.generate_token.outputs.token }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
@@ -23,6 +23,7 @@ jobs:
          CLAUDE_APP_ID: ${{ secrets.CLAUDE_APP_ID }}
          CLAUDE_APP_PRIVATE_KEY: ${{ secrets.CLAUDE_APP_PRIVATE_KEY }}
          CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+          CLAUDE_CODE_OAUTH_TOKEN_B: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN_B }}
          GH_TOKEN_VALUE: ${{ secrets.GH_TOKEN }}
          DOCS_GITHUB_TOKEN: ${{ secrets.SECRETS_ADMIN_PAT }}
        run: |
@@ -34,8 +35,11 @@ jobs:
          echo -n "$CLAUDE_APP_ID" | gh secret set CLAUDE_APP_ID --repo "$TARGET"
          echo -n "$CLAUDE_APP_PRIVATE_KEY" | gh secret set CLAUDE_APP_PRIVATE_KEY --repo "$TARGET"
          echo -n "$CLAUDE_CODE_OAUTH_TOKEN" | gh secret set CLAUDE_CODE_OAUTH_TOKEN --repo "$TARGET"
+          if [ -n "${CLAUDE_CODE_OAUTH_TOKEN_B:-}" ]; then
+            echo -n "$CLAUDE_CODE_OAUTH_TOKEN_B" | gh secret set CLAUDE_CODE_OAUTH_TOKEN_B --repo "$TARGET"
+          fi
          echo -n "$GH_TOKEN_VALUE" | gh secret set GH_TOKEN --repo "$TARGET"
          if [ -n "${DOCS_GITHUB_TOKEN:-}" ]; then
            echo -n "$DOCS_GITHUB_TOKEN" | gh secret set DOCS_GITHUB_TOKEN --repo "$TARGET"
          fi
-          echo "Synced CLAUDE_APP_ID, CLAUDE_APP_PRIVATE_KEY, CLAUDE_CODE_OAUTH_TOKEN, GH_TOKEN → $TARGET"
+          echo "Synced CLAUDE_APP_ID, CLAUDE_APP_PRIVATE_KEY, CLAUDE_CODE_OAUTH_TOKEN, CLAUDE_CODE_OAUTH_TOKEN_B (if set), GH_TOKEN → $TARGET"
@@ -31,6 +31,13 @@ env:
  PRAISONAI_ALLOW_NETWORK: '0'
  PRAISONAI_LIVE_TESTS: '0'
  PRAISONAI_TEST_TIER: main
+  # litellm reads this at import and, without it, downloads its model-price
+  # table from raw.githubusercontent.com before a single test runs -- which
+  # made the suite depend on that host being reachable. It has to be set in
+  # the environment rather than a conftest, because pytest imports plugins
+  # (and litellm with them) before any conftest is loaded.
+  LITELLM_LOCAL_MODEL_COST_MAP: 'True'
+  HF_HUB_OFFLINE: '1'

 jobs:
  test-core-collect:
@@ -109,6 +116,55 @@ jobs:
            extra_ignore: ""
            pythonpath: >-
              ${{ github.workspace }}/src/praisonai-agents:${{ github.workspace }}/src/praisonai-code
+          - shard: agents-core
+            workdir: src/praisonai-agents
+            # src/praisonai-agents/tests holds 727 files and 11,437 test
+            # functions, and until now NO workflow referenced the directory at
+            # all -- so the LLM layer, the provider adapters, the local-model
+            # package and the streaming paths were guarded by tests that never
+            # ran. Nearly every defect found in the recent local-model audits
+            # had a working test sitting unused beside it.
+            #
+            # The ten directories originally left out of this shard (knowledge,
+            # runtime, session, tools, bus, plugins, skills, agents, rag,
+            # streaming -- about 52 failures) have since been repaired, so they
+            # are all listed below: 1,567 tests in the seven added here alone.
+            # Leaving them out would have meant the repairs were never exercised.
+            #
+            # The whole tree in one invocation was hanging. The cause was NOT
+            # the pytest-rerunfailures run_server thread guessed at here
+            # earlier: it is test_botos_integration.py, where a supervised
+            # Bot.start() deliberately never returns and the mock adapter
+            # reports is_running=True forever (fixed on this branch). Naming
+            # directories explicitly is still the safer shape for a shard.
+            paths: >-
+              tests/unit/agent/ tests/unit/app/ tests/unit/approval/
+              tests/unit/auth/ tests/unit/background/ tests/unit/checkpoints/
+              tests/unit/cli/ tests/unit/compaction/ tests/unit/conditions/
+              tests/unit/config/ tests/unit/context/ tests/unit/embedding/
+              tests/unit/eval/ tests/unit/gateway/ tests/unit/goal/
+              tests/unit/hooks/ tests/unit/llm/ tests/unit/lsp/ tests/unit/mcp/
+              tests/unit/memory/ tests/unit/permissions/ tests/unit/policy/
+              tests/unit/process/ tests/unit/runs/ tests/unit/sandbox/
+              tests/unit/security/ tests/unit/server/ tests/unit/snapshot/
+              tests/unit/storage/ tests/unit/task/ tests/unit/telemetry/
+              tests/unit/trace/ tests/unit/utils/ tests/unit/workflows/
+              tests/unit/workspace/ tests/unit/knowledge/ tests/unit/session/
+              tests/unit/plugins/ tests/unit/runtime/ tests/unit/tools/
+              tests/unit/bus/ tests/unit/skills/ tests/unit/agents/
+              tests/unit/rag/ tests/unit/streaming/
+              tests/unit/test_adapter_registry.py
+            # Both tests previously deselected here now pass in combination.
+            # test_lazy_litellm_import was looking the loader module up by
+            # name and could get a different object than the one get_litellm()
+            # mutates; it reads the function's own __globals__ now.
+            # test_skills_discovery_uses_paths_module read the real home
+            # directory and asserted every discovered dir was named "skills",
+            # which is false once a remote skill has ever been cached; it runs
+            # against a temp HOME and asserts the paths module is used.
+            extra_ignore: ""
+            pythonpath: >-
+              ${{ github.workspace }}/src/praisonai-agents
          - shard: train
            workdir: src/praisonai-train
            # tests/unit/data/ (96 tests) was collected by no workflow at all.
@@ -42,7 +42,7 @@ jobs:
  # ============================================
  smoke:
    runs-on: ubuntu-latest
-    timeout-minutes: 12
+    timeout-minutes: 16
    steps:
    - uses: actions/checkout@v4
      with:
@@ -773,7 +773,9 @@ jobs:
    env:
      # Bump deliberately: this is the version everything in this job was
      # verified against. Latest at time of writing is v0.33.3.
-      OLLAMA_VERSION: v0.33.2
+      # No leading "v": install.sh passes ?version=$OLLAMA_VERSION and ollama.com
+      # redirects to GitHub tag v$version (vv0.33.2 → 404 if set to v0.33.2).
+      OLLAMA_VERSION: "0.33.2"
      # 522 MB, capabilities: completion, tools, thinking.
      OLLAMA_TEST_MODEL: qwen3:0.6b

@@ -122,3 +122,4 @@ jobs/*
 # Local Android SDK/NDK/JDK for building praisonai-mobile on a laptop without
 # Android Studio (how: .github/workflows/mobile-platforms.yml). Several GB.
 .android-sdk/
+.worktrees/
@@ -12,7 +12,6 @@
 <a href="https://github.com/MervinPraison/PraisonAI"><img src="https://static.pepy.tech/badge/PraisonAI" alt="Total Downloads" /></a>
 <a href="https://github.com/MervinPraison/PraisonAI"><img src="https://img.shields.io/github/v/release/MervinPraison/PraisonAI" alt="Latest Stable Version" /></a>
 <a href="https://github.com/MervinPraison/PraisonAI"><img src="https://img.shields.io/badge/License-MIT-yellow.svg" alt="License" /></a>
-<a href="https://registry.modelcontextprotocol.io/servers/io.github.MervinPraison/praisonai"><img src="https://img.shields.io/badge/MCP-Registry-blue" alt="MCP Registry" /></a>
 </p>

 <div align="center">
@@ -776,7 +776,7 @@ Methods:
 * <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">wait_for_completions</a>(timeout: Optional[float] = None, agent_ids: Optional[List[str]] = None) -> List[SubAgentCompletionEvent]</code>
 * <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">where_does_it_run</a>() -> str</code>
 * <code title="class AutoApproveBackend">AutoApproveBackend.<a href="./src/praisonai-agents/praisonaiagents/approval/backends.py">request_approval</a>(request: ApprovalRequest) -> ApprovalDecision</code>
-* <code title="class AutoApproveBackend">AutoApproveBackend.<a href="./src/praisonai-agents/praisonaiagents/approval/backends.py">request_approval_sync</a>(request: ApprovalRequest) -> ApprovalDecision</code>
+* <code title="class AutoApproveBackend">AutoApproveBackend.<a href="./src/praisonai-agents/praisonaiagents/approval/backends.py">request_approval_sync</a>(request: ApprovalRequest, timeout: Optional[float] = None) -> ApprovalDecision</code>
 * <code title="class RunOutcome">RunOutcome.<a href="./src/praisonai-agents/praisonaiagents/agent/run_outcome.py">completed</a>(output: Optional[str] = None) -> 'RunOutcome'</code>
 * <code title="class RunOutcome">RunOutcome.<a href="./src/praisonai-agents/praisonaiagents/agent/run_outcome.py">from_exception</a>(exc: BaseException, output: Optional[str] = None) -> 'RunOutcome'</code>
 * <code title="class RunOutcome">RunOutcome.<a href="./src/praisonai-agents/praisonaiagents/agent/run_outcome.py">succeeded</a>() -> bool</code>
@@ -1018,7 +1018,8 @@ export { // Bot types
  FailoverManager, // Enums
  SandboxStatus, // Gateway types
  type GatewayConfig, // Other types
-  type ProviderStatus, AutonomyLevel, GatewayEventType, RagRetrievalPolicy } from "./gateway";
+  type ProviderStatus, // Session-projection reducer (Issue #5324)
+  SessionProjection, AutonomyLevel, GatewayEventType, RagRetrievalPolicy } from "./gateway";
 export { LLMGuardrail, createLLMGuardrail } from "./guardrails/llm-guardrail";
 export { DisplayTypes, HooksManager, WorkflowHooksExecutor, clearAllCallbacks, clearApprovalCallback, createHooksManager, createLoggingOperationHooks, createLoggingWorkflowHooks, createTimingWorkflowHooks, createValidationOperationHooks, createWorkflowHooks, executeCallback, executeSyncCallback, getRegisteredDisplayTypes, hasApprovalCallback, registerApprovalCallback, registerDisplayCallback, requestApproval, unregisterDisplayCallback } from "./hooks";
 export { // Computer Use
@@ -1056,7 +1057,7 @@ export type { AsyncLearnProtocol, BaseStoreOptions, LearnEntryConvertible, Learn
 export { Memory, createMemory } from "./memory/memory";
 export type { MemoryConfig, MemoryEntry } from "./memory/memory";
 export { RulesManager, createRulesManager, createSafetyRules } from "./memory/rules-manager";
-export { DEFAULT_PROFILE, HarnessProfile, listHarnessProfiles, registerProfile, register_profile, resetHarnessRegistry, resolveHarness, resolve_harness } from "./model-harness";
+export { DEFAULT_PROFILE, HarnessProfile, ModelRequestBlocked, ScriptExhausted, ScriptedModel, allowModelRequests, listHarnessProfiles, modelRequestsAllowed, noModelRequests, registerProfile, register_profile, resetHarnessRegistry, resolveHarness, resolve_harness } from "./model-harness";
 export type { HarnessProfileConfig, HarnessRegistryEntry, HarnessResolverProtocol } from "./model-harness";
 export { // Adapters
  NoopObservabilityAdapter, // Constants
@@ -16,7 +16,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
    praisonai_tools \
-    "praisonai>=4.7.8" \
+    "praisonai>=4.7.12" \
    "praisonai[chat]" \
    "embedchain[github,youtube]"

@@ -20,7 +20,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
    praisonai_tools \
-    "praisonai>=4.7.8" \
+    "praisonai>=4.7.12" \
    "praisonai[ui]" \
    "praisonai[chat]" \
    "praisonai[realtime]" \
@@ -16,7 +16,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
    praisonai_tools \
-    "praisonai>=4.7.8" \
+    "praisonai>=4.7.12" \
    "praisonai[ui]" \
    "praisonai[crewai]"

@@ -140,4 +140,4 @@ Key fixes applied:
 4. String-based LLM guardrails are also supported

 The guardrail will automatically retry (up to max_retries times) if validation fails.
-""")
+""")
@@ -0,0 +1,28 @@
+"""OpenRouter video generation via VideoAgent.
+
+Requires OPENROUTER_API_KEY. Optionally set OPENROUTER_REFERER and
+OPENROUTER_APP_TITLE for OpenRouter dashboard attribution.
+
+The openrouter/ prefix is required so the request routes to OpenRouter.
+Discover exact model ids at https://openrouter.ai/api/v1/videos/models.
+"""
+
+from praisonaiagents import VideoAgent
+
+agent = VideoAgent(llm="openrouter/google/veo-3.1")
+
+video = agent.generate(
+    "A serene mountain landscape at sunset, locked camera",
+    duration=6,
+    resolution="720p",
+    aspect_ratio="16:9",
+    generate_audio=False,
+)
+print(f"Video ID: {video.id}")
+
+completed = agent.wait_for_completion(video.id)
+if completed.status == "completed":
+    agent.download(video.id, "openrouter_veo.mp4")
+    print("Saved openrouter_veo.mp4")
+else:
+    print(f"Generation ended with status: {completed.status}")
@@ -27,11 +27,10 @@ mcp:
      enabled: true
    
    time:
-      # Time and date utilities
-      command: npx
+      # Time and date utilities (published Python MCP server via uvx)
+      command: uvx
      args:
-        - "-y" 
-        - "@modelcontextprotocol/server-time"
+        - "mcp-server-time"
      env: {}
      enabled: true

@@ -1,9 +0,0 @@
-{
-  "short_term_limit": 100,
-  "long_term_limit": 1000,
-  "episodic_retention_days": 30,
-  "enable_summarization": false,
-  "summarization_threshold": 50,
-  "importance_threshold": 0.7,
-  "auto_promote": true
-}
@@ -1,21 +0,0 @@
-{
-  "root_path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp6_tn0h1t",
-  "total_files": 0,
-  "total_tokens": 11,
-  "levels": 3,
-  "built_at": "2026-02-02T17:47:19.241146",
-  "nodes": {
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp6_tn0h1t": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp6_tn0h1t",
-      "level": 3,
-      "summary": "Project with 0 folders and 0 root files: ",
-      "children": [],
-      "token_count": 11,
-      "created_at": "2026-02-02T17:47:19.241224",
-      "metadata": {
-        "folder_count": 0,
-        "root_file_count": 0
-      }
-    }
-  }
-}
@@ -1,35 +0,0 @@
-{
-  "root_path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp7wxr57wm",
-  "total_files": 1,
-  "total_tokens": 16,
-  "levels": 2,
-  "built_at": "2026-02-02T17:47:19.243732",
-  "nodes": {
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp7wxr57wm/src/app.py": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp7wxr57wm/src/app.py",
-      "level": 1,
-      "summary": "# Application code",
-      "children": [],
-      "token_count": 5,
-      "created_at": "2026-02-02T17:47:19.243833",
-      "metadata": {
-        "filename": "app.py",
-        "relative_path": "src/app.py"
-      }
-    },
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp7wxr57wm/src": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp7wxr57wm/src",
-      "level": 2,
-      "summary": "Folder contains 1 files: # Application code",
-      "children": [
-        "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp7wxr57wm/src/app.py"
-      ],
-      "token_count": 11,
-      "created_at": "2026-02-02T17:47:19.243852",
-      "metadata": {
-        "folder_name": "src",
-        "file_count": 1
-      }
-    }
-  }
-}
@@ -1,21 +0,0 @@
-{
-  "root_path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp8qwfp4zl",
-  "total_files": 0,
-  "total_tokens": 11,
-  "levels": 3,
-  "built_at": "2026-02-02T17:49:37.074095",
-  "nodes": {
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp8qwfp4zl": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp8qwfp4zl",
-      "level": 3,
-      "summary": "Project with 0 folders and 0 root files: ",
-      "children": [],
-      "token_count": 11,
-      "created_at": "2026-02-02T17:49:37.074201",
-      "metadata": {
-        "folder_count": 0,
-        "root_file_count": 0
-      }
-    }
-  }
-}
@@ -1,33 +0,0 @@
-{
-  "root_path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp9nw2kfoj",
-  "total_files": 2,
-  "total_tokens": 14,
-  "levels": 1,
-  "built_at": "2026-02-02T17:47:19.242486",
-  "nodes": {
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp9nw2kfoj/utils.py": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp9nw2kfoj/utils.py",
-      "level": 1,
-      "summary": "def helper(): return True",
-      "children": [],
-      "token_count": 7,
-      "created_at": "2026-02-02T17:47:19.242582",
-      "metadata": {
-        "filename": "utils.py",
-        "relative_path": "utils.py"
-      }
-    },
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp9nw2kfoj/main.py": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp9nw2kfoj/main.py",
-      "level": 1,
-      "summary": "def main(): print('Hello')",
-      "children": [],
-      "token_count": 7,
-      "created_at": "2026-02-02T17:47:19.242630",
-      "metadata": {
-        "filename": "main.py",
-        "relative_path": "main.py"
-      }
-    }
-  }
-}
@@ -1,35 +0,0 @@
-{
-  "root_path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp_dbottar",
-  "total_files": 1,
-  "total_tokens": 16,
-  "levels": 2,
-  "built_at": "2026-02-02T17:49:37.077857",
-  "nodes": {
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp_dbottar/src/app.py": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp_dbottar/src/app.py",
-      "level": 1,
-      "summary": "# Application code",
-      "children": [],
-      "token_count": 5,
-      "created_at": "2026-02-02T17:49:37.077985",
-      "metadata": {
-        "filename": "app.py",
-        "relative_path": "src/app.py"
-      }
-    },
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp_dbottar/src": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp_dbottar/src",
-      "level": 2,
-      "summary": "Folder contains 1 files: # Application code",
-      "children": [
-        "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmp_dbottar/src/app.py"
-      ],
-      "token_count": 11,
-      "created_at": "2026-02-02T17:49:37.078011",
-      "metadata": {
-        "folder_name": "src",
-        "file_count": 1
-      }
-    }
-  }
-}
@@ -1,33 +0,0 @@
-{
-  "root_path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpp5k48vv4",
-  "total_files": 2,
-  "total_tokens": 14,
-  "levels": 1,
-  "built_at": "2026-02-02T17:49:37.075925",
-  "nodes": {
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpp5k48vv4/utils.py": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpp5k48vv4/utils.py",
-      "level": 1,
-      "summary": "def helper(): return True",
-      "children": [],
-      "token_count": 7,
-      "created_at": "2026-02-02T17:49:37.076036",
-      "metadata": {
-        "filename": "utils.py",
-        "relative_path": "utils.py"
-      }
-    },
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpp5k48vv4/main.py": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpp5k48vv4/main.py",
-      "level": 1,
-      "summary": "def main(): print('Hello')",
-      "children": [],
-      "token_count": 7,
-      "created_at": "2026-02-02T17:49:37.076227",
-      "metadata": {
-        "filename": "main.py",
-        "relative_path": "main.py"
-      }
-    }
-  }
-}
@@ -1,33 +0,0 @@
-{
-  "root_path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwwb01831",
-  "total_files": 2,
-  "total_tokens": 16,
-  "levels": 1,
-  "built_at": "2026-02-02T17:49:37.079504",
-  "nodes": {
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwwb01831/javascript.js": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwwb01831/javascript.js",
-      "level": 1,
-      "summary": "JavaScript web development",
-      "children": [],
-      "token_count": 7,
-      "created_at": "2026-02-02T17:49:37.079612",
-      "metadata": {
-        "filename": "javascript.js",
-        "relative_path": "javascript.js"
-      }
-    },
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwwb01831/python.py": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwwb01831/python.py",
-      "level": 1,
-      "summary": "Python programming language code",
-      "children": [],
-      "token_count": 9,
-      "created_at": "2026-02-02T17:49:37.079658",
-      "metadata": {
-        "filename": "python.py",
-        "relative_path": "python.py"
-      }
-    }
-  }
-}
@@ -1,33 +0,0 @@
-{
-  "root_path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwzdvf3_7",
-  "total_files": 2,
-  "total_tokens": 16,
-  "levels": 1,
-  "built_at": "2026-02-02T17:47:19.245229",
-  "nodes": {
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwzdvf3_7/javascript.js": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwzdvf3_7/javascript.js",
-      "level": 1,
-      "summary": "JavaScript web development",
-      "children": [],
-      "token_count": 7,
-      "created_at": "2026-02-02T17:47:19.245344",
-      "metadata": {
-        "filename": "javascript.js",
-        "relative_path": "javascript.js"
-      }
-    },
-    "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwzdvf3_7/python.py": {
-      "path": "/var/folders/0b/t9c4hwvx6lg6jk58gw8w_dp80000gn/T/tmpwzdvf3_7/python.py",
-      "level": 1,
-      "summary": "Python programming language code",
-      "children": [],
-      "token_count": 9,
-      "created_at": "2026-02-02T17:47:19.245395",
-      "metadata": {
-        "filename": "python.py",
-        "relative_path": "python.py"
-      }
-    }
-  }
-}
@@ -18,76 +18,76 @@ This report compares **Python SDK feature categories** against **Python document
 | Category | Features | Docs | Lines |
 |----------|----------|------|-------|
 | ✅ AGUI | 1 | 2 | 511 |
-| ✅ Agent | 22 | 59 | 18844 |
+| ✅ Agent | 22 | 59 | 19076 |
 | ✅ Agent-to-Agent (A2A) | 1 | 7 | 2996 |
 | ✅ Agent-to-User (A2U) | 1 | 3 | 701 |
-| ✅ Approval | 1 | 8 | 3934 |
+| ✅ Approval | 1 | 8 | 4028 |
 | ✅ Audio | 2 | 12 | 1000 |
-| ✅ Auto Generation | 5 | 12 | 3879 |
-| ✅ Autonomy | 3 | 7 | 2794 |
-| ✅ Bots | 7 | 36 | 14678 |
+| ✅ Auto Generation | 5 | 12 | 3941 |
+| ✅ Autonomy | 3 | 7 | 2805 |
+| ✅ Bots | 7 | 38 | 15809 |
 | ✅ Budget | 1 | 1 | 287 |
-| ✅ CLI | 5 | 124 | 38873 |
-| ✅ Chunking | 2 | 2 | 424 |
+| ✅ CLI | 5 | 124 | 39199 |
+| ✅ Chunking | 2 | 2 | 435 |
 | ✅ Citations | 2 | 1 | 202 |
 | ✅ Code Execution | 2 | 14 | 5022 |
 | ✅ Conditions | 1 | 3 | 1386 |
 | ✅ Configuration | 3 | 8 | 3341 |
-| ✅ Context Management | 16 | 36 | 14275 |
+| ✅ Context Management | 16 | 36 | 14384 |
 | ✅ Deep Research | 8 | 2 | 591 |
 | ✅ Display | 6 | 3 | 844 |
 | ✅ Embeddings | 6 | 23 | 2051 |
-| ✅ Evaluation | 1 | 9 | 3537 |
-| ✅ Events | 1 | 2 | 767 |
-| ✅ Execution | 3 | 4 | 1693 |
-| ✅ Failover | 2 | 1 | 480 |
+| ✅ Evaluation | 1 | 9 | 3637 |
+| ✅ Events | 1 | 2 | 782 |
+| ✅ Execution | 3 | 4 | 1694 |
+| ✅ Failover | 2 | 1 | 537 |
 | ✅ Files | 2 | 7 | 2376 |
-| ✅ Flow | 1 | 3 | 788 |
-| ✅ Gateway | 7 | 87 | 33047 |
-| ✅ Guardrails | 5 | 4 | 2470 |
+| ✅ Flow | 1 | 3 | 811 |
+| ✅ Gateway | 7 | 97 | 37652 |
+| ✅ Guardrails | 5 | 4 | 2642 |
 | ✅ Handoffs | 11 | 6 | 2748 |
-| ✅ Hooks | 2 | 9 | 4990 |
-| ✅ Image | 1 | 11 | 1255 |
-| ✅ Knowledge | 4 | 19 | 5964 |
-| ✅ LLM | 3 | 15 | 5944 |
-| ✅ Loops | 4 | 5 | 1475 |
-| ✅ MCP | 1 | 63 | 14860 |
-| ✅ Memory | 6 | 21 | 8186 |
+| ✅ Hooks | 2 | 9 | 5230 |
+| ✅ Image | 1 | 11 | 1295 |
+| ✅ Knowledge | 4 | 20 | 6260 |
+| ✅ LLM | 3 | 15 | 5958 |
+| ✅ Loops | 4 | 5 | 1478 |
+| ✅ MCP | 1 | 64 | 15116 |
+| ✅ Memory | 6 | 21 | 8198 |
 | ✅ OCR | 2 | 1 | 237 |
-| ✅ Observability | 2 | 23 | 2842 |
+| ✅ Observability | 2 | 23 | 2859 |
 | ✅ Optimizer | 1 | 2 | 792 |
-| ✅ Output | 3 | 5 | 1273 |
+| ✅ Output | 3 | 5 | 1289 |
 | ✅ Parallel Execution | 3 | 2 | 528 |
-| ✅ Planning | 6 | 7 | 1825 |
-| ✅ Plugins | 8 | 5 | 3316 |
+| ✅ Planning | 6 | 7 | 1827 |
+| ✅ Plugins | 8 | 5 | 4112 |
 | ✅ Prompts | 2 | 9 | 2043 |
-| ✅ Providers | 1 | 57 | 8640 |
+| ✅ Providers | 1 | 58 | 8751 |
 | ✅ Query | 1 | 2 | 736 |
-| ✅ RAG | 5 | 15 | 3124 |
+| ✅ RAG | 5 | 15 | 3152 |
 | ✅ Realtime | 2 | 5 | 668 |
 | ✅ Reflection | 3 | 3 | 732 |
 | ✅ Retrieval | 2 | 5 | 1046 |
 | ✅ Routing | 1 | 2 | 430 |
-| ✅ Sandbox | 5 | 7 | 3690 |
-| ✅ Security | 1 | 3 | 3598 |
-| ✅ Sessions | 4 | 20 | 8049 |
-| ✅ Skills | 6 | 16 | 5984 |
+| ✅ Sandbox | 5 | 7 | 3720 |
+| ✅ Security | 1 | 3 | 3642 |
+| ✅ Sessions | 4 | 21 | 8388 |
+| ✅ Skills | 6 | 16 | 6045 |
 | ✅ Tasks | 2 | 8 | 3266 |
 | ✅ Telemetry | 1 | 2 | 660 |
-| ✅ Templates | 1 | 8 | 1667 |
-| ✅ Tools | 12 | 145 | 44509 |
+| ✅ Templates | 1 | 9 | 1859 |
+| ✅ Tools | 12 | 147 | 45378 |
 | ✅ Tracing | 3 | 2 | 139 |
 | ✅ Vector Store | 1 | 12 | 1219 |
-| ✅ Video | 2 | 6 | 728 |
+| ✅ Video | 2 | 10 | 1303 |
 | ✅ Vision | 2 | 1 | 329 |
-| ✅ Web | 3 | 10 | 2815 |
-| ✅ Workflows | 5 | 21 | 8851 |
+| ✅ Web | 3 | 10 | 2822 |
+| ✅ Workflows | 5 | 21 | 8880 |

 ## Documentation Without Features

 These docs exist but don't match any implemented feature category:

- ℹ️ Database (43 docs, 7657 lines)
+- ℹ️ Database (44 docs, 8292 lines)
 - ℹ️ Documents (1 docs, 810 lines)
 - ℹ️ Teams (1 docs, 214 lines)

@@ -6,8 +6,13 @@ Compares agent instantiation times across popular AI agent frameworks.

 Usage:
    python benchmarks/simple_benchmark.py
+    python benchmarks/simple_benchmark.py --skip-crewai --json
+    python benchmarks/simple_benchmark.py --json --fail-over-threshold
 """

+import json
+import os
+import sys
 import time
 from typing import Literal
 import argparse
@@ -16,6 +21,14 @@ import argparse
 ITERATIONS = 100
 WARMUP = 10

+THRESHOLDS_PATH = os.path.join(os.path.dirname(__file__), 'thresholds.json')
+
+
+def load_thresholds(path: str = THRESHOLDS_PATH) -> dict:
+    """Load benchmark thresholds from thresholds.json."""
+    with open(path) as f:
+        return json.load(f)
+

 def sample_tool(city: Literal['nyc', 'sf']):
    """Sample tool for benchmark testing."""
@@ -36,35 +49,44 @@ def measure_instantiation(create_fn, iterations=ITERATIONS, warmup=WARMUP):
    return sum(times) / len(times)


-def run_benchmark():
+def run_benchmark(skip_crewai: bool = False, quiet: bool = False,
+                  iterations: int = ITERATIONS, warmup: int = WARMUP):
    """Run the benchmark across all available frameworks (without tools)."""
    results = {}
-    
-    print('=' * 60)
-    print('PraisonAI Agents - Performance Benchmark')
-    print('=' * 60)
-    print(f'\nIterations: {ITERATIONS}')
-    print('Metric: Agent instantiation time (microseconds)\n')
+
+    def log(msg: str = ''):
+        if not quiet:
+            print(msg)
+
+    def measure(create_fn):
+        return measure_instantiation(create_fn, iterations=iterations, warmup=warmup)
+
+    log('=' * 60)
+    log('PraisonAI Agents - Performance Benchmark')
+    log('=' * 60)
+    log(f'\nIterations: {iterations}')
+    log(f'Warmup: {warmup}')
+    log('Metric: Agent instantiation time (microseconds)\n')
    
    # PraisonAI (without tools)
-    print("Testing PraisonAI...")
+    log("Testing PraisonAI...")
    from praisonaiagents import Agent as PraisonAgent
    
-    results['PraisonAI'] = measure_instantiation(
+    results['PraisonAI'] = measure(
        lambda: PraisonAgent(name='Test', model='gpt-4o-mini', output="silent")
    )
    
-    results['PraisonAI (LiteLLM)'] = measure_instantiation(
+    results['PraisonAI (LiteLLM)'] = measure(
        lambda: PraisonAgent(name='Test', model='openai/gpt-4o-mini', output="silent")
    )
    
    # Other frameworks for comparison
-    print("Testing other frameworks...")
+    log("Testing other frameworks...")
    
    try:
        from agno.agent import Agent as AgnoAgent
        from agno.models.openai import OpenAIChat
-        results['Agno'] = measure_instantiation(
+        results['Agno'] = measure(
            lambda: AgnoAgent(model=OpenAIChat(id='gpt-4o-mini'))
        )
    except ImportError:
@@ -72,7 +94,7 @@ def run_benchmark():
    
    try:
        from pydantic_ai import Agent as PydanticAgent
-        results['PydanticAI'] = measure_instantiation(
+        results['PydanticAI'] = measure(
            lambda: PydanticAgent('openai:gpt-4o-mini')
        )
    except ImportError:
@@ -80,7 +102,7 @@ def run_benchmark():
    
    try:
        from agents import Agent as OpenAIAgent
-        results['OpenAI Agents SDK'] = measure_instantiation(
+        results['OpenAI Agents SDK'] = measure(
            lambda: OpenAIAgent(name='Test', model='gpt-4o-mini')
        )
    except ImportError:
@@ -96,47 +118,50 @@ def run_benchmark():
            """Get weather info."""
            return 'sunny' if city == 'sf' else 'cloudy'
        
-        results['LangGraph'] = measure_instantiation(
+        results['LangGraph'] = measure(
            lambda: create_react_agent(model=ChatOpenAI(model='gpt-4o-mini'), tools=[get_weather_lg])
        )
    except ImportError:
        pass
    
-    try:
-        from crewai.agent import Agent as CrewAgent
-        from crewai.tools import tool as crewai_tool
-        
-        @crewai_tool("Weather Tool")
-        def get_weather_crew(city: Literal['nyc', 'sf']):
-            """Get weather info."""
-            return 'sunny' if city == 'sf' else 'cloudy'
-        
-        results['CrewAI'] = measure_instantiation(
-            lambda: CrewAgent(
-                role='Weather Agent',
-                goal='Provide weather info',
-                backstory='A weather expert',
-                tools=[get_weather_crew],
-                verbose=False
+    if skip_crewai:
+        log("Skipping CrewAI (--skip-crewai): optional framework import can hang on some platforms.")
+    else:
+        try:
+            from crewai.agent import Agent as CrewAgent
+            from crewai.tools import tool as crewai_tool
+
+            @crewai_tool("Weather Tool")
+            def get_weather_crew(city: Literal['nyc', 'sf']):
+                """Get weather info."""
+                return 'sunny' if city == 'sf' else 'cloudy'
+
+            results['CrewAI'] = measure(
+                lambda: CrewAgent(
+                    role='Weather Agent',
+                    goal='Provide weather info',
+                    backstory='A weather expert',
+                    tools=[get_weather_crew],
+                    verbose=False
+                )
            )
-        )
-    except ImportError:
-        pass
+        except ImportError:
+            pass
    
    # Print results
-    print('\n' + '=' * 60)
-    print('RESULTS')
-    print('=' * 60)
+    log('\n' + '=' * 60)
+    log('RESULTS')
+    log('=' * 60)
    
    baseline = results.get('PraisonAI', 1)
-    print(f"\n{'Framework':<25} {'Avg Time (μs)':<15} {'Relative':<10}")
-    print('-' * 50)
+    log(f"\n{'Framework':<25} {'Avg Time (μs)':<15} {'Relative':<10}")
+    log('-' * 50)
    
    for name, avg in sorted(results.items(), key=lambda x: x[1]):
        ratio = avg / baseline
-        print(f'{name:<25} {avg:<15.2f} {ratio:.2f}x')
+        log(f'{name:<25} {avg:<15.2f} {ratio:.2f}x')
    
-    print('\n' + '=' * 60)
+    log('\n' + '=' * 60)
    return results


@@ -163,8 +188,13 @@ def get_package_versions():
    return versions


-def save_results(results: dict, filename: str = 'BENCHMARK_RESULTS.md'):
-    """Save benchmark results to a markdown file and update README."""
+def save_results(results: dict, filename: str = 'BENCHMARK_RESULTS.md', *,
+                 iterations: int = ITERATIONS, log=print):
+    """Save benchmark results to a markdown file and update README.
+
+    Status messages are emitted via ``log`` (defaults to ``print``). Pass a
+    stderr logger when stdout must stay machine-readable (e.g. ``--json``).
+    """
    import os
    import re
    from datetime import datetime
@@ -191,7 +221,7 @@ def save_results(results: dict, filename: str = 'BENCHMARK_RESULTS.md'):
    with open(filepath, 'w') as f:
        f.write('# PraisonAI Agents - Benchmark Results\n\n')
        f.write(f'**Generated:** {datetime.now().strftime("%Y-%m-%d %H:%M:%S")}\n')
-        f.write(f'**Iterations:** {ITERATIONS}\n')
+        f.write(f'**Iterations:** {iterations}\n')
        f.write('**Test:** Agent instantiation (without tools)\n\n')
        f.write('## Results\n\n')
        f.write('| Framework | Avg Time (μs) | Relative |\n')
@@ -208,17 +238,17 @@ def save_results(results: dict, filename: str = 'BENCHMARK_RESULTS.md'):
        f.write('python benchmarks/simple_benchmark.py\n')
        f.write('```\n')
    
-    print(f'\nResults saved to: {filepath}')
+    log(f'\nResults saved to: {filepath}')
    
    # Also update the main README.md
    readme_path = os.path.join(os.path.dirname(__file__), '..', '..', '..', 'README.md')
    if os.path.exists(readme_path):
-        update_readme(readme_path, table_rows)
+        update_readme(readme_path, table_rows, log=log)
    
    return filepath


-def update_readme(readme_path: str, table_rows: list):
+def update_readme(readme_path: str, table_rows: list, *, log=print):
    """Update the performance section in README.md with latest results."""
    import re
    
@@ -239,17 +269,54 @@ def update_readme(readme_path: str, table_rows: list):
        with open(readme_path, 'w') as f:
            f.write(content)
        
-        print(f'README.md updated: {readme_path}')
+        log(f'README.md updated: {readme_path}')
+
+
+def main():
+    parser = argparse.ArgumentParser(description='PraisonAI Agents - Performance Benchmark')
+    parser.add_argument('--save', action='store_true', help='Save results to file')
+    parser.add_argument('--skip-crewai', action='store_true',
+                        help='Skip the optional CrewAI benchmark (its import can hang on some platforms)')
+    parser.add_argument('--json', action='store_true',
+                        help='Emit results as JSON to stdout (for CI baselines/artifacts)')
+    parser.add_argument('--fail-over-threshold', action='store_true',
+                        help='Exit non-zero if PraisonAI instantiation exceeds the budget in thresholds.json')
+    args = parser.parse_args()
+
+    thresholds = load_thresholds()
+    iterations = int(thresholds.get('iterations', ITERATIONS))
+    warmup = int(thresholds.get('warmup', WARMUP))
+
+    results = run_benchmark(
+        skip_crewai=args.skip_crewai,
+        quiet=args.json,
+        iterations=iterations,
+        warmup=warmup,
+    )
+
+    if args.save:
+        # Keep stdout machine-readable under --json by sending save status to stderr.
+        save_log = (lambda msg='': print(msg, file=sys.stderr)) if args.json else print
+        save_results(results, iterations=iterations, log=save_log)
+    elif not args.json:
+        print('\nResults not saved (use --save flag to save results to file)')
+
+    exit_code = 0
+    if args.fail_over_threshold:
+        max_us = thresholds['praisonai_instantiation_us_max']
+        actual_us = results.get('PraisonAI')
+        if actual_us is not None and actual_us > max_us:
+            print(
+                f"ERROR: PraisonAI instantiation {actual_us:.2f} us > threshold {max_us} us",
+                file=sys.stderr,
+            )
+            exit_code = 1
+
+    if args.json:
+        print(json.dumps({'results': results, 'iterations': iterations, 'warmup': warmup}))
+
+    return exit_code


 if __name__ == '__main__':
-    parser = argparse.ArgumentParser(description='PraisonAI Agents - Performance Benchmark')
-    parser.add_argument('--save', action='store_true', help='Save results to file')
-    args = parser.parse_args()
-    
-    results = run_benchmark()
-    
-    if args.save:
-        save_results(results)
-    else:
-        print('\nResults not saved (use --save flag to save results to file)')
+    sys.exit(main())
@@ -0,0 +1,5 @@
+{
+  "praisonai_instantiation_us_max": 450,
+  "iterations": 100,
+  "warmup": 10
+}
@@ -0,0 +1,17 @@
+"""Root conftest: environment that must be set before third-party imports.
+
+pytest loads plugins registered through entry points before it loads
+``tests/conftest.py``, and something in that chain imports litellm. litellm
+reads LITELLM_LOCAL_MODEL_COST_MAP at *its* import and, without it, fetches
+model_prices_and_context_window.json from raw.githubusercontent.com -- so the
+suite made an outbound request before a single test ran, and stalled or failed
+when offline. A rootdir conftest is loaded early enough to win that race.
+
+Anything that merely needs to be set before the tests themselves belongs in
+tests/conftest.py instead.
+"""
+
+import os
+
+os.environ.setdefault("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+os.environ.setdefault("HF_HUB_OFFLINE", "1")
@@ -183,7 +183,9 @@ banner. Key fragments:
    timeout-minutes: 15
    env:
      # Bump deliberately; the cache key and what we test both follow this.
-      OLLAMA_VERSION: v0.33.2
+      # No leading "v": install.sh passes ?version=$OLLAMA_VERSION and ollama.com
+      # redirects to GitHub tag v$version (vv0.33.2 → 404 if set to v0.33.2).
+      OLLAMA_VERSION: "0.33.2"
      OLLAMA_TEST_MODEL: qwen3:0.6b
    steps:
    - uses: actions/checkout@v4
@@ -204,9 +206,12 @@ banner. Key fragments:
      shell: bash
      run: |
        set -euo pipefail
-        curl -fsSL -o /tmp/ollama.tgz \
-          "https://github.com/ollama/ollama/releases/download/${OLLAMA_VERSION}/ollama-linux-amd64.tgz"
-        sudo tar -C /usr -xzf /tmp/ollama.tgz
+        # The official installer, VERSION-PINNED via OLLAMA_VERSION (which it
+        # honours). Since v0.33 the linux bundle ships as a .tar.zst with GPU
+        # runtimes bundled in, so a direct release-asset download is not a
+        # workable alternative; the installer handles the zst/tgz split.
+        sudo apt-get update -qq && sudo apt-get install -y -qq zstd
+        curl -fsSL https://ollama.com/install.sh | sh
        ollama --version
    - name: Start Ollama and pull model
      shell: bash
@@ -146,6 +146,10 @@ _LAZY_IMPORTS = {
    'TerminationReason': ('praisonaiagents.run_outcome', 'TerminationReason'),
    'termination_to_run_status': ('praisonaiagents.run_outcome', 'termination_to_run_status'),
    'validate_decision_string': ('praisonaiagents.run_outcome', 'validate_decision_string'),
+    'RunTerminal': ('praisonaiagents.run_outcome', 'RunTerminal'),
+    'merge_run_terminal': ('praisonaiagents.run_outcome', 'merge_run_terminal'),
+    'is_sticky': ('praisonaiagents.run_outcome', 'is_sticky'),
+    'collapse': ('praisonaiagents.run_outcome', 'collapse'),
    
    # Main display utilities (imports rich)
    'TaskOutput': ('praisonaiagents.main', 'TaskOutput'),
@@ -37,6 +37,22 @@ _cache_lock = threading.RLock()
 # Global module cache
 _module_cache: Dict[str, Any] = {}

+# Import roots of praisonaiagents' declared *core* runtime dependencies (see
+# pyproject.toml ``[project].dependencies``). A lazy import that fails because
+# one of these is missing/broken indicates a genuinely broken install, which
+# must be surfaced rather than masked as a graceful ``None``. Optional
+# integrations whose own optional SDKs are absent are NOT in this set, so they
+# keep the documented ``None`` fallback.
+_CORE_REQUIRED_DEPENDENCIES = frozenset({
+    "pydantic",
+    "rich",
+    "openai",
+    "posthog",
+    "aiohttp",
+    "yaml",       # PyYAML — core dep; import root differs from the package name
+    "croniter",
+})
+

 def lazy_import(
    module_path: str,
@@ -220,9 +236,25 @@ def create_lazy_getattr_with_fallback(
                _cache[name] = None
                return None
            except ImportError as exc:
-                # Optional module not available: preserve the graceful None
-                # fallback, but record why so real import failures remain
-                # diagnosable instead of surfacing later as NoneType errors.
+                # Distinguish "optional integration not installed" from
+                # "a required *core* dependency is broken". A missing/broken core
+                # dependency (e.g. ``pydantic``) must be surfaced loudly instead
+                # of being masked as ``None`` (which resurfaces later as a
+                # confusing ``NoneType`` error). But an optional integration whose
+                # own optional SDK is absent must keep the documented graceful
+                # ``None`` fallback — the failing module name there belongs to
+                # that optional SDK, not to us. We therefore only escalate when
+                # the failing module is one of praisonaiagents' declared *core*
+                # runtime dependencies; everything else stays a soft None.
+                failing = getattr(exc, "name", None)
+                failing_root = failing.split(".")[0] if failing else None
+                if failing_root in _CORE_REQUIRED_DEPENDENCIES:
+                    raise ImportError(
+                        f"Failed to load {name!r} because importing {module_path!r} "
+                        f"raised: {exc}. This usually means a required core dependency "
+                        f"({failing_root!r}) is missing or broken, not that {name!r} is "
+                        f"an optional feature you chose not to install."
+                    ) from exc
                _logger.debug(
                    "Lazy import of %r (%s.%s) failed: %s",
                    name, module_path, attr_name, exc,
@@ -28,8 +28,10 @@ def _configure_environment():
        "PYDANTIC_WARNINGS_ENABLED": "False",
    }
    
+    # Use setdefault so a value the host exported before launch is preserved
+    # (and not leaked/overwritten into subprocesses spawned afterwards).
    for key, value in env_vars.items():
-        os.environ[key] = value
+        os.environ.setdefault(key, value)


 # ========================================================================
@@ -48,7 +50,11 @@ def _get_all_noisy_loggers() -> List[str]:
        # Markdown
        "markdown_it", "rich.markdown",
        # System
-        "asyncio", "selector_events", "pydantic",
+        # NOTE: "asyncio" is intentionally NOT suppressed here — it is the host
+        # application's own logger, not a PraisonAI dependency's. Forcing it to
+        # CRITICAL hides unrelated host errors (e.g. "Task exception was never
+        # retrieved").
+        "selector_events", "pydantic",
        "praisonaiagents.telemetry.telemetry",
    ]

@@ -7,10 +7,7 @@ from praisonaiagents._logging import get_logger
 import asyncio
 import contextlib
 import threading
-import concurrent.futures
-import random
-import re
-from typing import List, Optional, Any, Dict, Union, Literal, TYPE_CHECKING, Callable, Generator
+from typing import List, Optional, Any, Dict, Union, Literal, TYPE_CHECKING, Callable
 from collections import OrderedDict
 import inspect

@@ -23,9 +20,12 @@ from .tool_execution import ToolExecutionMixin, BackoffPolicy
 from .chat_handler import ChatHandlerMixin
 from .session_manager import SessionManagerMixin
 from .async_safety import AsyncSafeState, DualLock
-# NOTE: UnifiedExecutionMixin is deprecated and unused by any production path
-# (Issue #2644). It is kept in the MRO for backward compatibility during the
-# deprecation cycle and will be removed afterwards.
+# NOTE: UnifiedExecutionMixin's *public* methods are deprecated and unused by
+# any production path (Issue #2644); it is kept in the MRO for backward
+# compatibility during the deprecation cycle and will be removed afterwards.
+# The one live helper it used to own (_run_async_in_sync_context) has been
+# relocated to async_safety.run_async_in_sync_context, so dropping this mixin
+# from the MRO is a true no-op.
 from .unified_execution_mixin import UnifiedExecutionMixin
 from .sandbox_mixin import SandboxMixin
 from .message_steering import SteeringMixin
@@ -880,6 +880,11 @@ class Agent(GoalLoopMixin, SteeringMixin, SandboxMixin, SkillReviewMixin, Unifie
                    "streaming moved into output=; use "
                    "output=OutputConfig(stream=True)."
                ),
+                # Same migration, different container: these two were flat
+                # kwargs before they were grouped into tool_config. Without a
+                # hint the user gets a bare "unexpected keyword argument" and
+                # has to search a 41-parameter signature for a name that is not
+                # in it.
                "tool_retry_policy": (
                    "tool retry moved into tool_config=; use "
                    "tool_config=ToolConfig(retry_policy=RetryPolicy(...))."
@@ -1365,12 +1370,17 @@ class Agent(GoalLoopMixin, SteeringMixin, SandboxMixin, SkillReviewMixin, Unifie
                array_mode=ArrayMode.PASSTHROUGH,
                default=None,
            )
+        # Per-agent event-hook registry (isolates BEFORE_TOOL/AFTER_AGENT/...
+        # hooks so they never fire on unrelated agents sharing the default
+        # registry). A bare HookRegistry passed as ``hooks=`` still works too.
+        _hooks_registry = None
        if _hooks_config is not None:
            if isinstance(_hooks_config, list):
                _hooks_list = _hooks_config
            elif isinstance(_hooks_config, HooksConfig):
                step_callback = _hooks_config.on_step
                _hooks_list = list(_hooks_config.middleware or [])
+                _hooks_registry = getattr(_hooks_config, "registry", None)
                # Route on_step / on_tool_call onto the SAME middleware chain
                # that already powers ``hooks=[...]`` (MiddlewareManager). Both
                # keys used to be stored-and-never-called, so the TypeError that
@@ -2093,7 +2103,11 @@ class Agent(GoalLoopMixin, SteeringMixin, SandboxMixin, SkillReviewMixin, Unifie
        # Lazy init for HookRunner and StreamEventEmitter (zero overhead when not used)
        self.__hook_runner = None  # Will be initialized on first access
        self.__stream_emitter = None  # Will be initialized on first access
-        self._hooks_registry_param = hooks  # Store for lazy init
+        # Prefer a per-agent registry declared via HooksConfig(registry=...) so
+        # this agent's event hooks stay isolated; else fall back to the raw
+        # ``hooks`` value (a bare HookRegistry still isolates, anything else
+        # resolves to the shared default registry in the _hook_runner property).
+        self._hooks_registry_param = _hooks_registry if _hooks_registry is not None else hooks
        
        # Handle llm= deprecation: model= is the preferred parameter name
        # llm= still works but shows deprecation warning
@@ -2296,10 +2310,22 @@ class Agent(GoalLoopMixin, SteeringMixin, SandboxMixin, SkillReviewMixin, Unifie
        # was sent to OpenAI under a model name that was really a repr -- the
        # opposite of what passing your own model means. Duck-typed on
        # ``get_response`` so any conforming backend works, not just ``LLM``.
+        # Detection must match what the tool loop actually invokes: the custom
+        # path calls ``llm_instance.get_response(**llm_kwargs)`` (and the async
+        # ``get_response_async``) with PraisonAI-internal kwargs -- tools,
+        # tool_choice, seed, cancel_token, steering_drain, and more. The
+        # ``chat``/``achat`` (LLMProviderProtocol) and
+        # ``chat_completion``/``achat_completion`` (UnifiedLLMProtocol) surfaces
+        # take a different signature and are NOT wired into that loop, so a
+        # backend exposing only those would be adopted here and then raise
+        # ``AttributeError`` on the first turn. Adopt only the surface the
+        # executor can drive; a translation adapter for the other protocols
+        # would be a heavy, unused implementation rather than a fix.
+        _MODEL_BACKEND_METHODS = ("get_response", "get_response_async")
        _is_model_instance = (
            llm is not None
            and not isinstance(llm, (str, dict))
-            and callable(getattr(llm, "get_response", None))
+            and any(callable(getattr(llm, m, None)) for m in _MODEL_BACKEND_METHODS)
        )
        if _is_model_instance:
            self._llm_instance = llm
@@ -2575,6 +2601,11 @@ class Agent(GoalLoopMixin, SteeringMixin, SandboxMixin, SkillReviewMixin, Unifie
        self._on_budget_exceeded = _on_budget_exceeded
        # Thread-safe cost/token tracking (Gap 1a fix)
        self._cost_lock = threading.Lock()
+        # Serializes the read-modify-write of _auto_save_last_index. Auto-save
+        # now runs via asyncio.to_thread, so two concurrent async runs sharing
+        # one Agent could otherwise read the same last-saved index in separate
+        # worker threads and each persist the same turns, duplicating history.
+        self._auto_save_lock = threading.Lock()
        self._total_cost = 0.0
        self._total_tokens_in = 0
        self._total_tokens_out = 0
@@ -2861,7 +2892,10 @@ Your Goal: {self.goal}
                    # Reuse the caller's manager as-is (no disk I/O side-effect).
                    _mgr = self._approval_permissions
                else:
-                    _mgr = PermissionManager(agent_name=self.name)
+                    _mgr = PermissionManager(
+                        agent_name=self.name,
+                        workspace_root=os.getcwd(),
+                    )
                    if isinstance(self._approval_permissions, dict):
                        _mgr.load_rules_from_config(self._approval_permissions)
                    elif isinstance(self._approval_permissions, (list, tuple)):
@@ -3004,6 +3038,7 @@ Your Goal: {self.goal}
        # turns on the same Agent instance don't corrupt each other's buffer.
        self._turn_tools_used = []
        self._turn_tools_lock = DualLock()
+        self._tool_executor_lock = DualLock()

        # Database persistence (lazy - no imports until used)
        self._db = db
@@ -3031,6 +3066,12 @@ Your Goal: {self.goal}
        # Agent-centric feature instances (lazy loaded for zero performance impact)
        self._auto_memory = auto_memory
        self._policy = policy
+        # Route enabled GUARDRAIL/POLICY/SKILL plugins into their typed
+        # subsystems now that guardrail (_setup_guardrail), _skills and _policy
+        # are all set. A plugin declared PluginType.GUARDRAIL/POLICY/SKILL thus
+        # participates in the GuardrailChain / PolicyEngine / SkillManager, not
+        # just a generic lifecycle hook. No-op when no such plugin is enabled.
+        self._merge_plugin_subsystems()
        self._output_style = output_style
        # Backward-compatible: `thinking_budget` property mirrors the legacy int
        # budget when supplied via the alias (Issue #4452); the unified effort is
@@ -3211,7 +3252,7 @@ Your Goal: {self.goal}
        for k, v in self.__dict__.items():
            if k in ("_Agent__cache_lock",):
                object.__setattr__(result, k, threading.RLock())
-            elif k == "_cost_lock":
+            elif k in ("_cost_lock", "_auto_save_lock"):
                object.__setattr__(result, k, threading.Lock())
            elif k == "_approvals_lock":
                object.__setattr__(result, k, asyncio.Lock())
@@ -3413,8 +3454,27 @@ Your Goal: {self.goal}
                        from ..config.feature_configs import MemoryConfig
                        clone_kwargs['memory'] = MemoryConfig(auto_save=value)
        
-        # Create new Agent instance
-        return self.__class__(**{k: v for k, v in clone_kwargs.items() if v is not None})
+        # Create new Agent instance. Plugin tools are shallow-copied above so
+        # callable identity and their ownership metadata can be carried over;
+        # otherwise the constructor's name-collision guard would treat a copied
+        # plugin tool as a caller-declared tool and disabling the plugin could
+        # leave the clone with an executable stale capability.
+        clone = self.__class__(**{k: v for k, v in clone_kwargs.items() if v is not None})
+        plugin_owners = getattr(self, "_plugin_tool_owners", None)
+        if plugin_owners:
+            inherited_owners = {
+                id(tool): plugin_owners[id(tool)]
+                for tool in (clone.tools if isinstance(clone.tools, (list, tuple)) else [])
+                if id(tool) in plugin_owners
+            }
+            if inherited_owners:
+                # Keep ownership recorded while the constructor merged any
+                # plugins that became enabled after the source was created.
+                # Those entries are needed to revoke newly attached tools.
+                current_owners = dict(getattr(clone, "_plugin_tool_owners", {}) or {})
+                current_owners.update(inherited_owners)
+                clone._plugin_tool_owners = current_owners
+        return clone

    @property
    def _telemetry(self):
@@ -4074,34 +4134,41 @@ Summary:"""
        * ``"safe"``  -> subprocess isolation, no tool access from the code.
        * ``"unsafe"`` -> same process, restricted builtins, and the
          ``code_tools_allow`` allow-list injected as callable tool proxies.
+        * ``"isolated"`` -> subprocess isolation AND the ``code_tools_allow``
+          allow-list, with each tool call bridged back to the parent under the
+          approval gate (the safe-and-tool-capable path).

-        Both expose one tool named ``execute_code``, which the approval registry
+        All expose one tool named ``execute_code``, which the approval registry
        classes as "critical", so it stays gated under every preset but "full".
        """
        from ..tools.python_tools import build_code_execution_tools

+        _tool_capable = ("unsafe", "isolated")
        code_tools = bool(getattr(exec_config, "code_tools", False))
        allowed = list(getattr(exec_config, "code_tools_allow", None) or [])
-        if code_tools and code_execution_mode != "unsafe":
+        if code_tools and code_execution_mode not in _tool_capable:
            import warnings
            warnings.warn(
-                "ExecutionConfig(code_tools=True) needs code_mode='unsafe': in "
-                "'safe' mode the code runs in a separate process and cannot "
-                "reach the agent's tools, so code_tools_allow is ignored.",
+                "ExecutionConfig(code_tools=True) needs code_mode='unsafe' or "
+                "'isolated': in 'safe' mode the code runs in a separate process "
+                "and cannot reach the agent's tools, so code_tools_allow is "
+                "ignored.",
                UserWarning,
                stacklevel=3,
            )
            code_tools = False

-        # In unsafe mode the allow-list must resolve against ONLY the tools this
-        # agent was granted, never the process-global registry (which can hold
-        # plugin/entry-point tools the agent was never given). Build a private
-        # registry from self.tools and pass it down so code-mode inherits the
-        # agent's exact tool boundary.
+        # In tool-capable modes the allow-list must resolve against ONLY the
+        # tools this agent was granted, never the process-global registry (which
+        # can hold plugin/entry-point tools the agent was never given). Build a
+        # private registry from self.tools and pass it down so code-mode
+        # inherits the agent's exact tool boundary.
        scoped_registry = None
-        if code_tools and code_execution_mode == "unsafe":
+        if code_tools and code_execution_mode in _tool_capable:
            from ..tools.registry import ToolRegistry
-            scoped_registry = ToolRegistry()
+            # discovery_enabled=False: an allow-listed name absent from THIS
+            # agent's tools must NOT resolve to an installed global plugin.
+            scoped_registry = ToolRegistry(discovery_enabled=False)
            for t in (self.tools or []):
                if callable(t) or hasattr(t, "name"):
                    try:
@@ -4109,12 +4176,44 @@ Summary:"""
                    except Exception:
                        pass

+        # Gate every code-mode tool call through the SAME PolicyEngine,
+        # permission deny-rules and per-tool ``input_guardrails`` the direct
+        # tool-call loop runs. The approval framework the proxy already honours
+        # does not cover these, so without this hook a model could reach an
+        # allow-listed tool from code and bypass a guardrail that blocks the
+        # same call on the normal path.
+        #
+        # Contract (mirrors the direct path in ``_check_tool_approval_sync``):
+        #   returns ``(denial_reason, arguments)`` where ``denial_reason`` is a
+        #   string to block or ``None`` to allow, and ``arguments`` is the
+        #   possibly-guardrail-rewritten kwargs dict the tool must run with. The
+        #   proxy uses the returned args so a sanitising guardrail's rewrite is
+        #   honoured (not discarded) exactly as on the direct path. BYPASS mode
+        #   skips the permission-deny gate, matching the direct contract, so a
+        #   call permitted directly is not rejected only in code mode. Only runs
+        #   in the tool-capable modes.
+        policy_hook = None
+        if code_tools and code_execution_mode in _tool_capable:
+            def policy_hook(name, arguments):
+                if not self._is_bypass_mode():
+                    deny = self._check_permission_manager_deny(name, arguments)
+                    if isinstance(deny, dict):
+                        return deny.get("error", "denied by permission policy"), arguments
+                verdict = self._check_tool_policy_and_guardrails(name, arguments)
+                if isinstance(verdict, dict):
+                    return verdict.get("error", "denied by policy"), arguments
+                # ``verdict`` is ``(None, rewritten_args)`` when allowed; hand the
+                # possibly-rewritten args back so the proxy dispatches with them.
+                _, rewritten = verdict
+                return None, rewritten
+
        # An unknown code_mode raises out of here rather than silently
        # producing nothing, which is the failure this whole change is about.
        new_tools = build_code_execution_tools(
            code_mode=code_execution_mode,
            allowed_tools=allowed if code_tools else [],
            registry=scoped_registry,
+            policy_hook=policy_hook,
        )

        existing = {getattr(t, "__name__", None) for t in (self.tools or [])}
@@ -5292,14 +5391,19 @@ Summary:"""
                        f"response_len={len(response_str)}"
                    )
                
-                # Auto-save session after each async iteration (memory integration)
-                self._auto_save_session()
+                # Auto-save session after each async iteration (memory integration).
+                # Offloaded to a worker thread so the synchronous FileLock/JSON
+                # write never stalls the shared event loop (mirrors the goal-
+                # completion-judge offload below).
+                await asyncio.to_thread(self._auto_save_session)

                # ─────────────────────────────────────────────────────────────
                # GOAL COMPLETION JUDGE (opt-in): independent acceptance-criteria
                # gate. No-op when no goal loop is active (behaviour unchanged).
+                # Offloaded to a worker thread (blocking judge LLM call / hooks)
+                # so it never stalls the shared event loop.
                # ─────────────────────────────────────────────────────────────
-                _goal_gate = self._goal_gate(response_str)
+                _goal_gate = await self._goal_gate_async(response_str)
                if _goal_gate is not None:
                    _outcome, _reason = _goal_gate
                    if _outcome == "done":
@@ -5431,8 +5535,8 @@ Summary:"""
                # Yield control to allow other async tasks to run
                await asyncio.sleep(0)
            
-            # Final auto-save before returning
-            self._auto_save_session()
+            # Final auto-save before returning (offloaded off the event loop).
+            await asyncio.to_thread(self._auto_save_session)
            
            # Max iterations reached
            return AutonomyResult(
@@ -5848,6 +5952,8 @@ Summary:"""
            List of available tools
        """
        if not self.plan_mode:
+            if getattr(self, "_plugin_tool_owners", None):
+                return [tool for tool in self.tools if self._is_plugin_tool_active(tool)]
            return self.tools
            
        # Filter to read-only tools only
@@ -5855,6 +5961,8 @@ Summary:"""
        
        filtered_tools = []
        for tool in self.tools:
+            if not self._is_plugin_tool_active(tool):
+                continue
            tool_name = getattr(tool, '__name__', str(tool)).lower()
            
            # Check if tool is in restricted list
@@ -7015,6 +7123,25 @@ Answer:"""
        """The most recent rejection message a guardrail sent back to the model."""
        return getattr(self, "_last_guardrail_error", None)

+    def _guardrail_blocked_message(self, error=None):
+        """Fixed, user-facing string returned when an output guardrail blocks a run.
+
+        A blocked response must never surface the rejected (potentially unsafe)
+        model text, but it must also never be ``None`` - callers expect a
+        ``str`` and would otherwise crash on ``.strip()`` (issue #5596, FR-001).
+
+        The returned string is intentionally **fixed** and carries no diagnostic
+        detail: a guardrail's own rejection reason (especially an LLM validator's
+        free-text) can echo the blocked output, so appending it would leak the
+        very content the guardrail rejected. The reason is logged instead, where
+        it stays auditable without reaching the caller. The rejection is also
+        exposed programmatically via :attr:`last_guardrail_error`.
+        """
+        reason = str(error) if error else getattr(self, "_last_guardrail_error", None)
+        if reason:
+            logging.info(f"Agent {self.name}: guardrail block reason: {reason}")
+        return "I'm sorry, but I can't share that response because it did not pass the safety guardrail."
+
    def _process_guardrail(self, task_output):
        """Process the guardrail validation for a task output.
        
@@ -7030,8 +7157,16 @@ Answer:"""
            return GuardrailResult(success=True, result=task_output)

        try:
-            # Call the guardrail function
-            result = self._guardrail_fn(task_output)
+            # Call the guardrail function, awaiting it if the caller supplied an
+            # async def. Calling a coroutine function synchronously would only
+            # build a coroutine object (never running its body), which then
+            # crashes GuardrailResult.from_tuple and is misreported as a
+            # validation failure.
+            if asyncio.iscoroutinefunction(self._guardrail_fn):
+                from ..utils.async_bridge import run_coroutine_from_any_context
+                result = run_coroutine_from_any_context(self._guardrail_fn(task_output))
+            else:
+                result = self._guardrail_fn(task_output)

            # Convert the result to a GuardrailResult
            return GuardrailResult.from_tuple(result)
@@ -7329,10 +7464,25 @@ Answer:"""
            return "empty"
        # Create a simple hash based on tool names
        tool_names = []
+        try:
+            from ..tools.hosted import is_hosted_tool
+        except ImportError:
+            is_hosted_tool = lambda _tool: False
+
        for tool in tools:
            if callable(tool) and hasattr(tool, '__name__'):
                tool_names.append(tool.__name__)
-            elif isinstance(tool, dict) and 'function' in tool and 'name' in tool['function']:
+            elif isinstance(tool, dict) and is_hosted_tool(tool):
+                try:
+                    hosted_key = json.dumps(tool, sort_keys=True)
+                except (TypeError, ValueError):
+                    hosted_key = str(id(tool))
+                tool_names.append(f"hosted:{hosted_key}")
+            elif (
+                isinstance(tool, dict)
+                and isinstance(tool.get('function'), dict)
+                and tool['function'].get('name')
+            ):
                tool_names.append(tool['function']['name'])
            elif isinstance(tool, str):
                tool_names.append(tool)
@@ -7928,16 +8078,28 @@ Answer:"""
        except Exception as e:
            logger.warning(f"Task cleanup failed: {e}")

-        # ThreadPoolExecutor cleanup
+        # ThreadPoolExecutor cleanup. Detach the executor under the same lock
+        # that guards timeout retirement/replacement (see tool_execution.py) so
+        # cleanup and a concurrent timeout can never act on different executor
+        # generations. Shut down outside the critical section to avoid holding
+        # the lock across a blocking call.
        try:
-            if hasattr(self, '_tool_executor') and self._tool_executor:
+            executor = None
+            lock = getattr(self, '_tool_executor_lock', None)
+            if lock is not None:
+                with lock.sync():
+                    executor = getattr(self, '_tool_executor', None)
+                    self._tool_executor = None
+            elif hasattr(self, '_tool_executor'):
+                executor = self._tool_executor
+                self._tool_executor = None
+            if executor is not None:
                # Use cancel_futures only if supported (Python 3.9+)
                import sys
                if sys.version_info >= (3, 9):
-                    self._tool_executor.shutdown(wait=False, cancel_futures=True)
+                    executor.shutdown(wait=False, cancel_futures=True)
                else:
-                    self._tool_executor.shutdown(wait=False)
-                delattr(self, '_tool_executor')
+                    executor.shutdown(wait=False)
        except Exception as e:
            logger.warning(f"ThreadPoolExecutor cleanup failed: {e}")

@@ -7971,6 +8133,41 @@ Answer:"""
                elif hasattr(self.memory, 'close_connections'):
                    self.memory.close_connections()
            
+            # LLM client cleanup - release live client pools asynchronously.
+            # Mirror close()'s targets but never touch the ``llm_instance``
+            # property (it would lazily *create* a client just to close it);
+            # only tear down an already-materialised ``_llm_instance``.
+            try:
+                llm_instance = getattr(self, '_llm_instance', None)
+                if llm_instance is not None:
+                    aclose = getattr(llm_instance, 'aclose', None)
+                    if aclose is not None:
+                        try:
+                            if asyncio.iscoroutinefunction(aclose):
+                                await aclose()
+                            else:
+                                aclose()
+                        except Exception:
+                            close = getattr(llm_instance, 'close', None)
+                            if callable(close):
+                                close()
+                    else:
+                        close = getattr(llm_instance, 'close', None)
+                        if callable(close):
+                            close()
+
+                openai_client = getattr(self, '_Agent__openai_client', None)
+                if openai_client is not None and hasattr(openai_client, 'close'):
+                    openai_client.close()
+
+                llm = getattr(self, 'llm', None)
+                if llm and not isinstance(llm, str):
+                    llm_client = getattr(llm, '_client', None)
+                    if llm_client and hasattr(llm_client, 'close'):
+                        llm_client.close()
+            except Exception as e:
+                logger.warning(f"LLM client cleanup failed: {e}")
+
            # Close MCP clients passed via tools=[MCP(...)]
            # (mirrors remove_mcp_server()'s best-effort shutdown)
            if isinstance(self.tools, list):
@@ -8005,22 +8202,35 @@ Answer:"""
                    except asyncio.CancelledError:
                        pass

-            # ThreadPoolExecutor cleanup (async-safe)
-            if hasattr(self, '_tool_executor') and self._tool_executor:
+            # ThreadPoolExecutor cleanup (async-safe). Detach the executor under
+            # the lock that guards timeout retirement/replacement so cleanup and
+            # a concurrent timeout can never act on different executor
+            # generations. Capture the reference before scheduling the shutdown
+            # so the deferred call cannot resolve a replacement pool created by
+            # another call.
+            executor = None
+            lock = getattr(self, '_tool_executor_lock', None)
+            if lock is not None:
+                async with lock.async_lock():
+                    executor = getattr(self, '_tool_executor', None)
+                    self._tool_executor = None
+            elif hasattr(self, '_tool_executor'):
+                executor = self._tool_executor
+                self._tool_executor = None
+            if executor is not None:
                import sys
                loop = asyncio.get_running_loop()
                # Use run_in_executor to avoid blocking the event loop
                if sys.version_info >= (3, 9):
                    await loop.run_in_executor(
-                        None, 
-                        lambda: self._tool_executor.shutdown(wait=False, cancel_futures=True)
+                        None,
+                        lambda: executor.shutdown(wait=False, cancel_futures=True)
                    )
                else:
                    await loop.run_in_executor(
-                        None, 
-                        lambda: self._tool_executor.shutdown(wait=False)
+                        None,
+                        lambda: executor.shutdown(wait=False)
                    )
-                delattr(self, '_tool_executor')

            # Approval scope cleanup (see close()).
            self._release_approval_scope()
@@ -12,6 +12,38 @@ from contextlib import contextmanager, asynccontextmanager
 from weakref import WeakKeyDictionary


+def run_async_in_sync_context(coro):
+    """
+    Run an async coroutine to completion from a sync context with proper
+    event loop handling.
+
+    Handles the common cases:
+    1. No event loop exists - use asyncio.run()
+    2. Event loop already running - avoid deadlock by running the coroutine
+       to completion in a dedicated thread with its own event loop.
+
+    This is a live production helper on the sync tool-calling path (see
+    ``tool_execution.py``): it bridges ``async def`` tool bodies into the sync
+    tool loop, otherwise a bare un-awaited coroutine would be handed to the
+    model as the tool result and the tool body would never run (silent data
+    loss).
+
+    Delegates to the single canonical bridge
+    (:func:`utils.async_bridge.run_coroutine_from_any_context`) so all sync→async
+    call sites share one implementation: the caller's ``contextvars`` (trace /
+    session / approval context) are carried into the coroutine, and the running
+    loop is handled on a dedicated worker thread rather than raising.
+
+    ``timeout=None`` is passed deliberately: this helper backs the sync
+    tool-calling path, where a tool's own timeout policy (or lack of one) already
+    governs how long a tool may run. Imposing the bridge's default 5-minute
+    deadline here would silently cancel long-running tools that previously had no
+    limit — a backward-incompatible regression — so the bridge adds none.
+    """
+    from ..utils.async_bridge import run_coroutine_from_any_context
+    return run_coroutine_from_any_context(coro, timeout=None)
+
+
 class DualLock:
    """
    A dual-lock abstraction that automatically selects threading.Lock or asyncio.Lock
@@ -37,9 +69,24 @@ class DualLock:
    """
    
    def __init__(self):
-        """Initialize with separate threading and asyncio locks."""
-        self._thread_lock = threading.RLock()  # Re-entrant lock to handle nested acquisitions
-        self._async_locks = WeakKeyDictionary()  # Per-event-loop async locks
+        """Initialize with a single underlying mutex shared by both call styles.
+
+        A single ``threading.Lock`` backs both ``sync()`` and ``async_lock()``
+        so that holding one genuinely blocks the other — mutual exclusion is
+        real across sync and async callers of the same instance. A plain
+        ``Lock`` (not ``RLock``) is used because the async path acquires on a
+        worker thread but releases on the event-loop thread, which ``RLock``'s
+        thread-ownership semantics forbid. Same-thread ``sync()`` reentrancy is
+        preserved explicitly via an owner/depth guard below. The per-event-loop
+        ``asyncio.Lock`` only serialises coroutines on the same loop so they
+        queue for the underlying mutex cooperatively instead of every coroutine
+        racing straight into the blocking acquire.
+        """
+        self._thread_lock = threading.Lock()  # Shared mutex (cross-thread release ok)
+        self._owner_lock = threading.Lock()    # Guards the reentrancy bookkeeping
+        self._owner_thread: Any = None         # Thread id currently holding for sync reentrancy
+        self._owner_depth = 0
+        self._async_locks = WeakKeyDictionary()  # Per-event-loop async gate

    def __deepcopy__(self, memo):
        """Return an unlocked primitive with no event-loop state sharing."""
@@ -49,9 +96,31 @@ class DualLock:
    
    @contextmanager
    def sync(self):
-        """Acquire lock in synchronous context using threading.Lock."""
-        with self._thread_lock:
+        """Acquire lock in synchronous context using the shared mutex.
+
+        Re-entrant for the same thread: a nested ``sync()`` on the thread that
+        already holds the mutex does not re-acquire (and cannot self-deadlock).
+        """
+        me = threading.get_ident()
+        with self._owner_lock:
+            reentrant = self._owner_thread == me and self._owner_depth > 0
+            if reentrant:
+                self._owner_depth += 1
+        if not reentrant:
+            self._thread_lock.acquire()
+            with self._owner_lock:
+                self._owner_thread = me
+                self._owner_depth = 1
+        try:
            yield
+        finally:
+            with self._owner_lock:
+                self._owner_depth -= 1
+                released = self._owner_depth == 0
+                if released:
+                    self._owner_thread = None
+            if released:
+                self._thread_lock.release()
            
    def _get_async_lock(self):
        """Get or create an asyncio.Lock for the current event loop."""
@@ -64,16 +133,64 @@ class DualLock:
            self._async_locks[loop] = asyncio.Lock()
        return self._async_locks[loop]

+    async def _acquire_shared_mutex(self, loop):
+        """Acquire the shared mutex off-loop, releasing it if we are cancelled.
+
+        ``run_in_executor`` can complete (mutex acquired) after the awaiting
+        task is cancelled. If we let the ``CancelledError`` propagate without
+        checking, the worker would strand the mutex forever. We therefore wait
+        on the acquisition, and on cancellation ensure the mutex is released
+        once the worker finishes so no caller waits indefinitely.
+        """
+        future = loop.run_in_executor(None, self._thread_lock.acquire)
+        try:
+            await asyncio.shield(future)
+        except asyncio.CancelledError:
+            # The worker may still be mid-acquire; wait for it to settle, then
+            # release if it managed to take the mutex so it is never orphaned.
+            try:
+                acquired = await future
+            except asyncio.CancelledError:
+                acquired = False
+            if acquired:
+                self._thread_lock.release()
+            raise
+
    @asynccontextmanager
    async def async_lock(self):
-        """Acquire lock in asynchronous context using asyncio.Lock.
-        
-        Uses a per-event-loop asyncio.Lock to ensure proper async coordination
-        without blocking the event loop or violating thread ownership semantics.
+        """Acquire lock in asynchronous context using the shared mutex.
+
+        Backs onto the *same* ``threading.Lock`` as ``sync()`` so a concurrent
+        sync caller is genuinely excluded (and vice-versa). The blocking
+        ``acquire`` is offloaded to a worker thread so the event loop is never
+        blocked while waiting; the matching ``release`` (allowed cross-thread on
+        a plain ``Lock``) happens on the loop thread. A per-event-loop
+        ``asyncio.Lock`` serialises coroutines on this loop first so they don't
+        all spin up executor hops at once.
+
+        While the mutex is held, the *event-loop thread* is registered as the
+        sync owner so a nested ``with lock.sync()`` on that thread (used by the
+        async handoff seeding path) is treated as re-entrant instead of
+        self-deadlocking against the mutex we already hold.
        """
-        async_lock = self._get_async_lock()
-        async with async_lock:
-            yield
+        loop = asyncio.get_running_loop()
+        async_gate = self._get_async_lock()
+        async with async_gate:
+            await self._acquire_shared_mutex(loop)
+            me = threading.get_ident()
+            with self._owner_lock:
+                self._owner_thread = me
+                self._owner_depth = 1
+            try:
+                yield
+            finally:
+                with self._owner_lock:
+                    self._owner_depth -= 1
+                    released = self._owner_depth == 0
+                    if released:
+                        self._owner_thread = None
+                if released:
+                    self._thread_lock.release()
            
    def is_async_context(self) -> bool:
        """Check if we're currently in an async context."""
@@ -147,6 +264,11 @@ class AsyncSafeState:
    @asynccontextmanager
    async def async_lock(self):
        """Acquire lock in async context."""
+        # Fail fast BEFORE blocking on the shared mutex: an in-progress
+        # synchronous deepcopy holds the mutex for the whole copy, so waiting on
+        # it here would block until the copy finished instead of surfacing the
+        # documented RuntimeError (and could stall the loop indefinitely).
+        self._raise_if_copying()
        async with self._lock.async_lock():
            self._begin_async_access()
            try:
@@ -154,6 +276,14 @@ class AsyncSafeState:
            finally:
                self._end_async_access()

+    def _raise_if_copying(self) -> None:
+        """Reject entry while a synchronous deepcopy of this state is active."""
+        with self._activity_lock:
+            if self._copying:
+                raise RuntimeError(
+                    "Cannot access AsyncSafeState during synchronous deepcopy"
+                )
+
    def _begin_async_access(self) -> None:
        """Register async access unless a synchronous copy is in progress."""
        with self._activity_lock:
@@ -170,20 +300,32 @@ class AsyncSafeState:
            
    def __enter__(self):
        """Support for synchronous context manager protocol (backward compatibility)."""
-        self._lock._thread_lock.acquire()
+        self._sync_guard = self._lock.sync()
+        self._sync_guard.__enter__()
        return self.value
        
    def __exit__(self, exc_type, exc_val, exc_tb):
        """Support for synchronous context manager protocol (backward compatibility)."""
-        self._lock._thread_lock.release()
-        return None
+        return self._sync_guard.__exit__(exc_type, exc_val, exc_tb)
        
    async def __aenter__(self):
        """Support for asynchronous context manager protocol."""
+        # Fail fast before blocking on the mutex an in-progress deepcopy holds.
+        self._raise_if_copying()
+        loop = asyncio.get_running_loop()
        async_lock = self._lock._get_async_lock()
        await async_lock.acquire()
        try:
-            self._begin_async_access()
+            # Hold the SAME underlying mutex used by the sync path so async and
+            # sync accessors are mutually exclusive (matches DualLock.async_lock).
+            # Cancellation-safe acquire so the mutex is never stranded if the
+            # awaiting task is cancelled mid-acquisition.
+            await self._lock._acquire_shared_mutex(loop)
+            try:
+                self._begin_async_access()
+            except Exception:
+                self._lock._thread_lock.release()
+                raise
        except Exception:
            async_lock.release()
            raise
@@ -195,7 +337,10 @@ class AsyncSafeState:
        try:
            self._end_async_access()
        finally:
-            async_lock.release()
+            try:
+                self._lock._thread_lock.release()
+            finally:
+                async_lock.release()
        return None
            
    def get(self) -> Any:
@@ -751,6 +751,43 @@ Your Goal: {self.goal}"""
        
        return None

+    def _coerce_structured_output(self, response_text, schema_model):
+        """Parse a string response into the requested Pydantic model.
+
+        When ``output_pydantic``/``output_json`` is a Pydantic model, callers
+        expect a validated instance rather than a raw string. This tolerates
+        markdown code fences around the JSON payload.
+
+        Returns the parsed model instance on success, or the original
+        ``response_text`` unchanged if parsing is not possible (preserves
+        backward compatibility for non-Pydantic schemas or malformed output).
+        """
+        if not schema_model or not isinstance(response_text, str):
+            return response_text
+        if not hasattr(schema_model, 'model_validate_json'):
+            return response_text
+
+        candidate = response_text.strip()
+        try:
+            return schema_model.model_validate_json(candidate)
+        except Exception:
+            pass
+
+        # Strip a surrounding markdown code fence (```json ... ```) if present
+        fenced = re.search(r"```(?:json)?\s*(.*?)\s*```", candidate, re.DOTALL)
+        if fenced:
+            try:
+                return schema_model.model_validate_json(fenced.group(1).strip())
+            except Exception:
+                pass
+
+        logging.warning(
+            f"Agent {self.name}: output_pydantic={getattr(schema_model, '__name__', schema_model)} "
+            f"was requested but the response could not be parsed into the model; "
+            f"returning raw string."
+        )
+        return response_text
+
    def _supports_native_structured_output(self):
        """Check if current model supports native structured output via response_format.
        
@@ -778,6 +815,35 @@ Your Goal: {self.goal}"""
            )
            return False

+    def _load_session_history(self):
+        """Persisted session history for this agent, or [] when unavailable.
+
+        Extracted so BOTH branches of _build_messages can use it. The injection
+        used to live only in the manual-message branch, so an agent configured
+        with memory="history" silently dropped its persisted history whenever
+        _openai_client was set -- which is the default path for OpenAI models.
+        """
+        if not self._history_enabled or self._session_store is None:
+            return []
+        try:
+            # Prefer compacted working history (summary + tail) on resume when a
+            # compaction checkpoint exists (Issue #2741); falls back to raw chat
+            # history for backward compatibility.
+            if hasattr(self._session_store, "get_working_history"):
+                session_history = self._session_store.get_working_history(
+                    self._history_session_id,
+                    max_messages=self._history_limit
+                )
+            else:
+                session_history = self._session_store.get_chat_history(
+                    self._history_session_id,
+                    max_messages=self._history_limit
+                )
+            return list(session_history) if session_history else []
+        except Exception as e:
+            logging.debug(f"Failed to load session history: {e}")
+            return []
+
    def _build_messages(self, prompt, temperature=1.0, output_json=None, output_pydantic=None, tools=None, use_native_format=False, restore_durable=True, memory_prefetch_context: str = ""):
        """Build messages list for chat completion.
        
@@ -803,7 +869,7 @@ Your Goal: {self.goal}"""
                    tools=tools,
                    memory_prefetch_context=memory_prefetch_context,
                ),
-                chat_history=self.chat_history,
+                chat_history=self._load_session_history() + list(self.chat_history or []),
                output_json=None if use_native_format else output_json,
                output_pydantic=None if use_native_format else output_pydantic
            )
@@ -817,25 +883,7 @@ Your Goal: {self.goal}"""
                messages.append({"role": "system", "content": system_prompt})
            
            # Inject session history if enabled (from persistent storage)
-            if self._history_enabled and self._session_store is not None:
-                try:
-                    # Prefer compacted working history (summary + tail) on
-                    # resume when a compaction checkpoint exists (Issue #2741);
-                    # falls back to raw chat history for backward compatibility.
-                    if hasattr(self._session_store, "get_working_history"):
-                        session_history = self._session_store.get_working_history(
-                            self._history_session_id,
-                            max_messages=self._history_limit
-                        )
-                    else:
-                        session_history = self._session_store.get_chat_history(
-                            self._history_session_id,
-                            max_messages=self._history_limit
-                        )
-                    if session_history:
-                        messages.extend(session_history)
-                except Exception as e:
-                    logging.debug(f"Failed to load session history: {e}")
+            messages.extend(self._load_session_history())
            
            # Add in-memory chat history (current conversation)
            if self.chat_history:
@@ -894,6 +942,12 @@ Your Goal: {self.goal}"""
        elif isinstance(tools, list) and len(tools) == 0:
            # Explicit empty list - return immediately to enforce boundary
            return []
+
+        # Plugin tools are copied into the Agent's local tool list at
+        # construction time. Re-check ownership before advertising them so a
+        # later plugin disable/unregister revokes the advertised capability.
+        if getattr(self, "_plugin_tool_owners", None) and isinstance(tools, (list, tuple)):
+            tools = [tool for tool in tools if self._is_plugin_tool_active(tool)]
        
        if not tools:
            return []
@@ -909,9 +963,18 @@ Your Goal: {self.goal}"""
            return cached_tools
            
        formatted_tools = []
+        # Provider-hosted tool specs have no local callable; the provider
+        # executes them after this formatter forwards the allowlisted dict.
+        try:
+            from ..tools.hosted import is_hosted_tool
+        except ImportError:
+            is_hosted_tool = lambda _tool: False
+
        for tool in tools:
+            if isinstance(tool, dict) and is_hosted_tool(tool):
+                formatted_tools.append(tool)
            # Handle pre-formatted OpenAI tools
-            if isinstance(tool, dict) and tool.get('type') == 'function':
+            elif isinstance(tool, dict) and tool.get('type') == 'function':
                # Validate nested dictionary structure before accessing
                if 'function' in tool and isinstance(tool['function'], dict) and 'name' in tool['function']:
                    formatted_tools.append(tool)
@@ -920,7 +983,9 @@ Your Goal: {self.goal}"""
            # Handle lists of tools
            elif isinstance(tool, list):
                for subtool in tool:
-                    if isinstance(subtool, dict) and subtool.get('type') == 'function':
+                    if isinstance(subtool, dict) and is_hosted_tool(subtool):
+                        formatted_tools.append(subtool)
+                    elif isinstance(subtool, dict) and subtool.get('type') == 'function':
                        # Validate nested dictionary structure before accessing
                        if 'function' in subtool and isinstance(subtool['function'], dict) and 'name' in subtool['function']:
                            formatted_tools.append(subtool)
@@ -1403,15 +1468,8 @@ Your Goal: {self.goal}"""
        from ..hooks import HookEvent as _HookEvent
        import logging
        
-        # Map policy strategy to compactor strategy
-        strategy_map = {
-            "truncate": LegacyStrategy.TRUNCATE,
-            "summarise": LegacyStrategy.SUMMARIZE,
-            "drop_oldest_tools": LegacyStrategy.PRUNE,
-            "sliding_window": LegacyStrategy.SLIDING,
-        }
-        
-        compactor.strategy = strategy_map.get(policy.strategy.value, LegacyStrategy.PRUNE)
+        # Map policy strategy to compactor strategy (single source of truth)
+        compactor.strategy = LegacyStrategy.from_policy_value(policy.strategy.value)
        
        # Execute hooks
        try:
@@ -1596,15 +1654,8 @@ Your Goal: {self.goal}"""
        from ..hooks import HookEvent as _HookEvent
        import logging
        
-        # Map policy strategy to compactor strategy
-        strategy_map = {
-            "truncate": LegacyStrategy.TRUNCATE,
-            "summarise": LegacyStrategy.SUMMARIZE,
-            "drop_oldest_tools": LegacyStrategy.PRUNE,
-            "sliding_window": LegacyStrategy.SLIDING,
-        }
-        
-        compactor.strategy = strategy_map.get(policy.strategy.value, LegacyStrategy.PRUNE)
+        # Map policy strategy to compactor strategy (single source of truth)
+        compactor.strategy = LegacyStrategy.from_policy_value(policy.strategy.value)
        
        # Execute hooks
        try:
@@ -3345,6 +3396,36 @@ Your Goal: {self.goal}"""
                    # Append formatted knowledge to the prompt
                    prompt = f"{prompt}\n\n{formatted_context}"

+                    # Sync llm_prompt with the retrieved context so the RAG block
+                    # actually reaches the model. llm_prompt (not prompt) is what
+                    # gets sent to the LLM in both the custom-LLM and OpenAI paths
+                    # below; without this the retrieved knowledge is appended to
+                    # `prompt` but never delivered, so the model answers from
+                    # parametric memory only. Append (rather than reassign) to
+                    # preserve any response-template instruction already baked into
+                    # llm_prompt.
+                    if isinstance(llm_prompt, str):
+                        llm_prompt = f"{llm_prompt}\n\n{formatted_context}"
+                    elif isinstance(llm_prompt, list):
+                        # Multimodal prompt: append the retrieved context to the
+                        # last text part so attachment-bearing turns still receive
+                        # the RAG block. Copy dicts before mutating to avoid
+                        # aliasing the caller's attachment structures; if no text
+                        # part exists, add one.
+                        appended = False
+                        for i in range(len(llm_prompt) - 1, -1, -1):
+                            item = llm_prompt[i]
+                            if isinstance(item, dict) and item.get("type") == "text":
+                                updated = dict(item)
+                                updated["text"] = f"{updated.get('text', '')}\n\n{formatted_context}"
+                                llm_prompt[i] = updated
+                                appended = True
+                                break
+                        if not appended:
+                            llm_prompt = list(llm_prompt) + [
+                                {"type": "text", "text": formatted_context}
+                            ]
+
        if self._using_custom_llm:
            # Track messages THIS turn appends so a failure rolls back only our
            # own messages, never a concurrent turn's (see memory_mixin).
@@ -3518,6 +3599,13 @@ Your Goal: {self.goal}"""
                    # Apply guardrail validation for custom LLM response
                    try:
                        validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools, task_name, task_description, task_id, cancel_token=cancel_token, messages=processed_history)
+                        # Coerce into the requested Pydantic model so callers
+                        # get a validated instance, not a raw JSON string.
+                        # Only output_pydantic changes the return type; output_json
+                        # must stay a string for backward-compatible downstream
+                        # JSON parsing (task parsing, workflow results).
+                        if output_pydantic:
+                            validated_response = self._coerce_structured_output(validated_response, output_pydantic)
                        # Execute callback and display after validation
                        self._execute_callback_and_display(prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
                        return self._trigger_after_agent_hook(prompt, validated_response, start_time)
@@ -3525,7 +3613,7 @@ Your Goal: {self.goal}"""
                        logging.error(f"Agent {self.name}: Guardrail validation failed for custom LLM: {e}")
                        # Rollback chat history on guardrail failure
                        self._rollback_chat_history_to(chat_history_length)
-                        return None
+                        return self._guardrail_blocked_message(e)
                except ToolExecutionError:
                    raise
                except Exception as e:
@@ -3562,41 +3650,48 @@ Your Goal: {self.goal}"""
            )
            

+            # Track messages THIS turn appends so a failure rolls back only our
+            # own messages, never a concurrent turn's (see memory_mixin). The
+            # default (OpenAI) path used to skip this, so its positional rollback
+            # ``del chat_history[n:]`` erased a concurrent turn's interleaved
+            # messages and corrupted the shared transcript.
+            _turn_token = self._begin_turn_tracking()
+
            # Store chat history length for potential rollback
            chat_history_length = len(self.chat_history)
-            
-            # Normalize original_prompt for consistent chat history storage
-            normalized_content = original_prompt
-            if isinstance(original_prompt, list):
-                # Extract text from multimodal prompts
-                normalized_content = next((item["text"] for item in original_prompt if item.get("type") == "text"), "")
-            
-            # Prevent duplicate messages
-            if not (self.chat_history and 
-                    self.chat_history[-1].get("role") == "user" and 
-                    self.chat_history[-1].get("content") == normalized_content):
-                # Add user message to chat history BEFORE LLM call so handoffs can access it
-                self._append_to_chat_history({"role": "user", "content": normalized_content})
-                # Persist user message to DB (OpenAI path)
-                self._persist_message("user", normalized_content)

-            reflection_count = 0
-            start_time = time.time()
-            
-            # Apply context management before LLM call (auto-compaction)
-            # Zero overhead when context=False
-            system_prompt_content = messages[0].get("content", "") if messages and messages[0].get("role") == "system" else ""
-            processed_messages, context_result = self._apply_context_management(
-                messages=messages,
-                system_prompt=system_prompt_content,
-                tools=tools,
-            )
-            # Use processed messages for the LLM call
-            messages = processed_messages
-            
-            
-            # Wrap entire while loop in try-except for rollback on any failure
+            # Wrap entire turn in try/finally so per-turn tracking is always
+            # ended (the finally reads the ownership list during rollback).
            try:
+                # Normalize original_prompt for consistent chat history storage
+                normalized_content = original_prompt
+                if isinstance(original_prompt, list):
+                    # Extract text from multimodal prompts
+                    normalized_content = next((item["text"] for item in original_prompt if item.get("type") == "text"), "")
+
+                # Add user message to chat history BEFORE LLM call so handoffs can
+                # access it. Use atomic check-then-act to prevent TOCTOU races: an
+                # unatomic read of chat_history[-1] followed by a separate append
+                # lets a concurrent turn's user message be mistaken for this turn's
+                # own duplicate, silently dropping this user turn from the record.
+                if self._add_to_chat_history_if_not_duplicate("user", normalized_content):
+                    # Persist user message to DB (OpenAI path)
+                    self._persist_message("user", normalized_content)
+
+                reflection_count = 0
+                start_time = time.time()
+
+                # Apply context management before LLM call (auto-compaction)
+                # Zero overhead when context=False
+                system_prompt_content = messages[0].get("content", "") if messages and messages[0].get("role") == "system" else ""
+                processed_messages, context_result = self._apply_context_management(
+                    messages=messages,
+                    system_prompt=system_prompt_content,
+                    tools=tools,
+                )
+                # Use processed messages for the LLM call
+                messages = processed_messages
+
                while True:
                    try:
                        if self.verbose:
@@ -3642,6 +3737,11 @@ Your Goal: {self.goal}"""
                            # Apply guardrail validation even for JSON output
                            try:
                                validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id, cancel_token=cancel_token, messages=messages)
+                                # Coerce into the requested Pydantic model so callers
+                                # get a validated instance, not a raw JSON string.
+                                # output_json intentionally stays a string (backward compat).
+                                if output_pydantic:
+                                    validated_response = self._coerce_structured_output(validated_response, output_pydantic)
                                # Execute callback after validation
                                self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
                                return self._trigger_after_agent_hook(original_prompt, validated_response, start_time)
@@ -3649,7 +3749,7 @@ Your Goal: {self.goal}"""
                                logging.error(f"Agent {self.name}: Guardrail validation failed for JSON output: {e}")
                                # Rollback chat history on guardrail failure
                                self._rollback_chat_history_to(chat_history_length)
-                                return None
+                                return self._guardrail_blocked_message(e)

                        if not self.self_reflect:
                            # User message already added before LLM call via _build_messages
@@ -3670,7 +3770,7 @@ Your Goal: {self.goal}"""
                                    logging.error(f"Agent {self.name}: Guardrail validation failed for reasoning content: {e}")
                                    # Rollback chat history on guardrail failure
                                    self._rollback_chat_history_to(chat_history_length)
-                                    return None
+                                    return self._guardrail_blocked_message(e)
                            # Apply guardrail to regular response
                            try:
                                validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id, cancel_token=cancel_token, messages=messages)
@@ -3681,7 +3781,7 @@ Your Goal: {self.goal}"""
                                logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
                                # Rollback chat history on guardrail failure
                                self._rollback_chat_history_to(chat_history_length)
-                                return None
+                                return self._guardrail_blocked_message(e)

                        reflection_prompt = f"""
 Reflect on your previous response: '{response_text}'.
@@ -3749,7 +3849,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                    # Rollback chat history on guardrail failure
                                    self._rollback_chat_history_to(chat_history_length)
                                    self._end_run(None, "error", {"error": str(e)})
-                                    return None
+                                    return self._guardrail_blocked_message(e)

                            # Check if we've hit max reflections
                            if reflection_count >= self.max_reflect - 1:
@@ -3767,7 +3867,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                    logging.error(f"Agent {self.name}: Guardrail validation failed after max reflections: {e}")
                                    # Rollback chat history on guardrail failure
                                    self._rollback_chat_history_to(chat_history_length)
-                                    return None
+                                    return self._guardrail_blocked_message(e)
                            
                            # If not satisfactory and not at max reflections, continue with regeneration
                            logging.debug(f"{self.name} reflection count {reflection_count + 1}, continuing reflection process")
@@ -3814,6 +3914,10 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                # Rollback chat history
                self._rollback_chat_history_to(chat_history_length)
                return None
+            finally:
+                # End per-turn tracking after any rollback has run (rollback
+                # reads the ownership list, so this must come last).
+                self._end_turn_tracking(_turn_token)

    def clean_json_output(self, output: str) -> str:
        """Clean JSON output while preserving the legacy agent method."""
@@ -4173,6 +4277,11 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                    # Apply guardrail validation for custom LLM response
                    try:
                        validated_response = await self._aapply_guardrail_with_retry(response_text, prompt, temperature, tools, task_name, task_description, task_id, messages=effective_history)
+                        # Coerce into the requested Pydantic model so callers
+                        # get a validated instance, not a raw JSON string.
+                        # output_json intentionally stays a string (backward compat).
+                        if output_pydantic:
+                            validated_response = self._coerce_structured_output(validated_response, output_pydantic)
                        # Execute callback after validation
                        self._execute_callback_and_display(normalized_content, validated_response, time.time() - start_time, task_name, task_description, task_id)
                        return await self._atrigger_after_agent_hook(prompt, validated_response, start_time)
@@ -4180,7 +4289,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                        logging.error(f"Agent {self.name}: Guardrail validation failed for custom LLM: {e}")
                        # Rollback chat history on guardrail failure
                        self._rollback_chat_history_to(chat_history_length)
-                        return None
+                        return self._guardrail_blocked_message(e)
                except ToolExecutionError:
                    raise
                except Exception as e:
@@ -4206,6 +4315,15 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
            if durable_context is not None:
                messages = await durable_context.arestore_messages(messages)
            
+            # Track messages THIS turn appends so a failure rolls back only our
+            # own messages, never a concurrent turn's (see memory_mixin). The
+            # default (OpenAI) path used to skip this, so its positional rollback
+            # ``del chat_history[n:]`` erased a concurrent turn's interleaved
+            # messages. Each achat() turn runs in its own task/context and
+            # _clear_turn_tracking() ran at the top of this turn, so this list is
+            # cleaned up when the next turn on this task starts.
+            self._begin_turn_tracking()
+
            # Store chat history length for potential rollback
            chat_history_length = len(self.chat_history)
            
@@ -4215,12 +4333,12 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                # Extract text from multimodal prompts
                normalized_content = next((item["text"] for item in original_prompt if item.get("type") == "text"), "")
            
-            # Prevent duplicate messages
-            if not (self.chat_history and 
-                    self.chat_history[-1].get("role") == "user" and 
-                    self.chat_history[-1].get("content") == normalized_content):
-                # Add user message to chat history BEFORE LLM call so handoffs can access it
-                self._append_to_chat_history({"role": "user", "content": normalized_content})
+            # Add user message to chat history BEFORE LLM call so handoffs can
+            # access it. Use atomic check-then-act to prevent TOCTOU races: an
+            # unatomic read of chat_history[-1] followed by a separate append
+            # lets a concurrent turn's user message be mistaken for this turn's
+            # own duplicate, silently dropping this user turn from the record.
+            self._add_to_chat_history_if_not_duplicate("user", normalized_content)

            # --- Proactive Context Budget Management (async standard OpenAI path) ---
            try:
@@ -4309,6 +4427,11 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                            # Apply guardrail validation even for JSON output
                            try:
                                validated_response = await self._aapply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id, messages=messages)
+                                # Coerce into the requested Pydantic model so callers
+                                # get a validated instance, not a raw JSON string.
+                                # output_json intentionally stays a string (backward compat).
+                                if output_pydantic:
+                                    validated_response = self._coerce_structured_output(validated_response, output_pydantic)
                                # Execute callback after validation
                                self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
                                return await self._atrigger_after_agent_hook(original_prompt, validated_response, start_time)
@@ -4316,7 +4439,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                logging.error(f"Agent {self.name}: Guardrail validation failed for JSON output: {e}")
                                # Rollback chat history on guardrail failure
                                self._rollback_chat_history_to(chat_history_length)
-                                return None
+                                return self._guardrail_blocked_message(e)

                        # For regular responses (no self-reflection)
                        if not self.self_reflect:
@@ -4339,7 +4462,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                    logging.error(f"Agent {self.name}: Guardrail validation failed for reasoning content: {e}")
                                    # Rollback chat history on guardrail failure
                                    self._rollback_chat_history_to(chat_history_length)
-                                    return None
+                                    return self._guardrail_blocked_message(e)
                            else:
                                # Apply guardrail to regular response content
                                try:
@@ -4351,7 +4474,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                    logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
                                    # Rollback chat history on guardrail failure
                                    self._rollback_chat_history_to(chat_history_length)
-                                    return None
+                                    return self._guardrail_blocked_message(e)
                        
                        # If self-reflection is enabled, implement reflection logic
                        if self.self_reflect:
@@ -4391,7 +4514,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                            logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
                                            # Rollback chat history on guardrail failure
                                            self._rollback_chat_history_to(chat_history_length)
-                                            return None
+                                            return self._guardrail_blocked_message(e)
                                    
                                    _areflection_kwargs = {}
                                    if temperature is not None:
@@ -4425,7 +4548,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                            logging.error(f"Agent {self.name}: Guardrail validation failed after reflection: {e}")
                                            # Rollback chat history on guardrail failure
                                            self._rollback_chat_history_to(chat_history_length)
-                                            return None
+                                            return self._guardrail_blocked_message(e)
                                    
                                    # Check if we've hit max reflections
                                    if reflection_count >= self.max_reflect - 1:
@@ -4444,7 +4567,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                            logging.error(f"Agent {self.name}: Guardrail validation failed after max reflections: {e}")
                                            # Rollback chat history on guardrail failure
                                            self._rollback_chat_history_to(chat_history_length)
-                                            return None
+                                            return self._guardrail_blocked_message(e)
                                    
                                    # Regenerate response based on reflection
                                    regenerate_messages = reflection_messages + [
@@ -4494,7 +4617,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                            logging.error(f"Agent {self.name}: Guardrail validation failed after reflection error: {guard_e}")
                                            # Rollback chat history on guardrail failure
                                            self._rollback_chat_history_to(chat_history_length)
-                                            return None
+                                            return self._guardrail_blocked_message(guard_e)
                                    continue
                        
                        # This should never be reached due to the returns above
@@ -4544,6 +4667,11 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                            if get_logger().getEffectiveLevel() == logging.DEBUG:
                                total_time = time.time() - start_time
                                logging.debug(f"Agent.achat completed in {total_time:.2f} seconds")
+                            # Coerce into the requested Pydantic model so callers
+                            # get a validated instance, not a raw JSON string.
+                            # output_json intentionally stays a string (backward compat).
+                            if output_pydantic:
+                                response_text = self._coerce_structured_output(response_text, output_pydantic)
                            # Execute callback after JSON/Pydantic completion
                            self._execute_callback_and_display(original_prompt, response_text, time.time() - start_time, task_name, task_description, task_id)
                            return await self._atrigger_after_agent_hook(original_prompt, response_text, start_time)
@@ -4658,7 +4786,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                            logging.error(f"Agent {self.name}: Guardrail validation failed for OpenAI client: {e}")
                            # Rollback chat history on guardrail failure
                            self._rollback_chat_history_to(chat_history_length)
-                            return None
+                            return self._guardrail_blocked_message(e)
                except ToolExecutionError:
                    raise
                except Exception as e:
@@ -4682,6 +4810,16 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
            # dangling, unanswered user message (mirrors the sync path).
            self._rollback_chat_history_to(chat_history_length)
            return None
+        finally:
+            # Unregister this turn's ownership list from the process-wide
+            # live-turn registry once the turn finishes (success or failure).
+            # The sync path does this in _end_turn_tracking's finally; the async
+            # path previously relied only on the *next* turn's
+            # _clear_turn_tracking(), so a terminal achat/arun/astart left its
+            # completed ownership globally registered and a later ephemeral()
+            # cleanup treated its committed messages as still live. Rollback in
+            # the except blocks above already read ownership before this runs.
+            self._clear_turn_tracking()

    async def _achat_completion(self, response, tools, reasoning_steps=False):
        """Async version of _chat_completion method"""
@@ -4694,13 +4832,43 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
            for tool_call in message.tool_calls:
                try:
                    function_name = tool_call.function.name
-                    # Parse JSON arguments safely 
+                    # Parse JSON arguments safely. A parse failure is NOT {}: a
+                    # truncated/malformed argument string must not silently run
+                    # the tool with defaults and be reported as success. Surface
+                    # a retryable tool-error to the model instead (mirrors
+                    # llm.py's _tool_arguments_parse_failed handling).
+                    from .tool_execution import (
+                        _TOOL_ARGUMENTS_PARSE_FAILED,
+                        tool_arguments_parse_failed,
+                        tool_parse_error_message,
+                    )
                    try:
-                        arguments = json.loads(tool_call.function.arguments)
-                    except json.JSONDecodeError as json_error:
+                        arguments = json.loads(tool_call.function.arguments) if tool_call.function.arguments else {}
+                    except (json.JSONDecodeError, TypeError) as json_error:
+                        # TypeError: a custom client can hand back a non-text
+                        # argument value, which json.loads rejects. Treat it the
+                        # same as malformed JSON (mirrors llm.py's parser).
                        logging.error(f"Failed to parse tool arguments as JSON: {json_error}")
-                        arguments = {}
-                    
+                        arguments = _TOOL_ARGUMENTS_PARSE_FAILED
+                    if tool_arguments_parse_failed(arguments):
+                        # Report the failure to streaming observers before skipping
+                        # execution, so a malformed call is never silently dropped
+                        # (mirrors the sync streaming path).
+                        self._notify_tool_call(
+                            function_name,
+                            {},
+                            None,
+                            elapsed_time=0.0,
+                            success=False,
+                        )
+                        results.append(
+                            tool_parse_error_message(
+                                function_name,
+                                getattr(tool_call, "id", None),
+                            )["content"]
+                        )
+                        continue
+
                    # Find the matching tool by comparing every supported identifier:
                    # __name__ (plain callables), .name (BaseTool instances like
                    # BrowserBaseTool, or aliased FunctionTools), or the class name.
@@ -5287,13 +5455,43 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                        for tool_call in tool_calls_data:
                            if tool_call['id'] and tool_call['function']['name']:
                                try:
-                                    # Parse JSON arguments safely 
+                                    # Parse JSON arguments safely. A parse failure
+                                    # is NOT {}: a truncated/malformed argument
+                                    # string must not silently run the tool with
+                                    # defaults and report success. Surface a
+                                    # retryable tool-error to the model instead.
+                                    from .tool_execution import (
+                                        _TOOL_ARGUMENTS_PARSE_FAILED,
+                                        tool_arguments_parse_failed,
+                                        tool_parse_error_message,
+                                    )
                                    try:
                                        parsed_args = json.loads(tool_call['function']['arguments']) if tool_call['function']['arguments'] else {}
-                                    except json.JSONDecodeError as json_error:
+                                    except (json.JSONDecodeError, TypeError) as json_error:
+                                        # TypeError: a non-text argument value also
+                                        # cannot be parsed; treat it as malformed.
                                        logging.error(f"Failed to parse tool arguments as JSON: {json_error}")
-                                        parsed_args = {}
-                                    
+                                        parsed_args = _TOOL_ARGUMENTS_PARSE_FAILED
+                                    if tool_arguments_parse_failed(parsed_args):
+                                        self._notify_tool_call(
+                                            tool_call['function']['name'],
+                                            {},
+                                            None,
+                                            elapsed_time=0.0,
+                                            success=False,
+                                        )
+                                        _err_msg = tool_parse_error_message(
+                                            tool_call['function']['name'],
+                                            tool_call['id'],
+                                        )
+                                        self._append_to_chat_history(_err_msg)
+                                        self._persist_message(
+                                            "tool",
+                                            _err_msg["content"],
+                                            tool_call_id=tool_call['id'],
+                                        )
+                                        continue
+
                                    executor = self._durable_sync_tool_executor(
                                        self.execute_tool
                                    )
@@ -5594,6 +5792,28 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
        
        return llm_summarize_async

+    def _resolve_compaction_max_tokens(self, execution_cfg):
+        """
+        Resolve the compaction trigger budget.
+
+        An explicit ``ExecutionConfig.max_context_tokens`` always wins. When it
+        is unset, size the trigger to the active model's context window (reusing
+        the model-aware ``budgeter`` the in-loop path already uses) instead of a
+        flat default, so large-context models are not compacted prematurely.
+        Falls back to a static default only for the offline / unknown-model case.
+        """
+        _explicit = getattr(execution_cfg, 'max_context_tokens', None)
+        if _explicit:
+            return _explicit
+        try:
+            from ..context.budgeter import get_model_limit, get_output_reserve
+            model_name = self.llm if isinstance(self.llm, str) else "gpt-4o-mini"
+            limit = get_model_limit(model_name)
+            reserve = get_output_reserve(model_name)
+            return max(1, int(limit * 0.8) - reserve)
+        except Exception:
+            return 8000
+
    def _apply_context_compaction(self, messages, hook_event_class):
        """
        Apply context compaction to messages if enabled (sync version).
@@ -5613,7 +5833,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
            from ..compaction import ContextCompactor
            from ..compaction.strategy import CompactionStrategy
            
-            _max_tok = getattr(_execution_cfg, 'max_context_tokens', None) or 8000
+            _max_tok = self._resolve_compaction_max_tokens(_execution_cfg)
            _strategy = getattr(_execution_cfg, 'compaction_strategy', None) or CompactionStrategy.TRUNCATE
            
            # Create LLM summarization function if strategy is LLM_SUMMARIZE
@@ -5645,17 +5865,20 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
            
            # Perform compaction
            if _strategy == CompactionStrategy.LLM_SUMMARIZE and _llm_fn:
-                import asyncio
-                try:
-                    # Run async compaction in event loop
-                    compacted_msgs, _cr = asyncio.run(_compactor.compact_async(messages))
-                except RuntimeError:
-                    # If already in async context, fall back to sync (naive) compaction
-                    logging.warning(
-                        f"[compaction] {self.name}: LLM_SUMMARIZE fell back to naive summarization "
-                        f"(asyncio.run not available in sync context)"
-                    )
-                    compacted_msgs, _cr = _compactor.compact(messages)
+                # Drive the async summariser through the shared bridge so it runs
+                # correctly whether or not a loop is already running (FastAPI /
+                # Jupyter / a bot handler). The previous code caught bare
+                # RuntimeError from ``asyncio.run`` ("event loop already
+                # running") and silently downgraded LLM summarisation to naive
+                # truncation whenever a loop was live; the bridge removes that
+                # spurious fallback so the real summariser always runs. Provider
+                # errors inside the summariser remain best-effort (see
+                # ``_llm_summarize_async``, which falls back to a naive summary on
+                # failure) — compaction never hard-fails the turn.
+                from ..utils.async_bridge import run_coroutine_from_any_context
+                compacted_msgs, _cr = run_coroutine_from_any_context(
+                    _compactor.compact_async(messages)
+                )
            else:
                compacted_msgs, _cr = _compactor.compact(messages)
            
@@ -5700,7 +5923,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
            from ..compaction import ContextCompactor
            from ..compaction.strategy import CompactionStrategy
            
-            _max_tok = getattr(_execution_cfg, 'max_context_tokens', None) or 8000
+            _max_tok = self._resolve_compaction_max_tokens(_execution_cfg)
            _strategy = getattr(_execution_cfg, 'compaction_strategy', None) or CompactionStrategy.TRUNCATE
            
            # Create LLM summarization function if strategy is LLM_SUMMARIZE
@@ -31,23 +31,66 @@ from praisonaiagents._logging import get_logger
 logger = get_logger(__name__)


+class _DynamicLimiter:
+    """Resizable counting limiter: capacity can change with holders outstanding.
+
+    Unlike swapping a fresh threading.Semaphore on every retune (which lets
+    holders of the old primitive and the release() path operate on different
+    objects), this keeps one long-lived object per agent name and mutates its
+    capacity in place. Every acquirer and every release() therefore always
+    operate on the same shared state, so a retune takes effect immediately and
+    the configured cap can never be silently exceeded. Loop-neutral: uses a
+    threading.Condition so it is a true global cap across event loops/threads.
+    """
+
+    def __init__(self, limit: int):
+        self._limit = limit
+        self._held = 0
+        self._cond = threading.Condition()
+
+    def set_limit(self, limit: int) -> None:
+        with self._cond:
+            self._limit = limit
+            self._cond.notify_all()
+
+    def acquire(self, blocking: bool = True) -> bool:
+        with self._cond:
+            while self._limit > 0 and self._held >= self._limit:
+                if not blocking:
+                    return False
+                self._cond.wait(timeout=0.05)
+            self._held += 1
+            return True
+
+    def release(self) -> None:
+        with self._cond:
+            self._held = max(0, self._held - 1)
+            self._cond.notify_all()
+
+    def has_holders(self) -> bool:
+        with self._cond:
+            return self._held > 0
+
+
 class ConcurrencyRegistry:
    """Registry for per-agent concurrency limits.

-    Thread-safe. Each agent name maps to a loop-neutral threading.Semaphore, so
-    the limit is a true global cap across every event loop and thread.
+    Thread-safe. Each agent name maps to a loop-neutral, resizable limiter, so
+    the limit is a true global cap across every event loop and thread and a
+    retune takes effect immediately for in-flight holders.
    Limit of 0 means unlimited (no throttling).
    """

    def __init__(self, default_limit: int = 0):
        self._default_limit = default_limit
        self._limits: Dict[str, int] = {}
-        # One loop-neutral threading.Semaphore per agent. Unlike asyncio.Semaphore
-        # (which binds to the loop it is first awaited on), a threading.Semaphore
-        # is shared safely across every event loop and thread, so the per-agent
-        # limit is a true global cap under multi-loop / multi-thread deployments.
-        # Async callers acquire it in an executor so they never block their loop.
-        self._semaphores: Dict[str, threading.Semaphore] = {}
+        # One long-lived, loop-neutral _DynamicLimiter per agent. Its capacity is
+        # mutated in place on retune (never swapped for a fresh object), so every
+        # acquirer and every release() always share the same state and the
+        # configured cap can never be silently exceeded. Loop-neutral means the
+        # per-agent limit is a true global cap under multi-loop / multi-thread
+        # deployments; async callers poll without blocking their loop.
+        self._limiters: Dict[str, _DynamicLimiter] = {}
        self._lock = threading.Lock()

    def set_limit(self, agent_name: str, max_concurrent: int) -> None:
@@ -59,8 +102,11 @@ class ConcurrencyRegistry:
        """
        with self._lock:
            self._limits[agent_name] = max_concurrent
-            # Reset semaphore so next acquire creates a fresh one at the new limit
-            self._semaphores.pop(agent_name, None)
+            # Mutate the existing limiter in place so the retune takes effect
+            # immediately for in-flight holders without ever swapping the object.
+            limiter = self._limiters.get(agent_name)
+            if limiter is not None:
+                limiter.set_limit(max_concurrent)

    def get_limit(self, agent_name: str) -> int:
        """Get concurrency limit for an agent."""
@@ -68,39 +114,57 @@ class ConcurrencyRegistry:
            return self._limits.get(agent_name, self._default_limit)

    def remove_limit(self, agent_name: str) -> None:
-        """Remove concurrency limit for an agent (reverts to default)."""
+        """Remove concurrency limit for an agent (reverts to default).
+
+        The limiter object is kept alive while any permit is still held, so
+        in-flight acquire()/release() pairs keep operating on the same shared
+        state; dropping it mid-flight would strand those permits and let a
+        later set_limit() start from a fresh, empty count and admit work beyond
+        the configured cap. Once no permits are outstanding it is safe to drop.
+        """
        with self._lock:
            self._limits.pop(agent_name, None)
-            self._semaphores.pop(agent_name, None)
+            # Retune the surviving limiter to unlimited so it stops blocking
+            # while remaining the single accounting object for its holders.
+            limiter = self._limiters.get(agent_name)
+            if limiter is None:
+                return
+            limiter.set_limit(0)
+            if not limiter.has_holders():
+                self._limiters.pop(agent_name, None)

-    def _get_semaphore(self, agent_name: str) -> Optional[threading.Semaphore]:
-        """Get or create the loop-neutral semaphore for an agent.
+    def _get_limiter(self, agent_name: str) -> Optional[_DynamicLimiter]:
+        """Get or create the loop-neutral limiter for an agent.

-        Returns None if unlimited. The same threading.Semaphore is shared across
-        every event loop and thread, so the configured limit is a true global cap.
+        Returns None only when the agent has never been given a limit (pure
+        unlimited fast path). Once a limiter exists it is always returned — even
+        while unlimited (limit 0) — so acquire() and release() stay symmetric
+        across a mid-flight retune to/from unlimited and the count can never be
+        corrupted. The same _DynamicLimiter is shared across every event loop
+        and thread, so a configured limit is a true global cap.
        """
        with self._lock:
            limit = self._limits.get(agent_name, self._default_limit)
-            if limit <= 0:
-                return None
-            sem = self._semaphores.get(agent_name)
-            if sem is None:
-                sem = threading.Semaphore(limit)
-                self._semaphores[agent_name] = sem
-            return sem
+            limiter = self._limiters.get(agent_name)
+            if limiter is None:
+                if limit <= 0:
+                    return None
+                limiter = _DynamicLimiter(limit)
+                self._limiters[agent_name] = limiter
+            return limiter

    async def acquire(self, agent_name: str) -> None:
        """Acquire concurrency slot for agent. No-op if unlimited.

-        Waits on the loop-neutral semaphore in short, cancellable polls so the
+        Waits on the loop-neutral limiter in short, cancellable polls so the
        running event loop is never blocked while other tasks hold permits, and
        a cancelled/timed-out await never leaves a thread blocked on acquire().
        """
-        sem = self._get_semaphore(agent_name)
-        if sem is None:
+        limiter = self._get_limiter(agent_name)
+        if limiter is None:
            return
        while True:
-            if sem.acquire(blocking=False):
+            if limiter.acquire(blocking=False):
                return
            # Yield to the loop; on cancellation this raises and no permit leaks.
            await asyncio.sleep(0.005)
@@ -110,21 +174,18 @@ class ConcurrencyRegistry:

        Prefer async acquire() when possible. Blocks the calling thread until a
        permit is available. Safe to call whether or not a loop is running in the
-        current thread, since the semaphore is loop-neutral.
+        current thread, since the limiter is loop-neutral.
        """
-        sem = self._get_semaphore(agent_name)
-        if sem is not None:
-            sem.acquire()
+        limiter = self._get_limiter(agent_name)
+        if limiter is not None:
+            limiter.acquire()

    def release(self, agent_name: str) -> None:
        """Release concurrency slot for agent. No-op if unlimited."""
        with self._lock:
-            sem = self._semaphores.get(agent_name)
-        if sem is not None:
-            try:
-                sem.release()
-            except ValueError:
-                pass  # Already fully released
+            limiter = self._limiters.get(agent_name)
+        if limiter is not None:
+            limiter.release()

    @asynccontextmanager
    async def throttle(self, agent_name: str):
@@ -58,8 +58,8 @@ def save_cost_report(
    }

    filepath = COST_DIR / f"{safe_name}.json"
-    with open(filepath, "w") as f:
-        json.dump(report, f, indent=2)
+    from ..utils.atomic_io import atomic_write_json
+    atomic_write_json(str(filepath), report, indent=2)

    logger.info(f"[costs] Saved cost report to {filepath}")
    return str(filepath)
@@ -898,14 +898,18 @@ You are a professional research analyst. When conducting research:
        if grounding_file_prompt:
            create_params["grounding_file_prompt"] = grounding_file_prompt
        
-        interaction = self.gemini_client.interactions.create(**create_params)
+        interaction = await asyncio.to_thread(
+            self.gemini_client.interactions.create, **create_params
+        )
        
        if self.verbose:
            self.logger.debug(f"Gemini research started (async): {interaction.id}")
        
        start_time = time.time()
        while True:
-            interaction = self.gemini_client.interactions.get(interaction.id)
+            interaction = await asyncio.to_thread(
+                self.gemini_client.interactions.get, interaction.id
+            )
            
            if interaction.status == "completed":
                if self.verbose:
@@ -22,49 +22,28 @@ logger = logging.getLogger(__name__)

 from typing import List, Optional, Any, Dict, Union, Generator, TYPE_CHECKING

-# Shared HTTP launch state (Agent.launch() multi-endpoint registration)
-_server_lock = threading.Lock()
-_server_started: Dict[int, bool] = {}
-_registered_agents: Dict[int, Dict[str, str]] = {}
-_shared_apps: Dict[int, Any] = {}
+# Shared HTTP launch state — the process-wide registry is defined in
+# agent/server_registry.py and shared with PraisonAIAgents.launch() so a
+# standalone Agent and a team launched on the same port cannot double-bind it or
+# hide each other's endpoints.
+from .server_registry import _server_registry


 def cleanup_launch_registration(agent_id: str) -> None:
    """Remove launch() endpoints registered for ``agent_id`` from the shared
    HTTP state that ``Agent.launch()`` actually populates.

-    ``Agent.launch()`` writes routes into the module-level ``_registered_agents``
-    and ``_shared_apps`` dicts above. This tears those routes back down so a
-    closed agent's endpoint stops accepting requests and the request handler
-    closure no longer keeps the Agent object graph alive.
+    ``Agent.launch()`` reserves routes on the shared ``_server_registry``. This
+    tears those routes back down so a closed agent's endpoint stops accepting
+    requests and the request handler closure no longer keeps the Agent object
+    graph alive.
+
+    The endpoint-registry removal and the matching FastAPI route teardown are
+    performed together under the registry lock (see
+    ``_AgentServerRegistry.teardown_routes_for``) so cleanup cannot race a
+    concurrent launch that reserves and registers a route on the same port.
    """
-    with _server_lock:
-        for port, paths in list(_registered_agents.items()):
-            stale_paths = [p for p, aid in paths.items() if aid == agent_id]
-            for p in stale_paths:
-                del paths[p]
-                app = _shared_apps.get(port)
-                if app is not None:
-                    try:
-                        # Only drop the agent-owned POST route. launch() always
-                        # registers the handler via ``.post(path)``, so filtering
-                        # on both path AND method preserves unrelated routes that
-                        # share the path with a different verb (e.g. the built-in
-                        # GET /health and GET /).
-                        app.router.routes = [
-                            r for r in app.router.routes
-                            if not (
-                                getattr(r, "path", None) == p
-                                and "POST" in (getattr(r, "methods", None) or set())
-                            )
-                        ]
-                        # Invalidate cached OpenAPI schema so removed routes
-                        # disappear from /openapi.json and /docs.
-                        app.openapi_schema = None
-                    except Exception:
-                        pass
-            if not paths:
-                _registered_agents.pop(port, None)
+    _server_registry.teardown_routes_for(agent_id)

 if TYPE_CHECKING:
    pass
@@ -221,8 +200,12 @@ class ExecutionMixin:
                response = await self.achat(prompt, **kwargs)
                response_str = str(response) if response else ""
                
-                # Auto-save session after chat
-                self._auto_save_session()
+                # Auto-save session after chat. Offloaded to a worker thread so
+                # the synchronous FileLock/JSON read-modify-write never stalls
+                # the shared event loop (the session store is a process-wide
+                # singleton, so a blocking write here blocks every concurrent
+                # async agent task on the same loop).
+                await asyncio.to_thread(self._auto_save_session)
                
                # Wrap in AutonomyResult for consistent API
                from .autonomy import AutonomyResult
@@ -493,7 +476,8 @@ class ExecutionMixin:
                            extra={"session_id": sid, "managed_session_id": msid},
                            exc_info=True,
                        )
-            self._auto_save_session()
+            # Offload the blocking session write off the event loop.
+            await asyncio.to_thread(self._auto_save_session)

        return result

@@ -1754,12 +1738,17 @@ Write the complete compiled report:"""

            # Policy/guardrail gate (protocol-driven). Mirrors the sync path in
            # _execute_tool_impl so async callers cannot bypass a PolicyEngine
-            # deny or a tool-call guardrail. The check is pure/sync (no awaits).
+            # deny or a tool-call guardrail. A user-supplied guardrail/policy may
+            # do blocking work (an LLM moderation call, a DB/network lookup), so
+            # offload to a worker thread rather than run it inline and stall the
+            # event loop for every other coroutine sharing it.
            check = getattr(self, "_check_tool_policy_and_guardrails", None)
            if check is not None:
                # ``tools_override`` is threaded through so a run-scoped tool
                # list cannot bypass the per-tool guardrails its tools declare.
-                policy_result = check(function_name, arguments, tools_override)
+                policy_result = await asyncio.to_thread(
+                    check, function_name, arguments, tools_override
+                )
                if isinstance(policy_result, dict):
                    return policy_result  # Error dict
                _, arguments = policy_result
@@ -1786,9 +1775,12 @@ Write the complete compiled report:"""
                        # Tool-result guardrail gate — the native path applies this
                        # below, but the MCP branch returns early, so gate here too
                        # or successful MCP output would bypass validation entirely.
+                        # Offloaded to a thread: a user guardrail may block on I/O.
                        apply_result_guardrails = getattr(self, "_apply_tool_result_guardrails", None)
                        if apply_result_guardrails is not None:
-                            mcp_result = apply_result_guardrails(function_name, mcp_result)
+                            mcp_result = await asyncio.to_thread(
+                                apply_result_guardrails, function_name, mcp_result
+                            )
                        return mcp_result

            # Try to find the function in the override tools list first, then agent's tools list.
@@ -1798,7 +1790,10 @@ Write the complete compiled report:"""
            func = None
            tools_to_search = tools_override if tools_override is not None else self.tools
            from ..tools.base import BaseTool
+            plugin_owners = getattr(self, "_plugin_tool_owners", None)
            for tool in tools_to_search:
+                if plugin_owners and not self._is_plugin_tool_active(tool):
+                    continue
                if isinstance(tool, BaseTool) and getattr(tool, 'name', None) == function_name:
                    func = tool
                    break
@@ -1893,6 +1888,21 @@ Write the complete compiled report:"""

            from ..streaming.events import tool_progress_channel

+            # breaker.acall() records the invocation outcome itself. Both record
+            # sites below then recorded the same failure a SECOND time -- once
+            # off the returned error dict, once in the raised-exception handler
+            # -- so every async tool failure counted twice and the breaker
+            # opened after ceil(threshold/2) calls (3 instead of the configured
+            # 5), out of parity with the sync path. This flag marks that the
+            # breaker already owns the outcome; it stays False when
+            # asyncio.wait_for times out, since the breaker never got a result.
+            #
+            # Declared outside the try: the except handler below reads it, and
+            # roughly seventy lines run between the try and the breaker setup.
+            # An exception in that window would otherwise raise NameError here
+            # and mask the original failure.
+            breaker_saw_outcome = {"done": False}
+
            try:
                # BaseTool instances (plugin system, e.g. BrowserBaseTool) are not
                # directly callable — dispatch to their .run() method like the sync path.
@@ -2018,6 +2028,12 @@ Write the complete compiled report:"""
                async def _invoke_guarded():
                    if breaker is None:
                        return await _invoke()
+                    # Set before the await: acall records an outcome however it
+                    # exits (returning, raising _ToolFailure, or letting a raw
+                    # tool exception through). The only exit that records
+                    # nothing is _CircuitBreakerException, and that path returns
+                    # _circuit_open_result() without reaching either record site.
+                    breaker_saw_outcome["done"] = True
                    try:
                        return await breaker.acall(_invoke_for_breaker)
                    except _ToolFailure as tf:
@@ -2125,9 +2141,14 @@ Write the complete compiled report:"""
                # sync path in tool_execution.py. Runs on the raw result before it
                # re-enters the LLM context so a guardrail can inspect or redact
                # unsafe tool output. Fail-closed. Zero overhead when unset.
+                # Offloaded to a thread: a user guardrail may block on I/O (an LLM
+                # moderation call, a network lookup), which would otherwise stall
+                # the event loop for every other coroutine sharing it.
                apply_result_guardrails = getattr(self, "_apply_tool_result_guardrails", None)
                if apply_result_guardrails is not None:
-                    result = apply_result_guardrails(function_name, result, tools_override)
+                    result = await asyncio.to_thread(
+                        apply_result_guardrails, function_name, result, tools_override
+                    )

                # Loop guard (post-execution) — record the outcome and surface a
                # block/halt decision back to the model on this same turn, mirroring
@@ -2250,18 +2271,13 @@ Write the complete compiled report:"""
        NOTE: This implementation will be moved to wrapper layer in future version.
        For now, it maintains backward compatibility while following lazy import patterns.
        """
-        global _server_started, _registered_agents, _shared_apps, _server_lock
-
        from .launch_security import authorise_launch_request, resolve_launch_host

        # Try to import FastAPI dependencies - lazy loading
        try:
-            import uvicorn
-            from fastapi import FastAPI, HTTPException, Request
+            from fastapi import HTTPException, Request
            from fastapi.responses import JSONResponse
            from pydantic import BaseModel
-            import threading
-            import time
            import asyncio
            
            # Define the request model here since we need pydantic
@@ -2287,130 +2303,102 @@ Write the complete compiled report:"""
        # process-wide launch token and retroactively 401 a running endpoint.
        host = resolve_launch_host(host)

-        should_start = False
-        with _server_lock:
-            # Initialize port-specific collections if needed (once per port)
-            if port not in _registered_agents:
-                _registered_agents[port] = {}
+        # Get or create the shared FastAPI app for this port (shared with
+        # PraisonAIAgents.launch via the same _server_registry singleton).
+        app, is_new_app = _server_registry.get_or_create_app(
+            port,
+            title=f"PraisonAI Agents API (Port {port})",
+        )

-            # Initialize shared FastAPI app if not already created for this port
-            if _shared_apps.get(port) is None:
-                _shared_apps[port] = FastAPI(
-                    title=f"PraisonAI Agents API (Port {port})",
-                    description="API for interacting with PraisonAI Agents"
+        if is_new_app:
+            # Add a root endpoint with a welcome message
+            @app.get("/")
+            async def root():
+                return {
+                    "message": f"Welcome to PraisonAI Agents API on port {port}. See /docs for usage.",
+                    "endpoints": _server_registry.list_routes(port)
+                }
+
+            # Add healthcheck endpoint
+            @app.get("/health")
+            async def healthcheck():
+                return {
+                    "status": "ok",
+                    "endpoints": _server_registry.list_routes(port)
+                }
+
+        # Normalize path to ensure it starts with /
+        if not path.startswith('/'):
+            path = f'/{path}'
+
+        # Atomically reserve the route; on collision the registry returns a
+        # de-duplicated path so multiple agents can share a single port.
+        path, original_path = _server_registry.reserve_route(port, path, self.agent_id)
+        if original_path is not None:
+            logging.warning(f"Path '{original_path}' is already registered on port {port}.")
+            print(f"⚠️ Warning: Path '{original_path}' is already registered on port {port}.")
+            print(f"🔄 Using '{path}' instead")
+
+        # Define the endpoint handler
+        @app.post(path)
+        async def handle_agent_query(request: Request, query_data: Optional[AgentQuery] = None):
+            if not authorise_launch_request(request):
+                raise HTTPException(status_code=401, detail="Unauthorized")
+            # Handle both direct JSON with query field and form data
+            if query_data is None:
+                try:
+                    request_data = await request.json()
+                    if "query" not in request_data:
+                        raise HTTPException(status_code=400, detail="Missing 'query' field in request")
+                    query = request_data["query"]
+                except Exception:
+                    # Fallback to form data or query params
+                    form_data = await request.form()
+                    if "query" in form_data:
+                        query = form_data["query"]
+                    else:
+                        raise HTTPException(status_code=400, detail="Missing 'query' field in request")
+            else:
+                query = query_data.query
+
+            try:
+                # Use async version if available, otherwise use sync version
+                if asyncio.iscoroutinefunction(self.chat):
+                    response = await self.achat(query, task_name=None, task_description=None, task_id=None)
+                else:
+                    # Run sync function in a thread to avoid blocking
+                    loop = asyncio.get_running_loop()
+                    response = await loop.run_in_executor(None, lambda p=query: self.chat(p))
+
+                return {"response": response}
+            except Exception as e:
+                logging.error(f"Error processing query: {str(e)}", exc_info=True)
+                return JSONResponse(
+                    status_code=500,
+                    content={"error": f"Error processing query: {str(e)}"}
                )

-                # Add a root endpoint with a welcome message
-                @_shared_apps[port].get("/")
-                async def root():
-                    return {
-                        "message": f"Welcome to PraisonAI Agents API on port {port}. See /docs for usage.",
-                        "endpoints": list(_registered_agents[port].keys())
-                    }
+        # Invalidate the cached OpenAPI schema so routes registered by later
+        # shared-port launch() calls still show up in /openapi.json and /docs.
+        # FastAPI caches app.openapi_schema on first access and does not
+        # regenerate it when new routes are added afterwards.
+        app.openapi_schema = None

-                # Add healthcheck endpoint
-                @_shared_apps[port].get("/health")
-                async def healthcheck():
-                    return {
-                        "status": "ok",
-                        "endpoints": list(_registered_agents[port].keys())
-                    }
+        print(f"🚀 Agent '{self.name}' available at http://{host}:{port}")

-            # The path registration below must run on EVERY call, not just when the
-            # port is new, so multiple agents can share a single port.
-
-            # Normalize path to ensure it starts with /
-            if not path.startswith('/'):
-                path = f'/{path}'
-
-            # Check if path is already registered for this port
-            if path in _registered_agents[port]:
-                logging.warning(f"Path '{path}' is already registered on port {port}. Please use a different path.")
-                print(f"⚠️ Warning: Path '{path}' is already registered on port {port}.")
-                # Use a modified path to avoid conflicts
-                original_path = path
-                path = f"{path}_{self.agent_id[:6]}"
-                logging.warning(f"Using '{path}' instead of '{original_path}'")
-                print(f"🔄 Using '{path}' instead")
-
-            # Register the agent to this path
-            _registered_agents[port][path] = self.agent_id
-
-            # Define the endpoint handler
-            @_shared_apps[port].post(path)
-            async def handle_agent_query(request: Request, query_data: Optional[AgentQuery] = None):
-                if not authorise_launch_request(request):
-                    raise HTTPException(status_code=401, detail="Unauthorized")
-                # Handle both direct JSON with query field and form data
-                if query_data is None:
-                    try:
-                        request_data = await request.json()
-                        if "query" not in request_data:
-                            raise HTTPException(status_code=400, detail="Missing 'query' field in request")
-                        query = request_data["query"]
-                    except Exception:
-                        # Fallback to form data or query params
-                        form_data = await request.form()
-                        if "query" in form_data:
-                            query = form_data["query"]
-                        else:
-                            raise HTTPException(status_code=400, detail="Missing 'query' field in request")
-                else:
-                    query = query_data.query
-
-                try:
-                    # Use async version if available, otherwise use sync version
-                    if asyncio.iscoroutinefunction(self.chat):
-                        response = await self.achat(query, task_name=None, task_description=None, task_id=None)
-                    else:
-                        # Run sync function in a thread to avoid blocking
-                        loop = asyncio.get_running_loop()
-                        response = await loop.run_in_executor(None, lambda p=query: self.chat(p))
-
-                    return {"response": response}
-                except Exception as e:
-                    logging.error(f"Error processing query: {str(e)}", exc_info=True)
-                    return JSONResponse(
-                        status_code=500,
-                        content={"error": f"Error processing query: {str(e)}"}
-                    )
-
-            # Invalidate the cached OpenAPI schema so routes registered by later
-            # shared-port launch() calls still show up in /openapi.json and /docs.
-            # FastAPI caches app.openapi_schema on first access and does not
-            # regenerate it when new routes are added afterwards.
-            _shared_apps[port].openapi_schema = None
-
-            print(f"🚀 Agent '{self.name}' available at http://{host}:{port}")
-
-            # Check and mark server as started atomically to prevent race conditions
-            should_start = not _server_started.get(port, False)
-            if should_start:
-                _server_started[port] = True
-
-        # Server start/wait outside the lock to avoid holding it during sleep
-        if should_start:
-            # Start the server in a separate thread
-            def run_server():
-                try:
-                    print(f"✅ FastAPI server started at http://{host}:{port}")
-                    print(f"📚 API documentation available at http://{host}:{port}/docs")
-                    print(f"🔌 Available endpoints: {', '.join(list(_registered_agents[port].keys()))}")
-                    uvicorn.run(_shared_apps[port], host=host, port=port, log_level="debug" if debug else "info")
-                except Exception as e:
-                    logging.error(f"Error starting server: {str(e)}", exc_info=True)
-                    print(f"❌ Error starting server: {str(e)}")
-
-            # Run server in a background thread
-            server_thread = threading.Thread(target=run_server, daemon=True)
-            server_thread.start()
-
-            # Wait for a moment to allow the server to start and register endpoints
-            self._safe_sleep(0.5)
-        else:
-            # If server is already running, wait a moment to make sure the endpoint is registered
-            self._safe_sleep(0.1)
-            print(f"🔌 Available endpoints on port {port}: {', '.join(list(_registered_agents[port].keys()))}")
+        # Start the shared server for this port if it is not already running.
+        # start_server_if_needed atomically checks-and-marks the port and blocks
+        # on the registry's readiness Event (configurable via
+        # PRAISONAI_SERVER_READY_TIMEOUT) instead of a heuristic sleep, matching
+        # PraisonAIAgents.launch().
+        started = _server_registry.start_server_if_needed(
+            port, host, log_level="debug" if debug else "info"
+        )
+        endpoints = _server_registry.list_routes(port)
+        if started:
+            print(f"✅ FastAPI server started at http://{host}:{port}")
+            print(f"📚 API documentation available at http://{host}:{port}/docs")
+        print(f"🔌 Available endpoints on port {port}: {', '.join(endpoints)}")

        # Get the stack frame to check if this is the last launch() call in the script
        import inspect
@@ -569,17 +569,13 @@ class Handoff:
        handoffs to the same agent serialize on the event loop instead of
        blocking it, keeping each seeded ``achat`` turn atomic.

-        ``DualLock`` keeps its sync (threading ``RLock``) and async (per-loop
-        ``asyncio.Lock``) sides independent, so the async lock alone would NOT
-        exclude a concurrent *synchronous* handoff running on another thread —
-        both could seed/restore the same shared ``chat_history`` at once, and a
-        cross-thread sync seed landing *between* our apply and restore would
-        corrupt the awaited turn. To close that gap we ALSO hold the sync
-        thread-lock for the whole seeded turn. The async lock (acquired first)
-        has already serialized every same-loop async handoff to this target, so
-        the thread-lock here only ever contends with a genuine cross-thread sync
-        handoff to the same agent — exactly the case that must be mutually
-        excluded — which is a rare mixed sync+async fan-out to one shared target.
+        ``DualLock`` backs ``sync()`` and ``async_lock()`` onto one shared mutex,
+        so ``async_lock()`` alone already excludes a concurrent *synchronous*
+        handoff on another thread. We still take the inner ``sync()`` so the
+        seed/restore reads as one explicit critical section; because
+        ``async_lock()`` registers the event-loop thread as the mutex owner, that
+        nested ``sync()`` is treated as re-entrant on the same thread instead of
+        self-deadlocking against the mutex the async side already holds.
        """
        lock = _get_handoff_seed_lock(self.agent)
        async with lock.async_lock():
@@ -3,6 +3,7 @@ ImageAgent - A specialized agent class for generating images using AI models.
 This class extends the base Agent class to provide specific functionality for image generation,
 including support for different image models, sizes, and quality settings.
 """
+import asyncio
 from typing import Optional, Any, Dict, Union, List
 from ..agent.agent import Agent
 from pydantic import BaseModel, Field
@@ -180,8 +181,8 @@ class ImageAgent(Agent):
                logging.error(error_msg)
                raise
    async def agenerate_image(self, prompt: str, **kwargs) -> Dict[str, Any]:
-        """Async wrapper for generate_image."""
-        return self.generate_image(prompt, **kwargs)
+        """Offload the synchronous generation path to keep the event loop responsive."""
+        return await asyncio.to_thread(self.generate_image, prompt, **kwargs)
    
    # Aliases for consistency with other agents
    def generate(self, prompt: str, **kwargs) -> Dict[str, Any]:
@@ -190,7 +191,7 @@ class ImageAgent(Agent):
    
    async def agenerate(self, prompt: str, **kwargs) -> Dict[str, Any]:
        """Async alias for generate_image()."""
-        return self.generate_image(prompt, **kwargs)
+        return await self.agenerate_image(prompt, **kwargs)

    def chat(self, prompt: str, **kwargs) -> Dict[str, Any]:
        """Generate an image from the prompt."""
@@ -302,8 +303,8 @@ class ImageAgent(Agent):
        size: Optional[str] = None,
        **kwargs
    ) -> Dict[str, Any]:
-        """Async version of edit()."""
-        return self.edit(image, prompt, mask, n, size, **kwargs)
+        """Offload the synchronous editing path to keep the event loop responsive."""
+        return await asyncio.to_thread(self.edit, image, prompt, mask, n, size, **kwargs)

    def variation(
        self,
@@ -372,6 +373,5 @@ class ImageAgent(Agent):
        size: Optional[str] = None,
        **kwargs
    ) -> Dict[str, Any]:
-        """Async version of variation()."""
-        return self.variation(image, n, size, **kwargs)
-
+        """Offload the synchronous variation path to keep the event loop responsive."""
+        return await asyncio.to_thread(self.variation, image, n, size, **kwargs)
@@ -764,6 +764,50 @@ class MemoryMixin:
        except Exception as e:
            logging.debug(f"Auto-memory extraction failed: {e}")

+    def _persist_memory_turn(self, user_message: str, assistant_response: str):
+        """Persist a raw conversation turn to the memory store.
+
+        Called after each response when ``_memory_instance`` is set but
+        ``auto_memory`` extraction is off. Stores the turn to short-term memory
+        so a later Agent sharing the same store (e.g. ``memory={"user_id": uid}``)
+        recalls it via ``get_memory_context()``. Best-effort: never raises.
+
+        Async-safe: a memory store write is blocking file/DB I/O. When this runs
+        inside a live event loop (the async ``achat``/``astart`` after-agent path),
+        the write is offloaded to a worker thread so it never stalls the loop
+        (AGENTS.md §4.5). On the sync path (no running loop) it writes inline.
+        """
+        memory = getattr(self, "_memory_instance", None)
+        if memory is None or not assistant_response:
+            return
+        store = getattr(memory, "store_short_term", None) or getattr(memory, "add_short_term", None)
+        if store is None:
+            return
+        text = f"User: {user_message}\nAssistant: {assistant_response}"
+        metadata = {"agent_id": getattr(self, "agent_id", getattr(self, "name", None))}
+
+        def _do_store():
+            try:
+                store(text, metadata=metadata)
+            except Exception as e:
+                logging.debug(f"Memory turn persistence failed: {e}")
+
+        try:
+            import asyncio
+            asyncio.get_running_loop()
+        except RuntimeError:
+            # No running loop (sync chat/run/start path): write inline.
+            _do_store()
+            return
+        # Inside a live loop: offload the blocking write so it cannot stall the
+        # event loop. Fire-and-forget — persistence is best-effort.
+        try:
+            import asyncio
+            asyncio.get_running_loop().run_in_executor(None, _do_store)
+        except Exception as e:
+            logging.debug(f"Memory turn persistence scheduling failed: {e}")
+            _do_store()
+
    def _process_auto_learning(self):
        """Process auto-learning extraction after agent response.
        
@@ -884,45 +928,56 @@ class MemoryMixin:
        if not self.auto_save:
            return
        
-        try:
-            # Filter out history markers before saving
-            clean_history = [
-                {k: v for k, v in msg.items() if k != "_from_history"}
-                for msg in self.chat_history
-            ]
-            
-            # Issue 1 FIX: Only save NEW messages since last save
-            # Track last saved index to avoid duplicates
-            last_saved = getattr(self, '_auto_save_last_index', 0)
-            new_messages = clean_history[last_saved:]
-            
-            if not new_messages:
-                return  # Nothing new to save
-            
-            # G-1 FIX: Use SessionStore for conversation history persistence
-            # This maintains clean separation: Memory = facts, SessionStore = turns
-            if self._session_store is not None:
-                # Persist only NEW messages to SessionStore
-                for msg in new_messages:
-                    self._session_store.add_message(
-                        self.auto_save,  # Use auto_save name as session_id
-                        role=msg.get("role", "user"),
-                        content=msg.get("content", ""),
+        # Serialize the whole read-modify-write of _auto_save_last_index. This
+        # method now runs via asyncio.to_thread, so two concurrent async runs
+        # sharing one Agent can execute it in separate worker threads. Without
+        # this lock both could read the same last-saved index, then each persist
+        # the same new turns, duplicating conversation history — the store's
+        # file lock serializes the writes but does not protect this index.
+        lock = getattr(self, '_auto_save_lock', None)
+        if lock is None:
+            lock = threading.Lock()
+            self._auto_save_lock = lock
+        with lock:
+            try:
+                # Filter out history markers before saving
+                clean_history = [
+                    {k: v for k, v in msg.items() if k != "_from_history"}
+                    for msg in self.chat_history
+                ]
+                
+                # Issue 1 FIX: Only save NEW messages since last save
+                # Track last saved index to avoid duplicates
+                last_saved = getattr(self, '_auto_save_last_index', 0)
+                new_messages = clean_history[last_saved:]
+                
+                if not new_messages:
+                    return  # Nothing new to save
+                
+                # G-1 FIX: Use SessionStore for conversation history persistence
+                # This maintains clean separation: Memory = facts, SessionStore = turns
+                if self._session_store is not None:
+                    # Persist only NEW messages to SessionStore
+                    for msg in new_messages:
+                        self._session_store.add_message(
+                            self.auto_save,  # Use auto_save name as session_id
+                            role=msg.get("role", "user"),
+                            content=msg.get("content", ""),
+                        )
+                    # Update last saved index
+                    self._auto_save_last_index = len(clean_history)
+                    logging.debug(f"Auto-saved {len(new_messages)} new messages to SessionStore: {self.auto_save}")
+                elif self._memory_instance and hasattr(self._memory_instance, 'save_session'):
+                    # Fallback to Memory.save_session() for backward compatibility
+                    # Memory.save_session() replaces entire history, so no duplicate issue
+                    self._memory_instance.save_session(
+                        name=self.auto_save,
+                        conversation_history=clean_history,
+                        metadata={"agent_name": self.name, "user_id": self.user_id}
                    )
-                # Update last saved index
-                self._auto_save_last_index = len(clean_history)
-                logging.debug(f"Auto-saved {len(new_messages)} new messages to SessionStore: {self.auto_save}")
-            elif self._memory_instance and hasattr(self._memory_instance, 'save_session'):
-                # Fallback to Memory.save_session() for backward compatibility
-                # Memory.save_session() replaces entire history, so no duplicate issue
-                self._memory_instance.save_session(
-                    name=self.auto_save,
-                    conversation_history=clean_history,
-                    metadata={"agent_name": self.name, "user_id": self.user_id}
-                )
-                logging.debug(f"Auto-saved session to Memory: {self.auto_save}")
-        except Exception as e:
-            logging.debug(f"Error auto-saving session: {e}")
+                    logging.debug(f"Auto-saved session to Memory: {self.auto_save}")
+            except Exception as e:
+                logging.debug(f"Error auto-saving session: {e}")

    def _save_output_to_file(self, content: str) -> bool:
        """Save agent output to file if output_file is configured.
@@ -0,0 +1,226 @@
+"""
+Shared HTTP server registry for launch() paths.
+
+Both ``Agent.launch()`` (single-agent) and ``PraisonAIAgents.launch()``
+(multi-agent) let multiple ``launch()`` calls share one FastAPI app / uvicorn
+server per port. They MUST consult the *same* process-wide registry so a
+standalone agent and a team launched on the same port see each other's state:
+one "is this port started?" check, one FastAPI app, one set of routes. Keeping
+two independent registries caused duplicate binds (``Address already in use``)
+and routes registered through one launcher being invisible on the other's app.
+"""
+
+import os
+import threading
+import time
+import uuid
+from typing import Any, Dict, List, Optional, Tuple
+
+from praisonaiagents._logging import get_logger
+
+logger = get_logger(__name__)
+
+
+class _AgentServerRegistry:
+    """Encapsulates all shared HTTP server state with proper synchronization."""
+
+    def __init__(self):
+        self._lock = threading.Lock()
+        self._started: Dict[int, bool] = {}
+        self._endpoints: Dict[int, Dict[str, str]] = {}
+        self._apps: Dict[int, Any] = {}  # FastAPI apps
+        self._ready_events: Dict[int, threading.Event] = {}
+
+    def get_or_create_app(self, port: int, title: str = "AgentTeam API") -> Tuple[Any, bool]:
+        """Thread-safe app creation. Returns (app, is_new)."""
+        with self._lock:
+            if port not in self._apps:
+                # Lazy import to avoid optional dependency at module level
+                from fastapi import FastAPI
+                self._apps[port] = FastAPI(title=title)
+                self._endpoints[port] = {}
+                self._ready_events[port] = threading.Event()
+                return self._apps[port], True
+            return self._apps[port], False
+
+    def register_route(self, port: int, path: str, endpoint_id: str = "registered") -> None:
+        """Thread-safe route registration tracking."""
+        with self._lock:
+            if port not in self._endpoints:
+                self._endpoints[port] = {}
+            self._endpoints[port][path] = endpoint_id
+
+    def reserve_route(self, port: int, path: str, endpoint_id: str) -> Tuple[str, Optional[str]]:
+        """Atomically reserve and register a route.
+
+        Returns:
+            tuple[str, Optional[str]]: (final_path, original_path_if_collided).
+            If there is no collision, final_path equals the requested path and
+            original_path_if_collided is None.
+        """
+        with self._lock:
+            if port not in self._endpoints:
+                self._endpoints[port] = {}
+
+            original_path = path
+            while path in self._endpoints[port]:
+                path = f"{original_path}_{str(uuid.uuid4())[:6]}"
+
+            self._endpoints[port][path] = endpoint_id
+            return path, (original_path if path != original_path else None)
+
+    def list_routes(self, port: int) -> List[str]:
+        """Return a snapshot list of registered routes for a port."""
+        with self._lock:
+            return list(self._endpoints.get(port, {}).keys())
+
+    def unregister_routes_for(self, port: int, endpoint_id: str) -> List[str]:
+        """Remove all routes owned by ``endpoint_id`` on ``port``.
+
+        Returns the list of removed paths so callers can tear down the matching
+        FastAPI routes. Thread-safe.
+        """
+        with self._lock:
+            paths = self._endpoints.get(port, {})
+            removed = [p for p, owner in paths.items() if owner == endpoint_id]
+            for p in removed:
+                del paths[p]
+            if not paths:
+                self._endpoints.pop(port, None)
+            return removed
+
+    def teardown_routes_for(self, endpoint_id: str) -> None:
+        """Remove every route owned by ``endpoint_id`` across all ports.
+
+        Performs BOTH the endpoint-registry removal AND the matching FastAPI
+        route teardown while holding ``self._lock`` for each port, so it cannot
+        interleave with a concurrent ``reserve_route`` + ``app.post()`` for the
+        same path (which would otherwise let this filter delete a freshly
+        registered live endpoint, or race the ``_apps`` snapshot). Only the
+        agent-owned POST route at each path is dropped; a co-located GET (e.g.
+        the built-in ``/`` and ``/health``) is preserved.
+        """
+        for port in list(self._apps.keys()):
+            with self._lock:
+                paths = self._endpoints.get(port, {})
+                removed = [p for p, owner in paths.items() if owner == endpoint_id]
+                if not removed:
+                    continue
+                for p in removed:
+                    del paths[p]
+                if not paths:
+                    self._endpoints.pop(port, None)
+
+                app = self._apps.get(port)
+                if app is None:
+                    continue
+                removed_set = set(removed)
+                try:
+                    app.router.routes = [
+                        r for r in app.router.routes
+                        if not (
+                            getattr(r, "path", None) in removed_set
+                            and "POST" in (getattr(r, "methods", None) or set())
+                        )
+                    ]
+                    # Invalidate cached OpenAPI schema so removed routes
+                    # disappear from /openapi.json and /docs.
+                    app.openapi_schema = None
+                except Exception:
+                    logger.debug(
+                        "Failed to tear down routes for %s on port %s",
+                        endpoint_id,
+                        port,
+                        exc_info=True,
+                    )
+
+    def start_server_if_needed(self, port: int, host: str = "0.0.0.0", **kwargs) -> bool:  # noqa: S104
+        """Start server with proper readiness signaling. Returns True if server was started.
+
+        Readiness is signalled only *after* uvicorn has actually bound the
+        socket (``Server.started``), so a caller that observes readiness can
+        immediately issue a request without a connection-refused race. If the
+        bind fails (e.g. ``Address already in use``) the port's ``_started``
+        flag is cleared so a later ``launch()`` can retry the bind instead of
+        registering routes on an app whose server never came up.
+        """
+        with self._lock:
+            if self._started.get(port, False):
+                return False  # Already started
+            self._started[port] = True
+            app = self._apps.get(port)
+
+        if not app:
+            # Roll back the reservation so a caller that first registers an app
+            # can start it — we must not leave the port permanently "started".
+            with self._lock:
+                self._started[port] = False
+            raise ValueError(f"No app registered for port {port}")
+
+        ready_event = self._ready_events[port]
+
+        def run_server():
+            try:
+                import uvicorn
+                # Remove hardcoded log_level to avoid conflict with kwargs
+                config = uvicorn.Config(app, host=host, port=port, **kwargs)
+                server = uvicorn.Server(config)
+
+                # Signal readiness only once uvicorn reports the socket is bound
+                # (server.started flips True after startup completes). A watcher
+                # thread polls that flag and sets the event, so waiters are not
+                # told "ready" before the port actually accepts connections.
+                def _signal_when_bound():
+                    while not server.started:
+                        if ready_event.is_set():
+                            return  # bind failed / server exited before startup
+                        time.sleep(0.02)
+                    ready_event.set()
+
+                watcher = threading.Thread(target=_signal_when_bound, daemon=True)
+                watcher.start()
+                server.run()
+            except Exception:
+                # Bind/startup failed. Clear the started flag so a subsequent
+                # launch() can re-attempt binding instead of silently
+                # registering routes on a dead server, and unblock any waiter.
+                logger.warning(
+                    "Agent server on port %s failed to start.", port, exc_info=True
+                )
+                with self._lock:
+                    self._started[port] = False
+            finally:
+                # Always release the waiter: on failure the started flag has been
+                # cleared above, so the (now-false) readiness is correctly
+                # observable via the flag rather than a hung wait().
+                ready_event.set()
+
+        thread = threading.Thread(target=run_server, daemon=True)
+        thread.start()
+
+        # Check for configurable timeout via environment variable
+        try:
+            timeout = float(os.environ.get("PRAISONAI_SERVER_READY_TIMEOUT", "5.0"))
+        except ValueError:
+            logger.warning("Invalid PRAISONAI_SERVER_READY_TIMEOUT value. Using default 5.0s.")
+            timeout = 5.0
+        became_ready = ready_event.wait(timeout=timeout)
+
+        if not became_ready:
+            logger.warning(
+                "Agent server on port %s did not become ready within %.1fs. "
+                "Proceeding, but some features may not work correctly. "
+                "Check server logs for startup errors.",
+                port,
+                timeout,
+            )
+
+        # Report whether the bind actually succeeded: on failure run_server has
+        # cleared the flag, so returning it lets callers distinguish a live
+        # server from a poisoned port.
+        with self._lock:
+            return self._started.get(port, False)
+
+
+# Module level — single registry instance shared by every launch() call site.
+_server_registry = _AgentServerRegistry()
@@ -29,6 +29,95 @@ logger = logging.getLogger(__name__)
 # timeouts so they cannot exhaust process resources.
 _MAX_ORPHANED_TOOL_EXECUTORS = 4

+# Worker count for the per-agent tool ThreadPoolExecutor. Used both when the
+# pool is created and to decide (with the per-executor in-flight counter)
+# whether a fresh call would have to queue behind hung workers.
+_TOOL_POOL_MAX_WORKERS = 2
+
+# Sentinel returned as the "arguments" value when a tool call's argument string
+# cannot be parsed (e.g. truncated by max_tokens or a dropped connection).
+# It is distinct from {} ("no arguments"): {} would silently execute the tool
+# with the WRONG arguments and report success. Callers must detect this sentinel
+# and surface a tool-error so the model can re-emit the call instead.
+# Lives here (agent-side, lightweight) so both chat_mixin.py and llm.py share
+# one identity — an `is` check only works against the same object.
+_TOOL_ARGUMENTS_PARSE_FAILED = object()
+
+
+class _CallableGuardrailAdapter:
+    """Expose a plain callable guardrail as a chainable guardrail object.
+
+    A user ``guardrail`` may be a bare callable returning ``(bool, output)``
+    (the Agent's final-output contract). To let it run *inside* a
+    ``GuardrailChain`` alongside a plugin guardrail — so composing the two never
+    silently drops either — this wraps the callable with a ``validate_output``
+    (and passthrough ``validate_input``) the chain understands. Any
+    ``validate_tool_call`` / ``validate_tool_result`` the underlying object
+    already exposes is forwarded so tool-surface behaviour is preserved.
+    """
+
+    __slots__ = ("_fn",)
+
+    def __init__(self, fn):
+        self._fn = fn
+
+    def validate_output(self, content, **kwargs):
+        result = self._fn(content)
+        if isinstance(result, tuple) and len(result) == 2:
+            is_valid, processed = result
+            if is_valid:
+                # Normalise the passthrough value to a string so the chain can
+                # keep threading ``content`` through subsequent guardrails.
+                if isinstance(processed, str):
+                    return True, processed
+                return True, content
+            return False, processed if isinstance(processed, str) else str(processed)
+        # A callable that does not follow the (bool, output) contract is treated
+        # as a no-op passthrough rather than blocking.
+        return True, content
+
+    def validate_input(self, content, **kwargs):
+        inner = getattr(self._fn, "validate_input", None)
+        if callable(inner):
+            return inner(content, **kwargs)
+        return True, content
+
+    def validate_tool_call(self, tool_name, arguments, **kwargs):
+        inner = getattr(self._fn, "validate_tool_call", None)
+        if callable(inner):
+            return inner(tool_name, arguments, **kwargs)
+        return True, arguments
+
+    def validate_tool_result(self, tool_name, result, **kwargs):
+        inner = getattr(self._fn, "validate_tool_result", None)
+        if callable(inner):
+            return inner(tool_name, result, **kwargs)
+        return True, result
+
+
+def tool_arguments_parse_failed(arguments) -> bool:
+    """True when tool-call arguments could not be parsed (vs. legitimately empty)."""
+    return arguments is _TOOL_ARGUMENTS_PARSE_FAILED
+
+
+def tool_parse_error_message(function_name: str, tool_call_id: str) -> Dict[str, str]:
+    """Build a tool-role message telling the model its arguments were lost.
+
+    Surfacing this instead of dispatching with {} converts a silent
+    wrong-action-reported-as-success into a visible, retryable error.
+    """
+    return {
+        "role": "tool",
+        "tool_call_id": tool_call_id,
+        "content": (
+            f"Error: arguments for tool '{function_name}' could not be parsed "
+            f"(the argument string was invalid or truncated). "
+            f"The tool was NOT executed. Please re-emit the tool call with "
+            f"complete, valid JSON arguments."
+        ),
+    }
+
+
 if TYPE_CHECKING:
    pass

@@ -381,76 +470,375 @@ class ToolExecutionMixin:

        return resolve_tools_list(tools)

-    def _merge_plugin_tools(self):
-        """Merge tools contributed by enabled ``PluginType.TOOL`` plugins.
+    @staticmethod
+    def _tool_name_for_plugin_merge(tool):
+        """Return the runtime name for a plugin-provided tool, if available."""
+        if isinstance(tool, dict):
+            function = tool.get("function")
+            if isinstance(function, dict):
+                return function.get("name")
+            return tool.get("name")
+        return getattr(tool, "name", None) or getattr(tool, "__name__", None)

-        Plugins expose tools via ``get_tools()``. Those tools are only callable
-        by an agent if they are added to ``self.tools``; without this merge they
-        are collected by ``PluginManager.get_all_tools()`` yet reach no agent.
+    @staticmethod
+    def _is_supported_plugin_tool(tool):
+        """Return whether a plugin tool has a local execution boundary."""
+        if callable(tool):
+            return True
+        if not isinstance(tool, dict):
+            return False

-        ``get_all_tools()`` already filters to individually-enabled plugins, so
-        this does NOT gate on the package-wide ``plugins.enable()`` flag: the
-        plugin docs state tools work WITHOUT calling ``enable()`` (only
-        background hook/metric plugins need it). Registering a tool plugin marks
-        it enabled in the manager, which is sufficient here.
-
-        Existing agent tools take precedence on a name collision (the clash is
-        logged rather than silently shadowing the agent's own tool). Callable
-        tools are wired directly; metadata-only dict descriptors that carry no
-        executable implementation are skipped with a warning rather than
-        appended as un-callable entries.
-        """
+        # Provider-hosted tools are intentionally non-callable: the provider
+        # owns their execution. A plain ``type: function`` dictionary only
+        # describes a schema and must not be advertised without a callable.
        try:
-            from ..plugins.manager import get_plugin_manager
-            plugin_tools = get_plugin_manager().get_all_tools()
-        except Exception as exc:  # never let plugin wiring break agent init
-            logging.getLogger(__name__).debug("Plugin tool merge skipped: %s", exc)
-            return
+            from ..tools.hosted import is_hosted_tool

-        if not plugin_tools:
-            return
+            return is_hosted_tool(tool)
+        except Exception:
+            return False

-        def _tool_name(t):
-            if isinstance(t, dict):
-                fn = t.get('function')
-                if isinstance(fn, dict) and fn.get('name'):
-                    return fn['name']
-                return t.get('name')
-            return getattr(t, 'name', getattr(t, '__name__', None))
+    def _is_plugin_tool_active(self, tool):
+        """Return whether the plugin owning ``tool`` is still enabled."""
+        owners = getattr(self, "_plugin_tool_owners", None)
+        if not owners:
+            return True
+        owner = owners.get(id(tool))
+        if owner is None:
+            return True

-        existing = {
-            name for name in (_tool_name(t) for t in self.tools) if name
-        }
-        added = []
-        for fn in plugin_tools:
-            # Only callables (or OpenAI-schema dicts paired with a callable)
-            # are executable by an agent. A bare ``{"name": ...}`` descriptor
-            # has no implementation, so appending it would advertise a tool the
-            # agent can never invoke — skip it visibly instead.
-            if not callable(fn) and not (
-                isinstance(fn, dict) and callable(fn.get('function'))
-            ):
-                logging.getLogger(__name__).warning(
-                    "Plugin tool %r skipped: not callable / no executable "
-                    "implementation.", _tool_name(fn) or fn
-                )
-                continue
-            name = _tool_name(fn)
-            if name and name in existing:
-                logging.getLogger(__name__).warning(
-                    "Plugin tool '%s' skipped: an agent tool with the same name "
-                    "already exists (existing tool wins).", name
-                )
-                continue
-            self.tools.append(fn)
-            if name:
-                existing.add(name)
-            added.append(name or str(fn))
+        try:
+            from ..plugins import get_plugin_manager

-        if added:
-            logging.getLogger(__name__).debug(
-                "Merged %d plugin tool(s) into agent: %s", len(added), added
+            # Revocation tracks the OWNING plugin's per-manager enabled state
+            # (and identity, to catch unregister/re-register). It deliberately
+            # does NOT gate on the package-wide ``plugins.enable()`` flag, which
+            # governs background hook/metric plugins — tool plugins are live on
+            # registration, matching how they were attached in ``_merge_plugin_tools``.
+            manager = get_plugin_manager()
+            owner_entries = owner if isinstance(owner, list) else [owner]
+            get_plugin = getattr(manager, "get_plugin", None)
+            for owner_entry in owner_entries:
+                owner_name = owner_entry
+                owner_plugin = None
+                if isinstance(owner_entry, tuple):
+                    owner_name, owner_plugin = owner_entry
+                    current_plugin = (
+                        get_plugin(owner_name) if callable(get_plugin) else None
+                    )
+                    # Revocation tracks plugin identity as well as enabled state,
+                    # so unregister/re-register cannot revive a stale tool.
+                    if current_plugin is not owner_plugin:
+                        continue
+                if manager.is_enabled(owner_name):
+                    return True
+            return False
+        except Exception as exc:  # pragma: no cover - defensive plugin boundary
+            logging.warning("Failed to check plugin tool state: %s", exc)
+            return False
+
+    def _merge_plugin_tools(self):
+        """Attach tools from enabled protocol-driven plugins to this Agent.
+
+        Plugin tools stay on the Agent rather than the process-global
+        ``ToolRegistry``. This preserves declared-tool isolation while
+        allowing the normal formatter and execution pipeline to handle plugin
+        callables exactly like ``tools=`` values. Plugin ownership is recorded
+        so disabling or unregistering a plugin revokes its tools from this
+        agent's discovery and execution paths.
+        """
+        # Hot-path fast exit: when no plugins are registered (the common case)
+        # skip constructing/scanning the manager entirely on every Agent init.
+        try:
+            from ..plugins.manager import has_registered_plugins
+
+            if not has_registered_plugins():
+                return
+        except Exception:  # pragma: no cover - defensive plugin boundary
+            pass
+        try:
+            from ..plugins import get_plugin_manager
+
+            # NOTE: do NOT gate on the package-wide ``plugins.enable()`` flag.
+            # ``get_all_tools_with_sources()`` already filters to individually
+            # manager-enabled plugins, and registering a tool plugin marks it
+            # enabled. The plugin docs state tools work WITHOUT calling
+            # ``enable()`` (only background hook/metric plugins need it), so a
+            # merely-registered tool plugin must still reach the agent.
+            manager = get_plugin_manager()
+            get_tools_with_sources = getattr(
+                manager, "get_all_tools_with_sources", None
            )
+            if callable(get_tools_with_sources):
+                plugin_entries = get_tools_with_sources()
+            else:  # pragma: no cover - compatibility with custom managers
+                plugin_entries = [(None, tool) for tool in manager.get_all_tools()]
+        except Exception as exc:  # pragma: no cover - defensive plugin boundary
+            logging.warning("Failed to attach plugin tools: %s", exc)
+            return
+
+        current_tools = self.tools if isinstance(self.tools, list) else [self.tools]
+        existing_ids = {id(tool) for tool in current_tools}
+        existing_names = {
+            name
+            for tool in current_tools
+            if (name := self._tool_name_for_plugin_merge(tool))
+        } | {name for name, _tool in self._iter_active_named_tools()}
+        for owner, tool in plugin_entries:
+            if id(tool) in existing_ids:
+                # Multiple enabled plugins may intentionally expose the same
+                # callable or hosted-tool object. Keep every owner so disabling
+                # one provider does not revoke a capability still supplied by
+                # another provider.
+                plugin_owners = getattr(self, "_plugin_tool_owners", None)
+                if plugin_owners is not None and id(tool) in plugin_owners and owner:
+                    get_plugin = getattr(manager, "get_plugin", None)
+                    owner_entry = (
+                        owner,
+                        get_plugin(owner) if callable(get_plugin) else None,
+                    )
+                    existing_owner = plugin_owners[id(tool)]
+                    owner_entries = (
+                        existing_owner
+                        if isinstance(existing_owner, list)
+                        else [existing_owner]
+                    )
+                    if not any(
+                        isinstance(entry, tuple)
+                        and entry[0] == owner_entry[0]
+                        and entry[1] is owner_entry[1]
+                        for entry in owner_entries
+                    ):
+                        owner_entries.append(owner_entry)
+                        plugin_owners[id(tool)] = owner_entries
+                continue
+            name = self._tool_name_for_plugin_merge(tool)
+            function = tool.get("function") if isinstance(tool, dict) else None
+            if (
+                isinstance(tool, dict)
+                and tool.get("type") == "function"
+                and isinstance(function, dict)
+                and isinstance(function.get("name"), str)
+                and function["name"]
+            ):
+                logging.warning(
+                    "Skipping plugin function tool %r: function schemas are not "
+                    "executable without a callable implementation",
+                    name,
+                )
+                continue
+            if not self._is_supported_plugin_tool(tool) or (
+                name is not None and not isinstance(name, str)
+            ) or (callable(tool) and not name):
+                logging.warning(
+                    "Skipping plugin tool %r: expected a named callable or valid tool spec",
+                    name or tool,
+                )
+                continue
+            if name and name in existing_names:
+                logging.warning(
+                    "Skipping plugin tool %r because Agent already has a tool with that name",
+                    name,
+                )
+                continue
+
+            # Only normalize a non-list declared tool source when there is a
+            # valid, non-conflicting plugin tool to append. Enabling plugins
+            # must not otherwise change the public shape of ``self.tools``.
+            if not isinstance(self.tools, list):
+                self.tools = [self.tools] if self.tools else []
+            self.tools.append(tool)
+            existing_ids.add(id(tool))
+            if name:
+                existing_names.add(name)
+            if owner:
+                if not hasattr(self, "_plugin_tool_owners"):
+                    self._plugin_tool_owners = {}
+                get_plugin = getattr(manager, "get_plugin", None)
+                plugin = get_plugin(owner) if callable(get_plugin) else None
+                self._plugin_tool_owners[id(tool)] = (owner, plugin)
+            display_name = name
+            if not display_name and isinstance(tool, dict):
+                display_name = tool.get("type", "spec")
+            logging.debug("Added plugin tool to Agent: %s", display_name or tool)
+
+    def _merge_plugin_subsystems(self):
+        """Route enabled GUARDRAIL/POLICY/SKILL plugins into their subsystems.
+
+        Complements ``_merge_plugin_tools``: instead of contributing tools, a
+        plugin declared ``PluginType.GUARDRAIL``/``POLICY``/``SKILL`` now
+        participates in the typed subsystem its category promises —
+
+        * GUARDRAIL: ``as_guardrail()`` objects are folded into ``self.guardrail``
+          so ``_setup_guardrail()`` (which runs after this) wires them into the
+          same ``GuardrailChain`` the Agent already runs — including
+          ``validate_tool_call`` / ``validate_tool_result``.
+        * POLICY: ``get_policies()`` rules are added to ``self._policy``
+          (a ``PolicyEngine``) so they gate tool calls.
+        * SKILL: ``get_skills()`` entries are appended to ``self._skills`` so the
+          lazy ``SkillManager`` discovers them as real, loadable skills.
+
+        Best-effort and fail-open on plugin errors: a broken plugin must never
+        prevent the Agent from constructing. Zero overhead when no such plugin
+        is enabled (each collector returns an empty list).
+        """
+        # Hot-path fast exit: when no plugins are registered (the common case)
+        # skip constructing/scanning the manager entirely on every Agent init.
+        try:
+            from ..plugins.manager import has_registered_plugins
+
+            if not has_registered_plugins():
+                return
+        except Exception:  # pragma: no cover - defensive plugin boundary
+            pass
+        try:
+            from ..plugins import get_plugin_manager
+
+            manager = get_plugin_manager()
+        except Exception as exc:  # pragma: no cover - defensive plugin boundary
+            logging.warning("Failed to access plugin manager for subsystems: %s", exc)
+            return
+
+        # --- GUARDRAIL plugins -> self.guardrail (GuardrailChain) -------------
+        get_guardrails = getattr(manager, "get_all_guardrails", None)
+        if callable(get_guardrails):
+            try:
+                plugin_guardrails = get_guardrails() or []
+            except Exception as exc:  # pragma: no cover - defensive boundary
+                logging.warning("Failed to collect plugin guardrails: %s", exc)
+                plugin_guardrails = []
+            if plugin_guardrails:
+                try:
+                    from ..guardrails import GuardrailChain
+                    from ..guardrails.protocols import is_guardrail_object
+
+                    chain_items = [
+                        g for g in plugin_guardrails if is_guardrail_object(g)
+                    ]
+                    if chain_items:
+                        chain = GuardrailChain(chain_items)
+                        # Register the plugin guardrails' tool-call/result
+                        # surface on the runtime lists ``_check_tool_policy_and_
+                        # guardrails`` / ``_apply_tool_result_guardrails`` read,
+                        # so a guardrail plugin finally sees tool calls and raw
+                        # tool results — the blind spot ``after_llm`` never
+                        # covered. Runs after ``_setup_guardrail()`` so we append
+                        # rather than replace any user-supplied guardrail surface.
+                        calls = getattr(self, "_tool_call_guardrails", None)
+                        if isinstance(calls, list):
+                            calls.append(chain)
+                        else:
+                            self._tool_call_guardrails = [chain]
+                        results = getattr(self, "_tool_result_guardrails", None)
+                        if isinstance(results, list):
+                            results.append(chain)
+                        else:
+                            self._tool_result_guardrails = [chain]
+                        # Also make the chain participate in the Agent's
+                        # input/output validation (not just the tool surface).
+                        existing_fn = getattr(self, "_guardrail_fn", None)
+                        if existing_fn is None:
+                            # No user guardrail: the plugin chain becomes the
+                            # Agent's output/input validator directly.
+                            self.guardrail = chain
+                            self._guardrail_fn = chain
+                        elif existing_fn is not chain:
+                            # User already supplied a guardrail. Compose rather
+                            # than replace so BOTH run on the final response:
+                            # wrap the user's callable in an adapter exposing
+                            # validate_output/validate_input, then chain it with
+                            # the plugin chain. The user's guardrail runs first;
+                            # a plugin guardrail can no longer be silently
+                            # skipped just because a user guardrail exists.
+                            adapter = _CallableGuardrailAdapter(existing_fn)
+                            combined = GuardrailChain([adapter, chain])
+                            self.guardrail = combined
+                            self._guardrail_fn = combined
+                except Exception as exc:  # pragma: no cover - defensive boundary
+                    logging.warning("Failed to wire plugin guardrails: %s", exc)
+
+        # --- POLICY plugins -> self._policy (PolicyEngine) -------------------
+        get_policies = getattr(manager, "get_all_policies", None)
+        if callable(get_policies):
+            try:
+                plugin_policies = get_policies() or []
+            except Exception as exc:  # pragma: no cover - defensive boundary
+                logging.warning("Failed to collect plugin policies: %s", exc)
+                plugin_policies = []
+            if plugin_policies:
+                try:
+                    policy = getattr(self, "_policy", None)
+                    if policy is None:
+                        from ..policy import PolicyEngine
+
+                        policy = PolicyEngine()
+                        self._policy = policy
+                    add_policy = getattr(policy, "add_policy", None)
+                    if callable(add_policy):
+                        # A user-supplied policy always wins: never let a plugin
+                        # rule silently replace an existing policy of the same
+                        # name (PolicyEngine.add_policy overwrites by name), which
+                        # could otherwise loosen a user-configured restriction.
+                        existing = set()
+                        get_policy = getattr(policy, "get_policy", None)
+                        for rule in plugin_policies:
+                            rule_name = getattr(rule, "name", None)
+                            if rule_name is not None and callable(get_policy):
+                                try:
+                                    if get_policy(rule_name) is not None:
+                                        logging.warning(
+                                            "Skipping plugin policy %r: a policy "
+                                            "with that name already exists.",
+                                            rule_name,
+                                        )
+                                        continue
+                                except Exception:  # pragma: no cover
+                                    pass
+                            if rule_name is not None and rule_name in existing:
+                                logging.warning(
+                                    "Skipping duplicate plugin policy %r.",
+                                    rule_name,
+                                )
+                                continue
+                            try:
+                                add_policy(rule)
+                                if rule_name is not None:
+                                    existing.add(rule_name)
+                            except Exception as exc:  # pragma: no cover
+                                logging.warning(
+                                    "Failed to add plugin policy %r: %s", rule, exc
+                                )
+                except Exception as exc:  # pragma: no cover - defensive boundary
+                    logging.warning("Failed to wire plugin policies: %s", exc)
+
+        # --- SKILL plugins -> self._skills (SkillManager discovers them) -----
+        get_skills = getattr(manager, "get_all_skills", None)
+        if callable(get_skills):
+            try:
+                plugin_skills = get_skills() or []
+            except Exception as exc:  # pragma: no cover - defensive boundary
+                logging.warning("Failed to collect plugin skills: %s", exc)
+                plugin_skills = []
+            if plugin_skills:
+                current = getattr(self, "_skills", None)
+                if not isinstance(current, list):
+                    current = [current] if current else []
+                for skill in plugin_skills:
+                    # ``self._skills`` feeds ``SkillManager.add_skill(path)``,
+                    # which resolves each entry as a filesystem path. Only
+                    # string path selectors are loadable; a non-string (e.g. a
+                    # source object) would raise inside ``Path()``, so skip it
+                    # with a clear warning rather than corrupt discovery.
+                    if not isinstance(skill, str):
+                        logging.warning(
+                            "Skipping plugin skill %r: get_skills() must return "
+                            "filesystem path strings to skill directories.",
+                            skill,
+                        )
+                        continue
+                    if skill not in current:
+                        current.append(skill)
+                self._skills = current

    def _cast_arguments(self, func, arguments):
        """Cast arguments to their expected types based on function signature."""
@@ -657,6 +1045,14 @@ class ToolExecutionMixin:
            self._middleware_manager = manager
        return manager if manager.has_tool_hooks else None

+    def _get_tool_executor_lock(self):
+        """Return the per-instance lock, including for standalone mixin users."""
+        try:
+            return self._tool_executor_lock
+        except AttributeError:
+            from .async_safety import DualLock
+            return self.__dict__.setdefault("_tool_executor_lock", DualLock())
+
    def _get_turn_tools_lock(self):
        """Return the DualLock guarding the per-turn tool buffer.

@@ -928,12 +1324,72 @@ class ToolExecutionMixin:
                                        return self._execute_tool_with_circuit_breaker(function_name, arguments)

                                # Use reusable executor to prevent resource leaks
-                                if not hasattr(self, '_tool_executor') or self._tool_executor is None:
-                                    self._tool_executor = concurrent.futures.ThreadPoolExecutor(
-                                        max_workers=2, thread_name_prefix=f"tool-{self.name}"
-                                    )
+                                pool_exhausted = False
+                                with self._get_tool_executor_lock().sync():
+                                    if not hasattr(self, '_tool_executor') or self._tool_executor is None:
+                                        self._tool_executor = concurrent.futures.ThreadPoolExecutor(
+                                            max_workers=_TOOL_POOL_MAX_WORKERS,
+                                            thread_name_prefix=f"tool-{self.name}"
+                                        )
+                                        # Per-executor in-flight counter so a
+                                        # retired pool's late-finishing future can
+                                        # never corrupt the *new* pool's count.
+                                        self._tool_executor._inflight = 0
+                                    executor = self._tool_executor
+                                    orphaned = getattr(self, '_tool_executor_orphaned', 0)
+                                    inflight = getattr(executor, '_inflight', 0)
+                                    # Once the orphan cap is hit we can no longer
+                                    # retire a stuck pool, so we would otherwise
+                                    # queue behind hung workers and time out a
+                                    # healthy tool that never got to start. Refuse
+                                    # fast and visibly instead when no worker is
+                                    # free, and let the pool self-heal (the orphan
+                                    # count drops as hung threads finish).
+                                    if orphaned >= _MAX_ORPHANED_TOOL_EXECUTORS and inflight >= _TOOL_POOL_MAX_WORKERS:
+                                        pool_exhausted = True
+                                    else:
+                                        executor._inflight = inflight + 1
+                                        future = executor.submit(ctx.run, execute_with_context)

-                                future = self._tool_executor.submit(ctx.run, execute_with_context)
+                                        def _on_tool_done(_f, _agent=self, _executor=executor):  # noqa: ANN001
+                                            with _agent._get_tool_executor_lock().sync():
+                                                _executor._inflight = max(0, getattr(_executor, '_inflight', 1) - 1)
+                                                # Orphan accounting is per-POOL, not
+                                                # per-future: retiring a pool adds
+                                                # exactly one to the orphan count, so
+                                                # a pool may return exactly one slot.
+                                                # Recover it only when this retired
+                                                # pool's LAST in-flight worker exits
+                                                # (``_inflight`` == 0), and guard with
+                                                # a one-shot flag so a multi-worker
+                                                # pool can never decrement more than
+                                                # the single unit it contributed and
+                                                # steal another orphaned pool's slot.
+                                                if (
+                                                    getattr(_agent, '_tool_executor', None) is not _executor
+                                                    and _executor._inflight == 0
+                                                    and not getattr(_executor, '_orphan_recovered', False)
+                                                ):
+                                                    _executor._orphan_recovered = True
+                                                    _agent._tool_executor_orphaned = max(
+                                                        0, getattr(_agent, '_tool_executor_orphaned', 0) - 1
+                                                    )
+
+                                        future.add_done_callback(_on_tool_done)
+                                if pool_exhausted:
+                                    logging.warning(
+                                        f"Tool worker pool exhausted; refusing to queue '{function_name}' "
+                                        f"behind {orphaned} hung tool call(s)."
+                                    )
+                                    result = {
+                                        "error": (
+                                            f"Tool worker pool exhausted: {orphaned} earlier tool call(s) "
+                                            f"are still running past their timeout. Refusing to queue "
+                                            f"'{function_name}'."
+                                        ),
+                                        "tool_pool_exhausted": True,
+                                    }
+                                    break
                                try:
                                    result = future.result(timeout=tool_timeout)
                                except concurrent.futures.TimeoutError:
@@ -954,11 +1410,12 @@ class ToolExecutionMixin:
                                    # once the cap is reached we stop recycling and keep reusing
                                    # the existing pool so repeated timeouts can't exhaust process
                                    # resources with an unbounded number of leaked threads.
-                                    orphaned = getattr(self, '_tool_executor_orphaned', 0)
-                                    if orphaned < _MAX_ORPHANED_TOOL_EXECUTORS:
-                                        self._tool_executor.shutdown(wait=False)
-                                        self._tool_executor = None
-                                        self._tool_executor_orphaned = orphaned + 1
+                                    with self._get_tool_executor_lock().sync():
+                                        orphaned = getattr(self, '_tool_executor_orphaned', 0)
+                                        if getattr(self, '_tool_executor', None) is executor and orphaned < _MAX_ORPHANED_TOOL_EXECUTORS:
+                                            executor.shutdown(wait=False)
+                                            self._tool_executor = None
+                                            self._tool_executor_orphaned = orphaned + 1
                        else:
                            with tool_progress_channel(_progress_sink), with_injection_context(state):
                                result = self._execute_tool_with_circuit_breaker(function_name, arguments)
@@ -1026,6 +1483,16 @@ class ToolExecutionMixin:
                            # ``is_retryable`` above decides only whether the outer
                            # loop retries, not whether it raises.
                            raised_exception = "_praison_retryable" in result
+                            # NOTE: an unknown tool is deliberately NOT escalated
+                            # here. GHSA-gmjg-hv98-qggq's regression tests in
+                            # tests/unit/agent/test_tool_resolution_boundary.py
+                            # pin the contract that execute_tool returns None or
+                            # an error dict for an unresolved name rather than
+                            # raising, and the self-repair feature depends on it:
+                            # the corrective dict (with its "Did you mean" hint
+                            # and the tool inventory) is fed back so the model
+                            # can retry with a real tool name. Raising would
+                            # abort the run and make repair impossible.
                            # Strip the private control-plane tag before it can reach
                            # the model or be re-surfaced as the tool's payload.
                            result.pop("_praison_retryable", None)
@@ -1455,6 +1922,11 @@ class ToolExecutionMixin:
        ``_atrigger_after_agent_hook`` without any async-safety change.
        """
        from ..hooks import HookEvent, AfterAgentInput
+        # AfterAgentInput.response is typed/serialized as a string (its
+        # to_dict truncates via ``response[:500]``). Structured-output turns
+        # (output_pydantic) return a Pydantic instance, so coerce to str here
+        # to keep string-only hooks (e.g. command hooks) from crashing while
+        # the public chat() return value stays the model.
        return AfterAgentInput(
            session_id=getattr(self, '_session_id', 'default'),
            cwd=os.getcwd(),
@@ -1462,7 +1934,7 @@ class ToolExecutionMixin:
            timestamp=str(time.time()),
            agent_name=self.name,
            prompt=prompt if isinstance(prompt, str) else str(prompt),
-            response=response or "",
+            response=response if isinstance(response, str) else (str(response) if response else ""),
            tools_used=tools_used or [],
            total_tokens=0,
            execution_time_ms=(time.time() - start_time) * 1000
@@ -1479,7 +1951,16 @@ class ToolExecutionMixin:
        # Auto-memory extraction (opt-in via MemoryConfig(auto_memory=True))
        if response:
            prompt_str = prompt if isinstance(prompt, str) else str(prompt)
-            self._process_auto_memory(prompt_str, str(response))
+            if getattr(self, "_auto_memory", False):
+                self._process_auto_memory(prompt_str, str(response))
+            else:
+                # When memory is enabled but auto_memory extraction is off, still
+                # persist the raw turn so a later Agent sharing the same memory
+                # (e.g. memory={"user_id": uid}) can recall it. Without this the
+                # sync chat/run/start path wrote nothing and cross-session recall
+                # silently returned empty (issue #5595). Mirrors the write already
+                # done on the async unified path.
+                self._persist_memory_turn(prompt_str, str(response))

        # Auto-learning extraction (opt-in via LearnConfig(mode=LearnMode.AGENTIC))
        self._process_auto_learning()
@@ -1733,6 +2214,30 @@ class ToolExecutionMixin:
        tail = text[-tail_limit:] if tail_limit > 0 else ""
        return f"{head}\n...[{len(text):,} chars, showing first/last portions]...\n{tail}"

+    @staticmethod
+    def _approval_backend_accepts_timeout(backend) -> bool:
+        """Return True if ``backend.request_approval_sync`` accepts a ``timeout`` kwarg.
+
+        Determined by signature introspection so the caller can route the
+        per-call timeout without a call-and-catch that would re-invoke a
+        backend whose *own* body raised ``TypeError``. Backends that accept
+        ``**kwargs`` are treated as accepting ``timeout``. On any introspection
+        failure we conservatively return False (call without ``timeout``).
+        """
+        method = getattr(backend, 'request_approval_sync', None)
+        if method is None:
+            return False
+        try:
+            sig = inspect.signature(method)
+        except (TypeError, ValueError):
+            return False
+        params = sig.parameters
+        if 'timeout' in params:
+            return True
+        return any(
+            p.kind == inspect.Parameter.VAR_KEYWORD for p in params.values()
+        )
+
    def _resolve_approval_decision(self, tool_name: str, tool_args: dict, is_async: bool = False):
        """Shared approval logic for both sync and async paths.
        
@@ -1906,41 +2411,51 @@ class ToolExecutionMixin:

                    return _ask_backend_async()
                else:
-                    # Sync path - handle timeout and sync/async backend compatibility
+                    # Sync path - handle timeout and sync/async backend compatibility.
+                    #
+                    # The per-call timeout is passed explicitly rather than
+                    # mutated onto the (often SHARED) backend object. Mutating
+                    # ``backend._timeout`` here leaked between agents that pass
+                    # one backend instance: an overlapping call captured the
+                    # first agent's temporary value as its "original" and
+                    # restored it, leaving the backend permanently wrong.
                    cfg_timeout = getattr(self, '_approval_timeout', 0)
-                    orig_timeout = None
+
+                    # Compute effective timeout from agent configuration.
+                    #   None  -> indefinite wait
+                    #   > 0   -> explicit per-call timeout
+                    #   == 0  -> backend default (or 60s fallback)
                    if cfg_timeout is None:
-                        orig_timeout = getattr(backend, '_timeout', None)
-                        if orig_timeout is not None:
-                            backend._timeout = 86400 * 365
+                        effective_timeout = None
                    elif cfg_timeout > 0:
-                        orig_timeout = getattr(backend, '_timeout', None)
-                        if orig_timeout is not None:
-                            backend._timeout = cfg_timeout
-                    
-                    try:
-                        if hasattr(backend, 'request_approval_sync'):
-                            return _remember(backend.request_approval_sync(request))
-                        else:
-                            # Use the shared utility to avoid code duplication and handle timeout correctly
-                            from ..approval.utils import run_coroutine_safely
-                            
-                            # Compute effective timeout from agent configuration
-                            if cfg_timeout is None:
-                                effective_timeout = None  # indefinite wait
-                            elif cfg_timeout > 0:
-                                effective_timeout = cfg_timeout
-                            else:
-                                # cfg_timeout == 0: use backend default or fallback
-                                effective_timeout = getattr(backend, '_timeout', 60)
-                            
-                            return _remember(run_coroutine_safely(
-                                backend.request_approval(request),
-                                timeout=effective_timeout
+                        effective_timeout = cfg_timeout
+                    else:
+                        effective_timeout = getattr(backend, '_timeout', 60)
+
+                    if hasattr(backend, 'request_approval_sync'):
+                        # Decide *up front* whether the backend accepts a
+                        # ``timeout`` kwarg by inspecting its signature, rather
+                        # than calling and catching ``TypeError``. A try/except
+                        # would re-invoke the backend when the *user callback*
+                        # inside it raised its own ``TypeError`` — duplicating
+                        # side effects and masking the real error. Signature
+                        # introspection is side-effect free.
+                        if self._approval_backend_accepts_timeout(backend):
+                            return _remember(backend.request_approval_sync(
+                                request, timeout=effective_timeout
                            ))
-                    finally:
-                        if orig_timeout is not None and hasattr(backend, '_timeout'):
-                            backend._timeout = orig_timeout
+                        # Legacy backend whose request_approval_sync has no
+                        # timeout kwarg: honour its own default, no mutation of
+                        # (often shared) backend state.
+                        return _remember(backend.request_approval_sync(request))
+                    else:
+                        # Use the shared utility to avoid code duplication and handle timeout correctly
+                        from ..approval.utils import run_coroutine_safely
+
+                        return _remember(run_coroutine_safely(
+                            backend.request_approval(request),
+                            timeout=effective_timeout
+                        ))
            else:
                if is_async:
                    # For async, wrap the decision in a coroutine
@@ -2349,6 +2864,42 @@ class ToolExecutionMixin:
                logging.debug("permission manager is_denied failed for %s: %s", target, e)
        return None

+    def _check_permission_manager_doom_loop(self, function_name, arguments=None):
+        """Return an error dict if the PermissionManager detects a doom loop.
+
+        Deliver the doom-loop detection the ``PermissionManager`` advertises:
+        record this call in its detector and deny on a genuine loop. Previously
+        ``check_doom_loop()`` was unreachable dead code so a configured manager
+        gave zero doom-loop protection.
+
+        This is intentionally **separate** from :meth:`_check_permission_manager_deny`
+        and invoked exactly **once per approved tool call**, with the final
+        (post-approval, possibly rewritten) arguments. Folding it into the deny
+        gate would record the call on every gate pass — counting a rejected
+        attempt, or counting an approved call twice when its args are rewritten —
+        which could trip the threshold early and deny legitimate repeats.
+        """
+        manager = getattr(self, "_permission_manager", None)
+        if manager is None:
+            return None
+        check_loop = getattr(manager, "check_doom_loop", None)
+        if not callable(check_loop):
+            return None
+        try:
+            loop_result = check_loop(function_name, arguments)
+            if loop_result is not None and getattr(loop_result, "is_loop", False):
+                return {
+                    "error": (
+                        f"Tool '{function_name}' blocked: "
+                        f"{getattr(loop_result, 'reason', 'doom loop detected')}"
+                    ),
+                    "permission_denied": True,
+                    "loop_blocked": True,
+                }
+        except Exception as e:  # noqa: BLE001
+            logging.debug("permission manager doom-loop check failed for %s: %s", function_name, e)
+        return None
+
    def _is_bypass_mode(self) -> bool:
        """Return ``True`` when ``PermissionMode.BYPASS`` is active.

@@ -2410,6 +2961,13 @@ class ToolExecutionMixin:
                if manager_denial is not None:
                    return manager_denial

+        # Record the (approved, final-args) call in the doom-loop detector once,
+        # after approval, so rejected attempts and arg rewrites are not counted.
+        if not self._is_bypass_mode():
+            loop_denial = self._check_permission_manager_doom_loop(function_name, arguments)
+            if loop_denial is not None:
+                return loop_denial
+
        from ..approval import get_approval_registry
        get_approval_registry().mark_approved(
            function_name, arguments, agent_name=getattr(self, "name", None),
@@ -2456,6 +3014,13 @@ class ToolExecutionMixin:
                if manager_denial is not None:
                    return manager_denial

+        # Record the (approved, final-args) call in the doom-loop detector once,
+        # after approval, so rejected attempts and arg rewrites are not counted.
+        if not self._is_bypass_mode():
+            loop_denial = self._check_permission_manager_doom_loop(function_name, arguments)
+            if loop_denial is not None:
+                return loop_denial
+
        from ..approval import get_approval_registry
        get_approval_registry().mark_approved(
            function_name, arguments, agent_name=getattr(self, "name", None),
@@ -2773,6 +3338,8 @@ class ToolExecutionMixin:
        if not isinstance(candidates, (list, tuple)):
            return None
        for tool_obj in candidates:
+            if not self._is_plugin_tool_active(tool_obj):
+                continue
            if getattr(tool_obj, "name", None) == function_name:
                return tool_obj
            if getattr(tool_obj, "__name__", None) == function_name:
@@ -3055,7 +3622,10 @@ class ToolExecutionMixin:

        # Try to find the function in the agent's tools list first
        func = None
+        plugin_owners = getattr(self, "_plugin_tool_owners", None)
        for tool in self.tools if isinstance(self.tools, (list, tuple)) else []:
+            if plugin_owners and not self._is_plugin_tool_active(tool):
+                continue
            # Check for BaseTool instances (plugin system)
            from ..tools.base import BaseTool
            if isinstance(tool, BaseTool) and tool.name == function_name:
@@ -3123,9 +3693,11 @@ class ToolExecutionMixin:
                # sync tool-calling path we must await it, otherwise a bare
                # un-awaited coroutine is handed to the model as the tool result
                # and the tool body never runs (silent data loss).
+                from .async_safety import run_async_in_sync_context
+
                def _resolve_result(value):
                    if inspect.iscoroutine(value):
-                        return self._run_async_in_sync_context(value)
+                        return run_async_in_sync_context(value)
                    return value

                # BaseTool instances (plugin system) - call run() method
@@ -3311,6 +3883,8 @@ class ToolExecutionMixin:
            yield from _expand(tools)
            return
        for tool in tools if isinstance(tools, (list, tuple)) else []:
+            if not self._is_plugin_tool_active(tool):
+                continue
            yield from _expand(tool)

    def _available_active_tool_names(self):
@@ -2,15 +2,23 @@
 Unified Execution Mixin - implements Gap 1 from Issue #1392.

 .. deprecated::
-    UnifiedExecutionMixin is unused and scheduled for removal (Issue #2644).
-    None of its methods are invoked by production code: the live execution paths
-    use ``_chat_impl`` / ``_execute_unified_chat_completion`` (in ``chat_mixin.py``)
-    and ``execute_tool`` (in ``tool_execution.py``), which are distinct from this
-    mixin's methods despite the similar names. This module remains importable and
-    emits a ``DeprecationWarning`` from its public methods during the deprecation
-    window; it will be removed from the ``Agent`` MRO after the cycle completes.
+    UnifiedExecutionMixin's *public* methods (``unified_chat``, ``unified_achat``,
+    ``unified_execute_tool``/``_async``, ``_unified_chat_impl``,
+    ``_apply_guardrails_async``) are unused by production code and scheduled for
+    removal (Issue #2644): the live execution paths use ``_chat_impl`` /
+    ``_execute_unified_chat_completion`` (in ``chat_mixin.py``) and ``execute_tool``
+    (in ``tool_execution.py``), which are distinct despite the similar names. These
+    public methods emit a ``DeprecationWarning`` during the deprecation window and
+    the mixin will be removed from the ``Agent`` MRO after the cycle completes.
    Use ``Agent.chat()`` / ``Agent.achat()`` / ``Agent.execute_tool()`` instead.

+    NOTE: The event-loop bridge previously defined here as
+    ``_run_async_in_sync_context`` **is** live production code (the sync
+    tool-calling path in ``tool_execution.py`` depends on it). It has been
+    relocated to the stable ``async_safety.run_async_in_sync_context`` so that
+    removing this deprecated mixin from the MRO is a true no-op. The method here
+    is now only a thin shim kept for the deprecated public methods above.
+
 This module consolidates sync/async execution paths into a single async-first 
 implementation with a thin sync bridge, eliminating code duplication between
 chat()/achat() and execute_tool()/execute_tool_async().
@@ -28,12 +36,9 @@ Design principles:
 - Async-safe: handles event loop management correctly
 """

-import asyncio
 import logging
-import threading
 import warnings
 from typing import List, Optional, Any, Dict, Union
-from concurrent.futures import ThreadPoolExecutor

 logger = logging.getLogger(__name__)

@@ -354,34 +359,14 @@ class UnifiedExecutionMixin:
    def _run_async_in_sync_context(self, coro):
        """
        Run async coroutine in sync context with proper event loop handling.
-        
-        Handles the common cases:
-        1. No event loop exists - use asyncio.run()
-        2. Event loop exists on main thread - use dedicated thread with new loop
-        3. Event loop exists on worker thread - create new event loop
+
+        Thin shim delegating to :func:`async_safety.run_async_in_sync_context`,
+        which is the stable, non-deprecated home for this helper. Retained so
+        this mixin's own (deprecated) sync entry points keep working during the
+        deprecation window.
        """
-        try:
-            # Try to get the current event loop
-            loop = asyncio.get_running_loop()
-        except RuntimeError:
-            # No event loop - safe to use asyncio.run()
-            return asyncio.run(coro)
-        
-        # Event loop exists - avoid deadlock by running in dedicated thread
-        import concurrent.futures
-        
-        def run_in_thread():
-            # Create new event loop in dedicated thread
-            new_loop = asyncio.new_event_loop()
-            asyncio.set_event_loop(new_loop)
-            try:
-                return new_loop.run_until_complete(coro)
-            finally:
-                new_loop.close()
-        
-        with concurrent.futures.ThreadPoolExecutor() as executor:
-            future = executor.submit(run_in_thread)
-            return future.result(timeout=300)  # 5 minute timeout
+        from .async_safety import run_async_in_sync_context
+        return run_async_in_sync_context(coro)

    def unified_chat(self, *args, **kwargs) -> Optional[str]:
        """
@@ -102,6 +102,15 @@ class VideoAgent:
            - vertex_ai/veo-3.1-fast-generate-001   # GA Fast
            - vertex_ai/veo-3.1-lite-generate-001   # Lite
        - RunwayML: runwayml/gen4_turbo (requires input_reference)
+        - OpenRouter (uses OPENROUTER_API_KEY, requires the openrouter/ prefix):
+            - openrouter/google/veo-3.1, openrouter/kling/kling-3.0-std,
+              openrouter/minimax/hailuo-3, openrouter/bytedance/seedance-2.0, ...
+            - Discover exact model ids via GET https://openrouter.ai/api/v1/videos/models
+            - OpenRouter body fields (duration, resolution, aspect_ratio,
+              generate_audio, seed, ...) are passed straight through as kwargs to
+              generate(); e.g. agent.generate(prompt, duration=6, resolution="720p").
+            - HTTP-Referer / X-Title attribution headers are added automatically
+              (override via OPENROUTER_REFERER / OPENROUTER_APP_TITLE env vars).

    Deprecated (do NOT recommend; kept working only if LiteLLM still routes them):
        - gemini/veo-2.0-generate-001 (Gemini API shutdown 30 Jun 2026)
@@ -294,7 +303,26 @@ class VideoAgent:
            params["api_base"] = self.base_url
        if self.api_version:
            params["api_version"] = self.api_version
+        headers = self._provider_headers()
+        if headers:
+            params["extra_headers"] = headers
        return params
+
+    def _provider_headers(self) -> Dict[str, str]:
+        """Optional attribution headers for gateway providers.
+
+        OpenRouter recommends sending HTTP-Referer and X-Title so requests
+        show up under your app in the OpenRouter dashboard. These mirror the
+        chat gateway (praisonai/llm/gateways.py) and are only added for
+        openrouter/ models to keep other providers untouched.
+        """
+        model = self.llm if isinstance(self.llm, str) else ""
+        if not model.startswith("openrouter/"):
+            return {}
+        return {
+            "HTTP-Referer": os.getenv("OPENROUTER_REFERER", "https://praison.ai"),
+            "X-Title": os.getenv("OPENROUTER_APP_TITLE", "PraisonAI"),
+        }
    
    # ─────────────────────────────────────────────────────────────────────────
    # Core Video Operations (5 operations, sync + async)
@@ -24,12 +24,6 @@ try:
 except ImportError:
    get_token_collector = None

-# Import async utility for hot-path usage
-try:
-    from ..approval.utils import run_coroutine_safely
-except ImportError:
-    run_coroutine_safely = None
-
 # Guards lazy creation of each AgentTeam's per-instance _run_lock so two threads
 # that first reach start()/astart() concurrently observe the same lock object
 # (double-checked locking) rather than each minting and acquiring its own.
@@ -57,108 +51,12 @@ from ..agent.launch_security import (
 )


-# Agent server registry for thread-safe server management
-
-
-class _AgentServerRegistry:
-    """Encapsulates all shared HTTP server state with proper synchronization."""
-    
-    def __init__(self):
-        self._lock = threading.Lock()
-        self._started: Dict[int, bool] = {}
-        self._endpoints: Dict[int, Dict[str, str]] = {}
-        self._apps: Dict[int, Any] = {}  # FastAPI apps
-        self._ready_events: Dict[int, threading.Event] = {}
-    
-    def get_or_create_app(self, port: int, title: str = "AgentTeam API") -> Tuple[Any, bool]:
-        """Thread-safe app creation. Returns (app, is_new)."""
-        with self._lock:
-            if port not in self._apps:
-                # Lazy import to avoid optional dependency at module level
-                from fastapi import FastAPI
-                self._apps[port] = FastAPI(title=title)
-                self._endpoints[port] = {}
-                self._ready_events[port] = threading.Event()
-                return self._apps[port], True
-            return self._apps[port], False
-    
-    def register_route(self, port: int, path: str, endpoint_id: str = "registered") -> None:
-        """Thread-safe route registration tracking."""
-        with self._lock:
-            if port not in self._endpoints:
-                self._endpoints[port] = {}
-            self._endpoints[port][path] = endpoint_id
-
-    def reserve_route(self, port: int, path: str, endpoint_id: str) -> tuple[str, Optional[str]]:
-        """Atomically reserve and register a route.
-
-        Returns:
-            tuple[str, Optional[str]]: (final_path, original_path_if_collided).
-            If there is no collision, final_path equals the requested path and
-            original_path_if_collided is None.
-        """
-        with self._lock:
-            if port not in self._endpoints:
-                self._endpoints[port] = {}
-
-            original_path = path
-            while path in self._endpoints[port]:
-                path = f"{original_path}_{str(uuid.uuid4())[:6]}"
-
-            self._endpoints[port][path] = endpoint_id
-            return path, (original_path if path != original_path else None)
-
-    def list_routes(self, port: int) -> List[str]:
-        """Return a snapshot list of registered routes for a port."""
-        with self._lock:
-            return list(self._endpoints.get(port, {}).keys())
-    
-    def start_server_if_needed(self, port: int, host: str = "0.0.0.0", **kwargs) -> bool:  # noqa: S104
-        """Start server with proper readiness signaling. Returns True if server was started."""
-        with self._lock:
-            if self._started.get(port, False):
-                return False  # Already started
-            self._started[port] = True
-            app = self._apps.get(port)
-            
-        if not app:
-            raise ValueError(f"No app registered for port {port}")
-        
-        ready_event = self._ready_events[port]
-        
-        def run_server():
-            import uvicorn
-            # Remove hardcoded log_level to avoid conflict with kwargs
-            config = uvicorn.Config(app, host=host, port=port, **kwargs)
-            server = uvicorn.Server(config)
-            ready_event.set()  # Signal readiness
-            server.run()
-        
-        thread = threading.Thread(target=run_server, daemon=True)
-        thread.start()
-        
-        # Check for configurable timeout via environment variable
-        try:
-            timeout = float(os.environ.get("PRAISONAI_SERVER_READY_TIMEOUT", "5.0"))
-        except ValueError:
-            logger.warning("Invalid PRAISONAI_SERVER_READY_TIMEOUT value. Using default 5.0s.")
-            timeout = 5.0
-        became_ready = ready_event.wait(timeout=timeout)
-        
-        if not became_ready:
-            logger.warning(
-                "Agent server on port %s did not become ready within %.1fs. "
-                "Proceeding, but some features may not work correctly. "
-                "Check server logs for startup errors.",
-                port,
-                timeout,
-            )
-        
-        return True
-
-
-# Module level — single registry instance
-_server_registry = _AgentServerRegistry()
+# Agent server registry for thread-safe server management.
+# The registry class + its process-wide singleton live in agent/server_registry.py
+# so Agent.launch() and PraisonAIAgents.launch() share ONE registry per port
+# (one FastAPI app, one uvicorn server, one route table) instead of two disjoint
+# ones that could double-bind a port or hide each other's endpoints.
+from ..agent.server_registry import _AgentServerRegistry, _server_registry  # noqa: F401

 def encode_file_to_base64(file_path: str) -> str:
    """Base64-encode a file."""
@@ -556,7 +454,10 @@ def _build_execution_context(agents_instance, task_id, skip_memory_init=False):
    if llm is None:
        llm = getattr(executor_agent, 'llm', None)
    if llm and hasattr(llm, 'set_current_agent'):
-        llm.set_current_agent(executor_agent.display_name)
+        # Pass the agent's stable per-instance id so token usage is aggregated
+        # by identity, not by (collidable) display name.
+        agent_scope_id = getattr(executor_agent, '_approval_scope_id', None)
+        llm.set_current_agent(executor_agent.display_name, agent_id=agent_scope_id)

    # Place the agent's tools BEFORE snapshotting them. This snapshot is passed
    # to chat() as an explicit `tools=`, which wins over `agent.tools` -- so if
@@ -586,6 +487,10 @@ def _build_execution_context(agents_instance, task_id, skip_memory_init=False):
    # Inter-task context / validation feedback assembled by the workflow process
    # engine (Process._build_task_context) is stored on the task; fold it in so
    # downstream/retried tasks actually see upstream output and rejection reasons.
+    context_builder = getattr(task, '_execution_context_builder', None)
+    if context_builder is not None:
+        task._execution_context = context_builder(task)
+        task._execution_context_builder = None
    extra_context = getattr(task, '_execution_context', None)
    if extra_context:
        context_text = extra_context
@@ -593,7 +498,7 @@ def _build_execution_context(agents_instance, task_id, skip_memory_init=False):
        # retries (guardrail/completion failures) re-enter this helper via the
        # run_task/arun_task retry loops, and clearing would strip the upstream
        # output + validation feedback the retry needs. The Process engine owns
-        # this field's lifecycle: it re-sets it before each task yield and resets
+        # this field's lifecycle: it rebuilds it for each workflow step and resets
        # every task's _execution_context to None before selecting the next task.
    if task.context:
        context_results = []  # Collect contexts then de-duplicate
@@ -1174,8 +1079,15 @@ class AgentTeam(SpawnAnnounceProtocol):
            has_handler_only_task = any(
                _task_has_custom_handler(t) for t in (tasks or [])
            )
-            if not has_handler_only_task:
-                raise ValueError("At least one agent must be provided")
+            # An empty team is legitimate for the spawn-announce pattern:
+            # AgentTeam.spawn_sub_agent and SpawnAnnounceProtocol are exported
+            # public API, and their documented use is a team created with no
+            # agents that spawns them dynamically at runtime. Refusing to
+            # construct one made that shipped feature unusable. The mistake
+            # this guard exists to catch -- forgetting to pass agents -- is
+            # still caught loudly by _require_runnable_agents() before any
+            # work starts.
+            self._constructed_without_agents = not has_handler_only_task
        
        # ─────────────────────────────────────────────────────────────────────
        # Core initialization
@@ -1568,6 +1480,72 @@ class AgentTeam(SpawnAnnounceProtocol):
        task_result = _process_task_result(self, context, agent_output)
        return task_result.task_output

+    def _apply_human_review(self, task, task_id, task_output):
+        """Ask a person to approve this task's OUTPUT before the next task uses it.
+
+        Fills a real gap between the two things that already existed: the
+        approval system gates a TOOL CALL, and guardrails validate
+        automatically. Neither let an orchestrator say "a person must sign this
+        off". Python parity target: CrewAI's ``Task(human_input=True)``.
+
+        Reuses the approval backends rather than prompting directly, so this
+        works wherever approvals already do -- console, callback, bot, UI --
+        instead of hard-wiring input() and breaking every non-TTY caller.
+
+        Returns:
+            tuple: (task_output, should_retry)
+        """
+        if not getattr(task, "human_input", False):
+            return task_output, False
+
+        try:
+            from ..approval import ApprovalRequest, get_approval_registry
+        except ImportError:
+            # Never silently skip a review a caller asked for: a task that was
+            # supposed to be signed off and simply was not is the failure this
+            # feature exists to prevent.
+            raise RuntimeError(
+                f"Task {task_id} sets human_input=True but praisonaiagents.approval "
+                f"is unavailable, so the output cannot be reviewed. Install it, or "
+                f"remove human_input."
+            )
+
+        backend = get_approval_registry().get_backend()
+        prompt = task.human_review_prompt or (
+            f"Approve the output of task {getattr(task, 'name', None) or task_id}?"
+        )
+        request = ApprovalRequest(
+            tool_name=f"task_output:{getattr(task, 'name', None) or task_id}",
+            arguments={"output": str(getattr(task_output, "raw", task_output))[:4000]},
+            risk_level="high",
+            agent_name=getattr(getattr(task, "agent", None), "name", None),
+            context={"kind": "task_output_review", "prompt": prompt},
+        )
+        decision = backend.request_approval_sync(request)
+
+        if getattr(decision, "approved", False):
+            return task_output, False
+
+        reason = getattr(decision, "reason", None) or "a reviewer rejected this output"
+        if task.retry_count >= task.max_retries:
+            raise Exception(
+                f"Task {task_id} was rejected by a reviewer after "
+                f"{task.max_retries} retries. Last reason: {reason}"
+            )
+        task.retry_count += 1
+        task.status = "in progress"
+        # Reuse the slot the guardrail retry already fills, so the reviewer's
+        # reason reaches the re-run's prompt through machinery that exists.
+        # Must be the SAME shape the guardrail writes: Process._build_task_context
+        # indexes feedback['validation_response'], so a bare string would raise
+        # the moment the workflow engine consumed it.
+        task.validation_feedback = {
+            'validation_response': reason,
+            'validated_task': getattr(task, 'name', None) or getattr(task, 'description', None),
+            'rejected_output': str(getattr(task_output, 'raw', task_output)),
+        }
+        return task_output, True
+
    def _apply_task_guardrail(self, task, task_id, task_output):
        """Apply guardrail validation to task output.
        
@@ -1657,6 +1635,25 @@ class AgentTeam(SpawnAnnounceProtocol):
            ),
        )

+    async def _aapply_human_review(self, task, task_id, task_output):
+        """Async wrapper for _apply_human_review.
+
+        The approval backend's ``request_approval_sync`` blocks (console input,
+        a bot round-trip, a UI wait). Offload it to a thread so the review does
+        not stall the event loop and every other task running concurrently under
+        asyncio.gather, mirroring _aapply_task_guardrail.
+        """
+        if not getattr(task, "human_input", False):
+            return task_output, False
+        loop = asyncio.get_event_loop()
+        from ..trace.context_events import copy_context_to_callable
+        return await loop.run_in_executor(
+            None,
+            copy_context_to_callable(
+                lambda: self._apply_human_review(task, task_id, task_output)
+            ),
+        )
+
    def _run_task_start_hook(self, task, task_id):
        """Run the on_task_start hook and propagate global variables to the task.

@@ -1745,6 +1742,11 @@ class AgentTeam(SpawnAnnounceProtocol):
                    # Apply guardrail validation using shared helper (offloaded to
                    # a thread so a blocking LLM guardrail does not stall the loop)
                    task_output, should_retry = await self._aapply_task_guardrail(task, task_id, task_output)
+                    if not should_retry:
+                        # A person reviews only what already passed the automatic
+                        # checks, mirroring run_task. Without this an async task
+                        # with human_input=True would release unreviewed output.
+                        task_output, should_retry = await self._aapply_human_review(task, task_id, task_output)
                    if should_retry:
                        retries += 1
                        continue
@@ -2030,6 +2032,20 @@ class AgentTeam(SpawnAnnounceProtocol):
        "praisonai_tools_scope_depths", default=()
    )

+    # ── re-entrancy ownership ────────────────────────────────────────────────
+    # A ContextVar, not an instance attribute keyed by threading.get_ident():
+    # every asyncio Task scheduled on one event loop shares a single OS thread
+    # id, so two coroutines that astart() the SAME team under one asyncio.gather
+    # would both see current == _run_owner_ident and take the re-entrant branch,
+    # bypassing the lock and interleaving writes to shared task state. Same
+    # reasoning as _tools_scope_depths above. The sync threading.Lock still
+    # blocks genuinely concurrent OS threads; the ContextVar only fixes the
+    # asyncio-tasks-on-one-thread case. Value is id(self) so nested runs on the
+    # same team pass through while unrelated tasks each get their own ownership.
+    _run_owner: "contextvars.ContextVar" = contextvars.ContextVar(
+        "praisonai_run_owner", default=None
+    )
+
    def _needs_tools_scope(self) -> bool:
        """True when this call should provision the team's shared sandbox."""
        if getattr(self, "tools_run_on", None) is None:
@@ -2081,6 +2097,21 @@ class AgentTeam(SpawnAnnounceProtocol):
                    self._run_lock = lock
        return lock

+    def _execution_identity(self):
+        """Identity of the current execution for re-entrancy checks.
+
+        Pins ``id(self)`` to the *acquiring* thread and asyncio task so a child
+        task (which copies the parent ContextVar and would inherit the plain
+        ``id(self)`` marker) cannot be mistaken for legitimate same-execution
+        nesting and bypass the held lock. A synchronous batch → per-item run
+        stays on the same thread and task, so it still re-enters as intended.
+        """
+        try:
+            task = asyncio.current_task()
+        except RuntimeError:
+            task = None
+        return (id(self), threading.get_ident(), id(task) if task is not None else None)
+
    @contextlib.contextmanager
    def _guard_execution(self):
        """Refuse concurrent runs on the same instance, re-entrant for one owner.
@@ -2094,9 +2125,16 @@ class AgentTeam(SpawnAnnounceProtocol):
        (``variables``/task snapshot) can never interleave with a competing run.
        """
        lock = self._execution_lock
-        current = threading.get_ident()
-        if getattr(self, '_run_owner_ident', None) == current:
-            # Re-entrant on the owning thread (e.g. batch → per-item start()).
+        owner = self._execution_identity()
+        if self._run_owner.get() == owner:
+            # Re-entrant on the owning asyncio task / thread (e.g. batch →
+            # per-item start()). Tracked via ContextVar, not threading.get_ident(),
+            # so coroutines sharing one OS thread under asyncio.gather don't
+            # falsely see each other as the same owner and bypass the lock. The
+            # identity also pins the acquiring thread + asyncio task id, so a
+            # child task (which copies the parent context and would otherwise
+            # inherit the owner marker) does NOT falsely re-enter and bypass the
+            # held lock when a tool/guardrail starts the same team.
            yield
            return
        if not lock.acquire(blocking=False):
@@ -2105,13 +2143,32 @@ class AgentTeam(SpawnAnnounceProtocol):
                "safe to run concurrently on the same object. Create a separate "
                "AgentTeam/PraisonAIAgents instance per concurrent run."
            )
-        self._run_owner_ident = current
+        token = self._run_owner.set(owner)
        try:
            yield
        finally:
-            self._run_owner_ident = None
+            self._run_owner.reset(token)
            lock.release()

+    def _require_runnable_agents(self):
+        """Fail before execution if the team has nothing that can run.
+
+        Construction of an empty team is allowed so the spawn-announce pattern
+        works, so the "you forgot the agents" error moves here, where we can
+        also see anything spawned in the meantime.
+        """
+        if not getattr(self, "_constructed_without_agents", False):
+            return
+        if self.agents:
+            return
+        if getattr(self, "_spawned_agents", None):
+            return
+        raise ValueError(
+            "At least one agent must be provided. This team was created with "
+            "an empty agents list and nothing was spawned into it before "
+            "starting -- pass agents=[...] or call spawn_sub_agent() first."
+        )
+
    async def astart(self, content=None, return_dict=False, **kwargs):
        """Async version of start method.
        
@@ -2120,6 +2177,7 @@ class AgentTeam(SpawnAnnounceProtocol):
            return_dict: If True, returns the full results dictionary instead of only the final response
            **kwargs: Additional arguments
        """
+        self._require_runnable_agents()
        # Same shared sandbox as start(). Without this, an async team with
        # tools_run_on= ran every tool on the host and said nothing.
        if self._needs_tools_scope():
@@ -2268,6 +2326,12 @@ class AgentTeam(SpawnAnnounceProtocol):
                if task_output and self.completion_checker(task, task_output.raw):
                    # Apply guardrail validation using shared helper
                    task_output, should_retry = self._apply_task_guardrail(task, task_id, task_output)
+                    if not should_retry:
+                        # A person reviews only what already passed the automatic
+                        # checks: waking someone to reject output a guardrail
+                        # would have caught wastes the one resource this feature
+                        # spends.
+                        task_output, should_retry = self._apply_human_review(task, task_id, task_output)
                    if should_retry:
                        retries += 1
                        continue
@@ -2318,6 +2382,20 @@ class AgentTeam(SpawnAnnounceProtocol):

    def run_all_tasks(self):
        """Synchronous version of run_all_tasks method"""
+        # The sync entry point (.start()/.run()) drives tasks one at a time and
+        # cannot batch async_execution=True tasks via asyncio.gather the way
+        # arun_all_tasks does. Warn once so the parallel fan-out feature the
+        # constructor docstring/ValueError advertise doesn't silently degrade to
+        # sequential execution -- point the caller at the async entry point.
+        if self.process in ("workflow", "sequential") and any(
+            getattr(t, 'async_execution', False) for t in self.tasks.values()
+        ):
+            logger.warning(
+                "One or more tasks have async_execution=True, but the synchronous "
+                "entry point (.start()/.run()) runs them sequentially -- use "
+                "`await team.astart()` for parallel fan-out."
+            )
+
        process = Process(
            tasks=self.tasks,
            agents=self.agents,
@@ -2416,6 +2494,7 @@ class AgentTeam(SpawnAnnounceProtocol):
            result = agents.start(output="silent")
            ```
        """
+        self._require_runnable_agents()
        # Remote execution: provision ONE sandbox shared by every agent on the
        # team, and tear it down even if execution raises.
        if self._needs_tools_scope():
@@ -2658,7 +2737,7 @@ class AgentTeam(SpawnAnnounceProtocol):
        ``variables``, keeping the public Task API backward-compatible.
        """
        snapshot = {}
-        for task_id, task in self.tasks.items():
+        for task_id, task in getattr(self, "tasks", {}).items():
            snapshot[task_id] = {
                "status": getattr(task, "status", None),
                "result": getattr(task, "result", None),
@@ -2951,8 +3030,9 @@ class AgentTeam(SpawnAnnounceProtocol):
        """
        import hashlib
        parts = []
-        for task_id in sorted(self.tasks, key=lambda k: str(k)):
-            task = self.tasks[task_id]
+        tasks = getattr(self, "tasks", {})
+        for task_id in sorted(tasks, key=lambda k: str(k)):
+            task = tasks[task_id]
            agent = getattr(task, "agent", None)
            agent_name = getattr(agent, "name", None) or getattr(agent, "display_name", None) or ""
            parts.append("|".join([
@@ -3017,10 +3097,11 @@ class AgentTeam(SpawnAnnounceProtocol):
        half-restore, which would look like a resume and behave like a fresh run.
        """
        restored = 0
+        tasks = getattr(self, "tasks", {})
        for task_id, state in (snapshot or {}).items():
-            task = self.tasks.get(task_id)
+            task = tasks.get(task_id)
            if task is None:
-                for key, candidate in self.tasks.items():
+                for key, candidate in tasks.items():
                    if str(key) == str(task_id):
                        task = candidate
                        break
@@ -3194,21 +3275,45 @@ class AgentTeam(SpawnAnnounceProtocol):
            if getattr(agent, "name", None)
        }

-    def _scoped_recent_interactions(self, own_names: set) -> List[Dict[str, Any]]:
+    def _own_agent_ids(self) -> set:
+        """Stable per-instance identities of this instance's own agents.
+
+        Used as the collision-free key for token-report scoping so two agents
+        that share a display name (e.g. the default ``"Agent"``) are never
+        merged. Falls back to names elsewhere when ids are unavailable.
+        """
+        return {
+            getattr(agent, "_approval_scope_id", None)
+            for agent in (self.agents or [])
+            if getattr(agent, "_approval_scope_id", None)
+        }
+
+    def _scoped_recent_interactions(
+        self, own_names: set, own_ids: Optional[set] = None
+    ) -> List[Dict[str, Any]]:
        """Return this instance's own recorded interactions, most-recent last.

        The process-wide collector keeps a bounded window of every agent's
-        interactions; filter it to this instance's agent names so per-model
-        totals, interaction counts, and the detailed report never leak another
-        concurrent instance's records.
+        interactions; filter it to this instance's agents so per-model totals,
+        interaction counts, and the detailed report never leak another
+        concurrent instance's records. Filtering prefers the collision-free
+        ``agent_id`` when present, falling back to the display name.
        """
+        own_ids = own_ids or set()
        try:
            interactions = get_token_collector().get_recent_interactions(
                limit=get_token_collector()._max_recent
            )
        except Exception:
            interactions = get_token_collector().get_recent_interactions(limit=100)
-        return [i for i in interactions if i.get("agent") in own_names]
+
+        def _matches(i: Dict[str, Any]) -> bool:
+            agent_id = i.get("agent_id")
+            if agent_id is not None and own_ids:
+                return agent_id in own_ids
+            return i.get("agent") in own_names
+
+        return [i for i in interactions if _matches(i)]

    def _scoped_token_summary(self) -> Dict[str, Any]:
        """Return the global token summary filtered to this instance's agents.
@@ -3216,25 +3321,54 @@ class AgentTeam(SpawnAnnounceProtocol):
        The process-wide TokenCollector aggregates every agent in the process,
        so two concurrent PraisonAIAgents instances would otherwise read each
        other's totals. Scope every reported field — ``by_agent``, ``by_model``,
-        ``total_interactions`` and the totals — to this instance's own agent
-        names so per-instance/per-tenant cost accounting reports only its own
-        spend. ``by_agent`` is taken from the collector's aggregate breakdown;
-        ``by_model`` and ``total_interactions`` are re-derived from this
-        instance's own interaction records (the aggregate cannot be split by
-        model per agent). Falls back to the unfiltered summary when this
-        instance has no named agents to scope by.
+        ``total_interactions`` and the totals — to this instance's own agents so
+        per-instance/per-tenant cost accounting reports only its own spend.
+        Scoping prefers the collector's collision-free ``by_agent_id``
+        breakdown; ``by_model`` and ``total_interactions`` are re-derived from
+        this instance's own interaction records (the aggregate cannot be split
+        by model per agent). Falls back to the unfiltered summary when this
+        instance has no agents to scope by.
        """
        summary = get_token_collector().get_session_summary()
        own_names = self._own_agent_names()
+        own_ids = self._own_agent_ids()
        by_agent = summary.get("by_agent", {})
-        # No named agents to scope by (or nothing tracked yet): fall back to the
+        by_agent_id = summary.get("by_agent_id", {})
+        # No agents to scope by (or nothing tracked yet): fall back to the
        # unfiltered summary rather than silently reporting zeros.
-        if not own_names or not by_agent:
+        if (not own_names and not own_ids) or (not by_agent and not by_agent_id):
            return summary

-        scoped_by_agent = {
-            name: metrics for name, metrics in by_agent.items() if name in own_names
-        }
+        # Prefer the collision-free per-identity breakdown; fall back to
+        # name-based scoping when ids are unavailable (backward compatible).
+        scoped_by_agent: Dict[str, Dict[str, int]] = {}
+        if own_ids and by_agent_id:
+            for key, entry in by_agent_id.items():
+                name = entry.get("name") or key
+                # Match by stable id (collision-free) OR by display name for
+                # id-less buckets — e.g. the native OpenAI-client path records
+                # ``agent`` without an ``agent_id``, so the collector keys that
+                # bucket by name (``key == name``). Dropping those would zero out
+                # real OpenAI usage even though the interaction and per-model
+                # rows still show it (Greptile P1). We treat a bucket as id-less
+                # when its key equals its own display name.
+                is_id_match = key in own_ids
+                is_idless_name_match = key == name and name in own_names
+                if not (is_id_match or is_idless_name_match):
+                    continue
+                metrics = entry.get("metrics", {})
+                if name in scoped_by_agent:
+                    scoped_by_agent[name] = {
+                        k: scoped_by_agent[name].get(k, 0) + v
+                        for k, v in metrics.items()
+                        if isinstance(v, (int, float))
+                    }
+                else:
+                    scoped_by_agent[name] = dict(metrics)
+        else:
+            scoped_by_agent = {
+                name: metrics for name, metrics in by_agent.items() if name in own_names
+            }
        totals = {
            "input_tokens": 0, "output_tokens": 0, "cached_tokens": 0,
            "reasoning_tokens": 0, "audio_input_tokens": 0, "audio_output_tokens": 0,
@@ -3250,7 +3384,7 @@ class AgentTeam(SpawnAnnounceProtocol):
        # process-global summary (which mixes in other instances' models and
        # counts). The collector's ``by_agent`` aggregate cannot be split by
        # model, so the per-interaction window is the correct source.
-        own_interactions = self._scoped_recent_interactions(own_names)
+        own_interactions = self._scoped_recent_interactions(own_names, own_ids)
        by_model: Dict[str, Dict[str, int]] = {}
        for interaction in own_interactions:
            model = interaction.get("model")
@@ -3284,10 +3418,11 @@ class AgentTeam(SpawnAnnounceProtocol):
        
        summary = self._scoped_token_summary()
        own_names = self._own_agent_names()
-        if own_names:
+        own_ids = self._own_agent_ids()
+        if own_names or own_ids:
            # Only surface this instance's own interactions so the detailed
            # report never exposes another concurrent instance's records.
-            recent = self._scoped_recent_interactions(own_names)[-20:]
+            recent = self._scoped_recent_interactions(own_names, own_ids)[-20:]
        else:
            recent = get_token_collector().get_recent_interactions(limit=20)
        
@@ -1,532 +0,0 @@
-"""
-Subagent Delegator for PraisonAI Agents.
-
-Provides delegation primitives for spawning subagents with scoped permissions
-and context. Inspired by Gemini CLI's delegate-to-agent pattern.
-"""
-
-from praisonaiagents._logging import get_logger
-import asyncio
-import warnings
-import threading
-from dataclasses import dataclass, field
-from typing import Any, Dict, List, Optional, Callable
-from enum import Enum
-
-from .profiles import AgentProfile, AgentMode, get_profile, BUILTIN_PROFILES
-
-logger = get_logger(__name__)
-
-class DelegationStatus(Enum):
-    """Status of a delegated task."""
-    PENDING = "pending"
-    RUNNING = "running"
-    COMPLETED = "completed"
-    FAILED = "failed"
-    CANCELLED = "cancelled"
-    TIMEOUT = "timeout"
-
-@dataclass
-class DelegationConfig:
-    """Configuration for subagent delegation."""
-    # Concurrency limits
-    max_concurrent_subagents: int = 3
-    max_total_subagents: int = 10
-    
-    # Timeouts
-    default_timeout_seconds: float = 300.0
-    max_timeout_seconds: float = 600.0
-    
-    # Resource limits
-    max_steps_per_subagent: int = 50
-    max_tokens_per_subagent: int = 50000
-    
-    # Permissions
-    inherit_permissions: bool = True
-    allow_nested_delegation: bool = False
-    
-    # Behavior
-    auto_cancel_on_parent_cancel: bool = True
-    collect_results: bool = True
-
-@dataclass
-class DelegationTask:
-    """A task delegated to a subagent."""
-    task_id: str
-    agent_name: str
-    objective: str
-    status: DelegationStatus = DelegationStatus.PENDING
-    
-    # Configuration
-    timeout_seconds: float = 300.0
-    max_steps: int = 50
-    
-    # Context
-    context: Dict[str, Any] = field(default_factory=dict)
-    parent_session_id: Optional[str] = None
-    
-    # Results
-    result: Optional[str] = None
-    error: Optional[str] = None
-    steps_taken: int = 0
-    tokens_used: int = 0
-    
-    # Metadata
-    metadata: Dict[str, Any] = field(default_factory=dict)
-
-@dataclass
-class DelegationResult:
-    """Result of a delegation operation."""
-    task_id: str
-    agent_name: str
-    success: bool
-    result: Optional[str] = None
-    error: Optional[str] = None
-    steps_taken: int = 0
-    tokens_used: int = 0
-    duration_seconds: float = 0.0
-
-class SubagentDelegator:
-    """
-    Manages delegation of tasks to subagents.
-    
-    Provides:
-    - Spawning subagents with scoped permissions
-    - Concurrency control
-    - Timeout management
-    - Result collection
-    - Cancellation support
-    
-    Example:
-        delegator = SubagentDelegator()
-        
-        # Delegate to explorer agent
-        result = await delegator.delegate(
-            agent_name="explorer",
-            objective="Find all authentication-related files",
-            context={"workspace": "/path/to/project"}
-        )
-        
-        # Delegate multiple tasks
-        results = await delegator.delegate_parallel([
-            ("explorer", "Find auth files"),
-            ("explorer", "Find database models"),
-        ])
-    """
-    
-    def __init__(
-        self,
-        config: Optional[DelegationConfig] = None,
-        agent_factory: Optional[Callable[[str], Any]] = None,
-        on_task_complete: Optional[Callable[[DelegationResult], None]] = None,
-    ):
-        """
-        Initialize the delegator.
-        
-        Args:
-            config: Delegation configuration
-            agent_factory: Factory function to create agents
-            on_task_complete: Callback when task completes
-        """
-        warnings.warn(
-            "SubagentDelegator is deprecated. Use Agent.handoff_to() or "
-            "parallel_handoffs() instead.",
-            DeprecationWarning,
-            stacklevel=2,
-        )
-        self.config = config or DelegationConfig()
-        self.agent_factory = agent_factory
-        self.on_task_complete = on_task_complete
-        
-        # State
-        self._tasks: Dict[str, DelegationTask] = {}
-        self._running_asyncio_tasks: Dict[str, "asyncio.Task"] = {}
-        self._running_count: int = 0
-        self._total_count: int = 0
-        self._task_counter: int = 0
-        
-        # Note: asyncio.Semaphore() is created lazily via _get_semaphore() to support
-        # Python 3.9 where Semaphore() requires an event loop at creation time.
-        self.__semaphore: Optional[asyncio.Semaphore] = None
-        self.__semaphore_lock: threading.Lock = threading.Lock()  # Thread-safe initialization
-    
-    def _get_semaphore(self) -> asyncio.Semaphore:
-        """Get the semaphore, creating it lazily if needed.
-        
-        This deferred creation is required for Python 3.9 compatibility
-        where asyncio.Semaphore() calls get_event_loop() at creation time.
-        """
-        if self.__semaphore is None:
-            with self.__semaphore_lock:  # Thread-safe initialization
-                if self.__semaphore is None:  # Double-check after acquiring lock
-                    self.__semaphore = asyncio.Semaphore(self.config.max_concurrent_subagents)
-        return self.__semaphore
-    
-    def get_available_agents(self) -> List[str]:
-        """Get list of available agent names for delegation."""
-        return [
-            name for name, profile in BUILTIN_PROFILES.items()
-            if profile.mode in (AgentMode.SUBAGENT, AgentMode.ALL)
-            and not profile.hidden
-        ]
-    
-    def get_agent_description(self, agent_name: str) -> str:
-        """Get description for an agent."""
-        profile = get_profile(agent_name)
-        if profile:
-            return profile.description
-        return f"Agent: {agent_name}"
-    
-    async def delegate(
-        self,
-        agent_name: str,
-        objective: str,
-        context: Optional[Dict[str, Any]] = None,
-        timeout_seconds: Optional[float] = None,
-        max_steps: Optional[int] = None,
-    ) -> DelegationResult:
-        """
-        Delegate a task to a subagent.
-        
-        Args:
-            agent_name: Name of the agent to delegate to
-            objective: Task objective/description
-            context: Optional context to pass to agent
-            timeout_seconds: Optional timeout override
-            max_steps: Optional max steps override
-            
-        Returns:
-            DelegationResult with task outcome
-        """
-        # Check limits
-        if self._total_count >= self.config.max_total_subagents:
-            return DelegationResult(
-                task_id="",
-                agent_name=agent_name,
-                success=False,
-                error=f"Max total subagents ({self.config.max_total_subagents}) reached",
-            )
-        
-        # Create task
-        task_id = self._generate_task_id()
-        task = DelegationTask(
-            task_id=task_id,
-            agent_name=agent_name,
-            objective=objective,
-            timeout_seconds=timeout_seconds or self.config.default_timeout_seconds,
-            max_steps=max_steps or self.config.max_steps_per_subagent,
-            context=context or {},
-        )
-        
-        self._tasks[task_id] = task
-        self._total_count += 1
-        
-        # Execute with concurrency control. Wrap in a real asyncio.Task so
-        # cancel_task/cancel_all can actually propagate cancellation to the
-        # awaited subagent coroutine instead of only flipping a status enum.
-        async with self._get_semaphore():
-            # If cancelled while queued on the semaphore, don't start the work.
-            if task.status == DelegationStatus.CANCELLED:
-                return DelegationResult(
-                    task_id=task.task_id,
-                    agent_name=task.agent_name,
-                    success=False,
-                    error="Task was cancelled",
-                )
-            exec_task = asyncio.ensure_future(self._execute_task(task))
-            self._running_asyncio_tasks[task_id] = exec_task
-            try:
-                return await exec_task
-            except asyncio.CancelledError:
-                # Distinguish a delegator-initiated cancel (via cancel_task/cancel_all,
-                # which sets status=CANCELLED) from an external/cooperative cancel of
-                # the caller. Only the former is swallowed into a result; an external
-                # cancel must propagate so wait_for/shutdown behave correctly.
-                if task.status == DelegationStatus.CANCELLED:
-                    return DelegationResult(
-                        task_id=task.task_id,
-                        agent_name=task.agent_name,
-                        success=False,
-                        error="Task was cancelled",
-                    )
-                raise
-            finally:
-                self._running_asyncio_tasks.pop(task_id, None)
-    
-    async def delegate_parallel(
-        self,
-        tasks: List[tuple],
-        collect_all: bool = True,
-    ) -> List[DelegationResult]:
-        """
-        Delegate multiple tasks in parallel.
-        
-        Args:
-            tasks: List of (agent_name, objective) or (agent_name, objective, context) tuples
-            collect_all: Wait for all tasks to complete
-            
-        Returns:
-            List of DelegationResults
-        """
-        # Create coroutines
-        coros = []
-        for task_spec in tasks:
-            if len(task_spec) == 2:
-                agent_name, objective = task_spec
-                context = None
-            else:
-                agent_name, objective, context = task_spec
-            
-            coros.append(self.delegate(agent_name, objective, context))
-        
-        # Execute
-        if collect_all:
-            results = await asyncio.gather(*coros, return_exceptions=True)
-            # Convert exceptions to failed results
-            processed = []
-            for i, result in enumerate(results):
-                if isinstance(result, Exception):
-                    agent_name = tasks[i][0] if tasks[i] else "unknown"
-                    processed.append(DelegationResult(
-                        task_id="",
-                        agent_name=agent_name,
-                        success=False,
-                        error=str(result),
-                    ))
-                else:
-                    processed.append(result)
-            return processed
-        else:
-            # Return as tasks complete
-            return await asyncio.gather(*coros)
-    
-    async def cancel_task(self, task_id: str) -> bool:
-        """
-        Cancel a running task.
-        
-        Args:
-            task_id: ID of task to cancel
-            
-        Returns:
-            True if task was cancelled
-        """
-        task = self._tasks.get(task_id)
-        # Only terminal tasks are un-cancellable; PENDING/RUNNING can be cancelled.
-        _terminal = (
-            DelegationStatus.COMPLETED,
-            DelegationStatus.FAILED,
-            DelegationStatus.CANCELLED,
-            DelegationStatus.TIMEOUT,
-        )
-        if not task or task.status in _terminal:
-            return False
-        
-        # Mark cancelled first so a PENDING task (still waiting on the semaphore)
-        # or a RUNNING coroutine both observe CANCELLED and refuse to complete.
-        task.status = DelegationStatus.CANCELLED
-        
-        # Propagate cancellation to the actual coroutine if it is already running.
-        asyncio_task = self._running_asyncio_tasks.get(task_id)
-        if asyncio_task is not None:
-            asyncio_task.cancel()
-        return True
-    
-    async def cancel_all(self) -> int:
-        """
-        Cancel all running tasks.
-        
-        Returns:
-            Number of tasks cancelled
-        """
-        cancelled = 0
-        for task_id in list(self._tasks.keys()):
-            if await self.cancel_task(task_id):
-                cancelled += 1
-        return cancelled
-    
-    def get_task_status(self, task_id: str) -> Optional[DelegationStatus]:
-        """Get status of a task."""
-        task = self._tasks.get(task_id)
-        return task.status if task else None
-    
-    def get_running_tasks(self) -> List[DelegationTask]:
-        """Get all running tasks."""
-        return [
-            t for t in self._tasks.values()
-            if t.status == DelegationStatus.RUNNING
-        ]
-    
-    def get_stats(self) -> Dict[str, Any]:
-        """Get delegation statistics."""
-        status_counts = {}
-        for task in self._tasks.values():
-            status = task.status.value
-            status_counts[status] = status_counts.get(status, 0) + 1
-        
-        return {
-            "total_tasks": self._total_count,
-            "running_tasks": self._running_count,
-            "status_counts": status_counts,
-            "max_concurrent": self.config.max_concurrent_subagents,
-            "max_total": self.config.max_total_subagents,
-        }
-    
-    async def _execute_task(self, task: DelegationTask) -> DelegationResult:
-        """Execute a delegation task."""
-        import time
-        start_time = time.time()
-        
-        # A cancel may have landed between scheduling and execution; honour it.
-        if task.status == DelegationStatus.CANCELLED:
-            return DelegationResult(
-                task_id=task.task_id,
-                agent_name=task.agent_name,
-                success=False,
-                error="Task was cancelled",
-            )
-        
-        task.status = DelegationStatus.RUNNING
-        self._running_count += 1
-        
-        try:
-            # Get agent profile
-            profile = get_profile(task.agent_name)
-            if not profile:
-                raise ValueError(f"Unknown agent: {task.agent_name}")
-            
-            # Create agent
-            agent = await self._create_agent(profile, task)
-            
-            # Execute with timeout
-            try:
-                result = await asyncio.wait_for(
-                    self._run_agent(agent, task),
-                    timeout=task.timeout_seconds,
-                )
-                task.result = result
-                # Don't clobber a cancellation that landed while we were running.
-                if task.status != DelegationStatus.CANCELLED:
-                    task.status = DelegationStatus.COMPLETED
-                
-            except asyncio.TimeoutError:
-                task.status = DelegationStatus.TIMEOUT
-                task.error = f"Task timed out after {task.timeout_seconds}s"
-            
-        except Exception as e:
-            task.status = DelegationStatus.FAILED
-            task.error = str(e)
-            logger.exception(f"Delegation task {task.task_id} failed")
-        
-        finally:
-            self._running_count -= 1
-            duration = time.time() - start_time
-        
-        # Build result
-        result = DelegationResult(
-            task_id=task.task_id,
-            agent_name=task.agent_name,
-            success=task.status == DelegationStatus.COMPLETED,
-            result=task.result,
-            error=task.error,
-            steps_taken=task.steps_taken,
-            tokens_used=task.tokens_used,
-            duration_seconds=duration,
-        )
-        
-        # Callback
-        if self.on_task_complete:
-            try:
-                self.on_task_complete(result)
-            except Exception as e:
-                logger.warning(f"Task complete callback error: {e}")
-        
-        return result
-    
-    async def _create_agent(self, profile: AgentProfile, task: DelegationTask) -> Any:
-        """Create an agent from profile using consolidated params."""
-        if self.agent_factory:
-            return self.agent_factory(profile.name)
-        
-        # Default: create Agent from praisonaiagents
-        try:
-            from ..agent.agent import Agent
-            
-            # Filter tools based on profile
-            tools = self._get_tools_for_profile(profile)
-            
-            # Use consolidated params (protocol-compliant)
-            # output='silent' replaces verbose=False
-            # execution={max_iter: N} replaces max_iter=N
-            agent = Agent(
-                name=profile.name,
-                role=profile.name.title(),
-                goal=task.objective,
-                backstory=profile.system_prompt,
-                tools=tools,
-                output='silent',
-                execution={'max_iter': min(task.max_steps, profile.max_steps)},
-            )
-            
-            return agent
-            
-        except ImportError:
-            raise RuntimeError("Could not import Agent class")
-    
-    async def _run_agent(self, agent: Any, task: DelegationTask) -> str:
-        """Run an agent on a task."""
-        # Build prompt with context
-        prompt = task.objective
-        
-        if task.context:
-            context_str = "\n".join(
-                f"- {k}: {v}" for k, v in task.context.items()
-            )
-            prompt = f"Context:\n{context_str}\n\nTask: {task.objective}"
-        
-        # Execute
-        if hasattr(agent, 'chat'):
-            result = agent.chat(prompt)
-        elif hasattr(agent, 'run'):
-            result = agent.run(prompt)
-        else:
-            raise RuntimeError("Agent has no chat or run method")
-        
-        return str(result) if result else ""
-    
-    def _get_tools_for_profile(self, profile: AgentProfile) -> List[Any]:
-        """Get tools for a profile, respecting restrictions."""
-        # For now, return empty list - tools should be injected
-        # In production, this would filter available tools
-        return []
-    
-    def _generate_task_id(self) -> str:
-        """Generate a unique task ID."""
-        self._task_counter += 1
-        return f"task_{self._task_counter}"
-
-# Convenience function for quick delegation
-async def delegate_to_agent(
-    agent_name: str,
-    objective: str,
-    context: Optional[Dict[str, Any]] = None,
-    timeout_seconds: float = 300.0,
-) -> DelegationResult:
-    """
-    Quick delegation to a subagent.
-    
-    Args:
-        agent_name: Name of agent to delegate to
-        objective: Task objective
-        context: Optional context
-        timeout_seconds: Timeout in seconds
-        
-    Returns:
-        DelegationResult
-    """
-    delegator = SubagentDelegator()
-    return await delegator.delegate(
-        agent_name=agent_name,
-        objective=objective,
-        context=context,
-        timeout_seconds=timeout_seconds,
-    )
@@ -213,7 +213,11 @@ class AllowedToolsFilter:
            "env_var_name": self.env_var_name,
            "env_value": self.env_value,
            "is_ci": self.is_ci,
-            "whitelist": list(self._whitelist) if self._whitelist else None,
+            # sorted, not list(): _whitelist is a set, so list() gave a
+            # different order on different runs. A diagnostic report you cannot
+            # diff between two runs is much less useful for the debugging it
+            # exists for.
+            "whitelist": sorted(self._whitelist) if self._whitelist else None,
            **self._diagnostics
        }
    
@@ -13,7 +13,7 @@ from __future__ import annotations
 import asyncio
 import re
 from praisonaiagents._logging import get_logger
-from typing import Any
+from typing import Any, Optional

 from .protocols import ApprovalDecision, ApprovalRequest

@@ -175,7 +175,7 @@ class AutoApproveBackend:
    async def request_approval(self, request: ApprovalRequest) -> ApprovalDecision:
        return ApprovalDecision(approved=True, reason="auto-approved", approver="system")

-    def request_approval_sync(self, request: ApprovalRequest) -> ApprovalDecision:
+    def request_approval_sync(self, request: ApprovalRequest, timeout: Optional[float] = None) -> ApprovalDecision:
        return ApprovalDecision(approved=True, reason="auto-approved", approver="system")

 class ConsoleBackend:
@@ -293,7 +293,7 @@ class ConsoleBackend:
            return (True, "always", suggested or None, None)
        return (True, "once", None, None)

-    def request_approval_sync(self, request: ApprovalRequest) -> ApprovalDecision:
+    def request_approval_sync(self, request: ApprovalRequest, timeout: Optional[float] = None) -> ApprovalDecision:
        """Synchronous approval via Rich console prompt."""
        try:
            approved, scope, scope_pattern, feedback = self._prompt_user(request)
@@ -474,10 +474,12 @@ class AgentApproval:
                reason=f"Agent approval error: {e}",
            )

-    def request_approval_sync(self, request: ApprovalRequest) -> ApprovalDecision:
+    def request_approval_sync(self, request: ApprovalRequest, timeout: Optional[float] = None) -> ApprovalDecision:
        """Synchronous wrapper."""
        from .utils import run_coroutine_safely
-        return run_coroutine_safely(self.request_approval(request), timeout=60)
+        return run_coroutine_safely(
+            self.request_approval(request), timeout=60 if timeout is None else timeout
+        )

 class CallbackBackend:
    """Wraps a legacy ``(function_name, arguments, risk_level) -> ApprovalDecision`` callback
@@ -489,7 +491,7 @@ class CallbackBackend:
    def __init__(self, callback):
        self._callback = callback

-    def request_approval_sync(self, request: ApprovalRequest) -> ApprovalDecision:
+    def request_approval_sync(self, request: ApprovalRequest, timeout: Optional[float] = None) -> ApprovalDecision:
        result = self._callback(request.tool_name, request.arguments, request.risk_level)
        if isinstance(result, ApprovalDecision):
            return result
@@ -189,6 +189,18 @@ class ApprovalProtocol(Protocol):
        """
        ...

+    # Backends MAY also implement an optional synchronous entry point:
+    #
+    #     def request_approval_sync(self, request, timeout=None) -> ApprovalDecision
+    #
+    # It is deliberately NOT declared on this runtime-checkable Protocol so an
+    # async-only backend still satisfies ``isinstance(b, ApprovalProtocol)``.
+    # When present, the per-call ``timeout`` (seconds; ``None`` = wait forever)
+    # is passed explicitly by the caller so shared backend state is never
+    # mutated — mutating a shared backend's timeout leaks between agents.
+    # Legacy sync backends without the ``timeout`` parameter are still
+    # supported; the caller falls back to a no-argument call on ``TypeError``.
+

@runtime_checkable
 class ApprovalStoreProtocol(Protocol):
@@ -31,16 +31,24 @@ class QwenCliAuth:
    name = "qwen-cli"

    def resolve_credentials(self) -> SubscriptionCredentials:
+        """Resolve Qwen CLI credentials.
+
+        The origin comes from the `PRAISONAI_QWEN_BASE_URL` environment variable when it is set
+        to a non-empty value, and otherwise from the Qwen default.
+        """
        tokens = _read_qwen_tokens()
        if not tokens:
            raise AuthError(
                "No Qwen CLI credentials at ~/.qwen/oauth_creds.json. "
                "Install Qwen CLI and run 'qwen login'."
            )
+        # Resolve once and use the same value for both: `headers_for` takes the origin the request
+        # will actually go to, which is not necessarily the Qwen default.
+        base_url = os.environ.get("PRAISONAI_QWEN_BASE_URL", "").strip() or _QWEN_BASE_URL
        return SubscriptionCredentials(
            api_key=tokens["access_token"],
-            base_url=os.environ.get("PRAISONAI_QWEN_BASE_URL", "").strip() or _QWEN_BASE_URL,
-            headers=self.headers_for(_QWEN_BASE_URL, ""),
+            base_url=base_url,
+            headers=self.headers_for(base_url, ""),
            auth_scheme="bearer",
            source="qwen-cli-file",
        )
@@ -69,6 +69,13 @@ class BackgroundRunner:
        """
        self.config = config or BackgroundConfig()
        self._tasks: Dict[str, BackgroundTask] = {}
+        # Guards every read/write of ``_tasks``. This runner is a process-wide
+        # singleton (``get_background_runner``) whose sync wrappers insert on a
+        # separate background-loop thread via ``run_coroutine_threadsafe`` while
+        # other surfaces read/mutate the dict on their own thread; without this
+        # a concurrent insert + iterate/delete can raise
+        # ``RuntimeError: dictionary changed size during iteration``.
+        self._tasks_lock: threading.Lock = threading.Lock()
        # Lazily create asyncio.Semaphore() to avoid Python 3.9 event loop issues
        self._semaphore: Optional[asyncio.Semaphore] = None
        self._semaphore_lock: threading.Lock = threading.Lock()  # Thread-safe initialization
@@ -86,17 +93,20 @@ class BackgroundRunner:
    @property
    def tasks(self) -> List[BackgroundTask]:
        """Get all tasks."""
-        return list(self._tasks.values())
+        with self._tasks_lock:
+            return list(self._tasks.values())
    
    @property
    def running_tasks(self) -> List[BackgroundTask]:
        """Get currently running tasks."""
-        return [t for t in self._tasks.values() if t.is_running]
+        with self._tasks_lock:
+            return [t for t in self._tasks.values() if t.is_running]
    
    @property
    def pending_tasks(self) -> List[BackgroundTask]:
        """Get pending tasks."""
-        return [t for t in self._tasks.values() if t.status == TaskStatus.PENDING]
+        with self._tasks_lock:
+            return [t for t in self._tasks.values() if t.status == TaskStatus.PENDING]
    
    async def submit(
        self,
@@ -133,7 +143,8 @@ class BackgroundRunner:
            }
        )
        
-        self._tasks[task.id] = task
+        with self._tasks_lock:
+            self._tasks[task.id] = task
        
        # Create the execution coroutine
        async def execute():
@@ -235,7 +246,8 @@ class BackgroundRunner:
    
    def get_task(self, task_id: str) -> Optional[BackgroundTask]:
        """Get a task by ID."""
-        return self._tasks.get(task_id)
+        with self._tasks_lock:
+            return self._tasks.get(task_id)
    
    async def cancel_task(self, task_id: str) -> bool:
        """
@@ -247,7 +259,8 @@ class BackgroundRunner:
        Returns:
            True if cancelled, False if not found or already completed
        """
-        task = self._tasks.get(task_id)
+        with self._tasks_lock:
+            task = self._tasks.get(task_id)
        if task is None:
            return False
        
@@ -276,10 +289,12 @@ class BackgroundRunner:
        Returns:
            List of completed tasks
        """
-        futures = [t._future for t in self._tasks.values() if t._future]
+        with self._tasks_lock:
+            futures = [t._future for t in self._tasks.values() if t._future]
        
        if not futures:
-            return list(self._tasks.values())
+            with self._tasks_lock:
+                return list(self._tasks.values())
        
        try:
            await asyncio.wait_for(
@@ -289,16 +304,18 @@ class BackgroundRunner:
        except asyncio.TimeoutError:
            pass
        
-        return list(self._tasks.values())
+        with self._tasks_lock:
+            return list(self._tasks.values())
    
    def clear_completed(self):
        """Remove completed tasks from tracking."""
-        completed_ids = [
-            task_id for task_id, task in self._tasks.items()
-            if task.is_completed
-        ]
-        for task_id in completed_ids:
-            del self._tasks[task_id]
+        with self._tasks_lock:
+            completed_ids = [
+                task_id for task_id, task in self._tasks.items()
+                if task.is_completed
+            ]
+            for task_id in completed_ids:
+                del self._tasks[task_id]
    
    def list_tasks(self, status: Optional[TaskStatus] = None) -> List[Dict[str, Any]]:
        """
@@ -310,7 +327,8 @@ class BackgroundRunner:
        Returns:
            List of task dictionaries
        """
-        tasks = self._tasks.values()
+        with self._tasks_lock:
+            tasks = list(self._tasks.values())
        
        if status:
            tasks = [t for t in tasks if t.status == status]
@@ -18,7 +18,10 @@ from .protocols import (
    BotMessage,
    BotUser,
    BotChannel,
+    QuotedRef,
    MessageType,
+    PlatformEvent,
+    PlatformEventKind,
    ChatCommandInfo,
    ChatCommandProtocol,
    BotOSProtocol,
@@ -135,10 +138,13 @@ __all__ = [
    "BotMessage",
    "BotUser",
    "BotChannel",
+    "QuotedRef",
    "BotConfig",
    "DisplayPolicy",
    "resolve_display_policy",
    "MessageType",
+    "PlatformEvent",
+    "PlatformEventKind",
    "ChatCommandInfo",
    "ChatCommandProtocol",
    "BotOSProtocol",
@@ -125,11 +125,29 @@ class BotConfig:
    
    # Edit interval for streaming responses in milliseconds (default: 700ms)
    stream_edit_interval_ms: int = 700
+
+    # Opt-in privacy-safe runtime footer on the final reply (default: False).
+    # When True, a short trailing line (e.g. ``— gpt-4o · 4.2s``) reporting the
+    # model and turn latency is appended to the final reply only. Privacy-safe:
+    # never surfaces tool args, URLs, paths or message content. Unavailable
+    # fields are omitted rather than shown as placeholders.
+    footer: bool = False
    
    # Intentional silence support
    allow_silence: bool = False
    silence_token: Optional[str] = None

+    # Bot-to-bot inbound (#5062). Off by default: a bot-authored inbound
+    # message is dropped as before. When ``allow_bots=True`` the channel accepts
+    # bot-authored messages and the runaway-loop guard (``BotLoopGuard``) is
+    # auto-active, tracking each participant pair and suppressing further replies
+    # once ``bot_loop_protection``'s sliding-window budget is exceeded.
+    # ``bot_loop_protection`` is a plain dict (enabled / max_events_per_window /
+    # window_seconds / cooldown_seconds) coerced into a ``BotLoopPolicy``; ``None``
+    # uses the default budget.
+    allow_bots: bool = False
+    bot_loop_protection: Optional[Dict[str, Any]] = None
+
    def __post_init__(self) -> None:
        if self.unknown_user_policy not in {"deny", "pair", "allow"}:
            raise ValueError(
@@ -174,8 +192,11 @@ class BotConfig:
            "owner_user_id": "***" if self.owner_user_id else None,
            "streaming": self.streaming,
            "stream_edit_interval_ms": self.stream_edit_interval_ms,
+            "footer": self.footer,
            "allow_silence": self.allow_silence,
            "silence_token": self.silence_token,
+            "allow_bots": self.allow_bots,
+            "bot_loop_protection": self.bot_loop_protection,
            "metadata": self.metadata,
        }
    
@@ -21,6 +21,7 @@ from typing import (
    Callable,
    Dict,
    List,
+    Literal,
    Mapping,
    Optional,
    Protocol,
@@ -496,6 +497,43 @@ class BotChannel:
        )


+@dataclass
+class QuotedRef:
+    """Resolved content of a message a user replied to or quoted.
+
+    Chat platforms carry only the *id* of a quoted message on the admission
+    path; the referenced text/media must be resolved and attached so the agent
+    can honour the referent ("do the second one", "why?"). This is that
+    resolved reference, rendered into the prompt as a compact quoted block.
+
+    Attributes:
+        message_id: Platform-specific id of the quoted message.
+        text: Resolved text of the quoted message ("" if unavailable).
+        author: ``"bot"`` when quoting the bot's own message, else ``"user"``.
+    """
+
+    message_id: str = ""
+    text: str = ""
+    author: str = ""  # "bot" | "user"
+
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "message_id": self.message_id,
+            "text": self.text,
+            "author": self.author,
+        }
+
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "QuotedRef":
+        """Create from dictionary."""
+        return cls(
+            message_id=data.get("message_id", ""),
+            text=data.get("text", ""),
+            author=data.get("author", ""),
+        )
+
+
@dataclass
 class BotMessage:
    """Represents a message in a messaging platform.
@@ -510,6 +548,11 @@ class BotMessage:
        reply_to: ID of message being replied to
        thread_id: Thread identifier (for threaded conversations)
        attachments: List of attachment URLs or data
+        quoted: Resolved reply/quote context (see :class:`QuotedRef`). When a
+            user replies to or quotes an earlier message the adapter resolves
+            the referenced content and attaches it here so the agent sees the
+            referent instead of answering context-blind. Rendered into the
+            prompt as a quoted block by :meth:`prompt_text`.
        metadata: Additional platform-specific metadata
        allow_control: Control-trust primitive. ``True`` (the default) only for
            interactive human turns; producers set it ``False`` for content of
@@ -530,6 +573,7 @@ class BotMessage:
    reply_to: Optional[str] = None
    thread_id: Optional[str] = None
    attachments: List[Dict[str, Any]] = field(default_factory=list)
+    quoted: Optional[QuotedRef] = None
    metadata: Dict[str, Any] = field(default_factory=dict)
    allow_control: bool = True
    
@@ -545,6 +589,7 @@ class BotMessage:
            "reply_to": self.reply_to,
            "thread_id": self.thread_id,
            "attachments": self.attachments,
+            "quoted": self.quoted.to_dict() if self.quoted else None,
            "metadata": self.metadata,
            "allow_control": self.allow_control,
        }
@@ -571,6 +616,7 @@ class BotMessage:
            reply_to=data.get("reply_to"),
            thread_id=data.get("thread_id"),
            attachments=data.get("attachments", []),
+            quoted=QuotedRef.from_dict(data["quoted"]) if data.get("quoted") else None,
            metadata=data.get("metadata", {}),
            allow_control=data.get("allow_control", True),
        )
@@ -581,6 +627,31 @@ class BotMessage:
        if isinstance(self.content, str):
            return self.content
        return self.content.get("text", "")
+
+    @property
+    def prompt_text(self) -> str:
+        """Text for the agent turn, with any resolved quote rendered inline.
+
+        When the user replied to or quoted an earlier message, the resolved
+        reference (:attr:`quoted`) is rendered as a compact quoted block above
+        the new text so the model can honour the referent, e.g.::
+
+            [In reply to: "Here are three options: 1) …  2) …  3) …"]
+            do the second one
+
+        With no quote this is identical to :attr:`text`. The quoted content is
+        injected as context only, never as a control frame — consistent with
+        the ``allow_control`` primitive.
+        """
+        body = self.text
+        quoted = self.quoted
+        if not quoted or not quoted.text.strip():
+            return body
+        snippet = " ".join(quoted.text.split())
+        if len(snippet) > 500:
+            snippet = snippet[:497] + "…"
+        block = f'[In reply to: "{snippet}"]'
+        return f"{block}\n{body}" if body else block
    
    @property
    def is_command(self) -> bool:
@@ -620,6 +691,76 @@ class BotMessage:
        return parts[1:] if len(parts) > 1 else []


+# Normalised inbound platform-event kinds. Mirrors how ``MessageType`` normalises
+# inbound messages: a small, portable vocabulary every adapter maps its native
+# event onto (Telegram ``MessageReactionUpdated`` / Discord ``on_raw_reaction_add``
+# / Slack ``reaction_added`` all become ``"reaction_added"``), so hooks/plugins
+# have one uniform surface to subscribe to (Issue #5161).
+PlatformEventKind = Literal[
+    "reaction_added",
+    "reaction_removed",
+    "message_edited",
+    "message_deleted",
+    "member_joined",
+    "member_left",
+    "thread_created",
+]
+
+
+@dataclass
+class PlatformEvent:
+    """A normalised inbound platform event that is not a text/media message.
+
+    Reactions, edits, deletions, membership changes and thread lifecycle are
+    delivered by the platform SDK but are not messages. This is the single
+    shared, protocol-only contract every adapter emits into and every hook or
+    plugin consumes — the inbound counterpart to the outbound ``ReactionResult``
+    / message-action surface the bot already ships.
+
+    Adapters build one of these from the native event and emit it via
+    ``MessageHookMixin.fire_platform_event`` (praisonai-bot), which routes it to
+    the matching :class:`~praisonaiagents.hooks.types.HookEvent`
+    (``REACTION_RECEIVED``/``MESSAGE_EDITED``/``MESSAGE_DELETED``/
+    ``MEMBER_JOINED``/``MEMBER_LEFT``/``THREAD_CREATED``). Platforms that cannot
+    deliver a given event simply never build one — graceful, capability-gated,
+    exactly like the outbound ``capabilities["reactions"]`` gate.
+
+    Attributes:
+        kind: Which normalised event this is (see :data:`PlatformEventKind`).
+        platform: Emitting platform name (``"telegram"``/``"discord"``/…).
+        chat_id: Channel/chat the event occurred in.
+        user_id: The user who caused the event (reactor, editor, joiner, …).
+        message_id: The message the event targets (reactions/edits/deletes).
+        emoji: The reaction emoji, for reaction add/remove.
+        new_text: The new content, for an edit.
+        thread_id: The thread identifier, for thread creation / threaded events.
+        raw: Escape hatch to the native SDK payload for adapter-specific needs.
+    """
+
+    kind: PlatformEventKind
+    platform: str
+    chat_id: str
+    user_id: str
+    message_id: Optional[str] = None
+    emoji: Optional[str] = None
+    new_text: Optional[str] = None
+    thread_id: Optional[str] = None
+    raw: Optional[Dict[str, Any]] = None
+
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary (native ``raw`` payload intentionally omitted)."""
+        return {
+            "kind": self.kind,
+            "platform": self.platform,
+            "chat_id": self.chat_id,
+            "user_id": self.user_id,
+            "message_id": self.message_id,
+            "emoji": self.emoji,
+            "new_text": self.new_text,
+            "thread_id": self.thread_id,
+        }
+
+
@runtime_checkable
 class BotMessageProtocol(Protocol):
    """Protocol for bot message handling."""
@@ -138,15 +138,31 @@ class BotLoopPolicy:
        """Build a policy from a config mapping (YAML/kwargs), tolerating None.

        Unknown keys are ignored so a forward-compatible config never breaks
-        construction. ``None`` yields the enabled default policy.
+        construction. ``None`` yields the enabled default policy. Malformed
+        numeric values (e.g. ``max_events_per_window: "abc"``) fall back to the
+        field default rather than raising, so a config typo cannot crash the
+        inbound path on the first bot message — the guard stays fail-safe.
        """
        if not data:
            return cls()
+
+        def _as_int(value: object, default: int) -> int:
+            try:
+                return int(value)  # type: ignore[arg-type]
+            except (TypeError, ValueError):
+                return default
+
+        def _as_float(value: object, default: float) -> float:
+            try:
+                return float(value)  # type: ignore[arg-type]
+            except (TypeError, ValueError):
+                return default
+
        return cls(
            enabled=bool(data.get("enabled", True)),
-            max_events_per_window=int(data.get("max_events_per_window", 20)),
-            window_seconds=float(data.get("window_seconds", 60.0)),
-            cooldown_seconds=float(data.get("cooldown_seconds", 60.0)),
+            max_events_per_window=_as_int(data.get("max_events_per_window", 20), 20),
+            window_seconds=_as_float(data.get("window_seconds", 60.0), 60.0),
+            cooldown_seconds=_as_float(data.get("cooldown_seconds", 60.0), 60.0),
        )


@@ -172,18 +172,16 @@ class EventBus:
        Returns:
            The published Event object
        """
-        # Fast path: if no subscribers and no durable sinks, return a minimal
-        # event without expensive operations.
-        if not self._subscribers and not self._sinks:
-            # Convert EventType enum to string
-            type_str = event_type.value if isinstance(event_type, EventType) else event_type
-            return Event(
-                type=type_str,
-                data=data or {},
-                source=source,
-                metadata=metadata or {},
-            )
-        
+        # There is no separate no-subscriber shortcut here. It built the same
+        # Event and returned it, so the only work it saved was the sink
+        # dispatch and the history append below -- and skipping the append
+        # meant an event published with nobody listening vanished from
+        # get_history(). History is a debugging record of what was published;
+        # whether anyone happened to be subscribed at the time is not something
+        # it should silently depend on. The append is O(1) under a lock already
+        # held and the list is capped at _max_history, so the saving was not
+        # worth the hole it left.
+
        # Convert EventType enum to string
        type_str = event_type.value if isinstance(event_type, EventType) else event_type
        
@@ -212,16 +210,19 @@ class EventBus:
        # the event even when there are no in-memory subscribers.
        self._dispatch_to_sinks(event)
        
-        # Fast path: if no subscribers, skip expensive work
-        if not self._subscribers:
-            return event
-        
-        # Store in history
+        # Record the event before considering subscribers, for the same reason
+        # _dispatch_to_sinks runs above: an event that was published happened,
+        # whether or not anyone was listening for it.
        with self._lock:
            self._event_history.append(event)
            if len(self._event_history) > self._max_history:
                self._event_history = self._event_history[-self._max_history:]
-            
+
+        # Fast path: with nothing subscribed there is no dispatch to do.
+        if not self._subscribers:
+            return event
+
+        with self._lock:
            # Get matching subscribers
            subscribers = [
                sub for sub in self._subscribers
@@ -266,17 +267,13 @@ class EventBus:
            The published Event object
        """
        type_str = event_type.value if isinstance(event_type, EventType) else event_type
-        
-        # Fast path: if no subscribers and no durable sinks, return a minimal
-        # event without expensive operations.
-        if not self._subscribers and not self._sinks:
-            return Event(
-                type=type_str,
-                data=data or {},
-                source=source,
-                metadata=metadata or {},
-            )
-        
+
+        # No pre-history no-subscriber shortcut here, matching publish_event's
+        # already-fixed shape: an event that was published happened, whether or
+        # not anyone was listening, and skipping the history append below would
+        # make it vanish from get_history(). The append is O(1) under a lock and
+        # the list is capped, so the only work worth skipping is subscriber
+        # matching/dispatch -- which the guard after the append still skips.
        event = Event(
            type=type_str,
            data=data or {},
@@ -289,15 +286,21 @@ class EventBus:
        # Durable persistence first (best-effort).
        self._dispatch_to_sinks(event)
        
-        # Fast path: if no subscribers, skip expensive work
-        if not self._subscribers:
-            return event
-        
-        # Store in history
+        # History is recorded BEFORE the no-subscriber fast path. Appending to a
+        # capped list is not the "expensive work" that path exists to skip --
+        # that is subscriber matching and dispatch. Returning first meant
+        # get_history() stayed empty unless something happened to be subscribed,
+        # so the record of what the system did depended on who was watching.
        with self._lock:
            self._event_history.append(event)
            if len(self._event_history) > self._max_history:
                self._event_history = self._event_history[-self._max_history:]
+
+        # Fast path: if no subscribers, skip subscriber matching and dispatch.
+        if not self._subscribers:
+            return event
+
+        with self._lock:
            
            subscribers = [
                sub for sub in self._subscribers
@@ -329,24 +332,20 @@ class EventBus:
        """
        Get recent event history.

-        Only events published while at least one subscriber was registered are
-        recorded. #2066 added a fast path that skips history (and the lock, and
-        Event construction) when ``has_subscribers`` is False, because memory
-        and sub-agent lifecycle publishes were paying uuid4 + lock + append on
-        every call with nothing listening.
-
-        The consequence is worth stating plainly, because it is surprising in
-        the case you would most want history -- debugging after the fact, with
-        nothing subscribed: publishing three events to a bus with no
-        subscribers and then calling this returns an empty list, not three
-        events. Subscribe (even a no-op) before publishing if you need a record.
+        Every event passed to ``publish``/``publish_event``/``publish_async`` is
+        recorded here (capped at ``_max_history``), regardless of whether any
+        subscriber was registered at publish time. History is a debugging record
+        of what the system did; it does not silently depend on who was watching.
+        Both the sync and async publish paths append before any no-subscriber
+        fast path runs, so publishing to a bus with nothing subscribed still
+        shows up here.

        Args:
            event_type: Optional filter by event type
            limit: Maximum number of events to return

        Returns:
-            List of recent events published while subscribers existed
+            List of recent events published (subscribed or not)
        """
        with self._lock:
            events = self._event_history.copy()
@@ -83,7 +83,8 @@ class CheckpointService:
        auto_checkpoint: bool = True,
        max_checkpoints: int = 100,
        user_name: str = "PraisonAI Checkpoints",
-        user_email: str = "[email protected]"
+        user_email: str = "[email protected]",
+        max_file_size: int = 2 * 1024 * 1024,
    ):
        """
        Initialize the checkpoint service.
@@ -96,6 +97,8 @@ class CheckpointService:
            max_checkpoints: Maximum checkpoints to keep
            user_name: Git user.name for commits (default: "PraisonAI Checkpoints")
            user_email: Git user.email for commits (default: "[email protected]")
+            max_file_size: Skip staging files larger than this many bytes
+                (default 2 MB); 0 disables the cap.
        """
        self.config = CheckpointConfig(
            workspace_dir=workspace_dir,
@@ -104,10 +107,14 @@ class CheckpointService:
            auto_checkpoint=auto_checkpoint,
            max_checkpoints=max_checkpoints,
            user_name=user_name,
-            user_email=user_email
+            user_email=user_email,
+            max_file_size=max_file_size,
        )
        
        self._initialized = False
+        # Save counter used to run `git gc --auto` on a bounded cadence so the
+        # shadow object store is periodically pruned during long sessions.
+        self._saves_since_gc = 0
        self._checkpoints: List[Checkpoint] = []
        self._event_handlers: Dict[CheckpointEvent, List[Callable]] = {
            event: [] for event in CheckpointEvent
@@ -184,6 +191,78 @@ class CheckpointService:
        
        return True
    
+    async def _get_project_objects_dir(self) -> Optional[str]:
+        """
+        Return the object database directory of the project's real git repo if
+        the workspace is inside one, else None.
+
+        Uses ``--git-common-dir`` so worktrees resolve to the shared object
+        store. Best-effort: any failure (no git, not a repo) returns None so the
+        shadow repo falls back to the standalone path unchanged.
+        """
+        try:
+            env = self._get_sanitized_env()
+            for flag in ("--git-common-dir", "--git-dir"):
+                process = await asyncio.create_subprocess_exec(
+                    "git", "rev-parse", flag,
+                    stdout=asyncio.subprocess.PIPE,
+                    stderr=asyncio.subprocess.PIPE,
+                    env=env,
+                    cwd=self.config.workspace_dir,
+                )
+                stdout, _ = await asyncio.wait_for(process.communicate(), timeout=30)
+                if process.returncode != 0:
+                    continue
+                git_dir = stdout.decode().strip()
+                if not git_dir:
+                    continue
+                if not os.path.isabs(git_dir):
+                    git_dir = os.path.join(self.config.workspace_dir, git_dir)
+                objects_dir = os.path.join(git_dir, "objects")
+                if os.path.isdir(objects_dir):
+                    return os.path.realpath(objects_dir)
+        except Exception as e:
+            logger.debug(f"Could not resolve project git objects dir: {e}")
+        return None
+
+    async def _apply_large_repo_config(self) -> None:
+        """
+        Apply large-repo git settings so index writes and untracked scans stay
+        fast. Each setting is best-effort: older git versions may not support
+        every key, and a failure must never break checkpointing.
+        """
+        settings = [
+            ("feature.manyFiles", "true"),
+            ("index.version", "4"),
+            ("core.untrackedCache", "true"),
+            ("core.fsmonitor", "true"),
+        ]
+        for key, value in settings:
+            try:
+                await self._run_git("config", key, value)
+            except Exception as e:
+                logger.debug(f"Skipping large-repo config {key}={value}: {e}")
+
+    async def _seed_from_project_objects(self, objects_dir: str) -> bool:
+        """
+        Point the shadow object store at the project's objects via a file-based
+        ``objects/info/alternates`` so new blobs resolve against existing
+        objects instead of being re-hashed and re-stored.
+
+        A file is used (not GIT_ALTERNATE_OBJECT_DIRECTORIES) because
+        ``_get_sanitized_env`` strips that env var. Returns True on success.
+        """
+        try:
+            info_dir = os.path.join(self.git_dir, "objects", "info")
+            os.makedirs(info_dir, exist_ok=True)
+            with open(os.path.join(info_dir, "alternates"), "w") as f:
+                f.write(f"{objects_dir}\n")
+            logger.debug(f"Seeded shadow git objects from {objects_dir}")
+            return True
+        except Exception as e:
+            logger.debug(f"Could not seed shadow git from project objects: {e}")
+            return False
+
    async def _init_shadow_git(self) -> bool:
        """Initialize the shadow git repository."""
        try:
@@ -196,7 +275,17 @@ class CheckpointService:
            
            # Set worktree to workspace
            await self._run_git("config", "core.worktree", self.config.workspace_dir)
-            
+
+            # Seed objects from the project's real repo (if any) so the first
+            # checkpoint only stores changed files, and tune for large repos.
+            # Both are best-effort: no real repo → unchanged standalone path.
+            objects_dir = await self._get_project_objects_dir()
+            if objects_dir and objects_dir != os.path.realpath(
+                os.path.join(self.git_dir, "objects")
+            ):
+                await self._seed_from_project_objects(objects_dir)
+            await self._apply_large_repo_config()
+
            # Create exclude file
            exclude_dir = os.path.join(self.git_dir, "info")
            os.makedirs(exclude_dir, exist_ok=True)
@@ -259,6 +348,64 @@ class CheckpointService:
            env.pop(var, None)
        
        return env
+
+    async def _unstage_oversized_files(self) -> None:
+        """
+        Remove files larger than ``config.max_file_size`` from the index so they
+        are not committed into the shadow store. Best-effort: any failure leaves
+        staging exactly as ``git add -A`` produced it and never breaks the turn.
+        """
+        max_size = self.config.max_file_size
+        if not max_size or max_size <= 0:
+            return
+        try:
+            # NUL-delimited so paths with spaces/newlines are handled safely.
+            staged = await self._run_git("diff", "--cached", "--name-only", "-z")
+            for path in staged.split("\0"):
+                if not path:
+                    continue
+                abs_path = os.path.join(self.config.workspace_dir, path)
+                try:
+                    # Use lstat so symlinks are measured by the link itself, not
+                    # their target: a small link to a large file must stay in the
+                    # checkpoint. Skip anything that isn't a regular file.
+                    if os.path.islink(abs_path) or not os.path.isfile(abs_path):
+                        continue
+                    if os.lstat(abs_path).st_size > max_size:
+                        # Keep the file on disk; only drop it from the index.
+                        await self._run_git("rm", "--cached", "--quiet", "--", path)
+                        logger.debug(f"Skipped oversized file from checkpoint: {path}")
+                except Exception as e:
+                    logger.debug(f"Could not evaluate size of {path}: {e}")
+        except Exception as e:
+            logger.debug(f"Oversized-file skip pass failed (continuing): {e}")
+
+    async def _maybe_gc(self, cadence: int = 20) -> None:
+        """
+        Every ``cadence`` saves, pack the shadow repo and make it self-contained.
+
+        We seed the shadow store from the project's objects via ``alternates``
+        for a fast first checkpoint, but a borrowed object can disappear if the
+        project repo later prunes or repacks it — which would make checkpoint
+        history unreadable. To keep ``/undo``/``/revert`` durable we periodically
+        run ``git repack -a -d`` (without ``--local``) so every reachable object,
+        including borrowed ones, is copied into a local pack the shadow store
+        owns. ``git gc --auto`` still handles pruning/packing of loose objects.
+
+        Best-effort: failures are swallowed so a gc issue never breaks a turn.
+        """
+        self._saves_since_gc += 1
+        if self._saves_since_gc < cadence:
+            return
+        self._saves_since_gc = 0
+        try:
+            # -a: pack all reachable objects; -d: drop redundant packs. Omitting
+            # --local pulls borrowed (alternate) objects into the local pack so
+            # the shadow store no longer depends on the project object database.
+            await self._run_git("repack", "-a", "-d", "--quiet")
+            await self._run_git("gc", "--auto", "--quiet")
+        except Exception as e:
+            logger.debug(f"Shadow-store gc skipped: {e}")
    
    async def save(
        self,
@@ -295,7 +442,13 @@ class CheckpointService:
        try:
            # Stage all changes
            await self._run_git("add", "-A")
-            
+
+            # Drop oversized files from the index so the shadow store never
+            # copies giant blobs (build outputs, model weights, media). This
+            # mirrors the existing "never fail the turn" contract: any error
+            # here is swallowed and staging proceeds as before.
+            await self._unstage_oversized_files()
+
            # Check if there are changes
            try:
                await self._run_git("diff", "--cached", "--quiet")
@@ -329,7 +482,12 @@ class CheckpointService:
            
            # Prune old checkpoints if needed
            await self._prune_checkpoints()
-            
+
+            # Periodically let git pack loose objects and drop unreachable ones
+            # so the shadow store doesn't leak disk over a long session. Bounded
+            # cadence keeps the per-turn cost near zero; best-effort only.
+            await self._maybe_gc()
+
            self._emit(CheckpointEvent.CHECKPOINT_CREATED, checkpoint)
            logger.info(f"Created checkpoint: {checkpoint.short_id} - {message}")
            
@@ -40,6 +40,11 @@ class CheckpointConfig:
    max_checkpoints: int = 100  # Maximum checkpoints to keep
    user_name: str = field(default_factory=lambda: os.getenv("PRAISONAI_GIT_USER_NAME", "PraisonAI Checkpoints"))
    user_email: str = field(default_factory=lambda: os.getenv("PRAISONAI_GIT_USER_EMAIL", "[email protected]"))
+    # Skip staging files larger than this (bytes); 0 disables the cap. Kept
+    # after the identity fields so existing positional construction of
+    # CheckpointConfig(workspace, storage, enabled, auto, max_checkpoints,
+    # user_name, user_email) is unaffected.
+    max_file_size: int = 2 * 1024 * 1024
    exclude_patterns: List[str] = field(default_factory=lambda: [
        ".git",
        ".praisonai",
@@ -14,3 +14,19 @@ class CompactionStrategy(str, Enum):
    SMART = "smart"
    LLM_SUMMARIZE = "llm_summarize"  # Use LLM for summarization
    PRUNE = "prune"  # Remove old tool outputs
+
+    @classmethod
+    def from_policy_value(cls, value: str) -> "CompactionStrategy":
+        """Map a context-policy strategy value to a compactor strategy.
+
+        The policy layer (``context.protocols.CompactionStrategy``) uses a
+        different member set than the compactor. This is the single source of
+        truth for translating between the two; keep it here rather than
+        duplicating the mapping at each call site.
+        """
+        return {
+            "truncate": cls.TRUNCATE,
+            "summarise": cls.SUMMARIZE,
+            "drop_oldest_tools": cls.PRUNE,
+            "sliding_window": cls.SLIDING,
+        }.get(value, cls.PRUNE)
@@ -33,11 +33,13 @@ Usage:

 from dataclasses import dataclass, field
 import math
-from typing import Dict, List, Any, Optional, Callable, Tuple, Union, FrozenSet
+from typing import Dict, List, Any, Optional, Callable, Tuple, Union, FrozenSet, TYPE_CHECKING
 from enum import Enum

-# Import AutonomyConfig from canonical location (no circular dep)
-from ..agent.autonomy import AutonomyConfig
+if TYPE_CHECKING:
+    # Type-only: the eager import pulled agent.autonomy and escalation.types onto
+    # every `from praisonaiagents import Agent` path (see resolve_autonomy).
+    from ..agent.autonomy import AutonomyConfig

 # Default tool output limit (16000 chars ≈ 4000 tokens)
 # Single source of truth shared by OutputConfig.tool_output_limit and
@@ -898,7 +900,7 @@ class ExecutionConfig:
    
    # Code execution (consolidated from allow_code_execution + code_execution_mode)
    code_execution: bool = False
-    code_mode: str = "safe"  # "safe" or "unsafe"
+    code_mode: str = "safe"  # "safe", "unsafe" or "isolated"
    
    # Code-execution-with-tools (code mode): when True, model-generated code may
    # call the agent's registered tools directly via injected proxies, enabling
@@ -963,6 +965,15 @@ class ExecutionConfig:
        # Validate the unified step budget early (before any early returns below).
        if self.max_steps is not None and self.max_steps < 1:
            raise ValueError("ExecutionConfig.max_steps must be >= 1 when set.")
+        # code_mode is a security-shaped switch: "safe" (subprocess, no tools),
+        # "unsafe" (same-process + tools, timeout not enforced), or "isolated"
+        # (subprocess + tools bridged back to the parent under the approval
+        # gate). Reject typos loudly rather than silently degrading isolation.
+        if self.code_mode not in ("safe", "unsafe", "isolated"):
+            raise ValueError(
+                "ExecutionConfig.code_mode must be 'safe', 'unsafe' or "
+                f"'isolated'; got {self.code_mode!r}."
+            )
        # Handle context_compaction serialization round-trip
        if isinstance(self.context_compaction, dict):
            from ..context.policy import ContextCompactionPolicy
@@ -1384,13 +1395,21 @@ class HooksConfig:
    
    # Middleware list
    middleware: List[Any] = field(default_factory=list)
-    
+
+    # Optional per-agent HookRegistry for event-based hooks (BEFORE_TOOL,
+    # AFTER_AGENT, ...). When None the agent uses the process-wide default
+    # registry (so plugins bridged into it still fire). Set this to a private
+    # ``HookRegistry()`` to keep event hooks registered for THIS agent from
+    # firing on unrelated agents that share the default registry.
+    registry: Optional[Any] = None
+
    def to_dict(self) -> Dict[str, Any]:
        """Convert to dictionary."""
        return {
            "on_step": str(self.on_step) if self.on_step else None,
            "on_tool_call": str(self.on_tool_call) if self.on_tool_call else None,
            "middleware": [str(m) for m in self.middleware],
+            "registry": type(self.registry).__name__ if self.registry is not None else None,
        }


@@ -1785,13 +1804,16 @@ def resolve_caching(value: CachingParam) -> Optional[CachingConfig]:
    return _resolve(value, CachingConfig)


-def resolve_autonomy(value: AutonomyParam) -> Optional[AutonomyConfig]:
+def resolve_autonomy(value: AutonomyParam) -> Optional["AutonomyConfig"]:
    """
    Resolve autonomy= parameter following precedence ladder.
    
    Delegates to the canonical resolver in param_resolver.py.
    Kept for backward compatibility with tests.
    """
+    # Lazy: agent.autonomy (and, through it, escalation.types) is only needed
+    # when a caller actually resolves an autonomy= value.
+    from ..agent.autonomy import AutonomyConfig
    from .param_resolver import resolve_autonomy as _resolve
    return _resolve(value, AutonomyConfig)

@@ -2052,10 +2074,18 @@ def __getattr__(name):

    ``ToolSearchConfig`` is resolved (and the tools subsystem imported) only on
    first access, so ``from praisonaiagents import Agent`` does not eagerly load
-    the entire tools package (issue #3191).
+    the entire tools package (issue #3191). ``AutonomyConfig`` follows the same
+    pattern (issue #5056).
    """
    if name == "ToolSearchConfig":
        cfg = _resolve_tool_search_config()
        globals()["ToolSearchConfig"] = cfg  # cache for subsequent access
        return cfg
+    if name == "AutonomyConfig":
+        # Same idea for AutonomyConfig: keep the historical import path
+        # (``from ...feature_configs import AutonomyConfig``) working without
+        # loading agent.autonomy / escalation.types at module import (#5056).
+        from ..agent.autonomy import AutonomyConfig
+        globals()["AutonomyConfig"] = AutonomyConfig
+        return AutonomyConfig
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
@@ -7,9 +7,16 @@ another tool's harness, and results can be diffed/aggregated across tools.

 Mapping (1:1 with EvalPort's ``Suite`` / ``ResultSet`` concepts):

-    EvalPackage <-> EvalPort Suite      (name/description/version + cases + thresholds)
-    EvalCase    <-> EvalPort test case  (input / expected_output / grader refs)
-    EvalReport   -> EvalPort ResultSet  (aggregate + per-case pass/score/latency)
+    EvalPackage <-> EvalPort Suite      (id/name/description + test_cases + graders)
+    EvalCase    <-> EvalPort TestCase   (id / input / expected_output / grader refs)
+    EvalReport   -> EvalPort ResultSet  (summary + per-case grader_results)
+
+Per-criterion scores map to native EvalPort ``grader_results`` (one per criterion,
+matching the criterion graders referenced by the Suite). PraisonAI-only fields
+with no EvalPort slot (package version, thresholds, seed, trajectory records, a
+nonpositive ``timeout_seconds`` that the spec's ``timeout_ms`` cannot represent,
+...) are kept under the ``praisonai`` key of the relevant ``metadata`` object so
+round trips stay lossless.

 Design: this module is intentionally **dependency-free**. It emits and consumes
 plain spec-shaped dicts (JSON), so the core SDK does not take on ``evalport-sdk``
@@ -27,53 +34,90 @@ Example:
    >>> pkg2.name == pkg.name
    True
 """
-from typing import Any, Dict, List
+import uuid
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Optional

 from .package import EvalCase, EvalPackage, EvalReport

-EVALPORT_SPEC_VERSION = "1.0"
+# EvalPort spec version the emitted documents conform to (semver, per spec).
+EVALPORT_SPEC_VERSION = "1.0.0-rc.5"
+
+# Grader that carries PraisonAI's own per-case verdict (EvalResult.passed/score).
+# EvalPort requires every test case to reference at least one grader.
+VERDICT_GRADER_ID = "praisonai_verdict"
+_HANDLER = "praisonaiagents.eval"
+_META_KEY = "praisonai"
+
+
+def _grader_def(grader_id: str, description: str) -> Dict[str, Any]:
+    return {
+        "id": grader_id,
+        "type": "custom",
+        "params": {"handler": _HANDLER},
+        "description": description,
+    }


 def _case_to_evalport(case: EvalCase) -> Dict[str, Any]:
-    """Map an ``EvalCase`` to an EvalPort test-case dict.
+    """Map an ``EvalCase`` to an EvalPort TestCase dict.

-    The native ``timeout_seconds`` is emitted as a dedicated top-level field so
-    it never collides with a user-supplied ``metadata["timeout_seconds"]`` key,
-    keeping round trips lossless for both.
+    ``criteria`` become grader references (defined at suite level); a case
+    without criteria references the verdict grader. ``timeout_seconds`` maps to
+    the spec's ``timeout_ms``; user ``metadata`` is passed through untouched.
    """
+    # A case's own criteria are its graders; the verdict grader is always
+    # referenced too so PraisonAI's pass/score verdict has a home in the
+    # ResultSet (spec requires >= 1 grader per case).
+    graders = [VERDICT_GRADER_ID] + [c for c in case.criteria if c != VERDICT_GRADER_ID]
    item: Dict[str, Any] = {
        "id": case.name,
        "input": case.input,
+        "graders": list(dict.fromkeys(graders)),
    }
    if case.expected is not None:
        item["expected_output"] = case.expected
-    if case.criteria:
-        item["graders"] = list(case.criteria)
-    if case.timeout_seconds is not None:
-        item["timeout_seconds"] = case.timeout_seconds
-    metadata = dict(case.metadata or {})
-    if metadata:
-        item["metadata"] = metadata
+    if case.metadata:
+        item["metadata"] = dict(case.metadata)
+    if case.timeout_seconds is not None and case.timeout_seconds > 0:
+        item["timeout_ms"] = max(1, int(round(case.timeout_seconds * 1000)))
+    elif case.timeout_seconds is not None:
+        # Nonpositive timeout is not spec-representable (timeout_ms >= 1); keep
+        # the raw value under the reserved namespace so import restores it
+        # instead of silently applying the 30s default.
+        meta = item.setdefault("metadata", {})
+        native = dict(meta.get(_META_KEY) or {})
+        native["timeout_seconds"] = case.timeout_seconds
+        meta[_META_KEY] = native
    return item


 def _case_from_evalport(item: Dict[str, Any]) -> EvalCase:
-    """Map an EvalPort test-case dict to an ``EvalCase``.
+    """Map an EvalPort TestCase dict to an ``EvalCase``.

-    Prefers the dedicated top-level ``timeout_seconds`` field; falls back to a
-    ``metadata["timeout_seconds"]`` value for suites emitted by other tools,
-    without mutating the caller's metadata.
+    Reads ``timeout_ms`` (spec); falls back to a legacy top-level or
+    ``metadata["timeout_seconds"]`` value, without mutating caller metadata.
    """
    metadata = dict(item.get("metadata") or {})
-    if "timeout_seconds" in item:
+    native = metadata.get(_META_KEY) or {}
+    if "timeout_ms" in item:
+        timeout = item["timeout_ms"] / 1000.0
+    elif "timeout_seconds" in item:
        timeout = item["timeout_seconds"]
+    elif "timeout_seconds" in native:
+        timeout = native["timeout_seconds"]  # nonpositive timeout preserved on export
    else:
        timeout = metadata.get("timeout_seconds", 30.0)
+    criteria = []
+    for g in item.get("graders") or item.get("criteria") or []:
+        gid = g.get("id") if isinstance(g, dict) else g
+        if gid and gid != VERDICT_GRADER_ID:
+            criteria.append(gid)
    return EvalCase(
        name=item.get("id") or item.get("name") or "case",
        input=item.get("input", ""),
        expected=item.get("expected_output", item.get("expected")),
-        criteria=list(item.get("graders") or item.get("criteria") or []),
+        criteria=criteria,
        metadata=metadata,
        timeout_seconds=timeout,
    )
@@ -88,16 +132,28 @@ def to_evalport(package: EvalPackage) -> Dict[str, Any]:
    Returns:
        A spec-shaped dict suitable for ``openeval.validate.validate_suite()``.
    """
-    return {
-        "evalport_version": EVALPORT_SPEC_VERSION,
-        "kind": "suite",
+    criteria = dict.fromkeys(c for case in package.cases for c in case.criteria)
+    criteria.pop(VERDICT_GRADER_ID, None)
+    graders = [_grader_def(VERDICT_GRADER_ID, "PraisonAI per-case verdict (EvalResult.passed/score)")]
+    graders += [_grader_def(c, f"PraisonAI criterion: {c}") for c in criteria]
+
+    suite: Dict[str, Any] = {
+        "version": EVALPORT_SPEC_VERSION,
+        "id": package.name,
        "name": package.name,
-        "description": package.description,
-        "version": package.version,
-        "cases": [_case_to_evalport(c) for c in package.cases],
-        "thresholds": dict(package.thresholds or {}),
-        "seed": package.seed,
+        "graders": graders,
+        "test_cases": [_case_to_evalport(c) for c in package.cases],
+        "metadata": {
+            _META_KEY: {
+                "version": package.version,
+                "thresholds": dict(package.thresholds or {}),
+                "seed": package.seed,
+            }
+        },
    }
+    if package.description:
+        suite["description"] = package.description
+    return suite


 def from_evalport(suite: Dict[str, Any]) -> EvalPackage:
@@ -110,22 +166,68 @@ def from_evalport(suite: Dict[str, Any]) -> EvalPackage:
        An ``EvalPackage`` whose cases can be run via ``EvalSuite`` /
        ``HarnessEvaluator`` as native ``EvalCase`` objects.
    """
-    cases: List[EvalCase] = [_case_from_evalport(c) for c in suite.get("cases", [])]
+    items = suite.get("test_cases", suite.get("cases", []))
+    cases: List[EvalCase] = [_case_from_evalport(c) for c in items]
+    meta = (suite.get("metadata") or {}).get(_META_KEY) or {}
+    legacy = "evalport_version" in suite  # documents emitted before spec alignment
    return EvalPackage(
-        name=suite.get("name", "suite"),
+        name=suite.get("name") or suite.get("id") or "suite",
        description=suite.get("description", ""),
-        version=suite.get("version", "1.0.0"),
+        version=meta.get("version") or (suite.get("version") if legacy else None) or "1.0.0",
        cases=cases,
-        thresholds=dict(suite.get("thresholds") or {}),
-        seed=suite.get("seed"),
+        thresholds=dict(meta.get("thresholds") or suite.get("thresholds") or {}),
+        seed=meta.get("seed", suite.get("seed")),
    )


-def report_to_evalport(report: EvalReport) -> Dict[str, Any]:
+def _unit_score(value: Any) -> Optional[float]:
+    if isinstance(value, bool) or not isinstance(value, (int, float)) or value != value:
+        return None
+    return min(1.0, max(0.0, float(value)))
+
+
+def _grader_result(grader_id: str, raw_score: Any, passed: bool) -> Dict[str, Any]:
+    """Build one spec-shaped ``grader_result`` entry.
+
+    Scores are clamped to the spec's [0, 1] range; the raw value is preserved
+    under ``metadata`` when clamping changes it. A null score means "not
+    verified", so ``passed`` is forced ``False`` in that case.
+    """
+    score = _unit_score(raw_score)
+    grader: Dict[str, Any] = {
+        "grader_id": grader_id,
+        "type": "custom",
+        "score": score,
+        "passed": bool(passed) if score is not None else False,
+    }
+    if score is not None and score != raw_score:
+        grader["metadata"] = {"openeval.raw_score": raw_score}
+    return grader
+
+
+def _runner_info() -> Dict[str, str]:
+    runner = {"name": "praisonaiagents"}
+    try:
+        from importlib.metadata import version
+
+        runner["version"] = version("praisonaiagents")
+    except Exception:
+        pass
+    return runner
+
+
+def report_to_evalport(
+    report: EvalReport,
+    *,
+    run_id: Optional[str] = None,
+    started_at: Optional[str] = None,
+) -> Dict[str, Any]:
    """Convert an ``EvalReport`` to an EvalPort ``ResultSet`` dict.

    Args:
        report: The aggregated report from running an eval package.
+        run_id: Unique run identifier (defaults to a random UUID).
+        started_at: RFC 3339 run start time (defaults to now, UTC).

    Returns:
        A spec-shaped dict suitable for
@@ -133,33 +235,52 @@ def report_to_evalport(report: EvalReport) -> Dict[str, Any]:
    """
    results: List[Dict[str, Any]] = []
    for r in report.results:
+        verdict = _grader_result(VERDICT_GRADER_ID, r.score, r.passed)
+        grader_results = [verdict]
+        # Surface each per-criterion score as its own grader_result so EvalPort
+        # consumers can diff criteria across tools (they match the criterion
+        # grader ids referenced by the exported Suite). ``passed`` is unknown at
+        # the criterion level, so mirror the case verdict.
+        for name, score in (r.criteria_scores or {}).items():
+            if name == VERDICT_GRADER_ID:
+                continue
+            grader_results.append(_grader_result(name, score, r.passed))
+
        item: Dict[str, Any] = {
-            "case_id": r.case_name,
-            "passed": r.passed,
-            "score": r.score,
-            "latency_ms": r.latency_ms,
+            "test_case_id": r.case_name,
+            "passed": verdict["passed"],
+            "grader_results": grader_results,
        }
        if r.actual_output is not None:
-            item["actual_output"] = r.actual_output
+            item["actual_output"] = str(r.actual_output)
+        if r.latency_ms is not None and r.latency_ms >= 0:
+            item["duration_ms"] = int(round(r.latency_ms))
        if r.error is not None:
-            item["error"] = r.error
-        if r.criteria_scores:
-            item["grader_scores"] = dict(r.criteria_scores)
+            item["error"] = {"type": "runner_error", "message": str(r.error)}
        if r.record is not None:
-            item["record"] = r.record
+            item["metadata"] = {_META_KEY: {"record": r.record}}
        results.append(item)

+    summary: Dict[str, Any] = {
+        "total": report.total_cases,
+        "passed": report.passed_cases,
+        "failed": report.failed_cases,
+        "pass_rate": report.pass_rate,
+    }
+    meta: Dict[str, Any] = {"thresholds_met": dict(report.thresholds_met or {})}
+    avg = _unit_score(report.average_score)
+    if avg is not None:
+        summary["avg_score"] = avg
+        if avg != report.average_score:
+            meta["average_score"] = report.average_score  # clamped
+
    return {
-        "evalport_version": EVALPORT_SPEC_VERSION,
-        "kind": "result_set",
-        "suite_name": report.package_name,
-        "summary": {
-            "total": report.total_cases,
-            "passed": report.passed_cases,
-            "failed": report.failed_cases,
-            "pass_rate": report.pass_rate,
-            "average_score": report.average_score,
-            "thresholds_met": dict(report.thresholds_met or {}),
-        },
+        "version": EVALPORT_SPEC_VERSION,
+        "suite_id": report.package_name,
+        "run_id": run_id or uuid.uuid4().hex,
+        "started_at": started_at or datetime.now(timezone.utc).isoformat(),
+        "runner": _runner_info(),
        "results": results,
+        "summary": summary,
+        "metadata": {_META_KEY: meta},
    }
@@ -12,265 +12,318 @@ Gap S2: WebSocketGateway is re-exported here for convenience but requires
 ``praisonai-bot`` (or ``pip install praisonai``) to be installed.
 """

-from .protocols import (
-    GatewayProtocol,
-    GatewaySessionProtocol,
-    GatewayClientProtocol,
-    GatewayEvent,
-    GatewayMessage,
-    EventType,
-    OperatorScope,
-    GatewayCloseCode,
-    # Declarative method -> required-scope registry (Issue #3206)
-    GatewayMethodDescriptor,
-    GATEWAY_METHODS,
-    register_gateway_method,
-    resolve_required_scope,
-    # Config hot-reload observability (Issue #3049)
-    ReloadStatus,
-    compute_config_revision,
-    # Saturation / back-pressure observability (Issue #4265)
-    HealthPressure,
-    evaluate_pressure,
-    # Push protocols and dataclasses
-    PushChannelProtocol,
-    PresenceProtocol,
-    DeliveryGuaranteeProtocol,
-    OutboundDeliveryProtocol,
-    IdempotencyStoreProtocol,
-    InMemoryIdempotencyStore,
-    ChannelInfo,
-    PresenceInfo,
-    # Home channel and delivery protocols
-    HomeChannelRegistryProtocol,
-    DeliveryResolverProtocol,
-    # Creation-time delivery-target pre-flight (Issue #3800)
-    DeliveryPreflightProtocol,
-    DeliveryValidation,
-    ScheduleTargetError,
-    # Remote-sandbox outbound media bridge (Issue #4951)
-    RemoteMediaResolver,
-    # Agent-facing outbound messaging
-    OutboundMessengerProtocol,
-    DeliveryResult,
-    TargetInfo,
-    # Agent-callable message reactions (Issue #3917)
-    ReactionResult,
-    ReactionStatus,
-    # Agent/gateway-callable thread creation (Issue #3987)
-    ThreadResult,
-    ThreadStatus,
-    # Agent-callable cross-conversation request/reply (Issue #3689)
-    ConversationReply,
-    ConversationReplyStatus,
-    ConversationRequestProtocol,
-    # Agent-facing live status/health (Issue #3688)
-    GatewayStatusProtocol,
-    GatewayStatus,
-    # Inbound route binding (Issue #2225)
-    RouteBinding,
-    RouteFacts,
-    RouteMatch,
-    resolve_route,
-    # Per-route, trust-tiered toolset scoping (Issue #2298)
-    ToolPolicy,
-    UNTRUSTED_DENY_SUBSTRINGS,
-    TRUST_TIERS,
-    # Outbound send-policy guard (Issue #2226)
-    SendDecision,
-    SendPolicyProtocol,
-    SendPolicy,
-    # Gateway idle-dormancy / scale-to-zero (Issue #2332)
-    IdleDecision,
-    GatewayIdlePolicyProtocol,
-    GatewayIdlePolicy,  # backward-compat alias
-    ScaleToZeroPolicy,
-    # Gateway freeze-thaw (involuntary host-suspend gap) recovery (Issue #4767)
-    ThawDecision,
-    ThawPolicyProtocol,
-    WallClockGapThawPolicy,
-    # Gateway graceful-drain on shutdown (Issue #2375)
-    DrainDecision,
-    GatewayDrainPolicyProtocol,
-    DrainTimeoutPolicy,
-    # Gateway inbound admission control (Issue #2454)
-    AdmissionDecision,
-    GatewayConcurrencyPolicyProtocol,
-    GatewayConcurrencyPolicy,  # backward-compat alias
-    ConcurrencyLimitPolicy,
-    # Gateway resource-pressure admission (Issue #3445)
-    ResourceSample,
-    ResourcePressurePolicyProtocol,
-    MemoryPressurePolicy,
-    # Gateway memory-pressure cache eviction (Issue #3804)
-    WarmSession,
-    MemoryPressureProtocol,
-    plan_pressure_evictions,
-    # Gateway rate-limit admission (Issue #2532)
-    RateLimitDecision,
-    RateLimitPolicyProtocol,
-    RateLimitPolicy,  # backward-compat alias
-    SlidingWindowRateLimitPolicy,
-    # Gateway cost-aware spend-budget admission (Issue #4894)
-    SpendBudgetPolicyProtocol,
-    SpendBudgetPolicy,  # backward-compat alias
-    WindowedSpendBudgetPolicy,
-    # Durable-queue dead-letter decision (Issue #3519)
-    PERMANENT_ERROR_CLASSES,
-    DeadLetterDecision,
-    DeadLetterPolicyProtocol,
-    AttemptAndAgeDeadLetterPolicy,
-    # Port-less, restart-safe external drain trigger (Issue #2390)
-    current_epoch,
-    DrainMarkerPolicy,
-    # Crash / shutdown forensics (Issue #2436)
-    ShutdownForensicsProtocol,
-    format_forensics_for_log,
-    is_supervised,
-    drain_timeout_has_headroom,
-    # Code-skew guard for hot operations (Issue #2460)
-    detect_code_skew,
-    # Restart-intent exit-code protocol (Issue #2437)
-    GATEWAY_OK_EXIT_CODE,
-    GATEWAY_RESTART_EXIT_CODE,
-    GATEWAY_FATAL_CONFIG_EXIT_CODE,
-    FatalConfigError,
-    classify_exit_reason,
-    RestartLoopGuard,
-    # Durable process-lifecycle record + unclean/OOM classification (Issue #4603)
-    LifecycleRecord,
-    classify_unclean_exit,
-    RestartStoreProtocol,
-    PersistentRestartLoopGuard,
-    classify_resource_pressure,
-    # Fleet-level crash-loop breaker for channel supervision (Issue #3840)
-    FleetSupervisionPolicy,
-    # Protocol version negotiation
-    PROTOCOL_VERSION,
-    MIN_PROTOCOL_VERSION,
-    MAX_PROTOCOL_VERSION,
-    ProtocolHello,
-    ProtocolHelloOk,
-    GapInfo,
-    ResumeSnapshot,
-    # Out-of-process platform-connector relay (Issue #2485)
-    CapabilityDescriptor,
-    RelayTransport,
-    # Gateway pipeline span-tracing seam (Issue #2716)
-    GATEWAY_TRACE_STAGES,
-    GatewayTraceHook,
-    NullGatewayTraceHook,
-    NULL_GATEWAY_TRACE_HOOK,
-    resolve_trace_hook,
-    # Gateway self-lifecycle command guardrail (Issue #2753)
-    LifecycleCommandDecision,
-    LifecycleCommandPolicyProtocol,
-    LifecycleCommandPolicy,  # backward-compat alias
-    LifecycleCommandGuardPolicy,
-    # Application-level connection liveness (Issue #2798)
-    LivenessDecision,
-    LivenessPolicyProtocol,
-    LivenessPolicy,
-    # Cluster-wide per-turn serialisation (Issue #3643)
-    TurnLeaseToken,
-    TurnLockProtocol,
-    LocalTurnLock,
-    # Per-session turn-execution isolation (Issue #4011)
-    WorkerWedgedError,
-    TurnPlacement,
-    TurnExecutorProtocol,
-    InProcessTurnExecutor,
-    # Global operator emergency-stop / pause brake (Issue #4220)
-    EmergencyStopState,
-    EmergencyStopProtocol,
-    NullEmergencyStop,
-    FileEmergencyStop,
-    # Schema-validated inbound frame codec (Issue #2831)
-    HelloParams,
-    HelloResult,
-    HelloError,
-    ConnectErrorCode,
-    ConnectRecoveryStep,
-    is_recoverable,
-    MessageParams,
-    LeaveParams,
-    JoinParams,
-    FrameDecodeError,
-    ClientFrame,
-    decode_client_frame,
-    # Weak / placeholder secret guard (Issue #3259)
-    KNOWN_WEAK_SECRETS,
-    WeakGatewaySecretError,
-    is_weak_secret,
-    assert_gateway_secret_strong,
-    # Per-platform identity canonicalization (Issue #3886)
-    IdentityCanonicalizerProtocol,
-)
-from .liveness import (
-    # Event-loop liveness watchdog (Issue #3385)
-    LoopWatchdogPolicy,
-    LoopWatchdog,
-)
-from .degraded_state import (
-    # Unified degraded-capability registry (Issue #3518)
-    DegradedOwner,
-    DegradedCapabilityProtocol,
-    DegradedCapabilityRegistry,
-    OWNER_KINDS,
-    DEGRADED_STATES,
-    # Fail-closed read of the degraded-owner contract (Issue #3640)
-    DegradedCapabilityLookupProtocol,
-    OwnerUnavailable,
-    assert_owner_available,
-)
-from .hooks import (
-    HookAction,
-    HookConfig,
-    InboundTriggerProtocol,
-    render_template,
-    compute_idempotency_key,
-    verify_webhook_signature,
-)
-from .config import (
-    GatewayConfig,
-    SessionConfig,
-    ApiConfig,
-    EmergencyStopConfig,
-    ChannelRouteConfig,
-    MultiChannelGatewayConfig,
-    # Config version stamp + doctor-driven migration (Issue #3841)
-    GATEWAY_CONFIG_VERSION,
-    ConfigVersionError,
-    LegacyConfigRule,
-    GATEWAY_CONFIG_RULES,
-    is_config_current,
-    migrate_config_with_doctor,
-    # Push config
-    PushConfig,
-    RedisConfig,
-    PresenceConfig,
-    DeliveryConfig,
-    PollingConfig,
-    LivenessConfig,
-    TurnLockConfig,
-    # Hot-reload registry (Issue #3378)
-    HOT_APPLIABLE_KEYS,
-    SupportsHotReload,
-    is_hot_appliable,
-    # Reload scope classification (Issue #3440)
-    ReloadScope,
-    classify_reload,
-)
+from .._lazy import create_lazy_getattr

-# Lazy loading cache
+# Lazy import map: symbol -> (submodule, attr). Deferring these keeps the
+# 283 KB / 7,201-line ``protocols.py`` (and siblings) out of package-import
+# time; each submodule is parsed only when one of its symbols is first
+# touched. Mirrors the ``mcp``/``skills`` packages (AGENTS.md §4.2).
+_LAZY_IMPORTS = {
+    # protocols.py (parsed lazily, on first touch)
+    "GatewayProtocol": ("praisonaiagents.gateway.protocols", "GatewayProtocol"),
+    "GatewaySessionProtocol": ("praisonaiagents.gateway.protocols", "GatewaySessionProtocol"),
+    "GatewayClientProtocol": ("praisonaiagents.gateway.protocols", "GatewayClientProtocol"),
+    "GatewayEvent": ("praisonaiagents.gateway.protocols", "GatewayEvent"),
+    "GatewayMessage": ("praisonaiagents.gateway.protocols", "GatewayMessage"),
+    "EventType": ("praisonaiagents.gateway.protocols", "EventType"),
+    "OperatorScope": ("praisonaiagents.gateway.protocols", "OperatorScope"),
+    # Multi-observer session sharing (Issue #5192)
+    "SessionVisibility": ("praisonaiagents.gateway.protocols", "SessionVisibility"),
+    "SessionSharingRole": ("praisonaiagents.gateway.protocols", "SessionSharingRole"),
+    "SessionObserverProtocol": ("praisonaiagents.gateway.protocols", "SessionObserverProtocol"),
+    # Session-projection reducer (Issue #5324)
+    "SessionProjection": ("praisonaiagents.gateway.session_projection", "SessionProjection"),
+    "SessionProjectionState": ("praisonaiagents.gateway.session_projection", "SessionProjectionState"),
+    "RunView": ("praisonaiagents.gateway.session_projection", "RunView"),
+    # Canonical session-address grammar (Issue #5383)
+    "SessionRef": ("praisonaiagents.gateway.addressing", "SessionRef"),
+    "build_session_path": ("praisonaiagents.gateway.addressing", "build_session_path"),
+    "parse_session_path": ("praisonaiagents.gateway.addressing", "parse_session_path"),
+    "derive_short_id": ("praisonaiagents.gateway.addressing", "derive_short_id"),
+    "slugify": ("praisonaiagents.gateway.addressing", "slugify"),
+    "RESERVED_NAMES": ("praisonaiagents.gateway.addressing", "RESERVED_NAMES"),
+    "SHORT_ID_LEN": ("praisonaiagents.gateway.addressing", "SHORT_ID_LEN"),
+    "GatewayCloseCode": ("praisonaiagents.gateway.protocols", "GatewayCloseCode"),
+    # Declarative method -> required-scope registry (Issue #3206)
+    "GatewayMethodDescriptor": ("praisonaiagents.gateway.protocols", "GatewayMethodDescriptor"),
+    "GATEWAY_METHODS": ("praisonaiagents.gateway.protocols", "GATEWAY_METHODS"),
+    "register_gateway_method": ("praisonaiagents.gateway.protocols", "register_gateway_method"),
+    "resolve_required_scope": ("praisonaiagents.gateway.protocols", "resolve_required_scope"),
+    "authorize_method": ("praisonaiagents.gateway.protocols", "authorize_method"),
+    "GatewayUnauthorized": ("praisonaiagents.gateway.protocols", "GatewayUnauthorized"),
+    # Config hot-reload observability (Issue #3049)
+    "ReloadStatus": ("praisonaiagents.gateway.protocols", "ReloadStatus"),
+    "compute_config_revision": ("praisonaiagents.gateway.protocols", "compute_config_revision"),
+    # Saturation / back-pressure observability (Issue #4265)
+    "HealthPressure": ("praisonaiagents.gateway.protocols", "HealthPressure"),
+    "evaluate_pressure": ("praisonaiagents.gateway.protocols", "evaluate_pressure"),
+    # Trusted-proxy ingress attribution (Issue #5312)
+    "IngressTrust": ("praisonaiagents.gateway.protocols", "IngressTrust"),
+    "IngressAttribution": ("praisonaiagents.gateway.protocols", "IngressAttribution"),
+    "resolve_ingress_attribution": ("praisonaiagents.gateway.protocols", "resolve_ingress_attribution"),
+    "parse_forwarded_for": ("praisonaiagents.gateway.protocols", "parse_forwarded_for"),
+    # Push protocols and dataclasses
+    "PushChannelProtocol": ("praisonaiagents.gateway.protocols", "PushChannelProtocol"),
+    "PresenceProtocol": ("praisonaiagents.gateway.protocols", "PresenceProtocol"),
+    "DeliveryGuaranteeProtocol": ("praisonaiagents.gateway.protocols", "DeliveryGuaranteeProtocol"),
+    "OutboundDeliveryProtocol": ("praisonaiagents.gateway.protocols", "OutboundDeliveryProtocol"),
+    "IdempotencyStoreProtocol": ("praisonaiagents.gateway.protocols", "IdempotencyStoreProtocol"),
+    "InMemoryIdempotencyStore": ("praisonaiagents.gateway.protocols", "InMemoryIdempotencyStore"),
+    "ChannelInfo": ("praisonaiagents.gateway.protocols", "ChannelInfo"),
+    "PresenceInfo": ("praisonaiagents.gateway.protocols", "PresenceInfo"),
+    # Home channel and delivery protocols
+    "HomeChannelRegistryProtocol": ("praisonaiagents.gateway.protocols", "HomeChannelRegistryProtocol"),
+    "DeliveryResolverProtocol": ("praisonaiagents.gateway.protocols", "DeliveryResolverProtocol"),
+    # Creation-time delivery-target pre-flight (Issue #3800)
+    "DeliveryPreflightProtocol": ("praisonaiagents.gateway.protocols", "DeliveryPreflightProtocol"),
+    "DeliveryValidation": ("praisonaiagents.gateway.protocols", "DeliveryValidation"),
+    "ScheduleTargetError": ("praisonaiagents.gateway.protocols", "ScheduleTargetError"),
+    # Remote-sandbox outbound media bridge (Issue #4951)
+    "RemoteMediaResolver": ("praisonaiagents.gateway.protocols", "RemoteMediaResolver"),
+    # Agent-facing outbound messaging
+    "OutboundMessengerProtocol": ("praisonaiagents.gateway.protocols", "OutboundMessengerProtocol"),
+    "DeliveryResult": ("praisonaiagents.gateway.protocols", "DeliveryResult"),
+    "TargetInfo": ("praisonaiagents.gateway.protocols", "TargetInfo"),
+    # Agent-callable message reactions (Issue #3917)
+    "ReactionResult": ("praisonaiagents.gateway.protocols", "ReactionResult"),
+    "ReactionStatus": ("praisonaiagents.gateway.protocols", "ReactionStatus"),
+    # Agent/gateway-callable thread creation (Issue #3987)
+    "ThreadResult": ("praisonaiagents.gateway.protocols", "ThreadResult"),
+    "ThreadStatus": ("praisonaiagents.gateway.protocols", "ThreadStatus"),
+    # Agent-callable message mutation: edit/delete (Issue #5054)
+    "MessageActionResult": ("praisonaiagents.gateway.protocols", "MessageActionResult"),
+    "MessageActionStatus": ("praisonaiagents.gateway.protocols", "MessageActionStatus"),
+    # Agent-callable cross-conversation request/reply (Issue #3689)
+    "ConversationReply": ("praisonaiagents.gateway.protocols", "ConversationReply"),
+    "ConversationReplyStatus": ("praisonaiagents.gateway.protocols", "ConversationReplyStatus"),
+    "ConversationRequestProtocol": ("praisonaiagents.gateway.protocols", "ConversationRequestProtocol"),
+    # Server→client interactive request/reply over the transport (Issue #5351)
+    "GatewayRequestKind": ("praisonaiagents.gateway.protocols", "GatewayRequestKind"),
+    "GatewayServerRequest": ("praisonaiagents.gateway.protocols", "GatewayServerRequest"),
+    "GatewayServerReply": ("praisonaiagents.gateway.protocols", "GatewayServerReply"),
+    "GatewayRequestChannelProtocol": ("praisonaiagents.gateway.protocols", "GatewayRequestChannelProtocol"),
+    # Agent-facing live status/health (Issue #3688)
+    "GatewayStatusProtocol": ("praisonaiagents.gateway.protocols", "GatewayStatusProtocol"),
+    "GatewayStatus": ("praisonaiagents.gateway.protocols", "GatewayStatus"),
+    # Inbound route binding (Issue #2225)
+    "RouteBinding": ("praisonaiagents.gateway.protocols", "RouteBinding"),
+    "RouteFacts": ("praisonaiagents.gateway.protocols", "RouteFacts"),
+    "RouteMatch": ("praisonaiagents.gateway.protocols", "RouteMatch"),
+    "resolve_route": ("praisonaiagents.gateway.protocols", "resolve_route"),
+    # Per-route, trust-tiered toolset scoping (Issue #2298)
+    "ToolPolicy": ("praisonaiagents.gateway.protocols", "ToolPolicy"),
+    "UNTRUSTED_DENY_SUBSTRINGS": ("praisonaiagents.gateway.protocols", "UNTRUSTED_DENY_SUBSTRINGS"),
+    "TRUST_TIERS": ("praisonaiagents.gateway.protocols", "TRUST_TIERS"),
+    # Outbound send-policy guard (Issue #2226)
+    "SendDecision": ("praisonaiagents.gateway.protocols", "SendDecision"),
+    "SendPolicyProtocol": ("praisonaiagents.gateway.protocols", "SendPolicyProtocol"),
+    "SendPolicy": ("praisonaiagents.gateway.protocols", "SendPolicy"),
+    # Gateway idle-dormancy / scale-to-zero (Issue #2332)
+    "IdleDecision": ("praisonaiagents.gateway.protocols", "IdleDecision"),
+    "GatewayIdlePolicyProtocol": ("praisonaiagents.gateway.protocols", "GatewayIdlePolicyProtocol"),
+    "GatewayIdlePolicy": ("praisonaiagents.gateway.protocols", "GatewayIdlePolicy"),
+    "ScaleToZeroPolicy": ("praisonaiagents.gateway.protocols", "ScaleToZeroPolicy"),
+    # Gateway freeze-thaw (involuntary host-suspend gap) recovery (Issue #4767)
+    "ThawDecision": ("praisonaiagents.gateway.protocols", "ThawDecision"),
+    "ThawPolicyProtocol": ("praisonaiagents.gateway.protocols", "ThawPolicyProtocol"),
+    "WallClockGapThawPolicy": ("praisonaiagents.gateway.protocols", "WallClockGapThawPolicy"),
+    # Gateway graceful-drain on shutdown (Issue #2375)
+    "DrainDecision": ("praisonaiagents.gateway.protocols", "DrainDecision"),
+    "GatewayDrainPolicyProtocol": ("praisonaiagents.gateway.protocols", "GatewayDrainPolicyProtocol"),
+    "DrainTimeoutPolicy": ("praisonaiagents.gateway.protocols", "DrainTimeoutPolicy"),
+    # Gateway inbound admission control (Issue #2454)
+    "AdmissionDecision": ("praisonaiagents.gateway.protocols", "AdmissionDecision"),
+    "GatewayConcurrencyPolicyProtocol": ("praisonaiagents.gateway.protocols", "GatewayConcurrencyPolicyProtocol"),
+    "GatewayConcurrencyPolicy": ("praisonaiagents.gateway.protocols", "GatewayConcurrencyPolicy"),
+    "ConcurrencyLimitPolicy": ("praisonaiagents.gateway.protocols", "ConcurrencyLimitPolicy"),
+    # Gateway resource-pressure admission (Issue #3445)
+    "ResourceSample": ("praisonaiagents.gateway.protocols", "ResourceSample"),
+    "ResourcePressurePolicyProtocol": ("praisonaiagents.gateway.protocols", "ResourcePressurePolicyProtocol"),
+    "MemoryPressurePolicy": ("praisonaiagents.gateway.protocols", "MemoryPressurePolicy"),
+    # Gateway memory-pressure cache eviction (Issue #3804)
+    "WarmSession": ("praisonaiagents.gateway.protocols", "WarmSession"),
+    "MemoryPressureProtocol": ("praisonaiagents.gateway.protocols", "MemoryPressureProtocol"),
+    "plan_pressure_evictions": ("praisonaiagents.gateway.protocols", "plan_pressure_evictions"),
+    # Gateway rate-limit admission (Issue #2532)
+    "RateLimitDecision": ("praisonaiagents.gateway.protocols", "RateLimitDecision"),
+    "RateLimitPolicyProtocol": ("praisonaiagents.gateway.protocols", "RateLimitPolicyProtocol"),
+    "RateLimitPolicy": ("praisonaiagents.gateway.protocols", "RateLimitPolicy"),
+    "SlidingWindowRateLimitPolicy": ("praisonaiagents.gateway.protocols", "SlidingWindowRateLimitPolicy"),
+    # Gateway cost-aware spend-budget admission (Issue #4894)
+    "SpendBudgetPolicyProtocol": ("praisonaiagents.gateway.protocols", "SpendBudgetPolicyProtocol"),
+    "SpendBudgetPolicy": ("praisonaiagents.gateway.protocols", "SpendBudgetPolicy"),
+    "WindowedSpendBudgetPolicy": ("praisonaiagents.gateway.protocols", "WindowedSpendBudgetPolicy"),
+    # Durable-queue dead-letter decision (Issue #3519)
+    "PERMANENT_ERROR_CLASSES": ("praisonaiagents.gateway.protocols", "PERMANENT_ERROR_CLASSES"),
+    "DeadLetterDecision": ("praisonaiagents.gateway.protocols", "DeadLetterDecision"),
+    "DeadLetterPolicyProtocol": ("praisonaiagents.gateway.protocols", "DeadLetterPolicyProtocol"),
+    "AttemptAndAgeDeadLetterPolicy": ("praisonaiagents.gateway.protocols", "AttemptAndAgeDeadLetterPolicy"),
+    # Port-less, restart-safe external drain trigger (Issue #2390)
+    "current_epoch": ("praisonaiagents.gateway.protocols", "current_epoch"),
+    "DrainMarkerPolicy": ("praisonaiagents.gateway.protocols", "DrainMarkerPolicy"),
+    # Crash / shutdown forensics (Issue #2436)
+    "ShutdownForensicsProtocol": ("praisonaiagents.gateway.protocols", "ShutdownForensicsProtocol"),
+    "format_forensics_for_log": ("praisonaiagents.gateway.protocols", "format_forensics_for_log"),
+    "is_supervised": ("praisonaiagents.gateway.protocols", "is_supervised"),
+    "drain_timeout_has_headroom": ("praisonaiagents.gateway.protocols", "drain_timeout_has_headroom"),
+    # Code-skew guard for hot operations (Issue #2460)
+    "detect_code_skew": ("praisonaiagents.gateway.protocols", "detect_code_skew"),
+    # Restart-intent exit-code protocol (Issue #2437)
+    "GATEWAY_OK_EXIT_CODE": ("praisonaiagents.gateway.protocols", "GATEWAY_OK_EXIT_CODE"),
+    "GATEWAY_RESTART_EXIT_CODE": ("praisonaiagents.gateway.protocols", "GATEWAY_RESTART_EXIT_CODE"),
+    "GATEWAY_FATAL_CONFIG_EXIT_CODE": ("praisonaiagents.gateway.protocols", "GATEWAY_FATAL_CONFIG_EXIT_CODE"),
+    "FatalConfigError": ("praisonaiagents.gateway.protocols", "FatalConfigError"),
+    "classify_exit_reason": ("praisonaiagents.gateway.protocols", "classify_exit_reason"),
+    "RestartLoopGuard": ("praisonaiagents.gateway.protocols", "RestartLoopGuard"),
+    # Durable process-lifecycle record + unclean/OOM classification (Issue #4603)
+    "LifecycleRecord": ("praisonaiagents.gateway.protocols", "LifecycleRecord"),
+    "classify_unclean_exit": ("praisonaiagents.gateway.protocols", "classify_unclean_exit"),
+    "RestartStoreProtocol": ("praisonaiagents.gateway.protocols", "RestartStoreProtocol"),
+    "PersistentRestartLoopGuard": ("praisonaiagents.gateway.protocols", "PersistentRestartLoopGuard"),
+    "classify_resource_pressure": ("praisonaiagents.gateway.protocols", "classify_resource_pressure"),
+    # Fleet-level crash-loop breaker for channel supervision (Issue #3840)
+    "FleetSupervisionPolicy": ("praisonaiagents.gateway.protocols", "FleetSupervisionPolicy"),
+    # Protocol version negotiation
+    "PROTOCOL_VERSION": ("praisonaiagents.gateway.protocols", "PROTOCOL_VERSION"),
+    "MIN_PROTOCOL_VERSION": ("praisonaiagents.gateway.protocols", "MIN_PROTOCOL_VERSION"),
+    "MAX_PROTOCOL_VERSION": ("praisonaiagents.gateway.protocols", "MAX_PROTOCOL_VERSION"),
+    "ProtocolHello": ("praisonaiagents.gateway.protocols", "ProtocolHello"),
+    "ProtocolHelloOk": ("praisonaiagents.gateway.protocols", "ProtocolHelloOk"),
+    "GapInfo": ("praisonaiagents.gateway.protocols", "GapInfo"),
+    "ResumeSnapshot": ("praisonaiagents.gateway.protocols", "ResumeSnapshot"),
+    # Out-of-process platform-connector relay (Issue #2485)
+    "CapabilityDescriptor": ("praisonaiagents.gateway.protocols", "CapabilityDescriptor"),
+    "RelayTransport": ("praisonaiagents.gateway.protocols", "RelayTransport"),
+    # Gateway pipeline span-tracing seam (Issue #2716)
+    "GATEWAY_TRACE_STAGES": ("praisonaiagents.gateway.protocols", "GATEWAY_TRACE_STAGES"),
+    "GatewayTraceHook": ("praisonaiagents.gateway.protocols", "GatewayTraceHook"),
+    "NullGatewayTraceHook": ("praisonaiagents.gateway.protocols", "NullGatewayTraceHook"),
+    "NULL_GATEWAY_TRACE_HOOK": ("praisonaiagents.gateway.protocols", "NULL_GATEWAY_TRACE_HOOK"),
+    "resolve_trace_hook": ("praisonaiagents.gateway.protocols", "resolve_trace_hook"),
+    # Gateway self-lifecycle command guardrail (Issue #2753)
+    "LifecycleCommandDecision": ("praisonaiagents.gateway.protocols", "LifecycleCommandDecision"),
+    "LifecycleCommandPolicyProtocol": ("praisonaiagents.gateway.protocols", "LifecycleCommandPolicyProtocol"),
+    "LifecycleCommandPolicy": ("praisonaiagents.gateway.protocols", "LifecycleCommandPolicy"),
+    "LifecycleCommandGuardPolicy": ("praisonaiagents.gateway.protocols", "LifecycleCommandGuardPolicy"),
+    # Application-level connection liveness (Issue #2798)
+    "LivenessDecision": ("praisonaiagents.gateway.protocols", "LivenessDecision"),
+    "LivenessPolicyProtocol": ("praisonaiagents.gateway.protocols", "LivenessPolicyProtocol"),
+    "LivenessPolicy": ("praisonaiagents.gateway.protocols", "LivenessPolicy"),
+    # Cluster-wide per-turn serialisation (Issue #3643)
+    "TurnLeaseToken": ("praisonaiagents.gateway.protocols", "TurnLeaseToken"),
+    "TurnLockProtocol": ("praisonaiagents.gateway.protocols", "TurnLockProtocol"),
+    "LocalTurnLock": ("praisonaiagents.gateway.protocols", "LocalTurnLock"),
+    # Per-session turn-execution isolation (Issue #4011)
+    "WorkerWedgedError": ("praisonaiagents.gateway.protocols", "WorkerWedgedError"),
+    "TurnPlacement": ("praisonaiagents.gateway.protocols", "TurnPlacement"),
+    "TurnExecutorProtocol": ("praisonaiagents.gateway.protocols", "TurnExecutorProtocol"),
+    "InProcessTurnExecutor": ("praisonaiagents.gateway.protocols", "InProcessTurnExecutor"),
+    # Global operator emergency-stop / pause brake (Issue #4220)
+    "EmergencyStopState": ("praisonaiagents.gateway.protocols", "EmergencyStopState"),
+    "EmergencyStopProtocol": ("praisonaiagents.gateway.protocols", "EmergencyStopProtocol"),
+    "NullEmergencyStop": ("praisonaiagents.gateway.protocols", "NullEmergencyStop"),
+    "FileEmergencyStop": ("praisonaiagents.gateway.protocols", "FileEmergencyStop"),
+    # Per-session stop scope (Issue #5129)
+    "StopScope": ("praisonaiagents.gateway.protocols", "StopScope"),
+    "StopResult": ("praisonaiagents.gateway.protocols", "StopResult"),
+    # Schema-validated inbound frame codec (Issue #2831)
+    "HelloParams": ("praisonaiagents.gateway.protocols", "HelloParams"),
+    "HelloResult": ("praisonaiagents.gateway.protocols", "HelloResult"),
+    "HelloError": ("praisonaiagents.gateway.protocols", "HelloError"),
+    "ConnectErrorCode": ("praisonaiagents.gateway.protocols", "ConnectErrorCode"),
+    "ConnectRecoveryStep": ("praisonaiagents.gateway.protocols", "ConnectRecoveryStep"),
+    "is_recoverable": ("praisonaiagents.gateway.protocols", "is_recoverable"),
+    "MessageParams": ("praisonaiagents.gateway.protocols", "MessageParams"),
+    # First-class attachment contract (Issue #5207)
+    "AttachmentRef": ("praisonaiagents.gateway.protocols", "AttachmentRef"),
+    "AttachmentStoreProtocol": ("praisonaiagents.gateway.protocols", "AttachmentStoreProtocol"),
+    "LeaveParams": ("praisonaiagents.gateway.protocols", "LeaveParams"),
+    "JoinParams": ("praisonaiagents.gateway.protocols", "JoinParams"),
+    "FrameDecodeError": ("praisonaiagents.gateway.protocols", "FrameDecodeError"),
+    "ClientFrame": ("praisonaiagents.gateway.protocols", "ClientFrame"),
+    "decode_client_frame": ("praisonaiagents.gateway.protocols", "decode_client_frame"),
+    # Weak / placeholder secret guard (Issue #3259)
+    "KNOWN_WEAK_SECRETS": ("praisonaiagents.gateway.protocols", "KNOWN_WEAK_SECRETS"),
+    "WeakGatewaySecretError": ("praisonaiagents.gateway.protocols", "WeakGatewaySecretError"),
+    "is_weak_secret": ("praisonaiagents.gateway.protocols", "is_weak_secret"),
+    "assert_gateway_secret_strong": ("praisonaiagents.gateway.protocols", "assert_gateway_secret_strong"),
+    # Per-platform identity canonicalization (Issue #3886)
+    "IdentityCanonicalizerProtocol": ("praisonaiagents.gateway.protocols", "IdentityCanonicalizerProtocol"),
+    # liveness.py — Event-loop liveness watchdog (Issue #3385)
+    "LoopWatchdogPolicy": ("praisonaiagents.gateway.liveness", "LoopWatchdogPolicy"),
+    "LoopWatchdog": ("praisonaiagents.gateway.liveness", "LoopWatchdog"),
+    # liveness.py — Startup-phase watchdog (Issue #5265)
+    "StartupWatchdog": ("praisonaiagents.gateway.liveness", "StartupWatchdog"),
+    # degraded_state.py — Unified degraded-capability registry (Issue #3518)
+    "DegradedOwner": ("praisonaiagents.gateway.degraded_state", "DegradedOwner"),
+    "DegradedCapabilityProtocol": ("praisonaiagents.gateway.degraded_state", "DegradedCapabilityProtocol"),
+    "DegradedCapabilityRegistry": ("praisonaiagents.gateway.degraded_state", "DegradedCapabilityRegistry"),
+    "OWNER_KINDS": ("praisonaiagents.gateway.degraded_state", "OWNER_KINDS"),
+    "DEGRADED_STATES": ("praisonaiagents.gateway.degraded_state", "DEGRADED_STATES"),
+    # Fail-closed read of the degraded-owner contract (Issue #3640)
+    "DegradedCapabilityLookupProtocol": ("praisonaiagents.gateway.degraded_state", "DegradedCapabilityLookupProtocol"),
+    "OwnerUnavailable": ("praisonaiagents.gateway.degraded_state", "OwnerUnavailable"),
+    "assert_owner_available": ("praisonaiagents.gateway.degraded_state", "assert_owner_available"),
+    # hooks.py — Inbound trigger / webhook contract (Issue #2281)
+    "HookAction": ("praisonaiagents.gateway.hooks", "HookAction"),
+    "HookConfig": ("praisonaiagents.gateway.hooks", "HookConfig"),
+    "InboundTriggerProtocol": ("praisonaiagents.gateway.hooks", "InboundTriggerProtocol"),
+    "render_template": ("praisonaiagents.gateway.hooks", "render_template"),
+    "compute_idempotency_key": ("praisonaiagents.gateway.hooks", "compute_idempotency_key"),
+    "verify_webhook_signature": ("praisonaiagents.gateway.hooks", "verify_webhook_signature"),
+    # config.py (always available)
+    "GatewayConfig": ("praisonaiagents.gateway.config", "GatewayConfig"),
+    "SessionConfig": ("praisonaiagents.gateway.config", "SessionConfig"),
+    "ApiConfig": ("praisonaiagents.gateway.config", "ApiConfig"),
+    "EmergencyStopConfig": ("praisonaiagents.gateway.config", "EmergencyStopConfig"),
+    # Attachment ceilings config (Issue #5207)
+    "AttachmentConfig": ("praisonaiagents.gateway.config", "AttachmentConfig"),
+    "ChannelRouteConfig": ("praisonaiagents.gateway.config", "ChannelRouteConfig"),
+    "MultiChannelGatewayConfig": ("praisonaiagents.gateway.config", "MultiChannelGatewayConfig"),
+    # Config version stamp + doctor-driven migration (Issue #3841)
+    "GATEWAY_CONFIG_VERSION": ("praisonaiagents.gateway.config", "GATEWAY_CONFIG_VERSION"),
+    "ConfigVersionError": ("praisonaiagents.gateway.config", "ConfigVersionError"),
+    "LegacyConfigRule": ("praisonaiagents.gateway.config", "LegacyConfigRule"),
+    "GATEWAY_CONFIG_RULES": ("praisonaiagents.gateway.config", "GATEWAY_CONFIG_RULES"),
+    "is_config_current": ("praisonaiagents.gateway.config", "is_config_current"),
+    "migrate_config_with_doctor": ("praisonaiagents.gateway.config", "migrate_config_with_doctor"),
+    # Push config
+    "PushConfig": ("praisonaiagents.gateway.config", "PushConfig"),
+    "RedisConfig": ("praisonaiagents.gateway.config", "RedisConfig"),
+    "PresenceConfig": ("praisonaiagents.gateway.config", "PresenceConfig"),
+    "DeliveryConfig": ("praisonaiagents.gateway.config", "DeliveryConfig"),
+    "PollingConfig": ("praisonaiagents.gateway.config", "PollingConfig"),
+    "LivenessConfig": ("praisonaiagents.gateway.config", "LivenessConfig"),
+    "TurnLockConfig": ("praisonaiagents.gateway.config", "TurnLockConfig"),
+    # Hot-reload registry (Issue #3378)
+    "HOT_APPLIABLE_KEYS": ("praisonaiagents.gateway.config", "HOT_APPLIABLE_KEYS"),
+    "SupportsHotReload": ("praisonaiagents.gateway.config", "SupportsHotReload"),
+    "is_hot_appliable": ("praisonaiagents.gateway.config", "is_hot_appliable"),
+    # Reload scope classification (Issue #3440)
+    "ReloadScope": ("praisonaiagents.gateway.config", "ReloadScope"),
+    "classify_reload": ("praisonaiagents.gateway.config", "classify_reload"),
+    # Candidate validation + rollback contract (Issue #5144)
+    "CandidateReport": ("praisonaiagents.gateway.config", "CandidateReport"),
+    "ReloadValidationProtocol": (
+        "praisonaiagents.gateway.config",
+        "ReloadValidationProtocol",
+    ),
+}
+
+# Lazy loading cache (shared with wrapper-backed implementations below)
 _lazy_cache = {}

+_lazy_getattr = create_lazy_getattr(_LAZY_IMPORTS, __name__, cache=_lazy_cache)
+

 def __getattr__(name: str):
-    """Lazy load heavy gateway implementations from praisonai-bot (C9).
+    """Lazy load gateway symbols and heavy implementations.

-    Gap S2: Re-export WebSocketGateway for convenience. Prefer
-    ``praisonai_bot.gateway`` or ``pip install praisonai`` for full stack.
+    Protocols/policies/config are served from their submodules on first
+    access via the ``_LAZY_IMPORTS`` map, so ``protocols.py`` is parsed only
+    when actually touched. ``WebSocketGateway``/``GatewaySession`` are
+    re-exported from ``praisonai-bot`` (C9); prefer ``praisonai_bot.gateway``
+    or ``pip install praisonai`` for the full stack (Gap S2).
    """
    if name in _lazy_cache:
        return _lazy_cache[name]
@@ -309,7 +362,25 @@ def __getattr__(name: str):
                "Install with: pip install praisonai-bot or pip install praisonai"
            )
    
-    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    # Everything else: lazy-load from the submodule map (protocols.py parsed
+    # only if a protocol/policy symbol is actually requested).
+    value = _lazy_getattr(name)
+    # Bind the resolved symbol into the module globals so it behaves exactly
+    # like an eager import: it becomes a real attribute (discoverable via
+    # ``dir``, deletable, and patchable via ``monkeypatch.delattr``/``setattr``),
+    # and future accesses skip ``__getattr__`` entirely.
+    globals()[name] = value
+    return value
+
+
+def __dir__():
+    """Expose the public API to ``dir()`` / IDE autocomplete without importing.
+
+    Returns the union of names already resolved into the module globals and the
+    declared public ``__all__``, so lazy exports remain discoverable before they
+    are first accessed (matching the eager-import behaviour).
+    """
+    return sorted(set(globals()) | set(__all__))


 __all__ = [
@@ -321,17 +392,40 @@ __all__ = [
    "GatewayMessage",
    "EventType",
    "OperatorScope",
+    # Multi-observer session sharing (Issue #5192)
+    "SessionVisibility",
+    "SessionSharingRole",
+    "SessionObserverProtocol",
+    # Session-projection reducer (Issue #5324)
+    "SessionProjection",
+    "SessionProjectionState",
+    "RunView",
+    # Canonical session-address grammar (Issue #5383)
+    "SessionRef",
+    "build_session_path",
+    "parse_session_path",
+    "derive_short_id",
+    "slugify",
+    "RESERVED_NAMES",
+    "SHORT_ID_LEN",
    "GatewayCloseCode",
    # Declarative method -> required-scope registry (Issue #3206)
    "GatewayMethodDescriptor",
    "GATEWAY_METHODS",
    "register_gateway_method",
    "resolve_required_scope",
+    "authorize_method",
+    "GatewayUnauthorized",
    # Config hot-reload observability (Issue #3049)
    "ReloadStatus",
    "compute_config_revision",
    "HealthPressure",
    "evaluate_pressure",
+    # Trusted-proxy ingress attribution (Issue #5312)
+    "IngressTrust",
+    "IngressAttribution",
+    "resolve_ingress_attribution",
+    "parse_forwarded_for",
    # Push protocols (always available)
    "PushChannelProtocol",
    "PresenceProtocol",
@@ -355,11 +449,18 @@ __all__ = [
    "ReactionResult",
    "ReactionStatus",
    "ThreadResult",
+    "MessageActionResult",
+    "MessageActionStatus",
    "ThreadStatus",
    # Agent-callable cross-conversation request/reply (Issue #3689)
    "ConversationReply",
    "ConversationReplyStatus",
    "ConversationRequestProtocol",
+    # Server→client interactive request/reply over the transport (Issue #5351)
+    "GatewayRequestKind",
+    "GatewayServerRequest",
+    "GatewayServerReply",
+    "GatewayRequestChannelProtocol",
    # Agent-facing live status/health (Issue #3688)
    "GatewayStatusProtocol",
    "GatewayStatus",
@@ -479,6 +580,8 @@ __all__ = [
    "EmergencyStopProtocol",
    "NullEmergencyStop",
    "FileEmergencyStop",
+    "StopScope",
+    "StopResult",
    # Schema-validated inbound frame codec (Issue #2831)
    "HelloParams",
    "HelloResult",
@@ -487,6 +590,9 @@ __all__ = [
    "ConnectRecoveryStep",
    "is_recoverable",
    "MessageParams",
+    # First-class attachment contract (Issue #5207)
+    "AttachmentRef",
+    "AttachmentStoreProtocol",
    "LeaveParams",
    "JoinParams",
    "FrameDecodeError",
@@ -502,6 +608,8 @@ __all__ = [
    # Event-loop liveness watchdog (Issue #3385)
    "LoopWatchdogPolicy",
    "LoopWatchdog",
+    # Startup-phase watchdog (Issue #5265)
+    "StartupWatchdog",
    # Unified degraded-capability registry (Issue #3518)
    "DegradedOwner",
    "DegradedCapabilityProtocol",
@@ -524,6 +632,8 @@ __all__ = [
    "SessionConfig",
    "ApiConfig",
    "EmergencyStopConfig",
+    # Attachment ceilings config (Issue #5207)
+    "AttachmentConfig",
    "ChannelRouteConfig",
    "MultiChannelGatewayConfig",
    # Config version stamp + doctor-driven migration (Issue #3841)
@@ -546,6 +656,8 @@ __all__ = [
    "is_hot_appliable",
    "ReloadScope",
    "classify_reload",
+    "CandidateReport",
+    "ReloadValidationProtocol",
    # Implementations (lazy loaded from praisonai wrapper)
    "WebSocketGateway",
    "GatewaySession",
@@ -0,0 +1,227 @@
+"""Canonical session-address grammar for the Gateway (Issue #5383).
+
+A pure, dependency-free grammar that maps between an opaque runtime
+``session_key`` (a UUID, or an ``agent:<id>:<uuid>``-style key) and a stable,
+human-readable, reconnect-survivable path — e.g. ``chat/<agentId>/<slug>-<shortId>``.
+It gives every surface (dashboard deep-links, bot channels, the CLI, the
+``praisonai-ts`` mirror, reconnecting clients) a single addressing contract to
+agree on, instead of each one carrying the raw UUID or re-inventing a resolver.
+
+This module lives in **core** because the grammar is a set of pure functions
+that define a protocol/contract every surface must share. It has **no**
+third-party dependencies and does no I/O — the concrete store-backed lookup
+(``session.resolve``) is implemented in the wrapper behind this contract.
+
+The grammar is deliberately lossy in one direction: ``build_session_path``
+derives a short id from the trailing UUID so the whole opaque instance id never
+appears in a URL, and an optional slug from the session label. ``parse_session_path``
+recovers a :class:`SessionRef` (agent id + short id + slug hint) that a resolver
+maps back to the concrete ``session_id``; it never assumes it can reconstruct the
+key byte-for-byte, except via the exact-key escape hatch (a ``!``-prefixed final
+segment) for sessions that must keep their literal key.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from typing import Optional
+from urllib.parse import quote, unquote
+
+__all__ = [
+    "SessionRef",
+    "SHORT_ID_LEN",
+    "RESERVED_NAMES",
+    "derive_short_id",
+    "slugify",
+    "build_session_path",
+    "parse_session_path",
+]
+
+# Length of the short id derived from the trailing hex of a session key. Eight
+# hex chars (32 bits) keeps URLs terse while staying collision-resistant enough
+# for a per-agent namespace, where a resolver disambiguates the rare clash.
+SHORT_ID_LEN = 8
+
+# Reserved final-segment names that denote a well-known / sentinel session
+# (e.g. the agent's ``main`` conversation) rather than a ``slug-shortId`` pair.
+# A resolver treats these as a friendly sentinel, not a short-id lookup.
+RESERVED_NAMES = frozenset({"main", "global", "default", "root"})
+
+# Trailing run of hex digits that terminates the key — the opaque UUID tail we
+# shorten from. Anchored at end-of-string so a *hex run followed by non-hex*
+# (e.g. ``report-deadbeef.js``) is NOT shortened and instead round-trips via the
+# literal-key escape hatch; only a genuine hex terminus (a UUID tail) is lossy.
+_HEX_TAIL_RE = re.compile(r"([0-9a-fA-F]{%d,})$" % SHORT_ID_LEN)
+
+# A ``slug-shortId`` final segment: optional slug, then a hex short id.
+_SLUG_SHORT_RE = re.compile(r"^(?:(?P<slug>.+)-)?(?P<short>[0-9a-fA-F]{%d})$" % SHORT_ID_LEN)
+
+# Characters allowed verbatim in a slug; everything else collapses to ``-``.
+_SLUG_STRIP_RE = re.compile(r"[^a-z0-9]+")
+
+# Exact-key escape hatch: a final segment prefixed with ``!`` carries the
+# percent-encoded literal key verbatim (for sessions that must not be shortened).
+_LITERAL_PREFIX = "!"
+
+
+@dataclass(frozen=True)
+class SessionRef:
+    """A parsed friendly reference to a session, for a resolver to map to an id.
+
+    Exactly one of the resolution hints is authoritative depending on how the
+    path was written:
+
+    - ``literal_key`` set  -> resolve by exact key (escape hatch).
+    - ``short_id`` set     -> resolve by short id (optionally aided by ``slug_hint``).
+    - ``slug_hint`` only, matching a reserved name -> resolve the sentinel session.
+
+    ``agent_id`` scopes the lookup so short ids only need to be unique per agent.
+    """
+
+    agent_id: str
+    short_id: Optional[str] = None
+    slug_hint: Optional[str] = None
+    literal_key: Optional[str] = None
+
+
+def derive_short_id(session_key: str) -> Optional[str]:
+    """Derive a stable short id from the terminating hex run of ``session_key``.
+
+    Returns the last :data:`SHORT_ID_LEN` hex chars of the key's *terminating*
+    hex run (e.g. a UUID tail), lower-cased. The run must end the key: a key
+    whose hex is followed by non-hex characters (e.g. ``report-deadbeef.js``)
+    is deliberately not shortened, so it round-trips via the literal-key escape
+    hatch instead of silently losing its suffix. Returns ``None`` when the key
+    has no terminating hex run long enough to shorten — the caller then falls
+    back to the literal key or reserved-name form.
+    """
+    if not session_key:
+        return None
+    match = _HEX_TAIL_RE.search(session_key)
+    if not match:
+        return None
+    hex_tail = match.group(1)
+    return hex_tail[-SHORT_ID_LEN:].lower()
+
+
+def slugify(display_name: Optional[str], max_len: int = 48) -> Optional[str]:
+    """Turn a human label into a lowercase, dash-separated URL slug.
+
+    Non-alphanumeric runs collapse to a single ``-``; leading/trailing dashes are
+    trimmed. Returns ``None`` for an empty/whitespace-only or unusable name so
+    the path omits the slug rather than emitting an empty segment.
+    """
+    if not display_name:
+        return None
+    slug = _SLUG_STRIP_RE.sub("-", display_name.strip().lower()).strip("-")
+    if not slug:
+        return None
+    if len(slug) > max_len:
+        slug = slug[:max_len].rstrip("-")
+    return slug or None
+
+
+def build_session_path(
+    agent_id: str,
+    session_key: str,
+    display_name: Optional[str] = None,
+    namespace: str = "chat",
+) -> Optional[str]:
+    """Build a stable, shareable path for a session.
+
+    Produces ``<namespace>/<agentId>/<slug>-<shortId>`` when a short id can be
+    derived (with the slug omitted when there is no usable ``display_name``), so
+    the whole opaque UUID never appears in the URL. When the key has no hex tail
+    to shorten, falls back to an exact-key segment (``!<percent-encoded-key>``)
+    so the address round-trips losslessly. Returns ``None`` if ``agent_id`` or
+    ``session_key`` is empty.
+
+    Path segments are percent-escaped, so a key that looks like a file
+    (e.g. ``report.js``) is never mistaken for a static asset.
+    """
+    if not agent_id or not session_key:
+        return None
+    ns = _seg(namespace) or "chat"
+    agent_seg = _seg(agent_id)
+    if not agent_seg:
+        return None
+
+    short_id = derive_short_id(session_key)
+    if short_id:
+        slug = slugify(display_name)
+        tail = f"{slug}-{short_id}" if slug else short_id
+    else:
+        # No hex tail to shorten: keep the literal key behind the escape hatch.
+        tail = _LITERAL_PREFIX + quote(session_key, safe="")
+    return f"{ns}/{agent_seg}/{tail}"
+
+
+def parse_session_path(path: str) -> Optional[SessionRef]:
+    """Parse a session path back into a :class:`SessionRef`.
+
+    Accepts the forms produced by :func:`build_session_path`, with or without a
+    leading namespace segment:
+
+    - ``chat/<agent>/<slug>-<shortId>`` / ``<agent>/<slug>-<shortId>``
+    - ``chat/<agent>/<shortId>`` / ``<agent>/<shortId>``
+    - ``chat/<agent>/main`` (reserved sentinel) / ``<agent>/main``
+    - ``chat/<agent>/!<percent-encoded-key>`` (exact-key escape hatch)
+
+    Returns ``None`` when the path has no agent segment or is otherwise
+    unparseable. A resolver combines the returned hints with its store to reach
+    the concrete ``session_id``.
+    """
+    if not path:
+        return None
+    parts = [p for p in path.strip().strip("/").split("/") if p]
+    # ``build_session_path`` percent-escapes the namespace and agent into single
+    # segments, so a canonical address is exactly ``<ns>/<agent>/<ref>`` (3) or
+    # ``<agent>/<ref>`` (2). Reject anything longer so a malformed link like
+    # ``chat/x/y/z`` never silently acquires a different meaning by using only
+    # its last two segments.
+    if len(parts) == 3:
+        # Drop the leading namespace segment; agent + final ref remain.
+        agent_id = unquote(parts[-2])
+        final = parts[-1]
+    elif len(parts) == 2:
+        agent_id = unquote(parts[0])
+        final = parts[1]
+    else:
+        return None
+
+    agent_id = agent_id.strip()
+    if not agent_id or not final:
+        return None
+
+    # Exact-key escape hatch.
+    if final.startswith(_LITERAL_PREFIX):
+        literal = unquote(final[len(_LITERAL_PREFIX):])
+        if not literal:
+            return None
+        return SessionRef(agent_id=agent_id, literal_key=literal)
+
+    final = unquote(final)
+
+    # Reserved sentinel (main/global/...): a friendly named session.
+    if final.lower() in RESERVED_NAMES:
+        return SessionRef(agent_id=agent_id, slug_hint=final.lower())
+
+    match = _SLUG_SHORT_RE.match(final)
+    if match:
+        return SessionRef(
+            agent_id=agent_id,
+            short_id=match.group("short").lower(),
+            slug_hint=match.group("slug") or None,
+        )
+
+    # No short id present: treat the whole segment as a slug hint the resolver
+    # can match against session labels.
+    return SessionRef(agent_id=agent_id, slug_hint=final)
+
+
+def _seg(value: Optional[str]) -> str:
+    """Percent-escape a single path segment (``/`` and dots are encoded)."""
+    if not value:
+        return ""
+    return quote(value.strip(), safe="")
@@ -297,6 +297,74 @@ class SupportsHotReload(Protocol):
        ...


+# ---------------------------------------------------------------------------
+# Candidate reload validation + rollback contract (Issue #5144)
+# ---------------------------------------------------------------------------
+#
+# ``classify_reload``/``ReloadScope`` decide *what* a config change touches; a
+# structural (``FULL``) change makes the runtime tear the live channels down
+# and bring the new config up. A config that passes the schema but fails at
+# *runtime* (unreachable model, an adapter that throws on load, bad wiring) is
+# not caught before that cutover, and there is no rollback — a single bad edit
+# can take a healthy always-on gateway offline.
+#
+# This is the small, canonical contract that makes a reload an *atomic,
+# validated cutover*: build/boot the candidate runtime in isolation and run the
+# existing readiness/turn pre-flight *before* the live one is withdrawn; only
+# swap when the candidate is proven healthy, otherwise keep the old config
+# serving. Core owns only the *shape* (a report + a protocol, no heavy
+# imports); the wrapper/bot gateway performs the concrete build+pre-flight.
+
+
+@dataclass
+class CandidateReport:
+    """Outcome of validating a candidate config *before* cutover.
+
+    Produced by :meth:`ReloadValidationProtocol.validate_candidate`: the
+    gateway builds/boots the candidate runtime in isolation (or a pre-flight)
+    and reports whether it is healthy enough to swap in. When ``ok`` is False
+    the live config keeps serving and the failures feed
+    :class:`~praisonaiagents.gateway.protocols.ReloadStatus.error` / ``health()``.
+
+    Attributes:
+        ok: Whether the candidate built and passed pre-flight — only then may
+            the live runtime be drained and the candidate swapped in.
+        failures: Human-readable reasons the candidate was rejected (empty when
+            ``ok``); joined into the reload ``error`` an operator sees.
+        candidate: Optional opaque handle to the already-built candidate runtime
+            the wrapper activates on success. Left untyped so core carries no
+            heavy runtime import; ``None`` when the runtime validates in place.
+    """
+
+    ok: bool
+    failures: "List[str]" = field(default_factory=list)
+    candidate: Optional[Any] = None
+
+
+@runtime_checkable
+class ReloadValidationProtocol(Protocol):
+    """Protocol a gateway implements for atomic, validated config cutover.
+
+    The invariant: **never drain the live runtime until**
+    :meth:`validate_candidate` **returns a report with** ``ok=True``. Owning
+    this contract in core keeps the "never cut over to an unvalidated
+    candidate" guarantee canonical so every gateway build honours it, rather
+    than being re-implemented per entry point (SIGHUP, ``gateway reload``,
+    ``gateway restart``).
+    """
+
+    async def validate_candidate(
+        self, new_config: Mapping[str, Any]
+    ) -> CandidateReport:
+        """Build/boot the candidate from ``new_config`` and pre-flight it.
+
+        Must not stop or mutate the live runtime; on any failure return a
+        report with ``ok=False`` and the reason(s) so the caller keeps the
+        previous config serving.
+        """
+        ...
+
+
@dataclass
 class SessionConfig:
    """Configuration for gateway sessions.
@@ -685,10 +753,17 @@ class ApiConfig:
            backed by the gateway's live agents and sessions.
        mcp: Serve an MCP JSON-RPC endpoint (``/mcp``) exposing the gateway's
            registered agents as callable tools.
+        stream: Emit true token-level deltas on the OpenAI-compatible
+            ``/v1/chat/completions`` SSE surface (``stream:true``) by wiring the
+            agent's existing ``stream_emitter`` through the gateway hot path.
+            Off by default: streaming responses stay byte-for-byte the buffered
+            single-chunk path, so enabling it never regresses non-streaming
+            latency or correctness.
    """

    openai: bool = False
    mcp: bool = False
+    stream: bool = False

    @property
    def enabled(self) -> bool:
@@ -697,7 +772,7 @@ class ApiConfig:

    def to_dict(self) -> Dict[str, Any]:
        """Convert to dictionary."""
-        return {"openai": self.openai, "mcp": self.mcp}
+        return {"openai": self.openai, "mcp": self.mcp, "stream": self.stream}

    @classmethod
    def from_dict(cls, data: Optional[Dict[str, Any]]) -> "ApiConfig":
@@ -707,6 +782,7 @@ class ApiConfig:
        return cls(
            openai=bool(data.get("openai", False)),
            mcp=bool(data.get("mcp", False)),
+            stream=bool(data.get("stream", False)),
        )


@@ -775,6 +851,163 @@ class EmergencyStopConfig:
        )


+@dataclass
+class AttachmentConfig:
+    """Attachment/media ceilings for the gateway wire protocol (Issue #5207).
+
+    The gateway is the control plane custom ``/ws`` clients build on, but the
+    ``message`` frame had no attachment contract and no per-attachment size/type
+    policy — only the whole-frame ``max_payload``. These ceilings are advertised
+    to clients in the ``hello`` handshake (folded into ``HelloResult.policy``) so
+    a well-behaved client self-limits — inlining small files, chunking larger
+    ones at ``chunk_bytes`` — instead of discovering the limit by being
+    disconnected. All fields are advisory limits; enforcement lives wherever the
+    gateway materialises attachments (a wrapper/bot concern).
+
+    Backward-compatible: attachments are opt-in per turn, and a client that
+    never sends one is unaffected by these values.
+
+    Attributes:
+        max_attachment_bytes: Largest single attachment (inline or stored) the
+            gateway accepts, in bytes. Default 10 MiB.
+        max_attachments: Maximum number of attachments on one ``message`` turn.
+        chunk_bytes: Advertised chunk size a client should use when streaming a
+            large file into the attachment store (reserve → put chunk → close).
+        allowed_types: Optional allow-list of MIME types / prefixes (e.g.
+            ``image/`` or ``application/pdf``). Empty (default) means no type
+            restriction is advertised.
+    """
+
+    max_attachment_bytes: int = 10 * 1024 * 1024  # 10 MiB
+    max_attachments: int = 10
+    chunk_bytes: int = 256 * 1024  # 256 KiB
+    allowed_types: List[str] = field(default_factory=list)
+
+    def __post_init__(self) -> None:
+        if self.max_attachment_bytes < 0:
+            raise ValueError(
+                "max_attachment_bytes must be >= 0 (use 0 to disable attachments)"
+            )
+        if self.max_attachments < 0:
+            raise ValueError("max_attachments must be >= 0")
+        if self.chunk_bytes <= 0:
+            raise ValueError("chunk_bytes must be > 0")
+
+    @property
+    def enabled(self) -> bool:
+        """Whether the gateway advertises any attachment capacity."""
+        return self.max_attachment_bytes > 0 and self.max_attachments > 0
+
+    def to_policy(self) -> Dict[str, Any]:
+        """Build the policy fragment advertised in ``HelloResult.policy``.
+
+        Only advertises the attachment ceilings when enabled, so a gateway with
+        attachments disabled advertises exactly today's policy shape. The type
+        allow-list is included only when configured.
+        """
+        if not self.enabled:
+            return {}
+        policy: Dict[str, Any] = {
+            "max_attachment_bytes": self.max_attachment_bytes,
+            "max_attachments": self.max_attachments,
+            "chunk_bytes": self.chunk_bytes,
+        }
+        if self.allowed_types:
+            policy["allowed_attachment_types"] = list(self.allowed_types)
+        return policy
+
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "max_attachment_bytes": self.max_attachment_bytes,
+            "max_attachments": self.max_attachments,
+            "chunk_bytes": self.chunk_bytes,
+            "allowed_types": list(self.allowed_types),
+        }
+
+    @classmethod
+    def from_dict(cls, data: Optional[Dict[str, Any]]) -> "AttachmentConfig":
+        """Create from a parsed ``gateway.attachments`` mapping (tolerant of None)."""
+        if not isinstance(data, dict):
+            return cls()
+        allowed = data.get("allowed_types")
+        return cls(
+            max_attachment_bytes=int(
+                data.get("max_attachment_bytes", 10 * 1024 * 1024)
+            ),
+            max_attachments=int(data.get("max_attachments", 10)),
+            chunk_bytes=int(data.get("chunk_bytes", 256 * 1024)),
+            allowed_types=[str(t) for t in allowed] if isinstance(allowed, (list, tuple)) else [],
+        )
+
+
+def _parse_drain_timeout(value: Any) -> Optional[float]:
+    """Coerce a ``gateway.drain_timeout`` YAML value to a float (or None).
+
+    Tolerant of env-substituted strings (e.g. ``"20"``); an unparseable /
+    negative value falls back to None (no drain) rather than raising, matching
+    the wrapper's fail-open handling of a malformed drain window (#5362).
+    """
+    if value is None:
+        return None
+    try:
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return None
+    return parsed if parsed >= 0 else None
+
+
+def _parse_watchdog_enabled(value: Any) -> bool:
+    """Whether a ``gateway.watchdog`` YAML value enables the loop watchdog.
+
+    Accepts both a bare truthy scalar and the documented block form
+    (``watchdog: { enabled: true, ... }``), mirroring the wrapper's watchdog
+    wiring so the public config reflects the same on/off decision (#5362).
+    """
+    if isinstance(value, dict):
+        flag = value.get("enabled", False)
+    else:
+        flag = value
+    if isinstance(flag, str):
+        return flag.strip().lower() in ("1", "true", "yes", "on")
+    return bool(flag)
+
+
+def _parse_watchdog_timeout(value: Any) -> Optional[float]:
+    """Derive the loop-stall budget (seconds) from a ``gateway.watchdog`` block.
+
+    Reads ``liveness_interval * liveness_strikes`` from the block form when
+    present so the public config exposes the effective wedge budget the CLI
+    ``--watchdog-timeout`` flag sets; returns None when unset/unparseable.
+    """
+    if not isinstance(value, dict):
+        return None
+    try:
+        interval = float(value.get("liveness_interval", 5.0))
+        strikes = int(value.get("liveness_strikes", 3)) or 3
+    except (TypeError, ValueError):
+        return None
+    budget = interval * strikes
+    return budget if budget > 0 else None
+
+
+def _parse_watchdog_timeout_scalar(value: Any) -> Optional[float]:
+    """Coerce an explicit ``gateway.watchdog_timeout`` scalar to a float.
+
+    Mirrors the ``watchdog_timeout > 0`` invariant (see ``GatewayConfig``):
+    tolerant of env-substituted strings (e.g. ``"15"``); an unparseable or
+    non-positive value falls back to None so a malformed timeout disables the
+    budget rather than raising (#5362).
+    """
+    if value is None:
+        return None
+    try:
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return None
+    return parsed if parsed > 0 else None
+
+
@dataclass
 class GatewayConfig:
    """Configuration for the gateway server.
@@ -807,6 +1040,13 @@ class GatewayConfig:
    host: str = "127.0.0.1"
    port: int = 8765
    bind_host: Optional[str] = None
+    # Issue #5312: CIDRs/IPs of upstream proxies/tunnels the operator declares
+    # trusted. Empty (default) means the gateway is directly exposed: forwarded
+    # headers are never believed and per-IP policy keys on the socket peer.
+    # When set, the real client IP is resolved by walking the forwarded chain
+    # across trusted hops (see ``resolve_ingress_attribution``); anything
+    # proxy-shaped but unattributable fails closed to the socket peer.
+    trusted_proxies: List[str] = field(default_factory=list)
    cors_origins: List[str] = field(default_factory=lambda: [])
    allowed_origins: List[str] = field(default_factory=lambda: [])
    auth_token: Optional[str] = None
@@ -830,6 +1070,18 @@ class GatewayConfig:
    max_concurrent_runs: int = 0  # Aggregate concurrency ceiling (0 = unlimited)
    queue_depth: int = 0  # Bounded wait queue when at the ceiling
    overflow_policy: str = "reject"  # reject | queue | shed_oldest
+    # Issue #5362: complete the back-pressure/lifecycle posture on the public
+    # config so a Python embedder / gateway.yaml can pin it declaratively —
+    # parity with the CLI ``--drain-timeout`` / ``--watchdog`` /
+    # ``--watchdog-timeout`` flags. ``None``/False keep today's behaviour
+    # (no drain, watchdog off); a CLI flag still overrides these at start.
+    drain_timeout: Optional[float] = None  # graceful-drain window on shutdown (s)
+    watchdog: bool = False  # event-loop liveness watchdog (opt-in)
+    watchdog_timeout: Optional[float] = None  # loop-stall budget before restart (s)
+    # Issue #5168: per-tenant/per-scope concurrency sub-limit within the global
+    # ceiling so one tenant cannot occupy every run slot. 0 disables the
+    # sub-limit (today's global-only behaviour).
+    max_concurrent_runs_per_scope: int = 0
    # Issue #2620: pre-auth edge protections for internet-exposed deployments.
    # Cap concurrent *unauthenticated* WebSocket connections per source IP so a
    # hostile client cannot park many half-open sockets up to max_connections
@@ -858,6 +1110,12 @@ class GatewayConfig:
    control: "EmergencyStopConfig" = field(
        default_factory=lambda: EmergencyStopConfig()
    )
+    # Issue #5207: first-class attachment ceilings for the gateway wire
+    # protocol, advertised to clients via HelloResult.policy so a custom /ws
+    # client can self-limit/chunk before sending a file. Defaults keep a
+    # sensible 10 MiB / 10-file / 256 KiB-chunk policy; attachments remain
+    # opt-in per turn so a client that never sends one is unaffected.
+    attachments: "AttachmentConfig" = field(default_factory=lambda: AttachmentConfig())
    # Issue #4766: per-session turn-execution seam. Selects *where* a session's
    # agent turn runs (see ``TurnExecutorProtocol``). ``None`` (the default)
    # resolves to ``InProcessTurnExecutor`` at runtime — today's on-loop
@@ -899,6 +1157,19 @@ class GatewayConfig:
            raise ValueError(
                "overflow_policy must be one of 'reject', 'queue', 'shed_oldest'"
            )
+        if self.drain_timeout is not None and self.drain_timeout < 0:
+            raise ValueError(
+                "drain_timeout must be >= 0 (use 0/None to disable the drain window)"
+            )
+        if self.watchdog_timeout is not None and self.watchdog_timeout <= 0:
+            raise ValueError(
+                "watchdog_timeout must be > 0 (leave unset to use the default budget)"
+            )
+        if self.max_concurrent_runs_per_scope < 0:
+            raise ValueError(
+                "max_concurrent_runs_per_scope must be >= 0 "
+                "(use 0 to disable the per-scope sub-limit)"
+            )
        if self.preauth_max_connections_per_ip < 0:
            raise ValueError(
                "preauth_max_connections_per_ip must be >= 0 (use 0 to disable "
@@ -966,6 +1237,7 @@ class GatewayConfig:
        return {
            "host": self.host,
            "port": self.port,
+            "trusted_proxies": list(self.trusted_proxies),
            "cors_origins": self.cors_origins,
            "allowed_origins": self.allowed_origins,
            "auth_token": "***" if self.auth_token else None,
@@ -981,6 +1253,10 @@ class GatewayConfig:
            "max_concurrent_runs": self.max_concurrent_runs,
            "queue_depth": self.queue_depth,
            "overflow_policy": self.overflow_policy,
+            "drain_timeout": self.drain_timeout,
+            "watchdog": self.watchdog,
+            "watchdog_timeout": self.watchdog_timeout,
+            "max_concurrent_runs_per_scope": self.max_concurrent_runs_per_scope,
            "preauth_max_connections_per_ip": self.preauth_max_connections_per_ip,
            "max_unauthorized_frames": self.max_unauthorized_frames,
            "push": self.push.to_dict(),
@@ -989,6 +1265,7 @@ class GatewayConfig:
            "liveness": self.liveness.to_dict(),
            "turn_lock": self.turn_lock.to_dict(),
            "control": self.control.to_dict(),
+            "attachments": self.attachments.to_dict(),
            # Report the active executor type so ``gateway doctor`` can surface
            # it; ``None`` means the in-process default (today's behaviour).
            "executor": (
@@ -1158,9 +1435,17 @@ class MultiChannelGatewayConfig:
                for tok, scopes in gw_data["auth_scopes"].items()
            }

+        raw_trusted = gw_data.get("trusted_proxies", [])
+        trusted_proxies = (
+            [str(p).strip() for p in raw_trusted if str(p).strip()]
+            if isinstance(raw_trusted, (list, tuple))
+            else []
+        )
+
        gateway_config = GatewayConfig(
            host=gw_data.get("host", "127.0.0.1"),
            port=gw_data.get("port", 8765),
+            trusted_proxies=trusted_proxies,
            cors_origins=gw_data.get("cors_origins", []),
            allowed_origins=gw_data.get("allowed_origins", []),
            auth_token=gw_data.get("auth_token"),
@@ -1177,6 +1462,20 @@ class MultiChannelGatewayConfig:
            max_concurrent_runs=int(gw_data.get("max_concurrent_runs", 0) or 0),
            queue_depth=int(gw_data.get("queue_depth", 0) or 0),
            overflow_policy=str(gw_data.get("overflow_policy", "reject") or "reject"),
+            drain_timeout=_parse_drain_timeout(gw_data.get("drain_timeout")),
+            watchdog=_parse_watchdog_enabled(gw_data.get("watchdog")),
+            # Prefer an explicit ``gateway.watchdog_timeout`` scalar (as emitted
+            # by ``to_dict`` and settable directly, matching the CLI
+            # ``--watchdog-timeout`` flag); otherwise derive the budget from the
+            # ``watchdog`` block so both round-trip and block forms work (#5362).
+            watchdog_timeout=(
+                _parse_watchdog_timeout_scalar(gw_data.get("watchdog_timeout"))
+                if gw_data.get("watchdog_timeout") is not None
+                else _parse_watchdog_timeout(gw_data.get("watchdog"))
+            ),
+            max_concurrent_runs_per_scope=int(
+                gw_data.get("max_concurrent_runs_per_scope", 0) or 0
+            ),
            preauth_max_connections_per_ip=int(
                gw_data.get("preauth_max_connections_per_ip", 32)
            ),
@@ -1188,6 +1487,7 @@ class MultiChannelGatewayConfig:
            liveness=LivenessConfig.from_dict(gw_data.get("liveness")),
            turn_lock=TurnLockConfig.from_dict(gw_data.get("turn_lock")),
            control=EmergencyStopConfig.from_dict(gw_data.get("control")),
+            attachments=AttachmentConfig.from_dict(gw_data.get("attachments")),
        )
        
        # Parse agents section (pass through as dicts)
@@ -52,6 +52,7 @@ from .protocols import GATEWAY_RESTART_EXIT_CODE
 __all__ = [
    "LoopWatchdogPolicy",
    "LoopWatchdog",
+    "StartupWatchdog",
 ]


@@ -437,3 +438,271 @@ class LoopWatchdog:
                    faulthandler.dump_traceback(file=fh, all_threads=True)
            except Exception:  # pragma: no cover - fail open
                pass
+
+
+class StartupWatchdog:
+    """Front-rung liveness guard for the pre-loop startup window (Issue #5265).
+
+    :class:`LoopWatchdog` only becomes meaningful once the asyncio event loop
+    is running — it probes the loop via ``loop.call_soon_threadsafe``, so it can
+    do nothing before a loop exists. That leaves the window between process
+    entry and "loop is live" (heavy imports, config/secret loading, the initial
+    adapter ``connect()``\\ s) uncovered. A wedge there — a hanging SDK import, a
+    blocking synchronous DNS/network call while minting credentials, a deadlock
+    inside ``connect()`` — leaves the process *alive but making no progress*: it
+    holds no serving loop, answers no probes, and never exits, so an external
+    supervisor never restarts it. It becomes a silent zombie with no diagnostics.
+
+    This primitive closes that gap. It arms a wall-clock **deadline** on a
+    dedicated daemon OS thread *before* the heavy work begins, so it keeps
+    working precisely when the main thread is wedged. If the deadline elapses
+    without :meth:`confirm_loop_live`, it dumps all-thread stacks via
+    :mod:`faulthandler` (so the hang is diagnosable) and hands the process back
+    to the supervisor via ``os._exit`` with a restart-intent code. A legitimately
+    slow-but-alive boot extends the deadline via :meth:`report_startup_progress`
+    (bounded, so a truly wedged phase still trips). :meth:`confirm_loop_live`
+    disarms it the instant the serving loop begins — at which point
+    :class:`LoopWatchdog` takes over.
+
+    Design mirrors :class:`LoopWatchdog`: **opt-in** (nothing arms it unless the
+    embedder calls :meth:`arm`), **fail-open** (any error inside the watchdog
+    must never kill a healthy startup), and **stdlib-only** (it must run when the
+    rest of the process cannot, so its correctness cannot depend on the code it
+    guards). Typical use at the gateway entry point::
+
+        sw = StartupWatchdog()
+        sw.arm(deadline_s=config.startup_watchdog_timeout)
+        load_config(config);   sw.report_startup_progress("config loaded")
+        connect_adapters();    sw.report_startup_progress("adapters connected")
+        loop = asyncio.new_event_loop()
+        LoopWatchdog().arm(loop)
+        sw.confirm_loop_live()  # disarm; loop watchdog now owns liveness
+        loop.run_forever()
+    """
+
+    def __init__(
+        self,
+        *,
+        on_expire: Literal["dump_and_exit", "dump_only"] = "dump_and_exit",
+        exit_code: int = GATEWAY_RESTART_EXIT_CODE,
+        dump_file: Optional[str] = None,
+        max_extensions: int = 10,
+    ) -> None:
+        if on_expire not in ("dump_and_exit", "dump_only"):
+            raise ValueError(
+                "on_expire must be 'dump_and_exit' or 'dump_only'"
+            )
+        if max_extensions < 0:
+            raise ValueError("max_extensions must be >= 0")
+        self.on_expire = on_expire
+        self.exit_code = exit_code
+        self.dump_file = dump_file
+        self.max_extensions = max_extensions
+        self._thread: Optional[threading.Thread] = None
+        # Each armed generation gets its own stop event handed straight to the
+        # thread, so a confirm/re-arm race can never let a stale generation
+        # observe a fresh generation's cleared flag and self-exit (mirrors the
+        # per-generation discipline of LoopWatchdog's deadline mode). The stop
+        # event, once set, stays set: a healthy confirm/disarm is monotonic.
+        self._stop = threading.Event()
+        # A separate wake event carries progress reports: it is briefly set to
+        # nudge the waiting thread to re-read the (extended) deadline, then
+        # cleared. Keeping it distinct from _stop means a progress nudge can
+        # never be mistaken for a disarm (and vice versa).
+        self._wake = threading.Event()
+        self._deadline_s = 0.0
+        self._extensions_used = 0
+        # Progress reports enqueue here; the watchdog thread drains this under
+        # the lock so budget is spent exactly once per report even when several
+        # reports batch between two wakes.
+        self._pending_progress = 0
+        self._expired = False
+        self._confirmed = False
+        self._lock = threading.Lock()
+
+    @property
+    def armed(self) -> bool:
+        return self._thread is not None and self._thread.is_alive()
+
+    @property
+    def expired(self) -> bool:
+        """True once the startup deadline expired (mainly for ``dump_only``)."""
+        return self._expired
+
+    @property
+    def confirmed(self) -> bool:
+        """True once :meth:`confirm_loop_live` has disarmed the watchdog."""
+        return self._confirmed
+
+    def arm(self, *, deadline_s: float, dump_stacks: bool = True) -> None:
+        """Arm a startup deadline on a dedicated daemon thread. Fail-open.
+
+        Idempotent: arming while already armed is a no-op. A non-finite or
+        non-positive ``deadline_s`` is ignored (nothing to guard). ``dump_stacks``
+        toggles the all-thread stack dump on expiry.
+        """
+        if not math.isfinite(deadline_s) or deadline_s <= 0:
+            return
+        with self._lock:
+            if self.armed:
+                return
+            stop = threading.Event()
+            wake = threading.Event()
+            self._stop = stop
+            self._wake = wake
+            self._deadline_s = deadline_s
+            self._extensions_used = 0
+            self._pending_progress = 0
+            self._expired = False
+            self._confirmed = False
+            self._dump_stacks_on_expire = dump_stacks
+            try:
+                thread = threading.Thread(
+                    target=self._run,
+                    args=(stop, wake),
+                    name="praisonai-startup-watchdog",
+                    daemon=True,
+                )
+                thread.start()
+            except Exception:  # pragma: no cover - fail open
+                self._thread = None
+                return
+            self._thread = thread
+
+    def report_startup_progress(self, phase: str) -> None:
+        """Signal forward progress, extending the deadline by one interval.
+
+        A legitimately slow boot (e.g. many adapters, cold DNS) calls this at
+        each phase so it is not killed. Extensions are bounded by
+        ``max_extensions`` so a phase that itself wedges still trips the
+        deadline. Fail-open and a no-op when not armed.
+
+        Budget accounting is done by the watchdog thread, not here: each report
+        merely enqueues a pending nudge. If several reports land before the
+        thread wakes, each still costs exactly one extension *and* each still
+        earns one fresh deadline window when the thread drains them — so a later
+        slow-but-progressing phase can never find the budget spent on resets it
+        never received (Issue #5265 review: "progress reports lose extensions").
+        """
+        if not self.armed:
+            return
+        with self._lock:
+            self._pending_progress += 1
+            wake = self._wake
+        # Wake the waiting thread so it drains the pending nudges, spends budget,
+        # and resets its window once per still-affordable report. The thread
+        # keeps waiting against its own (still-unset) stop event.
+        wake.set()
+
+    def confirm_loop_live(self) -> None:
+        """Disarm the startup watchdog; the event loop is now live.
+
+        Called immediately before :class:`LoopWatchdog` is armed. Safe to call
+        multiple times / when not armed. Sets the *current* generation's stop
+        event so that exact thread suppresses its self-exit even if it is
+        mid-dump, then drops the reference.
+        """
+        with self._lock:
+            self._confirmed = True
+            self._stop.set()
+            # Nudge the thread out of its wait immediately so the join is prompt.
+            self._wake.set()
+            thread = self._thread
+            self._thread = None
+        if thread is not None and thread is not threading.current_thread():
+            try:
+                thread.join(timeout=1.0)
+            except Exception:  # pragma: no cover - fail open
+                pass
+
+    def _run(self, stop: "threading.Event", wake: "threading.Event") -> None:
+        # Wait out the deadline against *this generation's* events. ``stop`` is
+        # set (and stays set) by a confirm/disarm — that must break us out and
+        # suppress the self-exit. ``wake`` is a transient progress nudge — on
+        # each wake we drain the queued progress reports, spending one extension
+        # of budget per report and resetting the window each time we can still
+        # afford it. A plain timeout with neither event set (and no affordable
+        # pending progress) means the startup genuinely wedged.
+        start = time.monotonic()
+        while not stop.is_set():
+            remaining = (start + self._deadline_s) - time.monotonic()
+            if remaining <= 0:
+                break
+            if wake.wait(remaining):
+                wake.clear()
+                if stop.is_set():
+                    return
+                # Drain queued progress reports. Each report costs one extension
+                # and, while budget remains, grants one fresh deadline window —
+                # so batched reports never lose the resets they paid for.
+                extended = False
+                with self._lock:
+                    pending = self._pending_progress
+                    self._pending_progress = 0
+                    for _ in range(pending):
+                        if self._extensions_used >= self.max_extensions:
+                            break
+                        self._extensions_used += 1
+                        extended = True
+                if extended:
+                    start = time.monotonic()
+                # Whether or not budget remained, loop back and re-check the
+                # (possibly unchanged) deadline against this generation's stop.
+                continue
+        self._on_expire(stop)
+
+    def _on_expire(self, stop: "threading.Event") -> None:
+        # Only fire if this generation was not disarmed. A confirm_loop_live()
+        # that raced in has set this exact event, so a stale generation stays
+        # inert and never self-exits a healthy startup.
+        if stop.is_set():
+            return
+        self._expired = True
+        if getattr(self, "_dump_stacks_on_expire", True):
+            try:
+                self._dump_stacks()
+            except Exception:  # pragma: no cover - fail open
+                pass
+        if self.on_expire == "dump_and_exit":
+            # If confirm_loop_live() raced in while we were dumping, honour it.
+            if stop.is_set():
+                return
+            try:
+                sys.stderr.flush()
+            except Exception:  # pragma: no cover - fail open
+                pass
+            # Final re-check the instant before the irreversible os._exit: the
+            # flush above can block (e.g. a full pipe) long enough for a
+            # concurrent confirm_loop_live() to land. A startup that has just
+            # confirmed a live loop must never be terminated (Issue #5265
+            # review: "confirmation cannot prevent late exit").
+            if stop.is_set():
+                return
+            os._exit(self.exit_code)
+
+    def _dump_stacks(self) -> None:
+        reason = (
+            f"gateway startup did not confirm a live event loop within "
+            f"~{self._deadline_s:.0f}s"
+        )
+        header = (
+            f"praisonai-startup-watchdog: {reason}; "
+            f"dumping all-thread stacks\n"
+        )
+        try:
+            sys.stderr.write(header)
+        except Exception:  # pragma: no cover - fail open
+            pass
+        try:
+            faulthandler.dump_traceback(file=sys.stderr, all_threads=True)
+        except Exception:  # pragma: no cover - fail open
+            pass
+        dump_file = self.dump_file
+        if dump_file:
+            try:
+                with open(dump_file, "a", encoding="utf-8") as fh:
+                    fh.write(header)
+                    fh.flush()
+                    faulthandler.dump_traceback(file=fh, all_threads=True)
+            except Exception:  # pragma: no cover - fail open
+                pass
@@ -0,0 +1,365 @@
+"""Session-projection reducer for Gateway clients (Issue #5324).
+
+A pure, dependency-free state machine that folds an initial snapshot plus a
+stream of :class:`~praisonaiagents.gateway.protocols.GatewayEvent`\\ s into a
+consistent, de-duplicated, memory-bounded view of a session. It turns *a
+stream of frames* into *a correct, resumable view* so every gateway client
+(the bundled Python client, the ``praisonai-ts`` mirror, the realtime
+dashboard, third-party integrations) stops re-implementing the same fiddly
+reconciliation logic:
+
+- **Snapshot + fold**: apply a snapshot, then fold live events into ``entries``.
+- **Identity de-duplication**: a final assistant message that arrives both as a
+  streamed delta sequence *and* as a persisted transcript row renders once.
+- **Optimistic reconciliation**: a locally-echoed outbound message is replaced
+  (not duplicated) by its durable counterpart, keyed on ``request_id``.
+- **Gap handling**: a sequence gap flips ``has_transport_gap``; after a resync
+  the reducer converges rather than appending duplicates.
+- **Bounded retention**: tracked runs are capped (LRU) so long-lived sessions
+  stay memory-bounded; a run with an active (unfinished) stream is never
+  evicted.
+
+This module lives in **core** because it is a pure reducer built on types core
+already owns (``GatewayEvent``/``GatewayMessage``/``EventType``); it must not be
+trapped in the transport client so all clients can share it. It has **no**
+third-party dependencies and does no I/O.
+"""
+
+from __future__ import annotations
+
+from collections import OrderedDict
+from dataclasses import dataclass, field, replace
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from .protocols import EventType, GatewayEvent, GatewayMessage
+
+__all__ = [
+    "RunView",
+    "SessionProjectionState",
+    "SessionProjection",
+]
+
+
+def _event_type_value(event: GatewayEvent) -> str:
+    """Return the wire string for an event type (enum or raw str)."""
+    etype = event.type
+    return etype.value if isinstance(etype, EventType) else str(etype)
+
+
+def _run_id_of(event: GatewayEvent) -> Optional[str]:
+    """Best-effort run/turn identifier for a streaming event."""
+    data = event.data or {}
+    for key in ("run_id", "turn_id", "message_id", "response_id"):
+        value = data.get(key)
+        if value:
+            return str(value)
+    return None
+
+
+@dataclass(frozen=True)
+class RunView:
+    """Immutable view of a single agent run/turn within a session.
+
+    Attributes:
+        run_id: The run/turn identifier the streamed deltas belong to.
+        text: The accumulated streamed text so far.
+        done: Whether the stream has ended (``STREAM_END``) for this run.
+        final_message_id: ``message_id`` of the persisted final row, once the
+            durable transcript row for this run has been reconciled in.
+    """
+
+    run_id: str
+    text: str = ""
+    done: bool = False
+    final_message_id: Optional[str] = None
+
+
+@dataclass(frozen=True)
+class SessionProjectionState:
+    """Immutable, de-duplicated, bounded view of a session.
+
+    Attributes:
+        entries: Ordered transcript messages, de-duplicated by identity.
+        runs: Per-run streaming views, keyed by ``run_id``.
+        has_transport_gap: ``True`` when a sequence gap was observed and a
+            resync has not yet reconciled the state.
+    """
+
+    entries: Tuple[GatewayMessage, ...] = ()
+    runs: Dict[str, RunView] = field(default_factory=dict)
+    has_transport_gap: bool = False
+
+
+class SessionProjection:
+    """Fold a snapshot + :class:`GatewayEvent`\\ s into a consistent view.
+
+    The reducer is idempotent: applying the same persisted message twice (e.g.
+    once from the live stream's ``STREAM_END`` and once as a durable transcript
+    row on reconnect) yields a single entry. Feed it with
+    :meth:`apply_snapshot` then :meth:`apply`; read :attr:`state` at any time.
+
+    Args:
+        max_tracked_runs: Cap on how many finished runs are tracked before the
+            oldest are evicted (LRU). Runs with an active/unfinished stream are
+            never evicted. Must be positive.
+    """
+
+    def __init__(self, *, max_tracked_runs: int = 200) -> None:
+        if max_tracked_runs <= 0:
+            raise ValueError("max_tracked_runs must be positive")
+        self._max_tracked_runs = max_tracked_runs
+        self._entries: List[GatewayMessage] = []
+        # message_id -> index into self._entries, for identity de-dup.
+        self._by_message_id: Dict[str, int] = {}
+        # request_id -> index into self._entries, for optimistic reconciliation.
+        self._by_request_id: Dict[str, int] = {}
+        # run_id -> RunView, ordered for LRU eviction.
+        self._runs: "OrderedDict[str, RunView]" = OrderedDict()
+        self._expected_sequence: Optional[int] = None
+        self._has_gap: bool = False
+
+    # -- public API -------------------------------------------------------
+
+    @property
+    def state(self) -> SessionProjectionState:
+        """Return the current immutable projection state."""
+        return SessionProjectionState(
+            entries=tuple(self._entries),
+            runs=dict(self._runs),
+            has_transport_gap=self._has_gap,
+        )
+
+    def apply_snapshot(
+        self, snapshot: Union[Dict[str, Any], Any]
+    ) -> SessionProjectionState:
+        """Replace all tracked state from a full snapshot.
+
+        A snapshot resets gap tracking and rebuilds the transcript from the
+        persisted rows it carries, so a resync converges rather than appending
+        duplicates. Accepts a mapping (``{"messages"/"entries": [...],
+        "cursor"/"sequence": N}``) or any object exposing ``messages``/
+        ``entries`` and an optional ``sequence``/``cursor`` attribute.
+        """
+        self._entries = []
+        self._by_message_id = {}
+        self._by_request_id = {}
+        self._runs = OrderedDict()
+        self._has_gap = False
+        self._expected_sequence = None
+
+        rows = _snapshot_rows(snapshot)
+        for row in rows:
+            message = _coerce_message(row)
+            if message is not None:
+                self._upsert_message(message)
+
+        seq = _snapshot_sequence(snapshot)
+        if seq is not None:
+            self._expected_sequence = seq + 1
+        return self.state
+
+    def apply(self, event: GatewayEvent) -> SessionProjectionState:
+        """Fold a single event into the state and return the new state.
+
+        Idempotent for persisted messages (de-duped by ``message_id``) and for
+        optimistic echoes (reconciled by ``request_id``). Streaming deltas are
+        accumulated per run; ``STREAM_END`` closes the run. Sequence gaps set
+        :attr:`SessionProjectionState.has_transport_gap`.
+        """
+        self._track_sequence(event)
+
+        etype = _event_type_value(event)
+        if etype in _STREAM_DELTA_TYPES:
+            self._apply_delta(event)
+        elif etype == EventType.STREAM_END.value:
+            self._apply_stream_end(event)
+        elif etype in _MESSAGE_TYPES:
+            self._apply_message(event)
+        # Other event types (health, presence, typing, ...) carry no transcript
+        # state; they are intentionally ignored by the projection.
+        return self.state
+
+    # -- sequence / gap tracking -----------------------------------------
+
+    def _track_sequence(self, event: GatewayEvent) -> None:
+        seq = event.sequence
+        if seq is None:
+            return
+        if self._expected_sequence is not None and seq > self._expected_sequence:
+            # A hole in the monotonic sequence: mark the view as gapped until a
+            # snapshot/resync reconciles it.
+            self._has_gap = True
+        # Advance the watermark to the highest sequence seen.
+        if self._expected_sequence is None or seq >= self._expected_sequence:
+            self._expected_sequence = seq + 1
+
+    # -- message handling -------------------------------------------------
+
+    def _apply_message(self, event: GatewayEvent) -> None:
+        message = _coerce_message(event.data)
+        if message is None:
+            return
+        request_id = _request_id_of(event, message)
+        self._upsert_message(message, request_id=request_id)
+
+    def _upsert_message(
+        self, message: GatewayMessage, *, request_id: Optional[str] = None
+    ) -> None:
+        # Identity de-dup: same persisted row seen twice renders once.
+        existing = self._by_message_id.get(message.message_id)
+        if existing is not None:
+            self._entries[existing] = message
+            self._reindex(existing, message, request_id)
+            return
+
+        # Optimistic reconciliation: a durable row replaces its local echo.
+        if request_id is not None:
+            echoed = self._by_request_id.get(request_id)
+            if echoed is not None:
+                stale_id = self._entries[echoed].message_id
+                self._entries[echoed] = message
+                # Drop the echo's provisional message_id so a re-delivered echo
+                # cannot later overwrite the durable row via identity lookup.
+                if stale_id != message.message_id:
+                    self._by_message_id.pop(stale_id, None)
+                self._reindex(echoed, message, request_id)
+                return
+
+        self._entries.append(message)
+        idx = len(self._entries) - 1
+        self._by_message_id[message.message_id] = idx
+        if request_id is not None:
+            self._by_request_id[request_id] = idx
+
+    def _reindex(
+        self,
+        idx: int,
+        message: GatewayMessage,
+        request_id: Optional[str],
+    ) -> None:
+        self._by_message_id[message.message_id] = idx
+        if request_id is not None:
+            self._by_request_id[request_id] = idx
+
+    # -- streaming handling ----------------------------------------------
+
+    def _apply_delta(self, event: GatewayEvent) -> None:
+        run_id = _run_id_of(event)
+        if run_id is None:
+            return
+        chunk = _delta_text_of(event)
+        current = self._runs.get(run_id)
+        if current is None:
+            current = RunView(run_id=run_id, text=chunk)
+        else:
+            current = replace(current, text=current.text + chunk)
+        self._runs[run_id] = current
+        self._runs.move_to_end(run_id)
+        self._evict_runs()
+
+    def _apply_stream_end(self, event: GatewayEvent) -> None:
+        run_id = _run_id_of(event)
+        if run_id is None:
+            return
+        current = self._runs.get(run_id) or RunView(run_id=run_id)
+        final_id = None
+        data = event.data or {}
+        final_id = data.get("message_id") or data.get("final_message_id")
+        current = replace(
+            current,
+            done=True,
+            final_message_id=str(final_id) if final_id else current.final_message_id,
+        )
+        self._runs[run_id] = current
+        self._runs.move_to_end(run_id)
+        self._evict_runs()
+
+    def _evict_runs(self) -> None:
+        if len(self._runs) <= self._max_tracked_runs:
+            return
+        # Evict oldest finished runs first; never evict an active stream.
+        for run_id in list(self._runs.keys()):
+            if len(self._runs) <= self._max_tracked_runs:
+                break
+            if self._runs[run_id].done:
+                del self._runs[run_id]
+
+
+# Streaming delta event types whose text is accumulated per run.
+_STREAM_DELTA_TYPES = frozenset(
+    {
+        EventType.TOKEN_STREAM.value,
+        "delta_text",
+    }
+)
+
+# Event types carrying a persisted transcript message.
+_MESSAGE_TYPES = frozenset(
+    {
+        EventType.MESSAGE.value,
+        "message_persisted",
+    }
+)
+
+
+def _delta_text_of(event: GatewayEvent) -> str:
+    data = event.data or {}
+    for key in ("delta", "text", "content", "token", "chunk"):
+        value = data.get(key)
+        if isinstance(value, str):
+            return value
+    return ""
+
+
+def _request_id_of(
+    event: GatewayEvent, message: GatewayMessage
+) -> Optional[str]:
+    data = event.data or {}
+    rid = data.get("request_id") or message.metadata.get("request_id")
+    return str(rid) if rid else None
+
+
+def _coerce_message(row: Any) -> Optional[GatewayMessage]:
+    if isinstance(row, GatewayMessage):
+        return row
+    if isinstance(row, dict):
+        payload = row
+        # A "message" event wraps the row under a "message" key in some frames.
+        if "message" in row and isinstance(row["message"], (dict, GatewayMessage)):
+            inner = row["message"]
+            return inner if isinstance(inner, GatewayMessage) else _coerce_message(inner)
+        if "content" not in payload and "message_id" not in payload:
+            return None
+        try:
+            return GatewayMessage.from_dict(payload)
+        except Exception:
+            return None
+    return None
+
+
+def _snapshot_rows(snapshot: Any) -> List[Any]:
+    if isinstance(snapshot, dict):
+        for key in ("messages", "entries", "transcript"):
+            rows = snapshot.get(key)
+            if isinstance(rows, list):
+                return rows
+        return []
+    for attr in ("messages", "entries", "transcript"):
+        rows = getattr(snapshot, attr, None)
+        if isinstance(rows, list):
+            return rows
+    return []
+
+
+def _snapshot_sequence(snapshot: Any) -> Optional[int]:
+    if isinstance(snapshot, dict):
+        for key in ("sequence", "cursor", "seq"):
+            value = snapshot.get(key)
+            if isinstance(value, int):
+                return value
+        return None
+    for attr in ("sequence", "cursor", "seq"):
+        value = getattr(snapshot, attr, None)
+        if isinstance(value, int):
+            return value
+    return None
@@ -193,6 +193,41 @@ class GoalLoopMixin:
        self._persist_goal_state()
        return "continue", reason

+    async def _goal_gate_async(self, response: str) -> Optional[Tuple[str, str]]:
+        """Async counterpart of :meth:`_goal_gate`.
+
+        Both the configured verification hooks (which may shell out) and the
+        completion judge's ``litellm.completion(...)`` call are blocking I/O, so
+        the whole gate is offloaded to a worker thread. Mirrors
+        ``_verification_gate_async``'s existing pattern for the same reason: a
+        slow judge/hook must never stall the event loop shared by other
+        concurrent agents.
+
+        Cancellation safety: the gate mutates shared goal state
+        (``_goal_state``, verdict counters, journal, persisted metadata). A bare
+        ``asyncio.to_thread`` worker keeps running after the awaiting coroutine
+        is cancelled, so its late writes could race with ``run_goal_async``'s
+        teardown (which nulls ``_goal_state``) or a subsequent run reusing the
+        same agent. We therefore join the worker before propagating the
+        cancellation, guaranteeing the gate finishes against the goal state it
+        started on and no orphaned thread mutates a newer run.
+        """
+        state = getattr(self, "_goal_state", None)
+        if state is None or state.status != "active":
+            return None
+        import asyncio
+        task = asyncio.ensure_future(asyncio.to_thread(self._goal_gate, response))
+        try:
+            return await task
+        except asyncio.CancelledError:
+            # Let the in-flight gate finish so its state mutations complete
+            # before run_goal_async's finally clears _goal_state, then re-raise.
+            try:
+                await asyncio.shield(task)
+            except Exception:
+                pass
+            raise
+
    # -- public API ------------------------------------------------------------

    def _resume_or_new_goal_state(
@@ -80,6 +80,7 @@ __all__ = [
    "MessageSendingInput",
    "MessageSentInput",
    "MessageUndeliveredInput",
+    "PlatformEventInput",
    # Gateway & schedule lifecycle event inputs
    "GatewayStartInput",
    "GatewayStopInput",
@@ -161,6 +162,7 @@ _LAZY_GROUPS = {
        'MessageSendingInput': ('praisonaiagents.hooks.events', 'MessageSendingInput'),
        'MessageSentInput': ('praisonaiagents.hooks.events', 'MessageSentInput'),
        'MessageUndeliveredInput': ('praisonaiagents.hooks.events', 'MessageUndeliveredInput'),
+        'PlatformEventInput': ('praisonaiagents.hooks.events', 'PlatformEventInput'),
        'GatewayStartInput': ('praisonaiagents.hooks.events', 'GatewayStartInput'),
        'GatewayStopInput': ('praisonaiagents.hooks.events', 'GatewayStopInput'),
        'ScheduleTriggerInput': ('praisonaiagents.hooks.events', 'ScheduleTriggerInput'),
@@ -388,6 +388,42 @@ class MessageUndeliveredInput(HookInput):
        return base


+@dataclass
+class PlatformEventInput(HookInput):
+    """Input for inbound platform-event hooks (Issue #5161).
+
+    Carries a normalised, non-message platform event — a reaction, edit,
+    deletion, membership change or thread creation — to
+    ``REACTION_RECEIVED`` / ``MESSAGE_EDITED`` / ``MESSAGE_DELETED`` /
+    ``MEMBER_JOINED`` / ``MEMBER_LEFT`` / ``THREAD_CREATED`` hooks. Fields mirror
+    :class:`~praisonaiagents.bots.protocols.PlatformEvent`; the ``kind`` field
+    disambiguates reaction add vs remove within ``REACTION_RECEIVED``. Emitted
+    by ``MessageHookMixin.fire_platform_event`` (praisonai-bot).
+    """
+    kind: str = ""
+    platform: str = ""
+    chat_id: str = ""
+    user_id: str = ""
+    message_id: Optional[str] = None
+    emoji: Optional[str] = None
+    new_text: Optional[str] = None
+    thread_id: Optional[str] = None
+
+    def to_dict(self) -> Dict[str, Any]:
+        base = super().to_dict()
+        base.update({
+            "kind": self.kind,
+            "platform": self.platform,
+            "chat_id": self.chat_id,
+            "user_id": self.user_id,
+            "message_id": self.message_id,
+            "emoji": self.emoji,
+            "new_text": self.new_text[:500] if self.new_text else None,
+            "thread_id": self.thread_id,
+        })
+        return base
+
+
@dataclass
 class GatewayStartInput(HookInput):
    """Input for GATEWAY_START hooks (gateway/BotOS started)."""
@@ -53,6 +53,9 @@ class HookRegistry:
        }
        self._enabled = enabled
        self._global_timeout = 60.0
+        # Monotonic counter giving each registered hook a stable secondary sort
+        # key so equal-priority hooks keep deterministic registration order.
+        self._seq_counter = 0
        # Guard mutation/iteration of the per-event hook lists. The default
        # registry is a cross-agent process-wide singleton, so a plugin
        # (un)registering a hook on one thread can race a get_hooks() iteration
@@ -80,6 +83,8 @@ class HookRegistry:
            The hook ID
        """
        with self._lock:
+            hook._seq = self._seq_counter
+            self._seq_counter += 1
            self._hooks[hook.event].append(hook)
        logger.debug(f"Registered hook '{hook.name}' for event '{hook.event.value}'")
        return hook.id
@@ -93,7 +98,8 @@ class HookRegistry:
        description: Optional[str] = None,
        sequential: bool = False,
        timeout: float = 60.0,
-        is_async: bool = False
+        is_async: bool = False,
+        priority: int = 100
    ) -> str:
        """
        Register a Python function as a hook.
@@ -107,6 +113,7 @@ class HookRegistry:
            sequential: Whether to run sequentially
            timeout: Timeout in seconds
            is_async: Whether the function is async
+            priority: Execution order (lower runs earlier). Default 100
            
        Returns:
            The hook ID
@@ -119,7 +126,8 @@ class HookRegistry:
            description=description,
            sequential=sequential,
            timeout=timeout,
-            is_async=is_async
+            is_async=is_async,
+            priority=priority
        )
        return self.register(hook)
    
@@ -133,7 +141,8 @@ class HookRegistry:
        sequential: bool = False,
        timeout: float = 60.0,
        env: Optional[Dict[str, str]] = None,
-        shell: bool = True
+        shell: bool = True,
+        priority: int = 100
    ) -> str:
        """
        Register a shell command as a hook.
@@ -148,6 +157,7 @@ class HookRegistry:
            timeout: Timeout in seconds
            env: Additional environment variables
            shell: Whether to run in shell mode
+            priority: Execution order (lower runs earlier). Default 100
            
        Returns:
            The hook ID
@@ -161,7 +171,8 @@ class HookRegistry:
            sequential=sequential,
            timeout=timeout,
            env=env or {},
-            shell=shell
+            shell=shell,
+            priority=priority
        )
        return self.register(hook)
    
@@ -170,7 +181,8 @@ class HookRegistry:
        event: HookEvent,
        matcher: Optional[str] = None,
        sequential: bool = False,
-        timeout: float = 60.0
+        timeout: float = 60.0,
+        priority: int = 100
    ) -> Callable:
        """
        Decorator to register a function as a hook.
@@ -180,6 +192,7 @@ class HookRegistry:
            matcher: Optional regex pattern to match targets
            sequential: Whether to run sequentially
            timeout: Timeout in seconds
+            priority: Execution order (lower runs earlier). Default 100
            
        Returns:
            Decorator function
@@ -199,7 +212,8 @@ class HookRegistry:
                matcher=matcher,
                sequential=sequential,
                timeout=timeout,
-                is_async=is_async
+                is_async=is_async,
+                priority=priority
            )
            
            @wraps(func)
@@ -253,6 +267,10 @@ class HookRegistry:
        with self._lock:
            hooks = list(self._hooks.get(event, []))
        
+        # Sort by priority (lower runs earlier); ties fall back to registration
+        # order via the monotonic _seq for deterministic, reproducible ordering.
+        hooks.sort(key=lambda h: (h.priority, h._seq))
+        
        if target is None:
            return [h for h in hooks if h.enabled]
        
@@ -284,20 +302,25 @@ class HookRegistry:
            Dictionary mapping event names to hook info
        """
        result = {}
-        for event in HookEvent:
-            hooks = self._hooks[event]
-            if hooks:
-                result[event.value] = [
-                    {
-                        "id": h.id,
-                        "name": h.name,
-                        "type": "command" if isinstance(h, CommandHook) else "function",
-                        "matcher": h.matcher,
-                        "enabled": h.enabled,
-                        "sequential": h.sequential
-                    }
-                    for h in hooks
-                ]
+        with self._lock:
+            for event in HookEvent:
+                hooks = list(self._hooks[event])
+                if hooks:
+                    # Present hooks in effective execution order so an operator
+                    # can see the sequence get_hooks() will run them in.
+                    hooks.sort(key=lambda h: (h.priority, h._seq))
+                    result[event.value] = [
+                        {
+                            "id": h.id,
+                            "name": h.name,
+                            "type": "command" if isinstance(h, CommandHook) else "function",
+                            "matcher": h.matcher,
+                            "enabled": h.enabled,
+                            "sequential": h.sequential,
+                            "priority": h.priority
+                        }
+                        for h in hooks
+                    ]
        return result
    
    def enable_hook(self, hook_id: str) -> bool:
@@ -359,7 +382,7 @@ def set_default_registry(registry: HookRegistry):
 def add_hook(
    event: Union[str, HookEvent],
    callback: Optional[Callable[[HookInput], HookResult]] = None,
-    priority: int = 10,
+    priority: int = 100,
    matcher: Optional[str] = None
 ) -> Union[str, Callable]:
    """Register a hook callback. Simplified API.
@@ -376,7 +399,9 @@ def add_hook(
    Args:
        event: Hook event name ('before_tool', 'after_llm', etc.) or HookEvent enum
        callback: Function to call when hook fires (optional when using as decorator)
-        priority: Execution order (lower = earlier). Default 10. (Reserved for future use)
+        priority: Execution order (lower = earlier). Default 100 — same default
+            bucket as ``register_function``/``on``, so hooks added through either
+            API keep registration order unless an explicit priority is given.
        matcher: Optional regex pattern to match specific targets (e.g., tool names)
        
    Returns:
@@ -402,7 +427,8 @@ def add_hook(
            get_default_registry().register_function(
                event=event,
                func=func,
-                matcher=matcher
+                matcher=matcher,
+                priority=priority
            )
            return func
        return decorator
@@ -410,7 +436,8 @@ def add_hook(
    return get_default_registry().register_function(
        event=event,
        func=callback,
-        matcher=matcher
+        matcher=matcher,
+        priority=priority
    )

 def remove_hook(hook_id: str) -> bool:
@@ -102,25 +102,41 @@ class HookRunner:
        if not hooks:
            return []
        
-        # Separate sequential and parallel hooks
-        sequential_hooks = [h for h in hooks if h.sequential]
-        parallel_hooks = [h for h in hooks if not h.sequential]
-        
-        results = []
-        
-        # Execute parallel hooks first
-        if parallel_hooks:
-            parallel_results = await self._execute_parallel(
-                parallel_hooks, event, input_data
-            )
-            results.extend(parallel_results)
-        
-        # Execute sequential hooks
-        if sequential_hooks:
-            sequential_results = await self._execute_sequential(
-                sequential_hooks, event, input_data
-            )
-            results.extend(sequential_results)
+        # The registry hands us hooks already sorted by (priority, _seq), so the
+        # list is in the exact order the operator asked them to run. Preserve
+        # that global order across parallel/sequential boundaries: walk the list
+        # and run each *contiguous* run of same-mode hooks together. A redact
+        # (sequential, priority 10) that must precede a trace (parallel,
+        # priority 20) therefore still runs first — splitting all parallel hooks
+        # to the front would have leaked the unredacted input to the tracer.
+        # Adjacent parallel hooks are still gathered concurrently; only their
+        # position relative to a lower-priority sequential hook is respected.
+        results: List[HookExecutionResult] = []
+        current_input = input_data
+        i = 0
+        n = len(hooks)
+        while i < n:
+            j = i + 1
+            is_seq = hooks[i].sequential
+            while j < n and hooks[j].sequential == is_seq:
+                j += 1
+            run = hooks[i:j]
+            if is_seq:
+                run_results = await self._execute_sequential(run, event, current_input)
+                # A sequential run may have blocked; stop the whole chain so a
+                # later parallel group never runs after a denial (matches the
+                # sequential-hook break semantics).
+                results.extend(run_results)
+                if any(
+                    r.output and r.output.is_denied() for r in run_results
+                ):
+                    break
+            else:
+                run_results = await self._execute_parallel(
+                    run, event, current_input
+                )
+                results.extend(run_results)
+            i = j
        
        return results
    
@@ -81,7 +81,22 @@ class HookEvent(str, Enum):
    # the event that is actually emitted.
    BEFORE_MESSAGE = "message_received"
    AFTER_MESSAGE = "message_sending"
-    
+
+    # Inbound platform events (non-message): the gateway hearing more than text.
+    # A user reacting/editing/deleting a message, joining/leaving a channel, or a
+    # thread being created are delivered by the platform SDK but previously
+    # dropped before reaching the agent or a hook. Adapters normalise the native
+    # event into a ``PlatformEvent`` and emit it via
+    # ``MessageHookMixin.fire_platform_event`` (praisonai-bot); platforms that
+    # cannot deliver a given event simply never emit it (capability-gated,
+    # opt-in). Payload: ``PlatformEventInput`` (Issue #5161).
+    REACTION_RECEIVED = "reaction_received"
+    MESSAGE_EDITED = "message_edited"
+    MESSAGE_DELETED = "message_deleted"
+    MEMBER_JOINED = "member_joined"
+    MEMBER_LEFT = "member_left"
+    THREAD_CREATED = "thread_created"
+
    # Gateway lifecycle
    GATEWAY_START = "gateway_start"
    GATEWAY_STOP = "gateway_stop"
@@ -301,6 +316,8 @@ class HookDefinition:
    name: Optional[str] = None
    description: Optional[str] = None
    timeout: float = 60.0  # Timeout in seconds
+    priority: int = 100  # Execution order (lower runs earlier)
+    _seq: int = field(default=0, repr=False)  # Registration order for stable ties
    
    def matches(self, target: str) -> bool:
        """Check if this hook matches the target (tool name, etc.)."""
@@ -182,7 +182,11 @@ class ChromaKnowledgeAdapter:
        carrying api_base when the embedder points at a local server, so the
        request cannot escape to a remote provider.
        """
-        block = self._embedder or {}
+        # getattr, not attribute access: the adapter is legitimately built
+        # without __init__ in places (and in tests), and an embedding call must
+        # not die on a missing attribute -- it should fall back to the
+        # environment default exactly as it did before the embedder was wired.
+        block = getattr(self, "_embedder", None) or {}
        provider = (block.get("provider") or "").strip().lower()
        cfg = block.get("config") or {}
        model = cfg.get("model")
@@ -207,6 +211,8 @@ class ChromaKnowledgeAdapter:
               **kwargs: Any):
        """Search for relevant content in ChromaDB."""
        from ..models import SearchResult, SearchResultItem
+        from ..protocols import require_scope
+        require_scope(user_id, agent_id, run_id, "search", backend="chroma")
        
        # Get embedding for query
        embedding_model, embedding_kwargs = self._embedding_call()
@@ -232,9 +238,15 @@ class ChromaKnowledgeAdapter:
        if run_id is not None:
            where_filter["run_id"] = run_id
        
-        # Merge with additional filters
+        # Merge with additional filters, but never let a caller-supplied filter
+        # overwrite the tenant-scope constraint: filters={"user_id": "bob"} must
+        # not widen an already-scoped query back onto another tenant's data
+        # (issue #5319).
        if filters:
-            where_filter.update(filters)
+            for key, value in filters.items():
+                if key in where_filter:
+                    continue
+                where_filter[key] = value
        
        # Search ChromaDB
        try:
@@ -285,6 +297,8 @@ class ChromaKnowledgeAdapter:
            **kwargs: Any):
        """Add content to ChromaDB."""
        from ..models import AddResult
+        from ..protocols import require_scope
+        require_scope(user_id, agent_id, run_id, "add", backend="chroma")
        
        # Convert content to string
        content_str = str(content)
@@ -371,6 +385,8 @@ class ChromaKnowledgeAdapter:
                run_id: Optional[str] = None, limit: int = 100, **kwargs: Any):
        """Get all items from ChromaDB."""
        from ..models import SearchResult, SearchResultItem
+        from ..protocols import require_scope
+        require_scope(user_id, agent_id, run_id, "get_all", backend="chroma")
        
        try:
            # ChromaDB doesn't support get_all with filters easily, so we'll use peek
@@ -506,8 +522,11 @@ class SQLiteKnowledgeAdapter:
               **kwargs: Any):
        """Search for relevant content in SQLite."""
        from ..models import SearchResult, SearchResultItem
+        from ..protocols import require_scope
        import json
        
+        require_scope(user_id, agent_id, run_id, "search", backend="sqlite")
+        
        conn = self._get_conn()
        
        # Build query with filters
@@ -524,6 +543,19 @@ class SQLiteKnowledgeAdapter:
            sql += " AND run_id = ?"
            params.append(run_id)
        
+        # Apply metadata filters against the stored JSON metadata column so a
+        # filtered search returns the same subset chroma/mem0 already return
+        # instead of silently discarding the filters (issue #5319).
+        if filters:
+            for key, value in filters.items():
+                # Quote the key in the JSON path so a top-level metadata key that
+                # itself contains dots (e.g. "source.type") is matched literally
+                # instead of being interpreted as a nested path (issue #5319).
+                escaped_key = key.replace('"', '""')
+                sql += " AND json_extract(metadata, ?) = ?"
+                params.append(f'$."{escaped_key}"')
+                params.append(value)
+        
        sql += " LIMIT ?"
        params.append(limit)
        
@@ -547,9 +579,12 @@ class SQLiteKnowledgeAdapter:
            **kwargs: Any):
        """Add content to SQLite."""
        from ..models import AddResult
+        from ..protocols import require_scope
        import json
        import time
        
+        require_scope(user_id, agent_id, run_id, "add", backend="sqlite")
+        
        content_str = str(content)
        doc_id = str(time.time_ns())
        metadata_json = json.dumps(metadata or {})
@@ -592,8 +627,11 @@ class SQLiteKnowledgeAdapter:
                run_id: Optional[str] = None, limit: int = 100, **kwargs: Any):
        """Get all items from SQLite."""
        from ..models import SearchResult, SearchResultItem
+        from ..protocols import require_scope
        import json
        
+        require_scope(user_id, agent_id, run_id, "get_all", backend="sqlite")
+        
        conn = self._get_conn()
        
        # Build query with filters
@@ -137,26 +137,15 @@ class Mem0Adapter:
    def _safe_mem0_search(self, **kwargs):
        """
        Defensive wrapper for mem0.search() to handle MongoDB vector store compatibility.
-        
-        Catches TypeError about unexpected 'vectors' kwarg and returns safe fallback.
-        This addresses the upstream mem0 bug: https://github.com/mem0ai/mem0/issues/3185
+
+        Delegates to the shared ``safe_mem0_search`` helper so the workaround for
+        the upstream mem0 bug (https://github.com/mem0ai/mem0/issues/3185) lives in
+        a single place and cannot drift from the memory-layer copy. The shared
+        helper returns an empty ``list`` on the fallback branch, which
+        ``_normalize_mem0_results`` already handles alongside the dict shape.
        """
-        try:
-            return self.memory.search(**kwargs)
-        except TypeError as e:
-            error_msg = str(e).lower()
-            if "unexpected keyword argument" in error_msg and "vectors" in error_msg:
-                logger.warning(
-                    "Detected mem0 MongoDB vector store compatibility issue. "
-                    "This is a known upstream bug: https://github.com/mem0ai/mem0/issues/3185. "
-                    "The MongoDB vector store requires Atlas and has signature mismatches. "
-                    "Returning empty results. Consider using Qdrant or Chroma as vector store backends."
-                )
-                # Return empty results structure that matches expected mem0 format
-                return {"results": []}
-            else:
-                # Re-raise if it's a different TypeError
-                raise
+        from ...memory.adapters.factories import safe_mem0_search
+        return safe_mem0_search(self.memory, **kwargs)
    
    def _normalize_mem0_results(self, raw: Any) -> SearchResult:
        """
@@ -169,7 +169,12 @@ class MongoDBKnowledgeAdapter:
            
        Returns:
            Document ID of inserted content
+
+        Raises:
+            ScopeRequiredError: If no scope identifier is provided
        """
+        from ..protocols import require_scope
+        require_scope(user_id, agent_id, run_id, "add", backend="mongodb")
        try:
            # Generate embedding if model is available
            embedding = None
@@ -210,6 +215,7 @@ class MongoDBKnowledgeAdapter:
        user_id: Optional[str] = None,
        agent_id: Optional[str] = None,
        run_id: Optional[str] = None,
+        filters: Optional[Dict[str, Any]] = None,
        **kwargs
    ) -> List[Dict[str, Any]]:
        """
@@ -221,11 +227,17 @@ class MongoDBKnowledgeAdapter:
            user_id: Optional user scope for tenant isolation
            agent_id: Optional agent scope for tenant isolation
            run_id: Optional run scope for tenant isolation
+            filters: Optional metadata filters applied to the query
            **kwargs: Additional search parameters
            
        Returns:
            List of matching documents
+
+        Raises:
+            ScopeRequiredError: If no scope identifier is provided
        """
+        from ..protocols import require_scope
+        require_scope(user_id, agent_id, run_id, "search", backend="mongodb")
        try:
            results = []

@@ -237,6 +249,13 @@ class MongoDBKnowledgeAdapter:
                    "run_id": run_id,
                }.items() if v is not None
            }
+
+            # Fold metadata filters in alongside the scope so a filtered search
+            # returns the same subset chroma/mem0 already return instead of
+            # silently discarding the filters (issue #5319).
+            if filters:
+                for key, value in filters.items():
+                    scope_filter[f"metadata.{key}"] = value
            
            # Try vector search first if available
            if self.use_vector_search and self.embedding_model and self._is_atlas_connection():
@@ -332,6 +351,127 @@ class MongoDBKnowledgeAdapter:
            logger.error(f"Failed to delete document {document_id}: {e}")
            return False
    
+    def get(self, item_id: str, **kwargs):
+        """Get a specific document by ID from the knowledge store.
+
+        Returns a ``SearchResultItem`` (or ``None``) so the result shape matches
+        the shared KnowledgeStoreProtocol and the other adapters — a caller that
+        reads ``.text``/``.metadata`` or calls ``.to_dict()`` must not break when
+        it switches to the MongoDB backend (issue #5319).
+        """
+        from ..models import SearchResultItem
+        try:
+            from bson import ObjectId
+            doc = self.collection.find_one({"_id": ObjectId(item_id)})
+            if not doc:
+                return None
+            return SearchResultItem(
+                id=str(doc["_id"]),
+                text=doc.get("content", ""),
+                metadata=doc.get("metadata", {}) or {},
+                score=1.0,
+            )
+        except Exception as e:
+            logger.error(f"Failed to get document {item_id}: {e}")
+            return None
+
+    def get_all(
+        self,
+        *,
+        user_id: Optional[str] = None,
+        agent_id: Optional[str] = None,
+        run_id: Optional[str] = None,
+        limit: int = 100,
+        **kwargs
+    ):
+        """Get all documents matching the given scope.
+
+        Returns a ``SearchResult`` to match the shared KnowledgeStoreProtocol
+        and the other adapters.
+
+        Raises:
+            ScopeRequiredError: If no scope identifier is provided
+        """
+        from ..models import SearchResult, SearchResultItem
+        from ..protocols import require_scope
+        require_scope(user_id, agent_id, run_id, "get_all", backend="mongodb")
+        try:
+            scope_filter = {
+                k: v for k, v in {
+                    "user_id": user_id,
+                    "agent_id": agent_id,
+                    "run_id": run_id,
+                }.items() if v is not None
+            }
+            items = []
+            for doc in self.collection.find(scope_filter).limit(limit):
+                items.append(SearchResultItem(
+                    id=str(doc["_id"]),
+                    text=doc.get("content", ""),
+                    metadata=doc.get("metadata", {}) or {},
+                    score=1.0,
+                ))
+            return SearchResult(results=items)
+        except Exception as e:
+            logger.error(f"Failed to get_all from MongoDB: {e}")
+            return SearchResult(results=[])
+
+    def update(self, item_id: str, content: Any, **kwargs):
+        """Update an existing document's content/metadata by ID.
+
+        Returns an ``AddResult`` to match the shared KnowledgeStoreProtocol and
+        the other adapters (issue #5319).
+        """
+        from ..models import AddResult
+        try:
+            from bson import ObjectId
+            update_fields: Dict[str, Any] = {"content": str(content)}
+            if "metadata" in kwargs and kwargs["metadata"] is not None:
+                update_fields["metadata"] = kwargs["metadata"]
+            result = self.collection.update_one(
+                {"_id": ObjectId(item_id)},
+                {"$set": update_fields}
+            )
+            if result.matched_count > 0:
+                return AddResult(success=True, id=item_id)
+            return AddResult(success=False, message=f"No document matched id {item_id}")
+        except Exception as e:
+            logger.error(f"Failed to update document {item_id}: {e}")
+            return AddResult(success=False, message=str(e))
+
+    def delete_all(
+        self,
+        *,
+        user_id: Optional[str] = None,
+        agent_id: Optional[str] = None,
+        run_id: Optional[str] = None,
+        **kwargs
+    ) -> bool:
+        """Delete all documents matching the given scope.
+
+        Enforces the same tenant-scope contract as the mem0 backend: an
+        unscoped call raises ``ScopeRequiredError`` instead of silently wiping
+        every tenant's documents from the (typically shared) collection.
+
+        Raises:
+            ScopeRequiredError: If no scope identifier is provided
+        """
+        from ..protocols import require_scope
+        require_scope(user_id, agent_id, run_id, "delete_all", backend="mongodb")
+        try:
+            scope_filter = {
+                k: v for k, v in {
+                    "user_id": user_id,
+                    "agent_id": agent_id,
+                    "run_id": run_id,
+                }.items() if v is not None
+            }
+            self.collection.delete_many(scope_filter)
+            return True
+        except Exception as e:
+            logger.error(f"Failed to delete_all from MongoDB: {e}")
+            return False
+
    def close(self):
        """Close MongoDB connection."""
        if hasattr(self, 'client'):
@@ -1,6 +1,7 @@
 from typing import List, Union, Optional, Dict, Any
 from functools import cached_property
 import importlib
+import importlib.util

 #: Names accepted for a chunking strategy that are not chunker ids themselves.
 #: ``ChunkingStrategy.FIXED`` and ``.PARAGRAPH`` had no chunker at all, so they
@@ -125,15 +126,32 @@ class Chunking:
        self._chunker = None
        self._embeddings = None
        
+    @staticmethod
+    def _import_auto_embeddings():
+        """Import chonkie's AutoEmbeddings with an actionable error message.
+
+        Only a genuinely missing ``chonkie`` package is translated into the
+        install hint. If ``chonkie`` is installed but its import fails for a
+        different reason (e.g. a missing transitive dependency), the original
+        error is re-raised unchanged so the real cause is not masked.
+        """
+        try:
+            from chonkie.embeddings import AutoEmbeddings
+        except ImportError as e:
+            if importlib.util.find_spec("chonkie") is None:
+                raise ImportError(
+                    "chonkie package not found. Please install it using: pip install 'praisonaiagents[knowledge]'"
+                ) from e
+            raise
+        return AutoEmbeddings
+
    @cached_property
    def embedding_model(self):
        """Lazy load the embedding model."""
        if self._embedding_model is None and self.chunker_type in ['semantic', 'sdpm', 'late']:
-            from chonkie.embeddings import AutoEmbeddings
-            return AutoEmbeddings.get_embeddings("all-MiniLM-L6-v2")
+            return self._import_auto_embeddings().get_embeddings("all-MiniLM-L6-v2")
        elif isinstance(self._embedding_model, str):
-            from chonkie.embeddings import AutoEmbeddings
-            return AutoEmbeddings.get_embeddings(self._embedding_model)
+            return self._import_auto_embeddings().get_embeddings(self._embedding_model)
        return self._embedding_model

    def _get_chunker_params(self) -> Dict[str, Any]:
@@ -1034,7 +1034,12 @@ class Knowledge:

        # Store corpus stats for later retrieval
        self._corpus_stats = result.corpus_stats
-        
+
+        # success defaults to True on the dataclass and was never assigned, so
+        # index() reported success even when every file failed to embed and
+        # result.errors was full. Derive it from the errors already collected.
+        result.success = not result.errors
+
        return result
    
    def get_corpus_stats(self):
@@ -7,6 +7,7 @@ NO heavy imports - only stdlib and typing.
 Implementations are provided by praisonai-tools or wrapper layer.
 """

+import threading
 from dataclasses import dataclass
 from typing import Any, Callable, Dict, List, Optional, Protocol, runtime_checkable
 from praisonaiagents._logging import get_logger
@@ -81,11 +82,16 @@ class RerankerRegistry:
    """
    
    _instance: Optional["RerankerRegistry"] = None
+    _lock = threading.Lock()
    
    def __new__(cls) -> "RerankerRegistry":
        if cls._instance is None:
-            cls._instance = super().__new__(cls)
-            cls._instance._rerankers: Dict[str, Callable[..., RerankerProtocol]] = {}
+            with cls._lock:
+                # Double-checked locking for multi-agent safety
+                if cls._instance is None:
+                    instance = super().__new__(cls)
+                    instance._rerankers: Dict[str, Callable[..., RerankerProtocol]] = {}
+                    cls._instance = instance
        return cls._instance
    
    def register(
@@ -93,30 +99,35 @@ class RerankerRegistry:
        name: str, 
        factory: Callable[..., RerankerProtocol]
    ) -> None:
-        """Register a reranker factory."""
-        self._rerankers[name] = factory
+        """Register a reranker factory. Thread-safe."""
+        with self._lock:
+            self._rerankers[name] = factory
    
    def get(
        self, 
        name: str, 
        **kwargs
    ) -> Optional[RerankerProtocol]:
-        """Get a reranker by name."""
-        if name in self._rerankers:
-            try:
-                return self._rerankers[name](**kwargs)
-            except Exception as e:
-                logger.warning(f"Failed to initialize reranker '{name}': {e}")
-                return None
-        return None
+        """Get a reranker by name. Thread-safe."""
+        with self._lock:
+            factory = self._rerankers.get(name)
+        if factory is None:
+            return None
+        try:
+            return factory(**kwargs)
+        except Exception as e:
+            logger.warning(f"Failed to initialize reranker '{name}': {e}")
+            return None
    
    def list_rerankers(self) -> List[str]:
-        """List all registered reranker names."""
-        return list(self._rerankers.keys())
+        """List all registered reranker names. Thread-safe."""
+        with self._lock:
+            return list(self._rerankers.keys())
    
    def clear(self) -> None:
-        """Clear all registered rerankers."""
-        self._rerankers.clear()
+        """Clear all registered rerankers. Thread-safe."""
+        with self._lock:
+            self._rerankers.clear()

 def get_reranker_registry() -> RerankerRegistry:
    """Get the global reranker registry instance."""
@@ -7,6 +7,7 @@ NO heavy imports - only stdlib and typing.
 Implementations are provided by the wrapper layer.
 """

+import threading
 from dataclasses import dataclass, field
 from typing import Any, Callable, Dict, List, Optional, Protocol, runtime_checkable
 from enum import Enum
@@ -99,11 +100,16 @@ class RetrieverRegistry:
    """
    
    _instance: Optional["RetrieverRegistry"] = None
+    _lock = threading.Lock()
    
    def __new__(cls) -> "RetrieverRegistry":
        if cls._instance is None:
-            cls._instance = super().__new__(cls)
-            cls._instance._retrievers: Dict[str, Callable[..., RetrieverProtocol]] = {}
+            with cls._lock:
+                # Double-checked locking for multi-agent safety
+                if cls._instance is None:
+                    instance = super().__new__(cls)
+                    instance._retrievers: Dict[str, Callable[..., RetrieverProtocol]] = {}
+                    cls._instance = instance
        return cls._instance
    
    def register(
@@ -111,30 +117,35 @@ class RetrieverRegistry:
        name: str, 
        factory: Callable[..., RetrieverProtocol]
    ) -> None:
-        """Register a retriever factory."""
-        self._retrievers[name] = factory
+        """Register a retriever factory. Thread-safe."""
+        with self._lock:
+            self._retrievers[name] = factory
    
    def get(
        self, 
        name: str, 
        **kwargs
    ) -> Optional[RetrieverProtocol]:
-        """Get a retriever by name."""
-        if name in self._retrievers:
-            try:
-                return self._retrievers[name](**kwargs)
-            except Exception as e:
-                logger.warning(f"Failed to initialize retriever '{name}': {e}")
-                return None
-        return None
+        """Get a retriever by name. Thread-safe."""
+        with self._lock:
+            factory = self._retrievers.get(name)
+        if factory is None:
+            return None
+        try:
+            return factory(**kwargs)
+        except Exception as e:
+            logger.warning(f"Failed to initialize retriever '{name}': {e}")
+            return None
    
    def list_retrievers(self) -> List[str]:
-        """List all registered retriever names."""
-        return list(self._retrievers.keys())
+        """List all registered retriever names. Thread-safe."""
+        with self._lock:
+            return list(self._retrievers.keys())
    
    def clear(self) -> None:
-        """Clear all registered retrievers."""
-        self._retrievers.clear()
+        """Clear all registered retrievers. Thread-safe."""
+        with self._lock:
+            self._retrievers.clear()

 def get_retriever_registry() -> RetrieverRegistry:
    """Get the global retriever registry instance."""
@@ -87,6 +87,10 @@ def __getattr__(name):
            from .model_capabilities import supports_streaming_with_tools
            _lazy_cache[name] = supports_streaming_with_tools
            return supports_streaming_with_tools
+        elif name == "max_output_tokens":
+            from .model_capabilities import max_output_tokens
+            _lazy_cache[name] = max_output_tokens
+            return max_output_tokens
        elif name == "ModelRouter":
            from .model_router import ModelRouter
            _lazy_cache[name] = ModelRouter
@@ -243,6 +247,7 @@ __all__ = [
    "process_stream_chunks",
    "supports_structured_outputs",
    "supports_streaming_with_tools",
+    "max_output_tokens",
    "ModelRouter",
    "ModelProfile",
    "TaskComplexity",
@@ -10,8 +10,70 @@ and integrates with Gap 2 (parallel tool execution).

 from ..protocols import LLMProviderAdapterProtocol
 import json
+import logging
+import re
 from typing import Dict, Any, List, Optional

+_logger = logging.getLogger(__name__)
+
+
+def resolve_ollama_chained_arguments(
+    arguments: Dict[str, Any],
+    tool_result_mapping: Dict[str, Any],
+    *,
+    correlation_id: Optional[str] = None,
+) -> Dict[str, Any]:
+    """Resolve exact Ollama function-name references without coercing values.
+
+    Weak local models sometimes emit the name of an earlier function as a
+    later argument.  Keep the complete result (including mappings, lists,
+    negative numbers, and decimals) so the destination tool receives exactly
+    what the first tool returned.
+    """
+    if not tool_result_mapping:
+        return arguments
+    resolved = dict(arguments)
+    for arg_name, arg_value in resolved.items():
+        if isinstance(arg_value, str) and arg_value in tool_result_mapping:
+            replacement = tool_result_mapping[arg_value]
+            resolved[arg_name] = replacement
+            extra = {"correlation_id": correlation_id} if correlation_id else {}
+            _logger.debug(
+                "[OLLAMA_FIX] Replaced %s with a prior result in %s arguments",
+                arg_value,
+                arg_name,
+                extra=extra,
+            )
+    return resolved
+
+
+def record_ollama_tool_result(
+    tool_result_mapping: Dict[str, Any],
+    function_name: str,
+    tool_result: Any,
+) -> None:
+    """Record an Ollama tool result exactly as returned by the callback."""
+    if not function_name:
+        return
+    # Some execution paths return a ToolResult-like object rather than the
+    # callback's raw payload.  Never expose its failure diagnostic as a
+    # successful same-turn value.
+    if getattr(tool_result, "error", None) is not None:
+        return
+    # The agent executor represents failures as an ``error`` mapping (or a
+    # one-item list containing one). Never feed that diagnostic back as a
+    # successful value to a dependent call in the same turn.
+    if isinstance(tool_result, dict) and tool_result.get("error"):
+        return
+    if (
+        isinstance(tool_result, list)
+        and tool_result
+        and isinstance(tool_result[0], dict)
+        and tool_result[0].get("error")
+    ):
+        return
+    tool_result_mapping[function_name] = tool_result
+

 def collapse_union_param_types(tools):
    """Rewrite tool schemas a local server cannot parse.
@@ -55,42 +117,147 @@ def collapse_union_param_types(tools):
    return [fix(tool) for tool in tools]


+def _advertised_tool_names(tools: List[Dict[str, Any]]) -> set:
+    """The set of tool names the model was actually offered.
+
+    Salvage must never invent a tool: a text-format call is only promoted when
+    its name matches something advertised, so a model hallucinating a plausible
+    name in prose cannot trigger an unexpected execution.
+    """
+    names = set()
+    for tool in tools or []:
+        if isinstance(tool, dict):
+            name = tool.get("function", {}).get("name") or tool.get("name")
+            if name:
+                names.add(name)
+    return names
+
+
+_FENCE_RE = re.compile(
+    r"(?P<fence>`{3,}|~{3,})"   # opening fence: >=3 backticks or tildes
+    r"[\s\S]*?"                  # body (lazy)
+    r"(?:(?P=fence)|\Z)",       # matching closing fence OR end-of-string
+    re.MULTILINE,
+)
+
+
+def _mask_protected_ranges(text: str) -> str:
+    """Blank out fenced code blocks so a call *shown* as an example is not run.
+
+    Only the span content is replaced (with spaces of equal length) -- offsets
+    are preserved, so nothing outside a fence shifts. Legitimate prose that
+    merely contains angle brackets is untouched because it is not a tool-call
+    dialect; the fence guard specifically protects documentation of a call.
+
+    Both Markdown fence styles are recognised (``` and ~~~), and an *unclosed*
+    fence is treated as running to end-of-text: a truncated example that opens
+    a fence and names a real tool must not slip through as an executable call.
+    """
+    def blank(match):
+        return " " * len(match.group(0))
+    return _FENCE_RE.sub(blank, text)
+
+
+def _make_tool_call(name: str, arguments: Any, seed: str, idx: int) -> Dict[str, Any]:
+    return {
+        "id": f"call_{name}_{idx}_{hash(seed) % 10000}",
+        "type": "function",
+        "function": {
+            "name": name,
+            "arguments": json.dumps(arguments) if isinstance(arguments, (dict, list)) else str(arguments),
+        },
+    }
+
+
+# <tool_call>{"name": ..., "arguments": ...}</tool_call>
+_TOOL_CALL_TAG_RE = re.compile(r"<tool_call>\s*(\{[\s\S]*?\})\s*</tool_call>", re.IGNORECASE)
+# <function=name>{...}</function>
+_FUNCTION_TAG_RE = re.compile(r"<function=([a-zA-Z0-9_\-]+)>\s*(\{[\s\S]*?\})\s*</function>", re.IGNORECASE)
+
+
 def _recover_json_tool_calls(response_text: str, tools: List[Dict[str, Any]]) -> Optional[List[Dict[str, Any]]]:
-    """Recover a tool call a local model emitted as JSON text.
+    """Salvage a tool call a local model emitted as text instead of natively.

    Shared by every locally-served engine: small models routinely answer with
    the call as content instead of using the tool_calls field. Deliberately NOT
    on DefaultAdapter -- a hosted model returning JSON prose must never have it
    parsed as a tool call.
+
+    Three dialects are recovered, in-process, with no extra LLM round trip:
+      1. the whole content is a JSON object/array of ``{"name", "arguments"}``;
+      2. one or more ``<tool_call>{...}</tool_call>`` blocks embedded in prose;
+      3. one or more ``<function=name>{...}</function>`` blocks embedded in prose.
+
+    Every recovered name is checked against the advertised tool allowlist so a
+    call is never invented, and fenced code blocks are masked first so a call
+    shown as a documentation example is never executed.
    """
    if not response_text or not tools:
        return None
-    
+
+    allowed = _advertised_tool_names(tools)
+    if not allowed:
+        return None
+
+    # Fast path: a clean full-content JSON payload (the common case, unchanged).
    try:
-        import json
        response_json = json.loads(response_text.strip())
-        
-        # Normalize to list so both single and multi-tool payloads are supported
        if isinstance(response_json, dict):
            response_json = [response_json]
-
        if isinstance(response_json, list):
            tool_calls: List[Dict[str, Any]] = []
            for idx, tool_json in enumerate(response_json):
-                if isinstance(tool_json, dict) and "name" in tool_json:
-                    tool_calls.append({
-                        "id": f"call_{tool_json['name']}_{idx}_{hash(response_text) % 10000}",
-                        "type": "function",
-                        "function": {
-                            "name": tool_json["name"],
-                            "arguments": json.dumps(tool_json.get("arguments", {}))
-                        }
-                    })
-            return tool_calls if tool_calls else None
+                name = tool_json.get("name") if isinstance(tool_json, dict) else None
+                if name:
+                    tool_calls.append(_make_tool_call(
+                        name, tool_json.get("arguments", {}), response_text, idx))
+            if tool_calls:
+                return tool_calls
    except (json.JSONDecodeError, TypeError, KeyError):
        pass
-    
-    return None
+
+    # Embedded dialects: only scan when a marker is present, and only outside
+    # fenced code so an example call in documentation is never run. The marker
+    # probe is case-insensitive to match the tag regexes -- a model that emits
+    # <TOOL_CALL> or <FUNCTION=...> must recover just like the lowercase form.
+    lowered = response_text.lower()
+    if not any(marker in lowered for marker in ("<tool_call>", "<function=")):
+        return None
+
+    scan_text = _mask_protected_ranges(response_text)
+
+    # Scan both dialects in a single pass ordered by position in the text, so a
+    # model that interleaves <tool_call> and <function=> blocks yields calls in
+    # the order it actually wrote them -- a dependent or side-effecting call is
+    # never reordered ahead of one it relies on.
+    matches = []
+    for match in _TOOL_CALL_TAG_RE.finditer(scan_text):
+        try:
+            data = json.loads(match.group(1))
+        except (json.JSONDecodeError, TypeError):
+            continue
+        if isinstance(data, dict):
+            name = data.get("name") or data.get("tool")
+            if name in allowed:
+                args = data.get("arguments", data.get("parameters", {}))
+                matches.append((match.start(), name, args))
+
+    for match in _FUNCTION_TAG_RE.finditer(scan_text):
+        name = match.group(1)
+        if name in allowed:
+            try:
+                args = json.loads(match.group(2))
+            except (json.JSONDecodeError, TypeError):
+                continue
+            matches.append((match.start(), name, args))
+
+    matches.sort(key=lambda m: m[0])
+    salvaged = [
+        _make_tool_call(name, args, response_text, idx)
+        for idx, (_, name, args) in enumerate(matches)
+    ]
+
+    return salvaged if salvaged else None


 class DefaultAdapter:
@@ -100,6 +267,25 @@ class DefaultAdapter:
        """Hosted providers get the schema untouched -- they support all of it."""
        return tools

+    def resolve_chained_arguments(
+        self,
+        arguments: Dict[str, Any],
+        tool_result_mapping: Dict[str, Any],
+        *,
+        correlation_id: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        """Provider hook for same-turn tool-result references."""
+        return arguments
+
+    def record_tool_result(
+        self,
+        tool_result_mapping: Dict[str, Any],
+        function_name: str,
+        tool_result: Any,
+    ) -> None:
+        """Provider hook for recording results used by later tool calls."""
+        return None
+
    def supports_prompt_caching(self) -> bool:
        return False
    
@@ -178,6 +364,25 @@ class OllamaAdapter(DefaultAdapter):
    def supports_streaming_with_tools(self) -> bool:
        # Ollama doesn't reliably support streaming with tools
        return False
+
+    def resolve_chained_arguments(
+        self,
+        arguments: Dict[str, Any],
+        tool_result_mapping: Dict[str, Any],
+        *,
+        correlation_id: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        return resolve_ollama_chained_arguments(
+            arguments, tool_result_mapping, correlation_id=correlation_id
+        )
+
+    def record_tool_result(
+        self,
+        tool_result_mapping: Dict[str, Any],
+        function_name: str,
+        tool_result: Any,
+    ) -> None:
+        record_ollama_tool_result(tool_result_mapping, function_name, tool_result)
    
    
    
@@ -234,9 +439,19 @@ Now provide your final answer using this result. Summarize the information natur
        # Ollama 400s the whole request on a union-typed parameter.
        return collapse_union_param_types(tools)

-    # A JSON grammar and tools cannot coexist here: the grammar makes the
-    # tool-call tag unemittable and the model fabricates instead.
-    format_and_tools_conflict = True
+    # Read from the quirk catalogue rather than restating it. Hard-coding this
+    # created a second source of truth: correcting the quirk note would have
+    # left the behaviour unchanged, which is the drift the catalogue exists to
+    # prevent.
+    @property
+    def format_and_tools_conflict(self) -> bool:
+        try:
+            from ...local.quirktable import Quirk, quirks_for
+            from ...local.capabilities import ApiStyle, LocalEngine
+            return Quirk.FORMAT_AND_TOOLS_MUTUALLY_DESTRUCTIVE in quirks_for(
+                LocalEngine.OLLAMA, ApiStyle.OPENAI_CHAT)
+        except Exception:  # noqa: BLE001 -- the catalogue must never break a request
+            return True
    
    
    def get_default_settings(self) -> Dict[str, Any]:
@@ -588,12 +588,48 @@ def get_retry_delay(category: ErrorCategory, attempt: int = 1, base_delay: float
    return 0


-def extract_retry_after(error: Exception) -> Optional[float]:
+def _retry_after_http_date_delay(value: str) -> Optional[float]:
+    """Return seconds until an HTTP-date ``Retry-After`` value, else ``None``.
+
+    RFC 7231 allows ``Retry-After`` to be either delta-seconds or an absolute
+    HTTP-date. Delta-seconds is handled by the caller; this parses the date
+    form using the stdlib (no new dependency) and returns the non-negative
+    delay from now. A value in the past or unparseable yields ``None`` so the
+    caller falls through to its other signals.
+    """
+    try:
+        from email.utils import parsedate_to_datetime
+    except Exception:  # pragma: no cover - stdlib always present
+        return None
+    try:
+        when = parsedate_to_datetime(value.strip())
+    except (TypeError, ValueError):
+        return None
+    if when is None:
+        return None
+    import time as _time
+    from datetime import timezone
+
+    # An HTTP-date is UTC; a naive parse result is treated as UTC per the RFC.
+    if when.tzinfo is None:
+        when = when.replace(tzinfo=timezone.utc)
+    delay = when.timestamp() - _time.time()
+    return delay if delay > 0 else None
+
+
+def extract_retry_after(
+    error: Exception, cap_seconds: float = 300.0,
+) -> Optional[float]:
    """Extract Retry-After header value from rate limit errors.
-    
+
    Args:
        error: Exception potentially containing Retry-After info
-        
+        cap_seconds: Upper bound applied to the parsed value. Defaults to 300s
+            (5 minutes) for in-tick backoff callers. A scheduler *hold* passes a
+            larger cap so a job can be parked past a long provider reset window
+            (e.g. an hourly quota) instead of re-firing every tick — capping at
+            5 minutes there would defeat the park.
+
    Returns:
        Delay in seconds if found, None otherwise
    """
@@ -608,14 +644,21 @@ def extract_retry_after(error: Exception) -> Optional[float]:
            retry_after = None
        if retry_after is not None:
            try:
-                return min(float(retry_after), 300.0)  # Cap at 5 minutes
+                return min(float(retry_after), cap_seconds)
            except (ValueError, TypeError):
-                pass  # Not a plain number (could be an HTTP-date); fall through
+                # Not a plain delta-seconds value — RFC 7231 also permits an
+                # HTTP-date. Parse it so a provider that sends an absolute reset
+                # instant still yields a usable delay (else a long quota window
+                # would leave the job un-parked). A past/unparseable date yields
+                # no delay and falls through to the other signals.
+                delay = _retry_after_http_date_delay(str(retry_after))
+                if delay is not None:
+                    return min(delay, cap_seconds)

    # 2. Some SDKs expose a numeric ``retry_after`` attribute directly.
    retry_after_attr = getattr(error, "retry_after", None)
    if isinstance(retry_after_attr, (int, float)):
-        return min(float(retry_after_attr), 300.0)
+        return min(float(retry_after_attr), cap_seconds)

    error_str = str(error)
    
@@ -632,7 +675,7 @@ def extract_retry_after(error: Exception) -> Optional[float]:
        if match:
            try:
                delay = float(match.group(1))
-                return min(delay, 300.0)  # Cap at 5 minutes
+                return min(delay, cap_seconds)
            except (ValueError, IndexError):
                continue
    
@@ -359,10 +359,17 @@ class FailoverManager:
                if profile.cooldown_until and current_time >= profile.cooldown_until:
                    profile.reset()
            
-            # Find first available profile
-            for profile in self._profiles:
-                if profile.is_available:
-                    return profile
+            # Find available profiles (already priority-sorted by add_profile).
+            available = [p for p in self._profiles if p.is_available]
+            if available:
+                if self.config.rotate_on_success:
+                    # Round-robin among the current highest-priority tier only,
+                    # so equal-priority credentials configured to spread load
+                    # actually share traffic instead of one absorbing all of it.
+                    top_priority = available[0].priority
+                    tier = [p for p in available if p.priority == top_priority]
+                    return tier[self._current_index % len(tier)]
+                return available[0]
            
            # If none available, return the one with shortest remaining cooldown
            available_profiles = [p for p in self._profiles if p.status != ProviderStatus.DISABLED]
@@ -443,6 +450,10 @@ class FailoverManager:
            profile: The profile that succeeded
        """
        with self._lock:
+            # Advance the round-robin cursor so the next get_next_profile()
+            # hands work to a different equal-priority credential.
+            if self.config.rotate_on_success:
+                self._current_index += 1
            if profile.status != ProviderStatus.AVAILABLE:
                # Snapshot the cooldown this success is recovering from BEFORE
                # resetting, so we can avoid clobbering a newer fleet-wide bench.
@@ -38,6 +38,27 @@ def _base_model_name(model_name: str) -> str:
    return name


+def _coerce_output_token_limit(value) -> int | None:
+    """Return a positive integer output ceiling, or ``None`` when invalid."""
+    if isinstance(value, bool) or value is None:
+        return None
+    try:
+        coerced = int(value)
+    except (TypeError, ValueError, OverflowError):
+        return None
+    # ``int`` truncates fractional metadata (for example, ``8192.5``), which
+    # would turn an invalid ceiling into an apparently authoritative limit.
+    # Keep numeric strings such as ``"8192"`` compatible with LiteLLM's
+    # metadata while rejecting non-integral numeric values before conversion.
+    if not isinstance(value, str):
+        try:
+            if value != coerced:
+                return None
+        except Exception:
+            return None
+    return coerced if coerced > 0 else None
+
+
 def _fallback_supports_structured_outputs(model_name: str) -> bool:
    """Static heuristic used only when litellm is unavailable.

@@ -577,8 +598,8 @@ def max_output_tokens(model_name: str):
        try:
            info = litellm.get_model_info(model=model_name)
            if info:
-                out = info.get("max_output_tokens")
-                if out:
+                out = _coerce_output_token_limit(info.get("max_output_tokens"))
+                if out is not None:
                    return out
        except Exception:
            pass
@@ -594,7 +615,7 @@ def max_output_tokens(model_name: str):
            if info is None and "/" in model_name:
                info = model_cost.get(model_name.split("/")[-1].lower())
            if info:
-                return info.get("max_output_tokens")
+                return _coerce_output_token_limit(info.get("max_output_tokens"))
    except Exception:
        pass

@@ -617,4 +638,4 @@ def is_gemini_internal_tool(tool) -> bool:
    if isinstance(tool, dict) and len(tool) == 1:
        tool_name = next(iter(tool.keys()))
        return tool_name in GEMINI_INTERNAL_TOOLS
-    return False
+    return False