Release v4.6.68

Removed three unreferenced modules that were never imported or used:
- praisonaiagents/agent/prepared_turn_mixin.py (418 lines)
- praisonaiagents/agent/turn_execution_mixin.py (355 lines)
- praisonaiagents/telemetry/token_telemetry.py (87 lines)

These modules defined classes that were never wired into Agent or imported anywhere. Removing ~860 lines of dead code for better maintainability.

Fixes #2162

Co-authored-by: praisonai-triage-agent[bot] <272766704+praisonai-triage-agent[bot]@users.noreply.github.com>
Co-authored-by: MervinPraison <[email protected]>

.github/workflows/claude.yml

@@ -151,24 +151,40 @@ jobs:
         id: check_fork
         uses: actions/github-script@v7
         with:
+          github-token: ${{ steps.app-token.outputs.token }}
           script: |
-            const isPR = context.eventName === 'pull_request' || (context.payload.issue && context.payload.issue.pull_request);
-            if (isPR) {
-              const prNumber = context.payload.pull_request ? context.payload.pull_request.number : context.payload.issue.number;
-              const pr = await github.rest.pulls.get({
-                owner: context.repo.owner,
-                repo: context.repo.repo,
-                pull_number: prNumber
-              });
-              
-              core.setOutput('pr_branch', pr.data.head.ref);
-              
-              if (pr.data.head.repo && pr.data.head.repo.full_name !== context.repo.full_name) {
-                core.setOutput('is_fork', 'true');
-                core.setOutput('clone_url', pr.data.head.repo.clone_url);
-                core.setOutput('branch', pr.data.head.ref);
-                return;
-              }
+            const isPR = context.eventName === 'pull_request'
+              || context.eventName === 'pull_request_review'
+              || context.eventName === 'pull_request_review_comment'
+              || (context.payload.issue && context.payload.issue.pull_request);
+
+            if (!isPR) {
+              core.setOutput('is_pr', 'false');
+              core.setOutput('is_fork', 'false');
+              return;
+            }
+
+            const prNumber = context.payload.pull_request
+              ? context.payload.pull_request.number
+              : context.payload.issue.number;
+            const pr = await github.rest.pulls.get({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              pull_number: prNumber
+            });
+
+            core.setOutput('is_pr', 'true');
+            core.setOutput('pr_number', String(prNumber));
+            core.setOutput('pr_branch', pr.data.head.ref);
+            core.setOutput('head_repo', pr.data.head.repo.full_name);
+            core.setOutput('merge_state', (pr.data.mergeable_state || pr.data.merge_state_status || '').toUpperCase());
+
+            const baseRepo = `${context.repo.owner}/${context.repo.repo}`;
+            if (pr.data.head.repo && pr.data.head.repo.full_name !== baseRepo) {
+              core.setOutput('is_fork', 'true');
+              core.setOutput('clone_url', pr.data.head.repo.clone_url);
+              core.setOutput('branch', pr.data.head.ref);
+              return;
             }
             core.setOutput('is_fork', 'false');
 
@@ -180,17 +196,19 @@ jobs:
           token: ${{ steps.app-token.outputs.token }}
 
       - name: Fetch PR branch and Setup Remote
-        if: github.event.issue.pull_request
+        if: steps.check_fork.outputs.is_pr == 'true'
         env:
           PR_BRANCH: ${{ steps.check_fork.outputs.pr_branch }}
+          PR_NUMBER: ${{ steps.check_fork.outputs.pr_number }}
           IS_FORK: ${{ steps.check_fork.outputs.is_fork }}
         run: |
-          # Fetch PR head from base repo and put it into local branch
-          git fetch origin pull/${{ github.event.issue.number }}/head:"${PR_BRANCH}"
-          
+          git fetch origin "pull/${PR_NUMBER}/head:${PR_BRANCH}"
+          git checkout "${PR_BRANCH}"
+          git fetch origin main
+
           # If it's a fork, make origin point to local to trick claude-code-action's `git fetch origin <branch>`
           if [ "${IS_FORK}" = "true" ]; then
-            git remote set-url origin file://$(pwd)
+            git remote set-url origin "file://$(pwd)"
           fi
 
       - uses: anthropics/claude-code-action@beta
@@ -200,11 +218,13 @@ jobs:
           allowed_bots: 'praisonai-triage-agent[bot]'
           claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
           github_token: ${{ steps.app-token.outputs.token }}
+          model: opus
           trigger_phrase: "@claude"
           label_trigger: "claude"
 
           direct_prompt: |
-            ${{ steps.check_fork.outputs.is_fork == 'true' && 'CRITICAL: THIS IS A PULL REQUEST FROM A FORK. YOU DO NOT HAVE PUSH PERMISSIONS TO THIS REPOSITORY. ONLY READ FILES, VALIDATE CODE, AND PROVIDE YOUR COMMENTS/FEEDBACK DIRECTLY IN THIS PR VIA COMMENTS. DO NOT ATTEMPT TO PUSH OR CREATE PULL REQUESTS.\n\n' || 'NOTE: The branch is under MervinPraison/PraisonAI (not a fork). You are able to make modifications and push directly to this branch.\n\n' }}
+            ${{ steps.check_fork.outputs.is_fork == 'true' && 'CRITICAL: THIS IS A PULL REQUEST FROM A FORK. YOU DO NOT HAVE PUSH PERMISSIONS TO THIS REPOSITORY. ONLY READ FILES, VALIDATE CODE, AND PROVIDE YOUR COMMENTS/FEEDBACK DIRECTLY IN THIS PR VIA COMMENTS. DO NOT ATTEMPT TO PUSH OR CREATE PULL REQUESTS.\n\n' || (steps.check_fork.outputs.is_pr == 'true' && 'NOTE: The branch is under MervinPraison/PraisonAI (not a fork). You are able to make modifications and push directly to this branch.\n\n' || 'NOTE: This is an issue (not a PR). Create a new fix branch per STEP 3 below.\n\n') }}
+            ${{ steps.check_fork.outputs.is_pr == 'true' && format('CONTEXT: PR #{0} on branch {1}. head.repo={2}, is_fork={3}, merge_state={4}. NEVER treat cursor/*, claude/* branch names or bot authors as forks — only is_fork=true means fork.\n\n', steps.check_fork.outputs.pr_number, steps.check_fork.outputs.pr_branch, steps.check_fork.outputs.head_repo, steps.check_fork.outputs.is_fork, steps.check_fork.outputs.merge_state) || '' }}
             You are working on the PraisonAI SDK. Follow AGENTS.md strictly.
 
             STEP 0 — SETUP GIT IDENTITY & AUTH (GLOBAL — required for all repos):
@@ -233,15 +253,9 @@ jobs:
             - Is there a name collision with existing exports in __init__.py?
             If ANY of these conceptual checks fail (excluding routing), add a comment to the issue explaining why and close it. Do NOT create a PR.
 
-            STEP 3 — IMPLEMENT (for changes in THIS repo — PraisonAI):
-            - Create a fix branch: git checkout -b claude/issue-$ISSUE_NUMBER-$(date +%Y%m%d)
-            - Implement a minimal, focused fix
-            - Follow protocol-driven design: protocols in core SDK, heavy implementations in wrapper
-            - Keep changes small and backward-compatible
-            - Commit: git add -A && git commit -m "fix: <description> (fixes #$ISSUE_NUMBER)"
-            - Push: git push origin claude/issue-$ISSUE_NUMBER-$(date +%Y%m%d)
-            CRITICAL: After pushing, you MUST immediately run: gh pr create --title "fix: <title>" --body "Fixes #$ISSUE_NUMBER" --head claude/issue-$ISSUE_NUMBER-$(date +%Y%m%d)
-            Do NOT stop after pushing. Do NOT say "manually create a PR". Always create the PR automatically.
+            ${{ steps.check_fork.outputs.is_pr == 'true' && steps.check_fork.outputs.is_fork != 'true' && steps.check_fork.outputs.merge_state == 'DIRTY' && format('STEP 3 — RESOLVE MERGE CONFLICTS ON PR (branch: {0}, PR #{1}):\n- DO NOT create a new branch or PR\n- git fetch origin main && git rebase origin/main\n- Resolve conflicts: keep this PR intent, merge in main newer logic\n- git add -A && git rebase --continue (repeat until done)\n- git push --force-with-lease origin {0}\n- Run targeted pytest on changed paths\n- Comment files resolved and what you skipped\n', steps.check_fork.outputs.pr_branch, steps.check_fork.outputs.pr_number) || '' }}
+            ${{ steps.check_fork.outputs.is_pr == 'true' && steps.check_fork.outputs.is_fork != 'true' && steps.check_fork.outputs.merge_state != 'DIRTY' && format('STEP 3 — IMPLEMENT ON EXISTING PR (branch: {0}, PR #{1}):\n- DO NOT create a new branch or PR\n- git checkout {0}\n- Implement minimal, focused fixes from reviewer feedback\n- Follow protocol-driven design: protocols in core SDK, heavy implementations in wrapper\n- Keep changes small and backward-compatible\n- Commit: git add -A && git commit -m "fix: <description>"\n- Push: git push origin {0}\n- Comment a summary of exact files modified and what you skipped\n', steps.check_fork.outputs.pr_branch, steps.check_fork.outputs.pr_number) || '' }}
+            ${{ steps.check_fork.outputs.is_pr != 'true' && 'STEP 3 — IMPLEMENT (for changes in THIS repo — PraisonAI):\n- Create a fix branch: git checkout -b claude/issue-$ISSUE_NUMBER-$(date +%Y%m%d)\n- Implement a minimal, focused fix\n- Follow protocol-driven design: protocols in core SDK, heavy implementations in wrapper\n- Keep changes small and backward-compatible\n- Commit: git add -A && git commit -m "fix: <description> (fixes #$ISSUE_NUMBER)"\n- Push: git push origin claude/issue-$ISSUE_NUMBER-$(date +%Y%m%d)\nCRITICAL: After pushing, you MUST immediately run: gh pr create --title "fix: <title>" --body "Fixes #$ISSUE_NUMBER" --head claude/issue-$ISSUE_NUMBER-$(date +%Y%m%d)\nDo NOT stop after pushing. Do NOT say "manually create a PR". Always create the PR automatically.\n' || '' }}
 
             STEP 3-ALT — IMPLEMENT IN EXTERNAL REPO (for PraisonAI-Tools, PraisonAIDocs, PraisonAI-Plugins):
             When work must happen in a different repository, follow these steps EXACTLY. Do NOT attempt to use `cd`, as directory state is not preserved. Use `git -C` for all commands.
@@ -267,10 +281,7 @@ jobs:
             - For docs changes: verify files exist and are valid markdown/MDX
             - Ensure no regressions
 
-            STEP 5 — CREATE PR:
-            - Commit with descriptive message, push, and create PR using `gh pr create`
-            - For external repos, use `gh pr create -R MervinPraison/<REPO_NAME>`
-            CRITICAL: You MUST create the PR automatically using `gh pr create`. Do NOT just provide a link or say "manual push required".
+            ${{ steps.check_fork.outputs.is_pr == 'true' && 'STEP 5 — SKIP: Already on an open PR. Do NOT run gh pr create.\n' || 'STEP 5 — CREATE PR:\n- Commit with descriptive message, push, and create PR using `gh pr create`\n- For external repos, use `gh pr create -R MervinPraison/<REPO_NAME>`\nCRITICAL: You MUST create the PR automatically using `gh pr create`. Do NOT just provide a link or say "manual push required".\n' }}
             NOTE: If you worked on an external repo, the wrapper action will report "No commits" for the main repo at the end. This is expected, do not worry about it.
           allowed_tools: |
             Bash(git:*)

.github/workflows/docker-publish.yml

@@ -22,8 +22,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           persist-credentials: false
-        with:
-          persist-credentials: false
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3

.github/workflows/gemini-issue-automated-triage.yml

@@ -36,9 +36,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           persist-credentials: false
-        with:
-          persist-credentials: false
-        with:
           token: ${{ steps.generate_token.outputs.token }}
 
       - name: Run Gemini Issue Triage

.github/workflows/gemini-issue-review.yml

@@ -37,9 +37,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           persist-credentials: false
-        with:
-          persist-credentials: false
-        with:
           token: ${{ steps.generate_token.outputs.token }}
 
       - name: Set up Git

.github/workflows/gemini-pr-review.yml

@@ -48,9 +48,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           persist-credentials: false
-        with:
-          persist-credentials: false
-        with:
           token: ${{ steps.generate_token.outputs.token }}
           ref: ${{ steps.checkout_ref.outputs.ref }}
           fetch-depth: 0

.github/workflows/integration-cross-repo.yml

@@ -60,16 +60,16 @@ jobs:
     - name: Run PraisonAI integration tests
       run: |
         cd PraisonAI/src/praisonai
-        python -m pytest tests/integration/test_aiui_* -v --timeout=300
+        python -m pytest tests/integration/test_aiui_* -v
         
     - name: Run PraisonAIUI integration tests
       run: |
         cd PraisonAIUI
         if [ -f tests/integration/test_agentic_roundtrip.py ]; then
-          python -m pytest tests/integration/test_agentic_roundtrip.py -v --timeout=300
+          python -m pytest tests/integration/test_agentic_roundtrip.py -v
         fi
         if [ -f tests/test_feature_sdk_backends.py ]; then
-          python -m pytest tests/test_feature_sdk_backends.py -v --timeout=300
+          python -m pytest tests/test_feature_sdk_backends.py -v
         fi
         
     - name: Test Pattern C CLI
@@ -152,7 +152,7 @@ jobs:
       run: |
         cd PraisonAI/src/praisonai
         if [ -n "$OPENAI_API_KEY" ]; then
-          python -m pytest tests/integration/test_aiui_host_agentic.py -v --timeout=600
+          python -m pytest tests/integration/test_aiui_host_agentic.py -v
         else
           echo "OPENAI_API_KEY not available, skipping agentic tests"
         fi

.github/workflows/python-package.yml

@@ -15,8 +15,8 @@ jobs:
 
     steps:
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
     - name: Set up Python ${{ matrix.python-version }}
       uses: actions/setup-python@v5
       with:

.github/workflows/release.yml

@@ -26,11 +26,8 @@ jobs:
     # First checkout the upstream repo at the specified tag
     - name: Checkout upstream repo
       uses: actions/checkout@v4
-        with:
-          persist-credentials: false
-        with:
-          persist-credentials: false
       with:
+        persist-credentials: false
         repository: MervinPraison/PraisonAI
         ref: ${{ github.event.inputs.tag }}
         fetch-depth: 0
@@ -48,11 +45,8 @@ jobs:
     # Now checkout the fork at main branch
     - name: Checkout fork
       uses: actions/checkout@v4
-        with:
-          persist-credentials: false
-        with:
-          persist-credentials: false
       with:
+        persist-credentials: false
         repository: ${{ github.event.inputs.fork_owner }}/PraisonAI
         ref: main
         fetch-depth: 0
@@ -449,11 +443,8 @@ jobs:
     # First checkout the fork's main branch
     - name: Checkout fork repo
       uses: actions/checkout@v4
-        with:
-          persist-credentials: false
-        with:
-          persist-credentials: false
       with:
+        persist-credentials: false
         fetch-depth: 0
         token: ${{ secrets.PAT_TOKEN }}
         repository: ${{ github.event.inputs.fork_owner }}/PraisonAI

.github/workflows/sync-secrets-to-aiui.yml

@@ -0,0 +1,31 @@
+name: Sync Claude Secrets to PraisonAIUI
+
+on:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  sync:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Sync secrets to PraisonAIUI
+        env:
+          GH_TOKEN: ${{ secrets.SECRETS_ADMIN_PAT }}
+          TARGET: MervinPraison/PraisonAIUI
+          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 }}
+          GH_TOKEN_VALUE: ${{ secrets.GH_TOKEN }}
+        run: |
+          set -euo pipefail
+          if [ -z "${GH_TOKEN:-}" ]; then
+            echo "SECRETS_ADMIN_PAT is not configured in this repository"
+            exit 1
+          fi
+          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"
+          echo -n "$GH_TOKEN_VALUE" | gh secret set GH_TOKEN --repo "$TARGET"
+          echo "Synced CLAUDE_APP_ID, CLAUDE_APP_PRIVATE_KEY, CLAUDE_CODE_OAUTH_TOKEN, GH_TOKEN → $TARGET"

.github/workflows/test-comprehensive.yml

@@ -37,10 +37,8 @@ jobs:
     steps:
     - name: Checkout code
       uses: actions/checkout@v4
-        with:
-          persist-credentials: false
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
 
     - name: Set up Python ${{ matrix.python-version }}
       uses: actions/setup-python@v5

.github/workflows/test-core.yml

@@ -22,8 +22,8 @@ jobs:
 
     steps:
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
     - name: Set up Python 3.11
       uses: actions/setup-python@v5
       with:

.github/workflows/test-extended.yml

@@ -13,10 +13,8 @@ jobs:
     steps:
     - name: Checkout code
       uses: actions/checkout@v4
-        with:
-          persist-credentials: false
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
 
     - name: Set up Python
       uses: actions/setup-python@v5
@@ -81,10 +79,8 @@ jobs:
     steps:
     - name: Checkout code
       uses: actions/checkout@v4
-        with:
-          persist-credentials: false
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
 
     - name: Set up Python
       uses: actions/setup-python@v5

.github/workflows/test-frameworks.yml

@@ -33,10 +33,8 @@ jobs:
     steps:
     - name: Checkout code
       uses: actions/checkout@v4
-        with:
-          persist-credentials: false
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
 
     - name: Set up Python ${{ matrix.python-version }}
       uses: actions/setup-python@v5

.github/workflows/test-optimized.yml

@@ -45,8 +45,8 @@ jobs:
     timeout-minutes: 5
     steps:
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
     - name: Set up Python 3.11
       uses: actions/setup-python@v5
       with:
@@ -105,8 +105,8 @@ jobs:
     
     steps:
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
     - name: Set up Python ${{ matrix.python-version }}
       uses: actions/setup-python@v5
       with:
@@ -187,8 +187,8 @@ jobs:
         fi
     
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
       if: steps.check-key.outputs.has_key == 'true'
     
     - name: Set up Python 3.11
@@ -259,8 +259,8 @@ jobs:
         fi
     
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
       if: steps.check-key.outputs.has_key == 'true'
     
     - name: Set up Python 3.11
@@ -318,8 +318,8 @@ jobs:
         fi
     
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
       if: steps.check-key.outputs.has_key == 'true'
     
     - name: Set up Python 3.11
@@ -377,8 +377,8 @@ jobs:
         fi
     
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
       if: steps.check-key.outputs.has_key == 'true'
     
     - name: Set up Python 3.11
@@ -436,8 +436,8 @@ jobs:
         fi
     
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
       if: steps.check-key.outputs.has_key == 'true'
     
     - name: Set up Python 3.11
@@ -495,8 +495,8 @@ jobs:
         fi
     
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
       if: steps.check-key.outputs.has_key == 'true'
     
     - name: Set up Python 3.11

.github/workflows/test-real.yml

@@ -45,10 +45,8 @@ jobs:
         
     - name: Checkout code
       uses: actions/checkout@v4
-        with:
-          persist-credentials: false
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
 
     - name: Set up Python ${{ matrix.python-version }}
       uses: actions/setup-python@v5

.gitignore

@@ -88,6 +88,9 @@ issue_details.json
 .internal
 .praisonai
 
+# Backup files
+*.bak
+
 # Rust / Cargo
 target/
 Cargo.lock

AGENTS.md

@@ -0,0 +1,10 @@
+# Agent Instructions
+
+You are working on the PraisonAI project.
+
+## Project Guidelines
+- Follow the existing code style and conventions
+- Be concise and helpful in responses  
+- Test implementation thoroughly
+- Ensure backward compatibility with existing APIs
+- Follow protocol-driven design: core protocols in praisonaiagents/, heavy implementations in praisonai/

README.md

@@ -770,11 +770,11 @@ agent.start('Write a movie script about a robot in Mars');
 
 ## ⚡ Performance
 
-PraisonAI is built for speed, with agent instantiation in under 4μs. This reduces overhead, improves responsiveness, and helps multi-agent systems scale efficiently in real-world production workloads.
+PraisonAI is built for speed, with agent instantiation in around 14μs. This reduces overhead, improves responsiveness, and helps multi-agent systems scale efficiently in real-world production workloads.
 
 | Performance Metric | PraisonAI |
 |--------------------|-----------|
-| Avg Instantiation Time | **3.77 μs** |
+| Avg Instantiation Time | **14 μs** |
 
 ---
 

api.md

@@ -59,6 +59,8 @@ Methods:
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">handoff_to</a>(target_agent: 'Agent', prompt: str, context: Optional[Dict[str, Any]] = None, config: Optional['HandoffConfig'] = None) -> 'HandoffResult'</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">handoff_to_async</a>(target_agent: 'Agent', prompt: str, context: Optional[Dict[str, Any]] = None, config: Optional['HandoffConfig'] = None) -> 'HandoffResult'</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">is_closed</a>() -> bool</code>
+* <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">llm_instance</a>()</code>
+* <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">llm_instance</a>(value)</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">llm_model</a>() -> Optional[str]</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">output_style</a>() -> Optional[str]</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">output_style</a>(value: Optional[str]) -> None</code>
@@ -240,6 +242,7 @@ from praisonaiagents import Memory, MemoryBackend, MemoryConfig
 
 Methods:
 
+* <code title="class Memory">Memory.<a href="./src/praisonai-agents/praisonaiagents/memory/memory.py">build_cache_optimized_context</a>(task_descr: str, user_id: Optional[str] = None, additional: str = '', max_items: int = 3, include_cache_boundary: bool = True, include_in_output: Optional[bool] = None) -> Dict[str, str]</code>
 * <code title="class Memory">Memory.<a href="./src/praisonai-agents/praisonaiagents/memory/memory.py">build_context_for_task</a>(task_descr: str, user_id: Optional[str] = None, additional: str = '', max_items: int = 3, include_in_output: Optional[bool] = None) -> str</code>
 * <code title="class Memory">Memory.<a href="./src/praisonai-agents/praisonaiagents/memory/memory.py">calculate_quality_metrics</a>(output: str, expected_output: str, llm: Optional[str] = None, custom_prompt: Optional[str] = None) -> Dict[str, float]</code>
 * <code title="class Memory">Memory.<a href="./src/praisonai-agents/praisonaiagents/memory/memory.py">close_connections</a>()</code>
@@ -515,10 +518,12 @@ Methods:
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">add_memory</a>(text: str, memory_type: str = 'long', **metadata) -> None</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">chat</a>(message: str, **kwargs) -> str</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">clear_memory</a>(memory_type: str = 'all') -> None</code>
+* <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">close</a>() -> None</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">create_agent</a>(*args, **kwargs) -> 'Agent'</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">get_context</a>(query: str, max_items: int = 3) -> str</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">get_state</a>(key: str, default: Any = None) -> Any</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">increment_state</a>(key: str, increment: int = 1, default: int = 0) -> None</code>
+* <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">is_expired</a>() -> bool</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">knowledge</a>() -> 'Knowledge'</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">memory</a>() -> 'Memory'</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">restore_state</a>() -> Dict[str, Any]</code>
@@ -527,6 +532,7 @@ Methods:
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">search_memory</a>(query: str, memory_type: str = 'long', limit: int = 5) -> List[Dict[str, Any]]</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">send_message</a>(message: str, **kwargs) -> str</code>
 * <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">set_state</a>(key: str, value: Any) -> None</code>
+* <code title="class Session">Session.<a href="./src/praisonai-agents/praisonaiagents/session/api.py">time_to_expiry</a>() -> Optional[float]</code>
 
 # MCP
 
@@ -609,7 +615,7 @@ Methods:
 * <code title="class ContextManager">ContextManager.<a href="./src/praisonai-agents/praisonaiagents/context/manager.py">register_snapshot_callback</a>(callback: Callable[[SnapshotHookData], None]) -> None</code>
 * <code title="class ContextManager">ContextManager.<a href="./src/praisonai-agents/praisonaiagents/context/manager.py">reset</a>() -> None</code>
 * <code title="class ContextManager">ContextManager.<a href="./src/praisonai-agents/praisonaiagents/context/manager.py">set_tool_budget</a>(tool_name: str, max_tokens: int, protected: bool = False) -> None</code>
-* <code title="class ContextManager">ContextManager.<a href="./src/praisonai-agents/praisonaiagents/context/manager.py">truncate_tool_output</a>(tool_name: str, output: str) -> str</code>
+* <code title="class ContextManager">ContextManager.<a href="./src/praisonai-agents/praisonaiagents/context/manager.py">truncate_tool_output</a>(tool_name: str, output: str, tool_call_id: str = None, run_id: str = None) -> str</code>
 
 # UI
 
@@ -698,7 +704,7 @@ Methods:
 
 Types:
 ```python
-from praisonaiagents import AgentAppConfig, AgentAppProtocol, AgentFlow, AgentManager, AgentOSConfig, AgentOSProtocol, AgentTeam, AutoApproveBackend, EmbeddingResult, aembedding, aembeddings, embedding, embeddings, get_dimensions
+from praisonaiagents import AgentAppConfig, AgentAppProtocol, AgentFlow, AgentManager, AgentOSConfig, AgentOSProtocol, AgentTeam, AutoApproveBackend, EmbeddingResult, __version__, aembedding, aembeddings, embedding, embeddings, get_dimensions
 ```
 
 Methods:
@@ -789,11 +795,17 @@ Methods:
 
 * <code title="cli">praisonai <a href="./src/praisonai/praisonai/cli/main.py">--help</a></code>
 * <code title="cli">praisonai acp acp-main <a href="./src/praisonai/praisonai/cli/commands/acp.py">--help</a></code>
+* <code title="cli">praisonai agent list <a href="./src/praisonai/praisonai/cli/commands/agent.py">--help</a></code>
+* <code title="cli">praisonai agent show <a href="./src/praisonai/praisonai/cli/commands/agent.py">--help</a></code>
 * <code title="cli">praisonai agents create <a href="./src/praisonai/praisonai/cli/commands/agents.py">--help</a></code>
 * <code title="cli">praisonai agents info <a href="./src/praisonai/praisonai/cli/commands/agents.py">--help</a></code>
 * <code title="cli">praisonai agents list <a href="./src/praisonai/praisonai/cli/commands/agents.py">--help</a></code>
 * <code title="cli">praisonai app app <a href="./src/praisonai/praisonai/cli/commands/app.py">--help</a></code>
 * <code title="cli">praisonai audit agent-centric <a href="./src/praisonai/praisonai/cli/commands/audit.py">--help</a></code>
+* <code title="cli">praisonai auth list <a href="./src/praisonai/praisonai/cli/commands/auth.py">--help</a></code>
+* <code title="cli">praisonai auth login <a href="./src/praisonai/praisonai/cli/commands/auth.py">--help</a></code>
+* <code title="cli">praisonai auth logout <a href="./src/praisonai/praisonai/cli/commands/auth.py">--help</a></code>
+* <code title="cli">praisonai auth status <a href="./src/praisonai/praisonai/cli/commands/auth.py">--help</a></code>
 * <code title="cli">praisonai batch batch-run <a href="./src/praisonai/praisonai/cli/commands/batch.py">--help</a></code>
 * <code title="cli">praisonai batch list <a href="./src/praisonai/praisonai/cli/commands/batch.py">--help</a></code>
 * <code title="cli">praisonai batch report <a href="./src/praisonai/praisonai/cli/commands/batch.py">--help</a></code>
@@ -834,6 +846,8 @@ Methods:
 * <code title="cli">praisonai claw claw <a href="./src/praisonai/praisonai/cli/commands/claw.py">--help</a></code>
 * <code title="cli">praisonai code <a href="./src/praisonai/praisonai/cli/main.py">--help</a></code>
 * <code title="cli">praisonai code code-main <a href="./src/praisonai/praisonai/cli/commands/code.py">--help</a></code>
+* <code title="cli">praisonai command list <a href="./src/praisonai/praisonai/cli/commands/command.py">--help</a></code>
+* <code title="cli">praisonai command show <a href="./src/praisonai/praisonai/cli/commands/command.py">--help</a></code>
 * <code title="cli">praisonai commit commit-main <a href="./src/praisonai/praisonai/cli/commands/commit.py">--help</a></code>
 * <code title="cli">praisonai completion bash <a href="./src/praisonai/praisonai/cli/commands/completion.py">--help</a></code>
 * <code title="cli">praisonai completion completion-callback <a href="./src/praisonai/praisonai/cli/commands/completion.py">--help</a></code>
@@ -846,6 +860,9 @@ Methods:
 * <code title="cli">praisonai config path <a href="./src/praisonai/praisonai/cli/commands/config.py">--help</a></code>
 * <code title="cli">praisonai config reset <a href="./src/praisonai/praisonai/cli/commands/config.py">--help</a></code>
 * <code title="cli">praisonai config set <a href="./src/praisonai/praisonai/cli/commands/config.py">--help</a></code>
+* <code title="cli">praisonai config show <a href="./src/praisonai/praisonai/cli/commands/config.py">--help</a></code>
+* <code title="cli">praisonai config sources <a href="./src/praisonai/praisonai/cli/commands/config.py">--help</a></code>
+* <code title="cli">praisonai config validate <a href="./src/praisonai/praisonai/cli/commands/config.py">--help</a></code>
 * <code title="cli">praisonai context add <a href="./src/praisonai/praisonai/cli/commands/context.py">--help</a></code>
 * <code title="cli">praisonai context clear <a href="./src/praisonai/praisonai/cli/commands/context.py">--help</a></code>
 * <code title="cli">praisonai context compact <a href="./src/praisonai/praisonai/cli/commands/context.py">--help</a></code>
@@ -880,10 +897,12 @@ Methods:
 * <code title="cli">praisonai doctor db <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor doctor-callback <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor env <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
+* <code title="cli">praisonai doctor fix <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor mcp <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor network <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor packaging <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor performance <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
+* <code title="cli">praisonai doctor runtime <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor selftest <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor skills <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor tools <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
@@ -914,6 +933,9 @@ Methods:
 * <code title="cli">praisonai gateway install <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway logs <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway mint-link <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
+* <code title="cli">praisonai gateway pause <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
+* <code title="cli">praisonai gateway reconnect <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
+* <code title="cli">praisonai gateway resume <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway send <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway start <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway status <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
@@ -1004,6 +1026,14 @@ Methods:
 * <code title="cli">praisonai pairing list <a href="./src/praisonai/praisonai/cli/commands/pairing.py">--help</a></code>
 * <code title="cli">praisonai pairing revoke <a href="./src/praisonai/praisonai/cli/commands/pairing.py">--help</a></code>
 * <code title="cli">praisonai paths show <a href="./src/praisonai/praisonai/cli/commands/paths.py">--help</a></code>
+* <code title="cli">praisonai permissions allow <a href="./src/praisonai/praisonai/cli/commands/permissions.py">--help</a></code>
+* <code title="cli">praisonai permissions ask <a href="./src/praisonai/praisonai/cli/commands/permissions.py">--help</a></code>
+* <code title="cli">praisonai permissions deny <a href="./src/praisonai/praisonai/cli/commands/permissions.py">--help</a></code>
+* <code title="cli">praisonai permissions export <a href="./src/praisonai/praisonai/cli/commands/permissions.py">--help</a></code>
+* <code title="cli">praisonai permissions import-rules <a href="./src/praisonai/praisonai/cli/commands/permissions.py">--help</a></code>
+* <code title="cli">praisonai permissions list <a href="./src/praisonai/praisonai/cli/commands/permissions.py">--help</a></code>
+* <code title="cli">praisonai permissions remove <a href="./src/praisonai/praisonai/cli/commands/permissions.py">--help</a></code>
+* <code title="cli">praisonai permissions reset <a href="./src/praisonai/praisonai/cli/commands/permissions.py">--help</a></code>
 * <code title="cli">praisonai plugins create <a href="./src/praisonai/praisonai/cli/commands/plugins.py">--help</a></code>
 * <code title="cli">praisonai plugins disable <a href="./src/praisonai/praisonai/cli/commands/plugins.py">--help</a></code>
 * <code title="cli">praisonai plugins discover <a href="./src/praisonai/praisonai/cli/commands/plugins.py">--help</a></code>
@@ -1302,7 +1332,7 @@ export { // Python parity additions
 export { // Subagent Tool (agent-as-tool pattern)
   SubagentTool, BaseTool, FunctionTool, ToolRegistry, ToolResult, ToolValidationError, createDelegator, createSubagentTool, createSubagentTools, createTool, getRegistry, getTool, registerTool, tool, validateTool } from "./tools";
 export { airweaveSearch, bedrockBrowserClick, bedrockBrowserFill, bedrockBrowserNavigate, bedrockCodeInterpreter, codeExecution, codeMode, createCustomTool, exaSearch, firecrawlCrawl, firecrawlScrape, parallelSearch, perplexitySearch, registerCustomTool, registerLocalTool, registerNpmTool, superagentGuard, superagentRedact, superagentVerify, tavilyCrawl, tavilyExtract, tavilySearch, valyuBioSearch, valyuCompanyResearch, valyuEconomicsSearch, valyuFinanceSearch, valyuPaperSearch, valyuPatentSearch, valyuSecSearch, valyuWebSearch } from "./tools/builtins";
-export { BudgetExceededError, MissingDependencyError, MissingEnvVarError, ToolsRegistry, composeMiddleware, createLoggingMiddleware, createRateLimitMiddleware, createRedactionMiddleware, createRetryMiddleware, createTimeoutMiddleware, createToolsRegistry, createTracingMiddleware, createValidationMiddleware, getRegistry, getTool, getToolsRegistry, get_registry, get_tool, registerTool, register_tool, resetToolsRegistry, validateTool, validate_tool } from "./tools/registry";
+export { BudgetExceededError, MissingDependencyError, MissingEnvVarError, ToolsRegistry, composeMiddleware, createLoggingMiddleware, createRateLimitMiddleware, createRedactionMiddleware, createRetryMiddleware, createTimeoutMiddleware, createToolsRegistry, createTracingMiddleware, createValidationMiddleware, getToolsRegistry, get_registry, get_tool, register_tool, resetToolsRegistry, validate_tool } from "./tools/registry";
 export type { InstallHints, PraisonTool, RedactionHooks, RegisteredTool, ToolCapabilities, ToolExecutionContext, ToolExecutionResult, ToolFactory, ToolHooks, ToolInstallStatus, ToolLimits, ToolLogger, ToolMetadata, ToolMiddleware, ToolParameterProperty, ToolParameterSchema } from "./tools/registry";
 export { registerBuiltinTools, tools } from "./tools/tools";
 export { // Classes

docker/Dockerfile.chat

@@ -16,7 +16,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=4.6.55" \
+    "praisonai>=4.6.68" \
     "praisonai[chat]" \
     "embedchain[github,youtube]"
 

docker/Dockerfile.dev

@@ -20,7 +20,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=4.6.55" \
+    "praisonai>=4.6.68" \
     "praisonai[ui]" \
     "praisonai[chat]" \
     "praisonai[realtime]" \

docker/Dockerfile.ui

@@ -16,7 +16,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=4.6.55" \
+    "praisonai>=4.6.68" \
     "praisonai[ui]" \
     "praisonai[crewai]"
 

example_llm_config_usage.py

@@ -0,0 +1,78 @@
+#!/usr/bin/env python3
+"""Example of using LLMConfig with Agent for model fallback configuration."""
+
+import sys
+sys.path.insert(0, 'src/praisonai-agents')
+
+from praisonaiagents import Agent
+from praisonaiagents.config import LLMConfig
+
+# Example 1: Pass LLMConfig via llm parameter
+def example_llm_param():
+    """Using LLMConfig via the llm parameter."""
+    config = LLMConfig(
+        model="gpt-4o",
+        fallback_models=["claude-3-5-sonnet", "gpt-4o-mini"],
+        base_url="https://api.example.com",
+        api_key="test-key"
+    )
+    
+    agent = Agent(
+        name="Assistant",
+        instructions="You are a helpful assistant",
+        llm=config  # Pass LLMConfig object directly
+    )
+    
+    print(f"Primary model: {agent.llm}")
+    print(f"Fallback models: {agent.fallback_models}")
+    print(f"Base URL: {agent.base_url}")
+    print(f"API Key: {agent.api_key}")
+
+# Example 2: Pass LLMConfig via model parameter
+def example_model_param():
+    """Using LLMConfig via the model parameter (preferred)."""
+    config = LLMConfig(
+        model="gpt-4o", 
+        fallback_models=["claude-3-5-sonnet", "gpt-4o-mini"],
+        base_url="https://api.example.com",
+        api_key="test-key"
+    )
+    
+    agent = Agent(
+        name="Assistant",
+        instructions="You are a helpful assistant",
+        model=config  # Pass LLMConfig via model parameter
+    )
+    
+    print(f"Primary model: {agent.llm}")
+    print(f"Fallback models: {agent.fallback_models}")
+    print(f"Base URL: {agent.base_url}")
+    print(f"API Key: {agent.api_key}")
+
+# Example 3: Traditional usage still works
+def example_traditional():
+    """Traditional usage without LLMConfig (no fallback support)."""
+    agent = Agent(
+        name="Assistant",
+        instructions="You are a helpful assistant",
+        model="gpt-4o",
+        base_url="https://api.example.com",
+        api_key="test-key"
+    )
+    
+    print(f"Primary model: {agent.llm}")
+    print(f"Fallback models: {agent.fallback_models}")  # Will be empty
+    print(f"Base URL: {agent.base_url}")
+    print(f"API Key: {agent.api_key}")
+
+if __name__ == "__main__":
+    print("=== Example 1: LLMConfig via llm parameter ===")
+    example_llm_param()
+    print()
+    
+    print("=== Example 2: LLMConfig via model parameter ===")
+    example_model_param()
+    print()
+    
+    print("=== Example 3: Traditional usage (no fallback) ===")
+    example_traditional()

examples/consolidated_params/advanced_workflow_full_features.py

@@ -32,7 +32,7 @@ workflow = AgentFlow(
         ),
     ],
     # Workflow-level consolidated params
-    output=WorkflowOutputConfig(output="verbose", stream=True),
+    output=WorkflowOutputConfig(verbose=True, stream=True),
     planning=WorkflowPlanningConfig(enabled=True, reasoning=True),
     memory=WorkflowMemoryConfig(backend="file"),
     # Agent-like params

examples/consolidated_params/basic_web.py

@@ -5,12 +5,12 @@ Demonstrates web search with consolidated params.
 Presets: duckduckgo, tavily, google, bing, serper, search_only, fetch_only
 """
 
-from praisonaiagents import Agent
+from praisonaiagents import Agent, WebConfig
 
-# Basic: Enable web search with preset
+# Basic: Enable web search with preset (search only — fetch needs Claude)
 agent = Agent(
     instructions="You are a research assistant with web access.",
-    web="duckduckgo",  # Presets: duckduckgo, tavily, google, bing, serper
+    web=WebConfig(search=True, fetch=False, search_provider="duckduckgo"),
 )
 
 if __name__ == "__main__":

examples/python/agents/code-agent.py

@@ -6,10 +6,18 @@ It includes self-reflection for improving code quality and multiple specialized
 """
 
 from praisonaiagents import Agent, Task, AgentTeam
-from praisonaiagents.tools import CodeInterpreter
 
-# Initialize code execution tool
-code_interpreter = CodeInterpreter()
+def run_python_code(code: str) -> str:
+    """Execute Python code safely and return stdout or errors."""
+    import io
+    import contextlib
+    try:
+        buf = io.StringIO()
+        with contextlib.redirect_stdout(buf):
+            exec(code, {"__builtins__": __builtins__})
+        return buf.getvalue() or "Executed successfully (no output)"
+    except Exception as e:
+        return f"Execution error: {e}"
 
 # Custom code analysis tool
 def analyze_code_complexity(code: str) -> str:
@@ -32,9 +40,7 @@ code_writer = Agent(
     goal="Write clean, efficient, and well-documented Python code",
     backstory="You are a senior Python developer with 10+ years of experience in writing production-quality code.",
     instructions="Write Python code that follows PEP 8 standards, includes proper error handling, and has clear documentation.",
-    reflection=True,
-    
-    
+    output="silent",
 )
 
 code_reviewer = Agent(
@@ -52,7 +58,7 @@ code_executor = Agent(
     goal="Safely execute and test Python code",
     backstory="You are responsible for running code and reporting results or errors.",
     instructions="Execute the provided code safely and report the output or any errors encountered.",
-    tools=[code_interpreter]
+    tools=[run_python_code]
 )
 
 # Example tasks
@@ -87,7 +93,7 @@ def demonstrate_code_agent():
     workflow = AgentTeam(
         agents=[code_writer, code_reviewer, code_executor],
         tasks=[write_task, review_task, execute_task],
-        process="sequential", output="verbose"
+        process="sequential", output="silent"
     )
     
     # Run the workflow
@@ -106,10 +112,8 @@ general_code_agent = Agent(
 3. Explain code concepts
 4. Refactor code for better performance
 5. Add tests and documentation""",
-    tools=[code_interpreter, analyze_code_complexity],
-    reflection=True,
-    
-    
+    tools=[run_python_code, analyze_code_complexity],
+    output="silent",
 )
 
 if __name__ == "__main__":

examples/python/agents/math-agent.py

@@ -41,21 +41,25 @@ def calculate_statistics(numbers: list) -> str:
     """Calculate basic statistics for a list of numbers."""
     if not numbers:
         return "No numbers provided"
-    
-    n = len(numbers)
-    mean = sum(numbers) / n
-    sorted_nums = sorted(numbers)
+    try:
+        nums = [float(x) for x in numbers]
+    except (TypeError, ValueError):
+        return "Error: all values must be numbers"
+
+    n = len(nums)
+    mean = sum(nums) / n
+    sorted_nums = sorted(nums)
     median = sorted_nums[n//2] if n % 2 else (sorted_nums[n//2-1] + sorted_nums[n//2]) / 2
-    variance = sum((x - mean)**2 for x in numbers) / n
+    variance = sum((x - mean)**2 for x in nums) / n
     std_dev = math.sqrt(variance)
-    
+
     return f"""Statistics:
 - Count: {n}
 - Mean: {mean:.4f}
 - Median: {median:.4f}
 - Standard Deviation: {std_dev:.4f}
-- Min: {min(numbers)}
-- Max: {max(numbers)}"""
+- Min: {min(nums)}
+- Max: {max(nums)}"""
 
 # Create specialized math agents
 calculation_agent = Agent(
@@ -66,8 +70,7 @@ calculation_agent = Agent(
     instructions="Perform the requested calculations step by step, showing your work clearly.",
     tools=[basic_calculator, solve_quadratic, calculate_statistics],
     reflection=True,
-    
-    
+    output="silent",
 )
 
 problem_solver_agent = Agent(
@@ -82,8 +85,7 @@ problem_solver_agent = Agent(
 4. Show each step of the solution
 5. Verify the answer makes sense""",
     reflection=True,
-    
-    
+    output="silent",
 )
 
 math_tutor_agent = Agent(
@@ -91,7 +93,8 @@ math_tutor_agent = Agent(
     role="Mathematics tutor and explainer",
     goal="Explain mathematical concepts clearly and help users understand math",
     backstory="You are a patient math tutor who can explain concepts at any level from elementary to advanced.",
-    instructions="Explain mathematical concepts using simple language, provide examples, and check for understanding."
+    instructions="Explain mathematical concepts using simple language, provide examples, and check for understanding.",
+    output="silent",
 )
 
 # Example workflow for solving complex problems
@@ -126,7 +129,8 @@ def solve_math_problem_workflow():
     workflow = AgentTeam(
         agents=[problem_solver_agent, calculation_agent, math_tutor_agent],
         tasks=[understand_task, solve_task, verify_task],
-        process="sequential", output="verbose"
+        process="sequential",
+        output="silent",
     )
     
     return workflow.start()
@@ -148,8 +152,7 @@ math_assistant = Agent(
 Always show your work step by step.""",
     tools=[basic_calculator, solve_quadratic, calculate_statistics],
     reflection=True,
-    
-    
+    output="silent",
 )
 
 if __name__ == "__main__":

examples/python/bot_run_control_example.py

@@ -0,0 +1,238 @@
+#!/usr/bin/env python3
+"""
+Example: Bot In-Flight Run Control
+
+Demonstrates how to use SessionRunControl to provide better UX for bots
+during long-running agent operations. Shows busy feedback, pending message
+handling, and /stop command support.
+
+This example shows the "before" and "after" behavior described in issue #1914.
+"""
+
+import asyncio
+import time
+from typing import Optional
+
+# Mock agent for demonstration
+class MockAgent:
+    """Simple mock agent that simulates long-running tasks."""
+    
+    def __init__(self, name="assistant"):
+        self.name = name
+        self.chat_history = []
+        self._interrupt_controller = None
+    
+    def chat(self, prompt: str) -> str:
+        """Simulate a long-running chat response."""
+        print(f"[Agent] Starting to process: '{prompt[:50]}...'")
+        
+        # Simulate long processing with interrupt checking
+        for i in range(30):  # 3 second task (30 x 0.1s)
+            time.sleep(0.1)
+            
+            # Check for interruption
+            if (self._interrupt_controller and 
+                hasattr(self._interrupt_controller, 'is_set') and
+                self._interrupt_controller.is_set()):
+                reason = getattr(self._interrupt_controller, 'reason', 'unknown')
+                print(f"[Agent] ⚠️ Interrupted: {reason}")
+                raise InterruptedError(f"Task cancelled: {reason}")
+                
+        response = f"Completed analysis of: {prompt}"
+        print(f"[Agent] ✅ Finished: {response}")
+        return response
+
+
+async def demo_without_run_control():
+    """Demonstrate the old behavior (silent blocking)."""
+    print("\n" + "="*60)
+    print("🔴 BEFORE: Without Run Control (Silent Blocking)")
+    print("="*60)
+    
+    from praisonai.bots._session import BotSessionManager
+    
+    # Standard session manager (no run control)
+    session_mgr = BotSessionManager()
+    agent = MockAgent("research-agent")
+    user_id = "user123"
+    
+    print("\n1. User sends long task:")
+    print("   User: 'research quantum computing trends'")
+    
+    # Start first task (this will run for 3 seconds)
+    task1 = asyncio.create_task(
+        session_mgr.chat(agent, user_id, "research quantum computing trends")
+    )
+    
+    # Simulate user sending follow-up messages during execution
+    await asyncio.sleep(0.5)
+    print("\n2. User sends follow-up (this will block silently):")
+    print("   User: 'actually focus on business impact'")
+    
+    task2 = asyncio.create_task(
+        session_mgr.chat(agent, user_id, "actually focus on business impact")  
+    )
+    
+    await asyncio.sleep(0.5)
+    print("\n3. User tries to send /stop (this will also block):")
+    print("   User: '/stop'")
+    
+    task3 = asyncio.create_task(
+        session_mgr.chat(agent, user_id, "/stop")
+    )
+    
+    print("\n⏳ Waiting for all messages to complete...")
+    results = await asyncio.gather(task1, task2, task3)
+    
+    print("\n📊 Results (all processed sequentially with no feedback):")
+    for i, result in enumerate(results, 1):
+        print(f"   {i}. {result}")
+    
+    print("\n❌ Problems with this approach:")
+    print("   - No immediate feedback on follow-up messages")
+    print("   - No way to cancel long-running tasks") 
+    print("   - User doesn't know if bot is working or broken")
+
+
+async def demo_with_run_control():
+    """Demonstrate the new behavior with run control."""
+    print("\n" + "="*60)
+    print("🟢 AFTER: With Run Control (Better UX)")  
+    print("="*60)
+    
+    from praisonai.bots._run_control import SessionRunControl
+    from praisonai.bots._session import BotSessionManager
+    from praisonai.bots._commands import handle_stop_command
+    
+    # Session manager with run control
+    run_control = SessionRunControl(
+        busy_mode="queue",
+        busy_ack_template="⏳ {action} — will process after current task finishes"
+    )
+    session_mgr = BotSessionManager(run_control=run_control)
+    agent = MockAgent("research-agent")
+    user_id = "user123"
+    
+    print("\n1. User sends long task:")
+    print("   User: 'research quantum computing trends'")
+    
+    # Start first task
+    task1 = asyncio.create_task(
+        session_mgr.chat_with_run_control(agent, user_id, "research quantum computing trends")
+    )
+    
+    # Immediate follow-up messages get acknowledgments
+    await asyncio.sleep(0.5)
+    print("\n2. User sends follow-up (gets immediate feedback):")
+    print("   User: 'actually focus on business impact'")
+    
+    result2 = await session_mgr.chat_with_run_control(
+        agent, user_id, "actually focus on business impact"
+    )
+    print(f"   Bot: {result2['response']}")
+    print(f"   Metadata: {result2['metadata']}")
+    
+    await asyncio.sleep(0.5)
+    print("\n3. User decides to stop current task:")
+    print("   User: '/stop'")
+    
+    stop_response = await handle_stop_command(user_id, run_control)
+    print(f"   Bot: {stop_response}")
+    
+    # Wait for first task to complete (it should be cancelled)
+    try:
+        result1 = await task1
+        print(f"\n📊 Task 1 result: {result1['response']}")
+        print(f"   Metadata: {result1['metadata']}")
+    except asyncio.CancelledError:
+        print("\n📊 Task 1 was cancelled as expected")
+    
+    print("\n4. User sends fresh request after stopping:")
+    print("   User: 'what's the weather like?'")
+    
+    result4 = await session_mgr.chat_with_run_control(
+        agent, user_id, "what's the weather like?"
+    )
+    print(f"   Bot: {result4['response']}")
+    
+    print("\n✅ Benefits of this approach:")
+    print("   - Immediate feedback on all messages")
+    print("   - Can cancel long-running tasks with /stop")
+    print("   - Pending messages are queued and processed in order")
+    print("   - User always knows what's happening")
+
+
+async def demo_interrupt_mode():
+    """Demonstrate interrupt mode."""
+    print("\n" + "="*60)
+    print("⚡ BONUS: Interrupt Mode (Cancel and Restart)")
+    print("="*60)
+    
+    from praisonai.bots._run_control import SessionRunControl
+    from praisonai.bots._session import BotSessionManager
+    
+    # Use interrupt mode instead of queue mode
+    run_control = SessionRunControl(busy_mode="interrupt")
+    session_mgr = BotSessionManager(run_control=run_control)
+    agent = MockAgent("research-agent")
+    user_id = "user123"
+    
+    print("\n1. User starts long task:")
+    print("   User: 'research renewable energy for 30 pages'")
+    
+    task1 = asyncio.create_task(
+        session_mgr.chat_with_run_control(agent, user_id, "research renewable energy for 30 pages")
+    )
+    
+    await asyncio.sleep(0.5)
+    print("\n2. User changes mind (interrupts and restarts):")
+    print("   User: 'actually just summarize solar panel efficiency'")
+    
+    result2 = await session_mgr.chat_with_run_control(
+        agent, user_id, "actually just summarize solar panel efficiency"
+    )
+    print(f"   Bot: {result2['response']}")
+    
+    # Wait for original task (should be interrupted)
+    try:
+        result1 = await task1
+        print(f"\n📊 Original task result: {result1['response']}")
+        if result1['metadata'].get('interrupted'):
+            print("   ✅ Original task was properly interrupted")
+    except:
+        print("\n📊 Original task was cancelled")
+    
+    print("\n✅ Interrupt mode is great for:")
+    print("   - When users frequently change direction")
+    print("   - Interactive sessions where latest intent matters most")
+    print("   - Real-time collaboration scenarios")
+
+
+async def main():
+    """Run all demonstrations."""
+    print("🚀 Bot In-Flight Run Control Demo")
+    print("Solving the silent lock problem described in issue #1914")
+    
+    # Show the problem
+    await demo_without_run_control()
+    
+    # Show the solution  
+    await demo_with_run_control()
+    
+    # Show advanced feature
+    await demo_interrupt_mode()
+    
+    print("\n" + "="*60)
+    print("🎯 Summary")
+    print("="*60)
+    print("SessionRunControl solves the silent blocking problem by providing:")
+    print("1. Immediate feedback for mid-run messages")
+    print("2. /stop command support via InterruptController")
+    print("3. Configurable policies (queue/interrupt/steer)")
+    print("4. Run generation tracking to prevent race conditions")
+    print("5. Pending message merging and ordering")
+    print("\nThis makes bot interactions feel responsive and predictable!")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

examples/python/tools/jira_agent_example.py

@@ -0,0 +1,115 @@
+#!/usr/bin/env python3
+"""
+JIRA Agent Watch Example
+
+This example demonstrates how to use the JIRA watch tools with PraisonAI agents.
+The agent can monitor JIRA issues and projects for changes, search for issues,
+and get detailed issue information.
+
+Setup:
+1. Install JIRA library: pip install jira
+2. Set environment variables:
+   - JIRA_EMAIL=your_email@example.com (for cloud JIRA)
+   - JIRA_API_TOKEN=your_api_token
+   - Or use JIRA_USERNAME + JIRA_API_TOKEN for server JIRA
+
+Usage:
+    python jira_agent_example.py
+"""
+
+import os
+from praisonaiagents import Agent
+from praisonaiagents.tools import jira_tools
+
+
+def main():
+    """Demonstrate JIRA agent watch capabilities."""
+    
+    # Check if JIRA credentials are available
+    jira_email = os.getenv('JIRA_EMAIL')
+    jira_token = os.getenv('JIRA_API_TOKEN')
+    jira_username = os.getenv('JIRA_USERNAME')
+    
+    if not jira_token:
+        print("❌ JIRA_API_TOKEN environment variable not set")
+        print("Please set your JIRA API token:")
+        print("export JIRA_API_TOKEN='your_token_here'")
+        return
+        
+    if not (jira_email or jira_username):
+        print("❌ JIRA credentials incomplete")
+        print("Please set either:")
+        print("- JIRA_EMAIL (for cloud JIRA)")
+        print("- JIRA_USERNAME (for server JIRA)")
+        return
+
+    # Create agent with JIRA tools
+    agent = Agent(
+        name="JIRA Monitor Agent",
+        instructions="""
+        You are a JIRA monitoring agent. You can:
+        
+        1. Watch specific JIRA issues for changes
+        2. Monitor JIRA projects for new issues and updates  
+        3. Search for issues using JQL queries
+        4. Get detailed information about specific issues
+        
+        When using JIRA tools, always provide the full JIRA URL 
+        (e.g., https://yourcompany.atlassian.net).
+        
+        Be helpful and provide clear summaries of JIRA activity.
+        """,
+        tools=jira_tools(),  # Add all JIRA tools
+        llm="gpt-4o-mini"
+    )
+    
+    print("🎯 JIRA Agent Watch Example")
+    print("=" * 50)
+    print()
+    print("Available JIRA tools:")
+    print("- jira_watch_issue: Monitor a specific issue for changes")
+    print("- jira_watch_project: Monitor a project for new/updated issues")
+    print("- jira_get_issue_info: Get detailed info about an issue")  
+    print("- jira_search_issues: Search issues using JQL")
+    print()
+    
+    # Example interactions
+    example_prompts = [
+        "Get information about JIRA issue DEMO-1 from https://praisonai.atlassian.net",
+        "Search for open issues in project DEMO from https://praisonai.atlassian.net using JQL: 'project = DEMO AND status = Open'",
+        "Watch JIRA issue DEMO-1 from https://praisonai.atlassian.net for 2 minutes (check every 60 seconds, max 2 checks)",
+    ]
+    
+    print("📝 Example prompts you can try:")
+    for i, prompt in enumerate(example_prompts, 1):
+        print(f"{i}. {prompt}")
+    print()
+    
+    # Interactive mode
+    print("💬 Interactive JIRA Agent (type 'quit' to exit)")
+    print("-" * 50)
+    
+    while True:
+        try:
+            user_input = input("\n🧑 You: ").strip()
+            
+            if user_input.lower() in ['quit', 'exit', 'q']:
+                print("👋 Goodbye!")
+                break
+                
+            if not user_input:
+                continue
+                
+            print("🤖 Agent: ", end="", flush=True)
+            response = agent.start(user_input)
+            print(response)
+            
+        except KeyboardInterrupt:
+            print("\n👋 Goodbye!")
+            break
+        except Exception as e:
+            print(f"❌ Error: {e}")
+
+
+if __name__ == "__main__":
+    main()

examples/python/workflows/workflow_branching.py

@@ -6,7 +6,6 @@ to different next steps based on output content.
 """
 
 from praisonaiagents import AgentFlow, Task
-from praisonaiagents import AgentFlowManager
 
 # Create a workflow with branching
 workflow = AgentFlow(
@@ -36,16 +35,9 @@ workflow = AgentFlow(
 )
 
 if __name__ == "__main__":
-    # Create manager and register workflow
-    manager = WorkflowManager()
-    manager.workflows["Decision Workflow"] = workflow
-    
-    # Execute - will branch to success_handler since 42 is positive
-    result = manager.execute(
-        "Decision Workflow",
-        default_llm="gpt-4o-mini"
-    )
-    
+    result = workflow.run("", llm="gpt-4o-mini", verbose=True)
     print("Workflow completed!")
-    for step_result in result["results"]:
-        print(f"  {step_result['step']}: {step_result['status']}")
+    for step in result.get("steps", []):
+        name = step.get("step", step.get("name", "unknown"))
+        status = step.get("status", "completed")
+        print(f"  {name}: {status}")

examples/python/workflows/workflow_early_stop.py

@@ -6,9 +6,8 @@ functions that return StepResult with stop_workflow=True.
 """
 
 from praisonaiagents import (
-    Workflow, Task, WorkflowContext, StepResult
+    AgentFlow, Task, WorkflowContext, StepResult
 )
-from praisonaiagents import AgentFlowManager
 
 # Custom validator that can stop the workflow
 def validate_data(context: WorkflowContext) -> StepResult:
@@ -50,20 +49,13 @@ workflow = AgentFlow(
 )
 
 if __name__ == "__main__":
-    manager = WorkflowManager()
-    manager.workflows["Data Processing"] = workflow
-    
-    # Test with invalid data (will stop early)
     print("=== Testing with invalid data ===")
-    result = manager.execute("Data Processing", default_llm="gpt-4o-mini")
-    
-    for step_result in result["results"]:
-        print(f"  {step_result['step']}: {step_result['output']}")
-    
-    # Test with valid data (will complete)
+    result = workflow.run("", llm="gpt-4o-mini", verbose=True)
+    for step in result.get("steps", []):
+        print(f"  {step.get('step', 'step')}: {step.get('output', step.get('status'))}")
+
     print("\n=== Testing with valid data ===")
     workflow.variables["data"] = {"value": 42}
-    result = manager.execute("Data Processing", default_llm="gpt-4o-mini")
-    
-    for step_result in result["results"]:
-        print(f"  {step_result['step']}: {step_result['status']}")
+    result = workflow.run("", llm="gpt-4o-mini", verbose=True)
+    for step in result.get("steps", []):
+        print(f"  {step.get('step', 'step')}: {step.get('status', 'completed')}")

examples/python/workflows/workflow_loops.py

@@ -6,7 +6,6 @@ Each item is processed individually with the loop variable.
 """
 
 from praisonaiagents import AgentFlow, Task
-from praisonaiagents import AgentFlowManager
 
 # Create a workflow with loop
 workflow = AgentFlow(
@@ -30,18 +29,12 @@ workflow = AgentFlow(
 )
 
 if __name__ == "__main__":
-    # Create manager and register workflow
-    manager = WorkflowManager()
-    manager.workflows["Process Items"] = workflow
-    
-    # Execute - will process each fruit in the list
-    result = manager.execute(
-        "Process Items",
-        default_llm="gpt-4o-mini"
-    )
-    
+    result = workflow.run("", llm="gpt-4o-mini", verbose=True)
     print("Workflow completed!")
-    for step_result in result["results"]:
-        print(f"  {step_result['step']}: {step_result['status']}")
-        if isinstance(step_result.get("output"), list):
-            print(f"    Loop processed {len(step_result['output'])} items")
+    for step in result.get("steps", []):
+        name = step.get("step", step.get("name", "unknown"))
+        status = step.get("status", "completed")
+        print(f"  {name}: {status}")
+        out = step.get("output")
+        if isinstance(out, list):
+            print(f"    Loop processed {len(out)} items")

examples/python/workflows/workflow_mixed_steps.py

@@ -53,7 +53,7 @@ workflow = AgentFlow(
         # Step 4: Function - formats output
         format_output
     ],
-    default_llm="gpt-4o-mini"  # Used for action-based steps
+    llm="gpt-4o-mini"  # Used for action-based steps
 )
 
 if __name__ == "__main__":

examples/python/workflows/workflow_parallel.py

@@ -5,7 +5,7 @@ Demonstrates running multiple agents concurrently and
 combining their results with an aggregator agent.
 """
 
-from praisonaiagents import Agent, Workflow
+from praisonaiagents import Agent, AgentFlow
 from praisonaiagents.workflows import parallel
 
 # Create parallel research agents

examples/python/workflows/workflow_routing.py

@@ -5,7 +5,7 @@ Demonstrates decision-based routing where an agent classifier
 routes to specialized handler agents based on the request type.
 """
 
-from praisonaiagents import Agent, Workflow
+from praisonaiagents import Agent, AgentFlow
 from praisonaiagents.workflows import route
 
 # Create classifier agent

gateway_server_prototype.py

@@ -0,0 +1,201 @@
+"""
+Prototype implementation of Gateway Server with backpressure and flow control.
+
+This demonstrates how the bounded queues and slow consumer detection would be
+implemented in the wrapper layer (praisonai package) using the configuration
+from the core SDK (praisonaiagents).
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from dataclasses import dataclass, field
+from typing import Any, Dict, Optional
+
+# For the prototype, we'll define minimal config inline
+# from praisonaiagents.gateway import GatewayConfig
+
+@dataclass 
+class SessionConfig:
+    max_inbox: int = 256
+    
+@dataclass
+class GatewayConfig:
+    session_config: SessionConfig = field(default_factory=SessionConfig)
+    max_buffered_bytes: int = 1024 * 1024  # 1MB default
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class GatewaySession:
+    """Gateway session with bounded inbox for backpressure."""
+    
+    _session_id: str
+    _agent_id: str
+    _client_id: Optional[str] = None
+    _inbox: asyncio.Queue = field(default_factory=asyncio.Queue)
+    _max_inbox: int = 256
+    _is_executing: bool = False
+    
+    def setup_inbox(self, max_inbox: int = 256) -> None:
+        """Set up the inbox with proper bounds."""
+        self._max_inbox = max_inbox
+        if max_inbox > 0:
+            self._inbox = asyncio.Queue(maxsize=max_inbox)
+        else:
+            self._inbox = asyncio.Queue()  # Unbounded for backward compat if 0
+    
+    async def queue_message(self, message: str) -> bool:
+        """Queue a user message with backpressure handling.
+        
+        Returns:
+            True if message was queued, False if inbox is full.
+        """
+        if self._max_inbox > 0:
+            # Bounded queue - check if full
+            try:
+                self._inbox.put_nowait(message)
+                return True
+            except asyncio.QueueFull:
+                return False
+        else:
+            # Unbounded queue for backward compatibility
+            await self._inbox.put(message)
+            return True
+
+
+class GatewayServer:
+    """Gateway server with flow control."""
+    
+    def __init__(self, config: GatewayConfig):
+        self.config = config
+        self._sessions: Dict[str, GatewaySession] = {}
+        self._clients: Dict[str, Any] = {}  # client_id -> WebSocket
+        
+    def create_session(self, session_id: str, agent_id: str, client_id: str) -> GatewaySession:
+        """Create a new session with bounded inbox."""
+        session = GatewaySession(
+            _session_id=session_id,
+            _agent_id=agent_id,
+            _client_id=client_id,
+        )
+        # Set up inbox with configured bounds
+        session.setup_inbox(self.config.session_config.max_inbox)
+        self._sessions[session_id] = session
+        return session
+        
+    async def handle_message(self, session: GatewaySession, message: str, client_id: str) -> str:
+        """Handle incoming message with inbox overflow protection."""
+        # Try to queue the message
+        queued = await session.queue_message(message)
+        if not queued:
+            # Inbox is full, send error response
+            await self._send_to_client(
+                client_id,
+                {
+                    "type": "error",
+                    "code": "inbox_full",
+                    "message": "Message queue is full. Please wait for current messages to be processed."
+                }
+            )
+            return "Inbox full - message rejected."
+        
+        # Start processing if not already running
+        if not session._is_executing:
+            session._is_executing = True
+            asyncio.create_task(self._drain_session_queue(session, client_id))
+            return "Processing started."
+        
+        return "Message queued."
+    
+    async def _drain_session_queue(self, session: GatewaySession, client_id: str) -> None:
+        """Process all messages from the session's inbox queue."""
+        try:
+            while not session._inbox.empty():
+                msg = await session._inbox.get()
+                # Process message here (simplified for prototype)
+                logger.info(f"Processing message for session {session._session_id}: {msg}")
+                
+                # Simulate processing delay
+                await asyncio.sleep(0.1)
+                
+                # Send response to client (simplified)
+                await self._send_to_client(
+                    client_id,
+                    {
+                        "type": "response",
+                        "message": f"Processed: {msg}"
+                    }
+                )
+        except Exception as e:
+            logger.error(f"Error processing queue for session {session._session_id}: {e}")
+        finally:
+            session._is_executing = False
+    
+    async def _send_to_client(self, client_id: str, data: Dict[str, Any]) -> None:
+        """Send data to client with slow consumer detection."""
+        ws = self._clients.get(client_id)
+        if ws:
+            try:
+                # Check for slow consumer (buffered bytes)
+                if self.config.max_buffered_bytes > 0:
+                    # Try to get transport buffer size
+                    transport = getattr(ws, 'transport', None) or getattr(ws, '_transport', None)
+                    if transport and hasattr(transport, 'get_write_buffer_size'):
+                        buffered_bytes = transport.get_write_buffer_size()
+                        if buffered_bytes > self.config.max_buffered_bytes:
+                            # Check if this is a droppable event type
+                            event_type = data.get("type", "")
+                            droppable_types = {"presence", "typing", "status"}
+                            
+                            if event_type in droppable_types:
+                                # Drop the message for slow consumers
+                                logger.warning(f"Dropping {event_type} event for slow consumer {client_id} (buffered: {buffered_bytes} bytes)")
+                                return
+                            else:
+                                # Critical event - close the connection
+                                logger.warning(f"Closing slow consumer {client_id} (buffered: {buffered_bytes} bytes > {self.config.max_buffered_bytes})")
+                                await ws.close(code=1013, reason="slow consumer")
+                                return
+                
+                # Send the message
+                await ws.send_json(data)
+            except Exception as e:
+                logger.error(f"Error sending to client {client_id}: {e}", exc_info=True)
+
+
+# Example usage demonstrating the flow control
+async def demo():
+    """Demonstrate the flow control implementation."""
+    
+    # Create config with bounded inbox and slow consumer detection
+    config = GatewayConfig()
+    config.session_config.max_inbox = 10  # Small inbox for demo
+    config.max_buffered_bytes = 1024  # 1KB for demo
+    
+    # Create server and session
+    server = GatewayServer(config)
+    session = server.create_session("sess1", "agent1", "client1")
+    
+    # Simulate filling up the inbox
+    print(f"Inbox max size: {session._inbox.maxsize}")
+    
+    # Fill the inbox
+    for i in range(10):
+        success = await session.queue_message(f"Message {i}")
+        print(f"Queued message {i}: {success}")
+    
+    # Try to add one more - should fail
+    success = await session.queue_message("Overflow message")
+    print(f"Overflow message queued: {success}")
+    
+    # Process messages
+    while not session._inbox.empty():
+        msg = await session._inbox.get()
+        print(f"Processing: {msg}")
+
+
+if __name__ == "__main__":
+    asyncio.run(demo())

src/praisonai-agents/DOCS_PARITY.md

@@ -1,6 +1,6 @@
 # Documentation Parity Tracker (Python)
 
-> **Categories:** 60 | **Documented:** 57 | **Parity:** 95.0%
+> **Categories:** 64 | **Documented:** 61 | **Parity:** 95.3%
 
 This report compares **Python SDK feature categories** against **Python documentation** (docs/concepts, docs/features, etc.).
 
@@ -8,72 +8,76 @@ This report compares **Python SDK feature categories** against **Python document
 
 | Metric | Count |
 |--------|-------|
-| Feature Categories | 60 |
-| **Documented Categories** | **57** |
+| Feature Categories | 64 |
+| **Documented Categories** | **61** |
 | **Undocumented Categories** | **3** |
-| **Parity** | **95.0%** |
+| **Parity** | **95.3%** |
 
 ## Documented Categories
 
 | Category | Features | Docs | Lines |
 |----------|----------|------|-------|
 | ✅ AGUI | 1 | 2 | 420 |
-| ✅ Agent | 19 | 30 | 8367 |
-| ✅ Agent-to-Agent (A2A) | 1 | 3 | 625 |
-| ✅ Approval | 1 | 2 | 889 |
+| ✅ Agent | 21 | 34 | 9360 |
+| ✅ Agent-to-Agent (A2A) | 1 | 7 | 2880 |
+| ✅ Agent-to-User (A2U) | 1 | 3 | 547 |
+| ✅ Approval | 1 | 3 | 1283 |
 | ✅ Audio | 2 | 10 | 649 |
-| ✅ Auto Generation | 3 | 6 | 1386 |
-| ✅ Autonomy | 3 | 4 | 1146 |
-| ✅ Bots | 5 | 2 | 573 |
+| ✅ Auto Generation | 4 | 6 | 1724 |
+| ✅ Autonomy | 3 | 5 | 1919 |
+| ✅ Bots | 7 | 14 | 5228 |
+| ✅ Budget | 1 | 1 | 287 |
+| ✅ CLI | 5 | 98 | 24103 |
 | ✅ Chunking | 2 | 2 | 285 |
 | ✅ Citations | 2 | 1 | 202 |
-| ✅ Code Execution | 3 | 13 | 3395 |
+| ✅ Code Execution | 3 | 13 | 3551 |
 | ✅ Conditions | 1 | 3 | 1086 |
-| ✅ Configuration | 2 | 7 | 1753 |
-| ✅ Context Management | 13 | 29 | 7576 |
-| ✅ Database | 1 | 31 | 3037 |
+| ✅ Configuration | 3 | 7 | 1812 |
+| ✅ Context Management | 16 | 31 | 9294 |
+| ✅ Database | 1 | 34 | 4339 |
 | ✅ Display | 6 | 2 | 1175 |
 | ✅ Embeddings | 6 | 23 | 1966 |
-| ✅ Evaluation | 1 | 6 | 2538 |
-| ✅ Events | 1 | 2 | 699 |
-| ✅ Execution | 3 | 2 | 515 |
-| ✅ Files | 2 | 5 | 863 |
-| ✅ Flow | 1 | 1 | 284 |
-| ✅ Gateway | 6 | 1 | 256 |
-| ✅ Guardrails | 4 | 4 | 2279 |
-| ✅ Handoffs | 10 | 4 | 1674 |
-| ✅ Hooks | 2 | 6 | 1606 |
+| ✅ Evaluation | 1 | 6 | 2539 |
+| ✅ Events | 1 | 2 | 704 |
+| ✅ Execution | 3 | 3 | 1059 |
+| ✅ Files | 2 | 6 | 1100 |
+| ✅ Flow | 1 | 2 | 435 |
+| ✅ Gateway | 6 | 10 | 3924 |
+| ✅ Guardrails | 4 | 4 | 2008 |
+| ✅ Handoffs | 11 | 5 | 2155 |
+| ✅ Hooks | 2 | 6 | 1916 |
 | ✅ Image | 1 | 11 | 1096 |
-| ✅ Knowledge | 3 | 12 | 3262 |
-| ✅ LLM | 2 | 2 | 1049 |
-| ✅ Loops | 4 | 1 | 250 |
-| ✅ MCP | 2 | 47 | 9754 |
-| ✅ Memory | 6 | 12 | 4474 |
-| ✅ Output | 3 | 4 | 868 |
-| ✅ Parallel Execution | 2 | 2 | 347 |
-| ✅ Planning | 6 | 5 | 1506 |
-| ✅ Plugins | 8 | 3 | 1494 |
-| ✅ Prompts | 2 | 4 | 796 |
-| ✅ Providers | 2 | 50 | 5396 |
+| ✅ Knowledge | 3 | 12 | 3466 |
+| ✅ LLM | 3 | 5 | 2251 |
+| ✅ Loops | 4 | 3 | 774 |
+| ✅ MCP | 2 | 51 | 10960 |
+| ✅ Memory | 6 | 16 | 5855 |
+| ✅ Observability | 2 | 23 | 2590 |
+| ✅ Output | 3 | 4 | 903 |
+| ✅ Parallel Execution | 3 | 2 | 399 |
+| ✅ Planning | 6 | 5 | 1519 |
+| ✅ Plugins | 8 | 3 | 1502 |
+| ✅ Prompts | 2 | 7 | 1570 |
+| ✅ Providers | 2 | 51 | 5911 |
 | ✅ Query | 1 | 2 | 616 |
-| ✅ RAG | 5 | 14 | 2910 |
+| ✅ RAG | 5 | 15 | 3190 |
 | ✅ Realtime | 2 | 5 | 584 |
 | ✅ Reflection | 3 | 2 | 469 |
-| ✅ Retrieval | 2 | 5 | 941 |
+| ✅ Retrieval | 2 | 5 | 944 |
 | ✅ Routing | 1 | 2 | 597 |
-| ✅ Sandbox | 5 | 3 | 1002 |
-| ✅ Security | 1 | 1 | 875 |
-| ✅ Sessions | 2 | 8 | 2128 |
-| ✅ Skills | 5 | 5 | 2002 |
-| ✅ Tasks | 2 | 5 | 2590 |
-| ✅ Telemetry | 1 | 2 | 541 |
+| ✅ Sandbox | 5 | 6 | 2369 |
+| ✅ Security | 1 | 3 | 2518 |
+| ✅ Sessions | 4 | 9 | 3826 |
+| ✅ Skills | 5 | 9 | 3415 |
+| ✅ Tasks | 2 | 6 | 3064 |
+| ✅ Telemetry | 1 | 2 | 560 |
 | ✅ Templates | 1 | 7 | 1411 |
-| ✅ Tools | 4 | 97 | 17827 |
+| ✅ Tools | 12 | 113 | 23809 |
 | ✅ Tracing | 3 | 2 | 139 |
-| ✅ Video | 2 | 5 | 460 |
+| ✅ Video | 2 | 6 | 510 |
 | ✅ Vision | 2 | 1 | 283 |
-| ✅ Web | 4 | 6 | 1267 |
-| ✅ Workflows | 4 | 9 | 3579 |
+| ✅ Web | 4 | 6 | 1297 |
+| ✅ Workflows | 4 | 11 | 4302 |
 
 ## Stub Documentation (Need Content)
 
@@ -94,21 +98,18 @@ These categories have documentation files but < 50 lines (stubs):
 
 These docs exist but don't match any implemented feature category:
 
-- ℹ️ AI SDK (1 docs, 112 lines)
-- ℹ️ Agent-to-User (A2U) (3 docs, 770 lines)
-- ℹ️ CLI (91 docs, 21386 lines)
-- ℹ️ Callbacks (1 docs, 985 lines)
-- ℹ️ Chat (3 docs, 546 lines)
+- ℹ️ AI SDK (2 docs, 527 lines)
+- ℹ️ Callbacks (1 docs, 1122 lines)
+- ℹ️ Chat (3 docs, 634 lines)
 - ℹ️ Deep Research (2 docs, 437 lines)
-- ℹ️ Deployment (29 docs, 5931 lines)
-- ℹ️ Jobs (1 docs, 358 lines)
+- ℹ️ Deployment (29 docs, 6013 lines)
+- ℹ️ Jobs (2 docs, 961 lines)
 - ℹ️ Middleware (1 docs, 60 lines)
-- ℹ️ Observability (21 docs, 1580 lines)
-- ℹ️ Process (1 docs, 506 lines)
-- ℹ️ Scheduler (2 docs, 858 lines)
-- ℹ️ Storage (4 docs, 722 lines)
-- ℹ️ Streaming (3 docs, 485 lines)
-- ℹ️ Token Management (3 docs, 1020 lines)
+- ℹ️ Process (1 docs, 545 lines)
+- ℹ️ Scheduler (2 docs, 1084 lines)
+- ℹ️ Storage (5 docs, 957 lines)
+- ℹ️ Streaming (4 docs, 1163 lines)
+- ℹ️ Token Management (4 docs, 1383 lines)
 - ℹ️ Vector Store (11 docs, 733 lines)
 
 ---

src/praisonai-agents/benchmarks/BENCHMARK_RESULTS.md

@@ -1,35 +1,70 @@
 # PraisonAI Agents - Benchmark Results
 
-**Generated:** 2025-12-18 14:40:22
-**Iterations:** 100
-**Test:** Agent instantiation (without tools)
+**Generated:** 2026-06-17
+**Methodology:** 10 warmup runs + 1000 measured iterations (`performance_benchmark.py` config)
+**Config:** `Agent(name=..., instructions=..., model="gpt-4o-mini", tools=[get_weather], output="silent")`
 
-## Results
+## Agent Instantiation Time
 
-| Framework | Avg Time (μs) | Relative |
-|-----------|---------------|----------|
-| **PraisonAI** | **3.77** | **1.00x (fastest)** |
-| OpenAI Agents SDK | 5.26 | 1.39x |
-| Agno | 5.64 | 1.49x |
-| PraisonAI (LiteLLM) | 7.56 | 2.00x |
-| PydanticAI | 226.94 | 60.16x |
-| LangGraph | 4,558.71 | 1,209x |
-| CrewAI | 15,607.92 | 4,138x |
+| Framework | Avg Time (μs) | Min (μs) | Median (μs) | Relative |
+|-----------|---------------|----------|-------------|----------|
+| OpenAI Agents SDK | 5.05 | — | — | 1.00x (fastest) |
+| Agno | 5.27 | — | — | 1.04x |
+| **PraisonAI** | **13.62** | **12.58** | **12.96** | **2.69x** |
+| PraisonAI (LiteLLM) | 14.49 | 13.37 | 13.71 | 2.86x |
+
+### PraisonAI detail (1000 iter)
+
+| Variant | Avg (μs) | Min (μs) | Median (μs) | Std (μs) |
+|---------|----------|----------|-------------|----------|
+| With tools (`gpt-4o-mini`) | 13.62 | 12.58 | 12.96 | 2.73 |
+| With tools (`openai/gpt-4o-mini`) | 14.49 | 13.37 | 13.71 | 4.14 |
+| Minimal, no tools | 13.40 | — | — | — |
+
+## Historical baseline
+
+| Source | Config | Avg (μs) | Notes |
+|--------|--------|----------|-------|
+| `performance_benchmark.py` (Dec 2025) | tools + 1000 iter | **3.20** | Pre-init-bloat baseline |
+| `simple_benchmark.py` (Dec 2025) | no tools, 100 iter | **3.77** | Saved in prior `BENCHMARK_RESULTS.md` |
+| Current (Jun 2026) | tools + 1000 iter | **13.62** | After lazy-init optimisations (was ~31 μs pre-fix) |
 
 ## Package Versions
 
 | Package | Version |
-|---------|--------|
-| PraisonAI | 0.1.5 |
-| Agno | 2.3.14 |
-| PydanticAI | 1.35.0 |
-| OpenAI Agents SDK | 0.6.3 |
-| LangGraph | 1.0.5 |
-| CrewAI | 1.6.1 |
+|---------|---------|
+| PraisonAI | 1.6.60 |
+| Agno | 2.4.2 |
+| OpenAI Agents SDK | 0.7.0 |
 
 ## How to Reproduce
 
+PraisonAI-only (fast; avoids CrewAI/LangGraph import hang):
+
 ```bash
 cd praisonai-agents
-python benchmarks/simple_benchmark.py
+PYTHONWARNINGS='ignore::DeprecationWarning' python -c "
+import time, statistics
+from typing import Literal
+from praisonaiagents import Agent
+
+def get_weather(city: Literal['nyc', 'sf']):
+    return 'cloudy' if city == 'nyc' else 'sunny'
+
+def bench(fn, n=1000, w=10):
+    for _ in range(w): fn()
+    times = []
+    for _ in range(n):
+        s = time.perf_counter(); fn(); times.append(time.perf_counter()-s)
+    print(f'avg={statistics.mean(times)*1e6:.2f} μs')
+
+bench(lambda: Agent(name='Test', instructions='Hi', model='gpt-4o-mini', tools=[get_weather], output='silent'))
+"
+```
+
+Full multi-framework benchmark:
+
+```bash
+cd praisonai-agents
+PYTHONWARNINGS='ignore::DeprecationWarning' python benchmarks/performance_benchmark.py --save
 ```

src/praisonai-agents/benchmarks/INSTANTIATION_RCA.md

@@ -0,0 +1,62 @@
+# Agent Instantiation — Root Cause Analysis
+
+**Date:** 2026-06-17  
+**Version:** praisonaiagents 1.6.58 (dev)  
+**Methodology:** 10 warmup + 1000 measured iterations, `time.perf_counter()`
+
+## Summary
+
+Instantiation regressed from **~3.2 μs** (Dec 2025) to **~31 μs** (~10×). After fixes, steady-state is **~16 μs** (~2× historical). PraisonAI is competitive with Agno (~5 μs) on warmed benchmarks using `model=`.
+
+## Before vs After Fixes
+
+| Config | Historical | Pre-fix | Post-fix |
+|--------|------------|---------|----------|
+| Minimal (`name`, `output="silent"`) | 3.20 μs | 31.36 μs | **16.35 μs** |
+| `model="gpt-4o-mini"` | 3.77 μs | 32.19 μs | **16.86 μs** |
+| `llm="gpt-4o-mini"` (deprecated) | 3.77 μs | 46.48 μs | **~18 μs** (warn once) |
+| LiteLLM `openai/gpt-4o-mini` cold | 7.31 μs | **3.7 s** | **53.63 μs** |
+| LiteLLM warmed instantiate | — | 60 μs | **19.83 μs** |
+| Import `Agent` | — | 20.8 ms | **29.6 ms** |
+
+## Root Causes (validated)
+
+| ID | Cause | Status | Fix applied |
+|----|-------|--------|-------------|
+| H1 | Config loader cold start (+9 ms first init) | CONFIRMED | Fast path when no config defaults (`_defaults_has_any_values`) |
+| H2 | LoopGuard on every init (~2 μs) | CONFIRMED | Lazy `_ensure_loop_guard()` on first tool call |
+| H3 | Approval presets | REJECTED | No change needed |
+| H4 | `llm=` deprecation every init (+16 μs) | CONFIRMED | Warn once per process in `warn_deprecated_param` |
+| H5 | Init bloat / config resolution (~25 μs) | CONFIRMED | Partially addressed via loader fast path |
+| H6 | LiteLLM eager `LLM()` import (+3.7 s) | CONFIRMED | Deferred `_llm_init_params` + lazy `llm_instance` property |
+| H7 | `ExecutionConfig` deprecation every init | CONFIRMED | Fixed stack walk; internal SDK path cached |
+
+## cProfile — Post-fix (500 warmed `model=` inits)
+
+| Function | Time (ms) | Per call |
+|----------|-----------|----------|
+| `Agent.__init__` | 8 | ~16 μs |
+| `apply_config_defaults` | 1 | ~2 μs |
+| `ExecutionConfig.__post_init__` | 1 | ~2 μs (first call only walks stack) |
+| `_warnings.warn` | **0** | removed from hot path |
+
+## Fixes implemented
+
+1. **`praisonaiagents/config/feature_configs.py`** — Fix `ExecutionConfig.__post_init__` deprecation guard; cache internal SDK detection.
+2. **`praisonaiagents/utils/deprecation.py`** — Warn once per process for deprecated params.
+3. **`praisonaiagents/config/loader.py`** — Skip `apply_config_defaults` lookups when config has no defaults.
+4. **`praisonaiagents/agent/agent.py`** — Lazy LoopGuard, lazy LLM instance, deferred LiteLLM params.
+5. **`benchmarks/simple_benchmark.py`** — Use `model=`, add 10-iteration warmup.
+6. **`benchmarks/performance_benchmark.py`** — Use `model=` instead of deprecated `llm=`.
+
+## Remaining gap (~16 μs vs ~3 μs)
+
+The residual ~13 μs is primarily **`Agent.__init__` size** (~1200 lines of consolidated config resolution added since v0.1.5). Further optimisation requires a dedicated slim fast-path for `Agent(name=..., output="silent")` without full config object churn.
+
+## How to reproduce
+
+```bash
+cd praisonai-agents
+PYTHONWARNINGS='ignore::DeprecationWarning' python benchmarks/simple_benchmark.py
+PYTHONWARNINGS='ignore::DeprecationWarning' python benchmarks/performance_benchmark.py
+```

src/praisonai-agents/benchmarks/param_impact_benchmark.py

@@ -120,8 +120,8 @@ def run_benchmark(iterations: int = 100) -> List[ParamResult]:
         ('backstory', 'I am an AI assistant.', 'backstory="..."'),
         
         # LLM configuration
-        ('llm', 'gpt-4o-mini', 'llm="gpt-4o-mini"'),
-        ('llm', 'openai/gpt-4o-mini', 'llm="openai/gpt-4o-mini" (LiteLLM)'),
+        ('model', 'gpt-4o-mini', 'model="gpt-4o-mini"'),
+        ('model', 'openai/gpt-4o-mini', 'model="openai/gpt-4o-mini" (LiteLLM)'),
         
         # Tools
         ('tools', [sample_tool], 'tools=[sample_tool]'),
@@ -298,7 +298,7 @@ def save_results(results: List[ParamResult], filename: str = 'PARAM_IMPACT_RESUL
         f.write('### Best Practices\n\n')
         f.write('1. Use `output="silent"` for production workloads\n')
         f.write('2. Only enable `context=True` when needed for long conversations\n')
-        f.write('3. Use `llm="gpt-4o-mini"` (OpenAI SDK) for fastest instantiation\n')
+        f.write('3. Use `model="gpt-4o-mini"` for fastest instantiation\n')
         f.write('4. Avoid `memory=True` unless persistent memory is required\n')
         
         f.write('\n## How to Reproduce\n\n')

src/praisonai-agents/benchmarks/performance_benchmark.py

@@ -185,7 +185,7 @@ def instantiate_praisonai_agent():
     return PraisonAgent(
         name="Test Agent",
         instructions="Be concise, reply with one sentence.",
-        llm="gpt-4o-mini",
+        model="gpt-4o-mini",
         tools=tools,
         output="silent"
     )
@@ -196,7 +196,7 @@ def instantiate_praisonai_litellm_agent():
     return PraisonAgent(
         name="Test Agent",
         instructions="Be concise, reply with one sentence.",
-        llm="openai/gpt-4o-mini",
+        model="openai/gpt-4o-mini",
         tools=tools,
         output="silent"
     )

src/praisonai-agents/benchmarks/simple_benchmark.py

@@ -14,6 +14,7 @@ import argparse
 
 
 ITERATIONS = 100
+WARMUP = 10
 
 
 def sample_tool(city: Literal['nyc', 'sf']):
@@ -23,8 +24,10 @@ def sample_tool(city: Literal['nyc', 'sf']):
     return 'sunny'
 
 
-def measure_instantiation(create_fn, iterations=ITERATIONS):
+def measure_instantiation(create_fn, iterations=ITERATIONS, warmup=WARMUP):
     """Measure average instantiation time in microseconds."""
+    for _ in range(warmup):
+        create_fn()
     times = []
     for _ in range(iterations):
         start = time.perf_counter()
@@ -48,11 +51,11 @@ def run_benchmark():
     from praisonaiagents import Agent as PraisonAgent
     
     results['PraisonAI'] = measure_instantiation(
-        lambda: PraisonAgent(name='Test', llm='gpt-4o-mini', output="silent")
+        lambda: PraisonAgent(name='Test', model='gpt-4o-mini', output="silent")
     )
     
     results['PraisonAI (LiteLLM)'] = measure_instantiation(
-        lambda: PraisonAgent(name='Test', llm='openai/gpt-4o-mini', output="silent")
+        lambda: PraisonAgent(name='Test', model='openai/gpt-4o-mini', output="silent")
     )
     
     # Other frameworks for comparison

src/praisonai-agents/praisonaiagents/__init__.py

@@ -114,6 +114,9 @@ def _get_lazy_cache():
 # ============================================================================
 
 _LAZY_IMPORTS = {
+    # Version information (lazy loaded to avoid file I/O on import)
+    '__version__': ('praisonaiagents._version', '__version__'),
+    
     # Tools (moved from eager imports for lazy loading)
     'Tools': ('praisonaiagents.tools.tools', 'Tools'),
     'BaseTool': ('praisonaiagents.tools.base', 'BaseTool'),
@@ -473,7 +476,7 @@ _LAZY_IMPORTS = {
     'GatewayClientProtocol': ('praisonaiagents.gateway.protocols', 'GatewayClientProtocol'),
     'GatewayEvent': ('praisonaiagents.gateway.protocols', 'GatewayEvent'),
     'GatewayMessage': ('praisonaiagents.gateway.protocols', 'GatewayMessage'),
-    'EventType': ('praisonaiagents.gateway.protocols', 'EventType'),
+    'GatewayEventType': ('praisonaiagents.gateway.protocols', 'EventType'),
     'GatewayConfig': ('praisonaiagents.gateway.config', 'GatewayConfig'),
     'SessionConfig': ('praisonaiagents.gateway.config', 'SessionConfig'),
     
@@ -775,6 +778,9 @@ def warmup(include_litellm: bool = False, include_openai: bool = True) -> dict:
 # ============================================================================
 
 __all__ = [
+    # Version information
+    '__version__',
+    
     # Core classes - the essentials
     'Agent',
     'AgentTeam',  # Primary class for multi-agent coordination (v1.0+)

src/praisonai-agents/praisonaiagents/_config.py

@@ -111,32 +111,3 @@ def _get_plugins_list() -> list:
 
 PLUGINS_ENABLED = _get_plugins_enabled()
 PLUGINS_LIST = _get_plugins_list()
-
-
-def missing_dependency_error(
-    package_name: str,
-    extra_name: Optional[str] = None,
-    feature_name: Optional[str] = None
-) -> ImportError:
-    """
-    Create a standardized ImportError for missing optional dependencies.
-    
-    Args:
-        package_name: The name of the missing package (e.g., 'litellm')
-        extra_name: The pip extra to install (e.g., 'llm' for [llm])
-        feature_name: Human-readable feature name (e.g., 'LLM support')
-    
-    Returns:
-        ImportError with helpful install instructions
-    """
-    if extra_name:
-        install_cmd = f"pip install 'praisonaiagents[{extra_name}]'"
-    else:
-        install_cmd = f"pip install {package_name}"
-    
-    feature_desc = f" for {feature_name}" if feature_name else ""
-    
-    return ImportError(
-        f"{package_name} is required{feature_desc} but not installed. "
-        f"Please install with: {install_cmd}"
-    )

src/praisonai-agents/praisonaiagents/_resolver_helpers.py

@@ -1,49 +0,0 @@
-"""
-Resolver helper functions for consolidated parameters.
-
-These helpers wrap the canonical resolver for specific parameter types,
-handling special cases like callable validators for guardrails.
-"""
-
-from typing import Any, Type
-
-from .config.param_resolver import resolve, ArrayMode
-from .config.presets import GUARDRAIL_PRESETS
-
-
-def resolve_guardrails(value: Any, config_class: Type) -> Any:
-    """
-    Resolve guardrails parameter.
-    
-    Handles special case of callable validators which have highest precedence.
-    
-    Args:
-        value: The guardrails parameter value
-        config_class: GuardrailConfig class
-        
-    Returns:
-        Resolved config, callable, or string
-    """
-    # Handle None
-    if value is None:
-        return None
-    
-    # Handle False (disabled)
-    if value is False:
-        return None
-    
-    # Handle callable (highest precedence after None check)
-    # This includes functions, lambdas, and callable objects
-    if callable(value) and not isinstance(value, type):
-        return value
-    
-    # Use canonical resolver for everything else
-    return resolve(
-        value=value,
-        param_name="guardrails",
-        config_class=config_class,
-        presets=GUARDRAIL_PRESETS,
-        array_mode=ArrayMode.PRESET_OVERRIDE,
-        string_mode="llm_prompt",
-        default=None,
-    )

src/praisonai-agents/praisonaiagents/_version.py

@@ -0,0 +1,51 @@
+"""
+Version utilities for praisonaiagents package.
+
+This module provides version information by reading from pyproject.toml
+to avoid duplication and ensure the single source of truth.
+"""
+
+import re
+from pathlib import Path
+
+
+def get_version() -> str:
+    """
+    Get the version string from pyproject.toml.
+    
+    Returns:
+        str: Version string (e.g., "1.6.52")
+        
+    Raises:
+        RuntimeError: If version cannot be found or parsed
+    """
+    try:
+        # Get the package root (praisonaiagents directory)
+        package_root = Path(__file__).parent
+        # Go up to src/praisonai-agents directory
+        pyproject_path = package_root.parent / "pyproject.toml"
+        
+        if not pyproject_path.exists():
+            raise RuntimeError(f"pyproject.toml not found at {pyproject_path}")
+            
+        content = pyproject_path.read_text()
+        
+        # Look for version = "X.Y.Z" pattern
+        match = re.search(r'version\s*=\s*["\']([^"\']+)["\']', content)
+        if not match:
+            raise RuntimeError("Version not found in pyproject.toml")
+            
+        return match.group(1)
+        
+    except Exception as e:
+        # Fallback to "unknown" to avoid breaking imports
+        import warnings
+        warnings.warn(
+            f"Failed to read version from pyproject.toml: {e}. Using 'unknown'.",
+            RuntimeWarning
+        )
+        return "unknown"
+
+
+# Cache the version to avoid reading the file multiple times
+__version__ = get_version()

src/praisonai-agents/praisonaiagents/agent/agent.py

@@ -7,6 +7,7 @@ import asyncio
 import contextlib
 import threading
 import concurrent.futures
+import random
 from typing import List, Optional, Any, Dict, Union, Literal, TYPE_CHECKING, Callable, Generator
 from collections import OrderedDict
 import inspect
@@ -16,7 +17,7 @@ from .chat_mixin import ChatMixin
 from .execution_mixin import ExecutionMixin
 from .memory_mixin import MemoryMixin
 from .async_memory_mixin import AsyncMemoryMixin
-from .tool_execution import ToolExecutionMixin
+from .tool_execution import ToolExecutionMixin, BackoffPolicy
 from .chat_handler import ChatHandlerMixin
 from .session_manager import SessionManagerMixin
 from .async_safety import AsyncSafeState
@@ -313,6 +314,38 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
                     cls._default_model_checked = True
         return cls._default_model
     
+    def _ensure_llm_instance(self):
+        """Lazy-create LLM instance from deferred init params (avoids import at Agent())."""
+        if self._llm_instance is not None:
+            return self._llm_instance
+        if self._llm_init_params:
+            try:
+                from ..llm.llm import LLM
+                self._llm_instance = LLM(**self._llm_init_params)
+            except ImportError as e:
+                raise ImportError(
+                    "LLM features requested but dependencies not installed. "
+                    "Please install with: pip install \"praisonaiagents[llm]\""
+                ) from e
+        return self._llm_instance
+
+    @property
+    def llm_instance(self):
+        return self._ensure_llm_instance()
+
+    @llm_instance.setter
+    def llm_instance(self, value):
+        self._llm_instance = value
+        if value is not None:
+            self._llm_init_params = None
+
+    def _ensure_loop_guard(self):
+        """Lazy-create loop guard on first tool execution."""
+        if self._loop_guard is None:
+            from ..escalation.loop_guard import LoopGuard, LoopGuardConfig
+            self._loop_guard = LoopGuard(LoopGuardConfig(enabled=True))
+        return self._loop_guard
+    
     @classmethod
     def _configure_logging(cls):
         """Configure logging settings once for all agent instances."""
@@ -518,7 +551,7 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
         backstory: Optional[str] = None,
         instructions: Optional[str] = None,
         # LLM configuration
-        llm: Optional[Union[str, Any]] = None,
+        llm: Optional[Union[str, Any]] = None,  # Can be string, dict, or LLMConfig object
         model: Optional[Union[str, Any]] = None,  # Alias for llm=
         base_url: Optional[str] = None,  # Kept separate (connection/auth)
         api_key: Optional[str] = None,  # Kept separate (connection/auth)
@@ -556,7 +589,8 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
         tool_config: Optional[Union[bool, 'ToolConfig']] = None,  # Tool execution configuration (timeout, retry, parallel)
         learn: Optional[Union[bool, str, Dict[str, Any], 'LearnConfig']] = None,  # Continuous learning (peer to memory)
         backend: Optional[Any] = None,  # External managed agent backend (e.g., ManagedAgentIntegration)
-        cli_backend: Optional[Union[str, Any]] = None,  # CLI backend for delegating turns (e.g., "claude-code")
+        cli_backend: Optional[Union[str, Any]] = None,  # CLI backend for delegating turns (e.g., "claude-code") - DEPRECATED
+        runtime: Optional[Union[str, Dict[str, Any], 'AgentRuntimeConfig']] = None,  # Model-scoped runtime configuration
         interrupt_controller: Optional['InterruptController'] = None,  # G2: Cooperative cancellation
         tool_search: Optional[Union[bool, str, Dict[str, Any], 'ToolSearchConfig']] = False,  # Progressive tool disclosure
         message_steering: Optional[Union[bool, 'MessageSteeringProtocol']] = False,  # Real-time message steering during execution
@@ -570,9 +604,10 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
             goal: Primary objective the agent aims to achieve.
             backstory: Background context shaping personality and decisions.
             instructions: Direct instructions (overrides role/goal/backstory). Recommended for simple agents.
-            llm: Model name string ("gpt-4o", "anthropic/claude-3-sonnet") or LLM object.
+            llm: Model name string ("gpt-4o", "anthropic/claude-3-sonnet"), LLMConfig object, or custom LLM.
+                Can accept LLMConfig(model="gpt-4o", fallback_models=["claude-3-5-sonnet", "gpt-4o-mini"]).
                 Defaults to OPENAI_MODEL_NAME env var or "gpt-4o-mini".
-            model: Alias for llm parameter.
+            model: Alias for llm parameter. Also accepts LLMConfig objects.
             base_url: Custom LLM endpoint URL (e.g., for Ollama). Kept separate for auth.
             api_key: API key for LLM provider. Kept separate for auth.
             tools: List of tools, functions, callables, or MCP instances.
@@ -657,13 +692,15 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
                 - None: Use local execution (default)
                 When provided, agent can delegate execution to managed infrastructure
                 for long-running tasks or when local resources are constrained.
-            cli_backend: CLI backend for delegating full turns to external CLI tools. Accepts:
-                - str: Backend ID ("claude-code", "codex-cli", "gemini-cli")
-                - CliBackendProtocol: Custom CLI backend instance
-                - None: Use standard LLM execution (default)
-                When provided, agent delegates entire conversation turns to the CLI tool
-                instead of using the built-in LLM. Enables session continuity and 
-                tool integration through external AI coding assistants.
+            cli_backend: **DEPRECATED** CLI backend for delegating full turns. Use 'runtime' parameter instead.
+                This parameter is deprecated in favor of model-scoped runtime configuration.
+                Will emit deprecation warning when used.
+            runtime: Model-scoped runtime configuration for turn delegation. Accepts:
+                - str: Runtime ID ("claude-code", "praisonai") 
+                - Dict[str, Any]: Runtime config {"runtime": "claude-code", "config_overrides": {...}}
+                - AgentRuntimeConfig: Pre-configured runtime instance
+                - None: Use model-based resolution (default)
+                Enables turn delegation with per-model runtime selection and fail-closed behavior.
             tool_search: Progressive tool disclosure configuration. Accepts:
                 - bool: False=disabled (default), True=auto mode
                 - str: Mode ("auto", "on", "off")  
@@ -718,16 +755,16 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
         # Precedence: Explicit params > Config file > Built-in defaults
         # Only applies when param is None (not explicitly set)
         # ============================================================
-        from ..config.loader import apply_config_defaults, get_default
+        from ..config.loader import apply_config_defaults, get_default, _defaults_has_any_values
         
         # Apply config defaults for LLM if not explicitly set
-        if llm is None and model is None:
+        if llm is None and model is None and _defaults_has_any_values():
             config_model = get_default("model")
             if config_model:
                 llm = config_model
         
         # Apply config defaults for base_url if not explicitly set
-        if base_url is None:
+        if base_url is None and _defaults_has_any_values():
             config_base_url = get_default("base_url")
             if config_base_url:
                 base_url = config_base_url
@@ -989,7 +1026,7 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
             _on_budget_exceeded = 'stop'
             # Default to False when no ExecutionConfig provided
             parallel_tool_calls = False
-            # (already set from parameter, no need to override)
+            _exec_config = ExecutionConfig()  # Default config for non-execution case
         
         # ─────────────────────────────────────────────────────────────────────
         # Resolve TEMPLATES param - FAST PATH
@@ -1392,7 +1429,7 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
             # String could be LLM prompt - passthrough for later processing
             _guardrails_config = guardrails
         else:
-            from .._resolver_helpers import resolve_guardrails as _resolve_guardrails
+            from ..config.param_resolver import resolve_guardrails as _resolve_guardrails
             _guardrails_config = _resolve_guardrails(
                 value=guardrails,
                 config_class=GuardrailConfig,
@@ -1478,10 +1515,8 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
         # Initialize autonomy features (agent-centric escalation/doom-loop)
         self._init_autonomy(autonomy, verification_hooks=verification_hooks)
         
-        # Initialize loop guard for all execution modes (not just autonomous)
-        # Provides idempotency-aware guardrails for tool execution loops
-        from ..escalation.loop_guard import LoopGuard, LoopGuardConfig
-        self._loop_guard = LoopGuard(LoopGuardConfig(enabled=True))
+        # Initialize loop guard lazily on first tool execution (zero init overhead)
+        self._loop_guard = None
 
         # If instructions are provided, use them to set role, goal, and backstory
         if instructions:
@@ -1527,6 +1562,8 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
         self.interrupt_controller = interrupt_controller
         # Check for model name in environment variable if not provided
         self._using_custom_llm = False
+        self._llm_instance = None
+        self._llm_init_params = None
         # Flag to track if final result has been displayed to prevent duplicates
         self._final_display_shown = False
         
@@ -1541,7 +1578,22 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
         
         # Handle llm= deprecation: model= is the preferred parameter name
         # llm= still works but shows deprecation warning
-        if llm is not None and model is None:
+        fallback_models = None  # Initialize for internal use
+        
+        # Check if llm is an LLMConfig object
+        from ..config import LLMConfig
+        if isinstance(llm, LLMConfig):
+            # Extract values from LLMConfig
+            llm_config_obj = llm
+            llm = llm_config_obj.model
+            fallback_models = llm_config_obj.fallback_models
+            if base_url is None:
+                base_url = llm_config_obj.base_url
+            if api_key is None:
+                api_key = llm_config_obj.api_key
+            if auth is None:
+                auth = llm_config_obj.auth
+        elif llm is not None and model is None:
             warn_deprecated_param(
                 "llm",
                 since="1.0.0",
@@ -1549,9 +1601,23 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
                 alternative="use 'model' instead. Example: Agent(model='gpt-4o-mini')",
                 stacklevel=3
             )
+        
         # model= is the preferred parameter (no warning)
         if model is not None:
-            llm = model  # model= takes precedence
+            # Check if model is an LLMConfig object
+            if isinstance(model, LLMConfig):
+                # Extract values from LLMConfig (same as llm handling)
+                llm = model.model
+                if fallback_models is None:
+                    fallback_models = model.fallback_models
+                if base_url is None:
+                    base_url = model.base_url
+                if api_key is None:
+                    api_key = model.api_key
+                if auth is None:
+                    auth = model.auth
+            else:
+                llm = model  # model= takes precedence
         
         # Store rate limiter (optional, zero overhead when None)
         self._rate_limiter = rate_limiter
@@ -1565,102 +1631,74 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
         self.base_url = base_url
         self.api_key = api_key
 
-        # If base_url is provided, always create a custom LLM instance
+        # If base_url is provided, defer custom LLM instance creation
         if base_url:
-            try:
-                from ..llm.llm import LLM
-                # Handle different llm parameter types with base_url
-                if isinstance(llm, dict):
-                    # Merge base_url and api_key into the dict
-                    llm_config = llm.copy()
-                    llm_config['base_url'] = base_url
-                    if api_key:
-                        llm_config['api_key'] = api_key
-                    if auth:
-                        llm_config['auth'] = auth
-                    llm_config['metrics'] = metrics
-                    llm_config['max_iter'] = max_iter
-                    self.llm_instance = LLM(**llm_config)
-                    self.llm = llm.get('model', Agent._get_default_model())
-                else:
-                    # Create LLM with model string and base_url (cached for performance)
-                    model_name = llm or Agent._get_default_model()
-                    self.llm_instance = LLM(
-                        model=model_name,
-                        base_url=base_url,
-                        api_key=api_key,
-                        auth=auth,
-                        metrics=metrics,
-                        max_iter=max_iter,
-                        web_search=web_search,
-                        web_fetch=web_fetch,
-                        prompt_caching=prompt_caching,
-                        claude_memory=claude_memory
-                    )
-                    self.llm = model_name
-                self._using_custom_llm = True
-            except ImportError as e:
-                raise ImportError(
-                    "LLM features requested but dependencies not installed. "
-                    "Please install with: pip install \"praisonaiagents[llm]\""
-                ) from e
+            if isinstance(llm, dict):
+                llm_config = llm.copy()
+                llm_config['base_url'] = base_url
+                if api_key:
+                    llm_config['api_key'] = api_key
+                if auth:
+                    llm_config['auth'] = auth
+                llm_config['metrics'] = metrics
+                llm_config['max_iter'] = max_iter
+                self._llm_init_params = llm_config
+                self.llm = llm.get('model', Agent._get_default_model())
+            else:
+                model_name = llm or Agent._get_default_model()
+                self._llm_init_params = {
+                    'model': model_name,
+                    'base_url': base_url,
+                    'api_key': api_key,
+                    'auth': auth,
+                    'metrics': metrics,
+                    'max_iter': max_iter,
+                    'web_search': web_search,
+                    'web_fetch': web_fetch,
+                    'prompt_caching': prompt_caching,
+                    'claude_memory': claude_memory,
+                }
+                self.llm = model_name
+            self._using_custom_llm = True
         # If the user passes a dictionary (for advanced configuration)
         elif isinstance(llm, dict) and "model" in llm:
-            try:
-                from ..llm.llm import LLM
-                # Add api_key if provided and not in dict
-                if api_key and 'api_key' not in llm:
-                    llm = llm.copy()
-                    llm['api_key'] = api_key
-                # Add auth if provided and not in dict
-                if auth and 'auth' not in llm:
-                    llm = llm.copy()
-                    llm['auth'] = auth
-                # Add metrics parameter
-                llm = llm.copy()
-                llm['metrics'] = metrics
-                llm['max_iter'] = max_iter
-                self.llm_instance = LLM(**llm)  # Pass all dict items as kwargs
-                self._using_custom_llm = True
-                self.llm = llm.get('model', Agent._get_default_model())
-            except ImportError as e:
-                raise ImportError(
-                    "LLM features requested but dependencies not installed. "
-                    "Please install with: pip install \"praisonaiagents[llm]\""
-                ) from e
-        # If the user passes a string with a slash (provider/model)
+            llm_config = llm.copy()
+            if api_key and 'api_key' not in llm_config:
+                llm_config['api_key'] = api_key
+            if auth and 'auth' not in llm_config:
+                llm_config['auth'] = auth
+            llm_config['metrics'] = metrics
+            llm_config['max_iter'] = max_iter
+            self._llm_init_params = llm_config
+            self._using_custom_llm = True
+            self.llm = llm_config.get('model', Agent._get_default_model())
+        # If the user passes a string with a slash (provider/model) — defer LiteLLM import
         elif isinstance(llm, str) and "/" in llm:
-            try:
-                from ..llm.llm import LLM
-                # Pass the entire string so LiteLLM can parse provider/model
-                llm_params = {'model': llm}
-                if api_key:
-                    llm_params['api_key'] = api_key
-                if auth:
-                    llm_params['auth'] = auth
-                llm_params['metrics'] = metrics
-                llm_params['web_search'] = web_search
-                llm_params['web_fetch'] = web_fetch
-                llm_params['prompt_caching'] = prompt_caching
-                llm_params['claude_memory'] = claude_memory
-                llm_params['max_iter'] = max_iter
-                self.llm_instance = LLM(**llm_params)
-                self._using_custom_llm = True
-                self.llm = llm
-                
-                # Ensure tools are properly accessible when using custom LLM
-                if tools:
-                    logging.debug(f"Tools passed to Agent with custom LLM: {tools}")
-                    # Store the tools for later use
-                    self.tools = tools
-            except ImportError as e:
-                raise ImportError(
-                    "LLM features requested but dependencies not installed. "
-                    "Please install with: pip install \"praisonaiagents[llm]\""
-                ) from e
+            llm_params = {'model': llm}
+            if api_key:
+                llm_params['api_key'] = api_key
+            if auth:
+                llm_params['auth'] = auth
+            llm_params['metrics'] = metrics
+            llm_params['web_search'] = web_search
+            llm_params['web_fetch'] = web_fetch
+            llm_params['prompt_caching'] = prompt_caching
+            llm_params['claude_memory'] = claude_memory
+            llm_params['max_iter'] = max_iter
+            self._llm_init_params = llm_params
+            self._using_custom_llm = True
+            self.llm = llm
+            
+            if tools:
+                logging.debug(f"Tools passed to Agent with custom LLM: {tools}")
+                self.tools = tools
         # Otherwise, fall back to OpenAI environment/name (cached for performance)
         else:
             self.llm = llm or Agent._get_default_model()
+        
+        # Store fallback models for resilience (defensive copy to avoid external mutations)
+        self.fallback_models = list(fallback_models) if fallback_models else []
+        
         # Handle tools parameter - ensure it's always a list
         if callable(tools):
             # If a single function/callable is passed, wrap it in a list
@@ -1760,6 +1798,8 @@ class Agent(SteeringMixin, SandboxMixin, UnifiedExecutionMixin, ToolExecutionMix
         self.allow_code_execution = allow_code_execution
         self.max_retry_limit = max_retry_limit
         self.code_execution_mode = code_execution_mode
+        # Store execution config for guardrail retry backoff
+        self._execution_config = _exec_config
         self.embedder_config = embedder_config
         self.knowledge = knowledge
         self.use_system_prompt = use_system_prompt
@@ -1855,20 +1895,24 @@ Your Goal: {self.goal}
                 self._approval_backend = None
                 self._approve_all_tools = False
                 self._approval_timeout = 0
+                self._approval_permissions = None
             else:
                 # Unknown string — treat as no approval
                 self._approval_backend = None
                 self._approve_all_tools = False
                 self._approval_timeout = 0
+                self._approval_permissions = None
         elif approval is True:
             from ..approval.backends import AutoApproveBackend
             self._approval_backend = AutoApproveBackend()
             self._approve_all_tools = False
             self._approval_timeout = 0  # 0 = use backend default
+            self._approval_permissions = None
         elif approval is False or approval is None:
             self._approval_backend = None
             self._approve_all_tools = False
             self._approval_timeout = 0
+            self._approval_permissions = None
             # No explicit approval kwarg — honour PRAISONAI_TOOL_SAFETY.
             # Default preset "default" blocks only destructive ops
             # (delete_*, execute_command, execute_code, kill_process, move/copy)
@@ -1898,17 +1942,21 @@ Your Goal: {self.goal}
             self._approval_backend = approval.backend
             self._approve_all_tools = approval.all_tools
             self._approval_timeout = approval.timeout  # None = indefinite, 0 = backend default
+            # Store permissions if provided (for CI-safe declarative policies)
+            self._approval_permissions = getattr(approval, 'permissions', None)
         elif isinstance(approval, dict):
             # Dict config: convert to ApprovalConfig
             approval_config = ApprovalConfig(**approval)
             self._approval_backend = approval_config.backend
             self._approve_all_tools = approval_config.all_tools
             self._approval_timeout = approval_config.timeout
+            self._approval_permissions = getattr(approval_config, 'permissions', None)
         else:
             # Plain backend object — dangerous tools only, backend default timeout
             self._approval_backend = approval
             self._approve_all_tools = False
             self._approval_timeout = 0
+            self._approval_permissions = None
         
         # Per-agent autonomy→approval bridge (G-BRIDGE-1 fix)
         # If autonomy level is full_auto and no explicit approval was set,
@@ -2032,10 +2080,26 @@ Your Goal: {self.goal}
         # Backend - external managed agent backend for hybrid execution
         self.backend = backend
         
-        # CLI Backend - external CLI backend for delegating full turns
+        # CLI Backend - external CLI backend for delegating full turns (DEPRECATED)
         self._cli_backend = None
         if cli_backend is not None:
+            from ..utils.deprecation import warn_deprecated_param
+            warn_deprecated_param(
+                "cli_backend",
+                since="1.0.0",
+                removal="2.0.0",
+                alternative="use 'runtime=' instead for model-scoped runtime configuration. For migration assistance, run: praisonai doctor fix --execute",
+                stacklevel=3
+            )
             self._cli_backend = self._resolve_cli_backend(cli_backend)
+        
+        # Runtime Configuration - model-scoped runtime selection
+        self._runtime_config = None
+        if runtime is not None:
+            self._runtime_config = self._resolve_runtime_config(runtime)
+        
+        # Runtime resolver for per-turn resolution
+        self._runtime_resolver = None  # Will be initialized on first use
 
         # Telemetry - lazy initialized via property for performance
         self.__telemetry = None
@@ -2103,6 +2167,7 @@ Your Goal: {self.goal}
             
             # LLM configuration 
             'llm': self.llm,
+            'fallback_models': list(self.fallback_models) if hasattr(self, 'fallback_models') and self.fallback_models else None,
             'base_url': getattr(self, 'base_url', None),
             'api_key': getattr(self, 'api_key', None),
             'auth': getattr(self, 'auth', None),
@@ -4954,6 +5019,22 @@ Answer:"""
             retry_count += 1
             logging.warning(f"Agent {self.name}: Guardrail validation failed (retry {retry_count}/{self.max_guardrail_retries}): {error}")
             
+            # Add exponential backoff delay to avoid hammering the LLM
+            execution_config = getattr(self, '_execution_config', None)
+            if execution_config is not None:
+                total_delay = BackoffPolicy.delay(
+                    retry_count,
+                    execution_config.retry_initial_delay,
+                    execution_config.retry_backoff_factor,
+                    execution_config.retry_jitter
+                )
+            else:
+                # Fall back to simple backoff if no execution config
+                total_delay = 1.0 * (2.0 ** (retry_count - 1))
+            
+            logging.info(f"Agent {self.name}: Waiting {total_delay:.2f}s before guardrail retry")
+            time.sleep(total_delay)
+            
             # Regenerate response for retry
             try:
                 retry_prompt = f"{prompt}\n\nNote: Previous response failed validation due to: {error}. Please provide an improved response."
@@ -4991,6 +5072,22 @@ Answer:"""
             retry_count += 1
             logging.warning(f"Agent {self.name}: Guardrail validation failed (retry {retry_count}/{self.max_guardrail_retries}): {error}")
             
+            # Add exponential backoff delay to avoid hammering the LLM
+            execution_config = getattr(self, '_execution_config', None)
+            if execution_config is not None:
+                total_delay = BackoffPolicy.delay(
+                    retry_count,
+                    execution_config.retry_initial_delay,
+                    execution_config.retry_backoff_factor,
+                    execution_config.retry_jitter
+                )
+            else:
+                # Fall back to simple backoff if no execution config
+                total_delay = 1.0 * (2.0 ** (retry_count - 1))
+            
+            logging.info(f"Agent {self.name}: Waiting {total_delay:.2f}s before guardrail retry")
+            await asyncio.sleep(total_delay)
+            
             # Regenerate response for retry (async version)
             try:
                 retry_prompt = f"{prompt}\n\nNote: Previous response failed validation due to: {error}. Please provide an improved response."
@@ -5079,17 +5176,163 @@ Answer:"""
         
         raise TypeError(f"Invalid cli_backend type: {type(cli_backend)}. Expected CliBackendProtocol instance or factory callable.")
     
-    async def _chat_via_cli_backend(self, prompt: str, **kwargs) -> Optional[str]:
+    def _resolve_runtime_config(self, runtime):
+        """Resolve runtime parameter to AgentRuntimeConfig instance.
+        
+        Args:
+            runtime: Runtime configuration parameter
+            
+        Returns:
+            AgentRuntimeConfig instance
+            
+        Raises:
+            TypeError: If runtime type is invalid
+            ValueError: If runtime configuration is invalid
+        """
+        try:
+            from ..runtime import AgentRuntimeConfig
+        except ImportError:
+            raise ImportError(
+                "Runtime configuration features requested but not available. "
+                "This should not happen in a properly installed praisonaiagents package."
+            )
+        
+        # If already an AgentRuntimeConfig instance, return as-is
+        if isinstance(runtime, AgentRuntimeConfig):
+            return runtime
+        
+        # If string, create config with runtime ID
+        if isinstance(runtime, str):
+            return AgentRuntimeConfig.from_runtime_id(runtime)
+        
+        # If dictionary, create config from dict
+        if isinstance(runtime, dict):
+            return AgentRuntimeConfig.from_dict(runtime)
+        
+        raise TypeError(f"Invalid runtime type: {type(runtime)}. Expected str, dict, or AgentRuntimeConfig instance.")
+    
+    def _get_runtime_resolver(self):
+        """Get or create runtime resolver instance (lazy initialization)."""
+        if self._runtime_resolver is None:
+            try:
+                from ..runtime.resolver import RuntimeResolver
+                self._runtime_resolver = RuntimeResolver()
+            except ImportError:
+                raise ImportError(
+                    "Runtime resolution features requested but not available. "
+                    "This should not happen in a properly installed praisonaiagents package."
+                )
+        return self._runtime_resolver
+    
+    async def _resolve_turn_runtime(self):
+        """Resolve runtime for current turn based on model and configuration.
+        
+        Returns:
+            Runtime instance if resolution succeeds, None otherwise
+        """
+        try:
+            from ..runtime.resolver import RuntimeResolutionContext
+        except ImportError:
+            return None
+        
+        # Create resolution context with current model and provider info
+        context = RuntimeResolutionContext(
+            model_name=getattr(self, 'llm', None),
+            provider_name=self._extract_provider_from_model(getattr(self, 'llm', None)),
+            agent_config={'agent_id': getattr(self, 'agent_id', None)}
+        )
+        
+        try:
+            resolver = self._get_runtime_resolver()
+            result = resolver.resolve_runtime_instance(
+                context=context,
+                model_runtime_configs=getattr(self, '_model_runtime_configs', None),
+                provider_runtime_configs=getattr(self, '_provider_runtime_configs', None),
+                legacy_cli_backend=getattr(self, '_cli_backend', None)
+            )
+            
+            if result and result.runtime:
+                return result.runtime
+            
+        except (ValueError, RuntimeError) as e:
+            # Only fall back for registry not initialized, propagate configuration errors
+            if "not initialized" in str(e).lower():
+                # Registry not initialized - this is expected in SDK-only mode
+                pass
+            else:
+                # Configuration error (unknown runtime ID etc.) - fail closed
+                raise RuntimeError(
+                    f"Runtime resolution failed for agent={getattr(self, 'display_name', 'unknown')!r}, "
+                    f"model={getattr(self, 'llm', None)!r}: {e}. "
+                    "Fix the runtime ID/configuration or remove the runtime override."
+                ) from e
+        
+        return None
+    
+    def _extract_provider_from_model(self, model_name):
+        """Extract provider name from model name.
+        
+        Args:
+            model_name: Model name string (e.g., "anthropic/claude-3-sonnet")
+            
+        Returns:
+            Provider name if detectable, None otherwise
+        """
+        if not model_name or not isinstance(model_name, str):
+            return None
+        
+        # Common provider patterns
+        if '/' in model_name:
+            return model_name.split('/')[0]
+        
+        # Provider inference based on model name patterns
+        model_lower = model_name.lower()
+        if 'claude' in model_lower or 'anthropic' in model_lower:
+            return 'anthropic'
+        elif 'gpt' in model_lower or 'openai' in model_lower:
+            return 'openai'
+        elif 'gemini' in model_lower or 'google' in model_lower:
+            return 'google'
+        elif 'llama' in model_lower:
+            return 'meta'
+        
+        return None
+    
+    async def _chat_via_runtime(self, runtime_instance, prompt: str, **kwargs) -> Optional[str]:
+        """Chat implementation using resolved runtime instance.
+        
+        This is similar to _chat_via_cli_backend but uses the resolved runtime
+        from the model-scoped runtime resolution system.
+        
+        Args:
+            runtime_instance: Resolved runtime instance (CliBackendProtocol)
+            prompt: User prompt
+            **kwargs: Additional chat parameters
+            
+        Returns:
+            Runtime response content
+        """
+        if not runtime_instance:
+            raise RuntimeError("Runtime instance is None")
+        
+        # Delegate to CLI backend implementation with runtime instance
+        # Pass runtime instance directly to avoid mutating shared state
+        return await self._chat_via_cli_backend(prompt=prompt, cli_backend=runtime_instance, **kwargs)
+    
+    async def _chat_via_cli_backend(self, prompt: str, cli_backend: Any = None, **kwargs) -> Optional[str]:
         """Chat implementation using CLI backend delegation.
         
         Args:
             prompt: User prompt
+            cli_backend: Optional specific backend instance to use (for runtime delegation)
             **kwargs: Additional chat parameters (passed through as metadata)
             
         Returns:
             CLI backend response content
         """
-        if not self._cli_backend:
+        # Use provided backend or fall back to instance backend
+        backend = cli_backend or self._cli_backend
+        if not backend:
             raise RuntimeError("CLI backend not configured")
         
         try:
@@ -5128,7 +5371,7 @@ Answer:"""
                     images = None
             
             # Execute CLI backend
-            result = await self._cli_backend.execute(
+            result = await backend.execute(
                 prompt=prompt,
                 session=session_binding,
                 images=images,

src/praisonai-agents/praisonaiagents/agent/chat_mixin.py

@@ -87,12 +87,41 @@ Your Goal: {self.goal}"""
         
         # Add memory context if memory is enabled
         if self._memory_instance:
-            memory_context = self.get_memory_context()
-            if memory_context:
-                system_prompt += f"\n\n## Memory (Information you remember about the user)\n{memory_context}"
-                # Display memory info to user if verbose
-                if self.verbose:
-                    self._display_memory_info()
+            # Use cache-optimized context if model supports prompt caching and cache boundary is requested
+            use_cache_boundaries = hasattr(self, '_model_supports_prompt_caching') and self._model_supports_prompt_caching()
+            if use_cache_boundaries and hasattr(self._memory_instance, 'build_cache_optimized_context'):
+                try:
+                    # Get the goal/task description for context building
+                    task_desc = getattr(self, 'goal', '') or 'general assistance'
+                    cache_result = self._memory_instance.build_cache_optimized_context(
+                        task_descr=task_desc,
+                        include_cache_boundary=True
+                    )
+                    memory_context = cache_result.get('stable_prefix', '')
+                    cache_boundary = cache_result.get('cache_boundary', '')
+                    if memory_context:
+                        system_prompt += f"\n\n## Memory (Information you remember about the user)\n{memory_context}"
+                        if cache_boundary:
+                            system_prompt += cache_boundary
+                        # Display memory info to user if verbose
+                        if self.verbose:
+                            self._display_memory_info()
+                except (AttributeError, KeyError, TypeError) as e:
+                    # Fall back to standard memory context if cache optimization fails
+                    memory_context = self.get_memory_context()
+                    if memory_context:
+                        system_prompt += f"\n\n## Memory (Information you remember about the user)\n{memory_context}"
+                        # Display memory info to user if verbose
+                        if self.verbose:
+                            self._display_memory_info()
+            else:
+                # Standard memory context without cache boundaries
+                memory_context = self.get_memory_context()
+                if memory_context:
+                    system_prompt += f"\n\n## Memory (Information you remember about the user)\n{memory_context}"
+                    # Display memory info to user if verbose
+                    if self.verbose:
+                        self._display_memory_info()
             
             # Add learn context if learn is enabled (auto-inject when memory="learn")
             learn_context = self.get_learn_context()
@@ -106,6 +135,60 @@ Your Goal: {self.goal}"""
                 system_prompt += f"\n\n## Available Skills\n{skills_prompt}"
                 system_prompt += "\n\nWhen a skill is relevant to the task, read its SKILL.md file to get detailed instructions. If the skill has scripts in its scripts/ directory, you can execute them using the execute_code or run_script tool."
         
+        # Cache the base system prompt BEFORE adding session context
+        # Session context is per-turn and should not be cached
+        if cache_key:
+            self._cache_put(self._system_prompt_cache, cache_key, system_prompt)
+        
+        # Add session context (platform awareness) if available - AFTER caching
+        try:
+            from ..session.context import get_session_context
+            session_ctx = get_session_context()
+            
+            # Format session context into prompt if origin or targets are present
+            if session_ctx.origin or session_ctx.reachable_targets:
+                context_parts = []
+                
+                # Add origin information
+                if session_ctx.origin:
+                    origin = session_ctx.origin
+                    origin_str = f"You are replying on {origin.platform}"
+                    if origin.chat_type and origin.chat_type != "unknown":
+                        origin_str += f" ({origin.chat_type}"
+                        if origin.display_name:
+                            origin_str += f' "{origin.display_name}"'
+                        origin_str += ")"
+                    if origin.thread_id:
+                        origin_str += f" in thread {origin.thread_id}"
+                    context_parts.append(origin_str + ".")
+                
+                # Add reachable targets
+                if session_ctx.reachable_targets:
+                    target_descriptions = []
+                    for target in session_ctx.reachable_targets:
+                        desc = f"{target.name}"
+                        if target.kind == "home":
+                            desc += f" ({target.platform}, home channel)"
+                        elif target.kind == "alias":
+                            desc += f" ({target.platform}:{target.channel_id})"
+                        target_descriptions.append(desc)
+                    
+                    if target_descriptions:
+                        context_parts.append(
+                            f"Reachable delivery targets: {', '.join(target_descriptions)}."
+                        )
+                
+                if context_parts:
+                    system_prompt += "\n\n## Session Context\n" + "\n".join(context_parts)
+        except ImportError:
+            # Session context module not available, continue without it
+            pass
+        except Exception as e:
+            # Log unexpected errors but continue
+            import logging
+            logging.debug(f"Session context injection failed: {e}", exc_info=True)
+            pass
+        
         # Add tool usage instructions if tools are available
         # Use provided tools or fall back to self.tools
         tools_to_use = tools if tools is not None else self.tools
@@ -145,10 +228,7 @@ Your Goal: {self.goal}"""
         except ImportError:
             pass  # Trust module not available, skip security instructions
         
-        # Cache the generated system prompt (only if cache_key is set, i.e., memory not enabled)
-        # Use LRU eviction to prevent unbounded growth
-        if cache_key:
-            self._cache_put(self._system_prompt_cache, cache_key, system_prompt)
+        # Note: Caching is done BEFORE session context injection to avoid cross-user leakage
         return system_prompt
 
     def _build_response_format(self, schema_model):
@@ -452,6 +532,17 @@ Your Goal: {self.goal}"""
             else:
                 cleaned_tools.append(tool)
         
+        # Sort tools deterministically for cache stability
+        def sort_formatted_tools(tools_list):
+            """Sort already-formatted tool schemas by function name."""
+            def sort_key(tool):
+                if isinstance(tool, dict) and tool.get('type') == 'function':
+                    return str(tool.get('function', {}).get('name') or '')
+                return ''
+            return sorted(tools_list, key=sort_key)
+        
+        cleaned_tools = sort_formatted_tools(cleaned_tools)
+        
         # Cache the formatted tools with LRU eviction, including tool search metadata
         self._cache_put(
             self._formatted_tools_cache,
@@ -869,7 +960,22 @@ Your Goal: {self.goal}"""
         
         return CompactionRoute.TRUNCATE_TOOLS, truncated_msgs
 
-    def _chat_completion(self, messages, temperature=1.0, tools=None, stream=None, reasoning_steps=False, task_name=None, task_description=None, task_id=None, response_format=None, _retry_depth=0):
+    def _get_next_fallback_model(self, fallback_index):
+        """Get the next fallback model from the chain, if available.
+        
+        Args:
+            fallback_index: Current index in the fallback chain (0-based)
+            
+        Returns:
+            str or None: Next model name if available, None otherwise
+        """
+        if not hasattr(self, 'fallback_models') or not self.fallback_models:
+            return None
+        if fallback_index >= len(self.fallback_models):
+            return None
+        return self.fallback_models[fallback_index]
+
+    def _chat_completion(self, messages, temperature=1.0, tools=None, stream=None, reasoning_steps=False, task_name=None, task_description=None, task_id=None, response_format=None, _retry_depth=0, _fallback_index=0):
         start_time = time.time()
 
         # --- Proactive Context Budget Management (default-on) ---
@@ -1112,7 +1218,8 @@ Your Goal: {self.goal}"""
                         return self._chat_completion(
                             truncated_messages, temperature, tools, stream, 
                             reasoning_steps, task_name, task_description, task_id, response_format, 
-                            _retry_depth=_retry_depth + 1
+                            _retry_depth=_retry_depth + 1,
+                            _fallback_index=_fallback_index
                         )
                 except Exception as compression_error:
                     logging.error(f"[{self.name}] Context compression failed: {compression_error}")
@@ -1123,9 +1230,34 @@ Your Goal: {self.goal}"""
                 # Don't retry without actual credential rotation
                 
             elif classification.should_fallback_model:
-                # TODO: Implement model fallback when available
-                logging.warning(f"[{self.name}] {classification.user_message} (model fallback not yet implemented)")
-                # Don't retry without actual model fallback
+                # Try next model in the fallback chain
+                next_model = self._get_next_fallback_model(_fallback_index)
+                if next_model:
+                    current_model = self.llm if isinstance(self.llm, str) else str(self.llm)
+                    logging.info(f"[{self.name}] {current_model} unavailable — falling back to {next_model}")
+                    
+                    # Apply backoff if suggested
+                    if classification.backoff_seconds and classification.backoff_seconds > 0:
+                        time.sleep(classification.backoff_seconds)
+                    
+                    # Temporarily override the model and clear dispatcher cache for this call
+                    original_llm = self.llm
+                    original_dispatcher = getattr(self, '_unified_dispatcher', None)
+                    try:
+                        self.llm = next_model
+                        self._unified_dispatcher = None  # Force recreation with new model
+                        return self._chat_completion(
+                            messages, temperature, tools, stream,
+                            reasoning_steps, task_name, task_description, task_id, response_format,
+                            _retry_depth=_retry_depth + 1,
+                            _fallback_index=_fallback_index + 1
+                        )
+                    finally:
+                        self.llm = original_llm
+                        self._unified_dispatcher = original_dispatcher  # Restore original dispatcher
+                else:
+                    logging.warning(f"[{self.name}] {classification.user_message} (no more fallback models available)")
+                    # Continue to error handling without retry
                 
             elif classification.is_retryable and classification.backoff_seconds > 0:
                 if _retry_depth < 2:  # Limit retry attempts
@@ -1134,7 +1266,8 @@ Your Goal: {self.goal}"""
                     return self._chat_completion(
                         messages, temperature, tools, stream, 
                         reasoning_steps, task_name, task_description, task_id, response_format, 
-                        _retry_depth=_retry_depth + 1
+                        _retry_depth=_retry_depth + 1,
+                        _fallback_index=_fallback_index
                     )
             
             # Include remediation hints for unimplemented recovery actions
@@ -1142,7 +1275,10 @@ Your Goal: {self.goal}"""
             if classification.should_rotate_credential:
                 user_message += " Credential rotation is not yet implemented."
             if classification.should_fallback_model:
-                user_message += " Model fallback is not yet implemented."
+                if not self.fallback_models:
+                    user_message += " No fallback models configured."
+                elif _fallback_index >= len(self.fallback_models):
+                    user_message += f" All {len(self.fallback_models)} fallback models exhausted."
             
             # Create LLMError with classification context
             error = LLMError(
@@ -1180,7 +1316,8 @@ Your Goal: {self.goal}"""
         response_format=None,
         stream_callback=None,
         emit_events=True,
-        _retry_depth=0
+        _retry_depth=0,
+        _fallback_index=0
     ):
         """
         Handle LLM errors in async context using structured classification.
@@ -1245,7 +1382,7 @@ Your Goal: {self.goal}"""
                         return await self._handle_async_llm_error(
                             retry_error, truncated_messages, temperature, tools, stream,
                             reasoning_steps, task_name, task_description, task_id, response_format,
-                            stream_callback, emit_events, _retry_depth + 1
+                            stream_callback, emit_events, _retry_depth + 1, _fallback_index
                         )
             except Exception as compression_error:
                 logging.error(f"[{self.name}] Context compression failed: {compression_error}")
@@ -1256,9 +1393,39 @@ Your Goal: {self.goal}"""
             # Don't retry without actual credential rotation
             
         elif classification.should_fallback_model:
-            # TODO: Implement model fallback when available
-            logging.warning(f"[{self.name}] {classification.user_message} (model fallback not yet implemented)")
-            # Don't retry without actual model fallback
+            # Try next model in the fallback chain
+            next_model = self._get_next_fallback_model(_fallback_index)
+            if next_model:
+                current_model = self.llm if isinstance(self.llm, str) else str(self.llm)
+                logging.info(f"[{self.name}] {current_model} unavailable — falling back to {next_model}")
+                
+                # Apply backoff if suggested
+                if classification.backoff_seconds and classification.backoff_seconds > 0:
+                    await asyncio.sleep(classification.backoff_seconds)
+                
+                # Temporarily override the model and clear dispatcher cache for this call
+                original_llm = self.llm
+                original_dispatcher = getattr(self, '_unified_dispatcher', None)
+                try:
+                    self.llm = next_model
+                    self._unified_dispatcher = None  # Force recreation with new model
+                    return await self._execute_unified_achat_completion(
+                        messages, temperature, tools, stream,
+                        reasoning_steps, task_name, task_description, task_id, response_format,
+                        stream_callback, emit_events
+                    )
+                except Exception as retry_error:
+                    return await self._handle_async_llm_error(
+                        retry_error, messages, temperature, tools, stream,
+                        reasoning_steps, task_name, task_description, task_id, response_format,
+                        stream_callback, emit_events, _retry_depth + 1, _fallback_index + 1
+                    )
+                finally:
+                    self.llm = original_llm
+                    self._unified_dispatcher = original_dispatcher  # Restore original dispatcher
+            else:
+                logging.warning(f"[{self.name}] {classification.user_message} (no more fallback models available)")
+                # Continue to error handling without retry
             
         elif classification.is_retryable and classification.backoff_seconds > 0:
             if _retry_depth < 2:  # Limit retry attempts
@@ -1275,7 +1442,7 @@ Your Goal: {self.goal}"""
                     return await self._handle_async_llm_error(
                         retry_error, messages, temperature, tools, stream,
                         reasoning_steps, task_name, task_description, task_id, response_format,
-                        stream_callback, emit_events, _retry_depth + 1
+                        stream_callback, emit_events, _retry_depth + 1, _fallback_index
                     )
         
         # Include remediation hints for unimplemented recovery actions
@@ -1283,7 +1450,10 @@ Your Goal: {self.goal}"""
         if classification.should_rotate_credential:
             user_message += " Credential rotation is not yet implemented."
         if classification.should_fallback_model:
-            user_message += " Model fallback is not yet implemented."
+            if not self.fallback_models:
+                user_message += " No fallback models configured."
+            elif _fallback_index >= len(self.fallback_models):
+                user_message += f" All {len(self.fallback_models)} fallback models exhausted."
         
         # Create LLMError with classification context
         error = LLMError(
@@ -1327,7 +1497,7 @@ Your Goal: {self.goal}"""
         """
         Execute unified chat completion using composition instead of runtime class mutation.
         
-        This method provides the same functionality as UnifiedChatMixin but uses
+        This method provides a unified dispatch for all LLM providers using
         composition for safety and maintainability.
         """
         from ..llm import create_llm_dispatcher
@@ -1451,7 +1621,7 @@ Your Goal: {self.goal}"""
         """
         Execute unified async chat completion using composition instead of runtime class mutation.
         
-        This method provides the same functionality as UnifiedChatMixin but uses
+        This method provides a unified async dispatch for all LLM providers using
         composition for safety and maintainability.
         """
         from ..llm import create_llm_dispatcher
@@ -1836,7 +2006,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
             logging.warning(f"Context management error (continuing without): {e}")
             return messages, None
 
-    def _truncate_tool_output(self, tool_name: str, output: str) -> str:
+    def _truncate_tool_output(self, tool_name: str, output: str, tool_call_id: str | None = None) -> str:
         """
         Truncate tool output according to configured budget.
         
@@ -1845,6 +2015,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
         Args:
             tool_name: Name of the tool
             output: Raw tool output
+            tool_call_id: Optional ID for the tool call
             
         Returns:
             Truncated output if over budget, otherwise original
@@ -1853,7 +2024,8 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
             return output
         
         try:
-            return self.context_manager.truncate_tool_output(tool_name, output)
+            run_id = getattr(self, '_run_id', None)
+            return self.context_manager.truncate_tool_output(tool_name, output, tool_call_id, run_id)
         except Exception as e:
             logging.warning(f"Tool output truncation error: {e}")
             return output
@@ -2496,8 +2668,9 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                             # If not satisfactory and not at max reflections, continue with regeneration
                             logging.debug(f"{self.name} reflection count {reflection_count + 1}, continuing reflection process")
                             messages.append({"role": "user", "content": "Now regenerate your response using the reflection you made"})
-                            # For custom LLMs during reflection, always use non-streaming to ensure complete responses
-                            use_stream = self.stream if not self._using_custom_llm else False
+                            # For reflection, always use non-streaming to ensure compatibility with sync adapters
+                            # and to avoid streaming complexity during regeneration process
+                            use_stream = False
                             response = self._chat_completion(messages, temperature=temperature, tools=None, stream=use_stream, task_name=task_name, task_description=task_description, task_id=task_id)
                             content = response.choices[0].message.content
                             response_text = content.strip() if content else ""

src/praisonai-agents/praisonaiagents/agent/memory_mixin.py

@@ -317,8 +317,9 @@ class MemoryMixin:
             return
         
         try:
-            from ..session import get_default_session_store
-            self._session_store = get_default_session_store()
+            if self._session_store is None:
+                from ..session import get_default_session_store
+                self._session_store = get_default_session_store()
             
             # Restore chat history from previous session
             history = self._session_store.get_chat_history(self._session_id)

src/praisonai-agents/praisonaiagents/agent/sandbox_mixin.py

@@ -5,15 +5,10 @@ Provides safe code execution capabilities to agents via the sandbox framework.
 """
 
 import logging
-from typing import Optional, Dict, Any, Union
+from typing import Optional, Dict, Any, Union, TYPE_CHECKING
 
-from ..sandbox import (
-    SandboxConfig, 
-    SandboxManager, 
-    SandboxResult,
-    check_code_safety,
-    format_warnings,
-)
+if TYPE_CHECKING:
+    from ..sandbox import SandboxConfig, SandboxManager, SandboxResult
 
 logger = logging.getLogger(__name__)
 
@@ -26,27 +21,33 @@ class SandboxMixin:
     
     def __init__(self, *args, **kwargs):
         # Extract sandbox config from kwargs
-        self.sandbox_config: Optional[SandboxConfig] = kwargs.pop('sandbox', None)
+        sandbox = kwargs.pop('sandbox', None)
         
         # Convert boolean True to default config
-        if self.sandbox_config is True:
-            self.sandbox_config = SandboxConfig.subprocess()
+        if sandbox is True:
+            from ..sandbox import SandboxConfig
+            sandbox = SandboxConfig.subprocess()
+        elif sandbox is False:
+            sandbox = None
+        
+        self.sandbox_config: Optional['SandboxConfig'] = sandbox
         
         super().__init__(*args, **kwargs)
         
-        self._sandbox_manager: Optional[SandboxManager] = None
+        self._sandbox_manager: Optional['SandboxManager'] = None
     
     @property
     def has_sandbox(self) -> bool:
         """Check if sandbox is configured for this agent."""
         return self.sandbox_config is not None
     
-    def get_sandbox_manager(self) -> Optional[SandboxManager]:
+    def get_sandbox_manager(self) -> Optional['SandboxManager']:
         """Get or create sandbox manager."""
         if not self.has_sandbox:
             return None
         
         if self._sandbox_manager is None:
+            from ..sandbox import SandboxManager
             self._sandbox_manager = SandboxManager(self.sandbox_config)
         
         return self._sandbox_manager
@@ -57,7 +58,7 @@ class SandboxMixin:
         language: str = "python",
         check_security: bool = True,
         **kwargs
-    ) -> SandboxResult:
+    ) -> 'SandboxResult':
         """Execute code safely in configured sandbox.
         
         Args:
@@ -79,6 +80,7 @@ class SandboxMixin:
         
         # Security pre-check if enabled
         if check_security:
+            from ..sandbox import check_code_safety, format_warnings
             warnings = check_code_safety(code, language)
             if warnings:
                 warning_text = format_warnings(warnings)
@@ -97,7 +99,7 @@ class SandboxMixin:
         language: str = "python", 
         check_security: bool = True,
         **kwargs
-    ) -> SandboxResult:
+    ) -> 'SandboxResult':
         """Synchronous wrapper for execute_code.
         
         Args:
@@ -117,7 +119,7 @@ class SandboxMixin:
         command: Union[str, list],
         check_security: bool = True,
         **kwargs
-    ) -> SandboxResult:
+    ) -> 'SandboxResult':
         """Run a shell command in the sandbox.
         
         Args:
@@ -139,6 +141,7 @@ class SandboxMixin:
         
         # Security pre-check
         if check_security:
+            from ..sandbox import check_code_safety, format_warnings
             warnings = check_code_safety(command_str, "bash")
             if warnings:
                 warning_text = format_warnings(warnings)

src/praisonai-agents/praisonaiagents/agent/summarization.py

@@ -7,21 +7,10 @@ Reuses existing telemetry/token_collector for token tracking.
 
 import logging
 from praisonaiagents._logging import get_logger
-from typing import Optional, List, Dict, Any, Callable
-from dataclasses import dataclass, field
+from typing import Optional, List, Dict, Any
 
 logger = get_logger(__name__)
 
-@dataclass
-class SummarizationConfig:
-    """Configuration for auto-summarization."""
-    enabled: bool = False
-    threshold: float = 0.8  # 80% of context window
-    context_window: int = 128000  # Default for GPT-4
-    summary_model: Optional[str] = None  # Use smaller model for summaries
-    preserve_system: bool = True  # Keep system message
-    preserve_recent: int = 2  # Keep last N message pairs
-
 class SummarizationManager:
     """
     Manages automatic conversation summarization.
@@ -179,45 +168,3 @@ SUMMARY:"""
         """Number of times summarization has been performed."""
         return self._summarization_count
 
-async def summarize_conversation(
-    messages: List[Dict[str, str]],
-    llm_call: Callable,
-    manager: SummarizationManager,
-) -> List[Dict[str, str]]:
-    """
-    Summarize conversation using LLM.
-    
-    Args:
-        messages: Current message history
-        llm_call: Async function to call LLM
-        manager: SummarizationManager instance
-        
-    Returns:
-        New message history with summary
-    """
-    prompt = manager.generate_summary_prompt(messages)
-    
-    # Call LLM to generate summary
-    summary = await llm_call(prompt)
-    
-    # Prepare new history
-    new_history = manager.prepare_summarized_history(messages, summary)
-    
-    # Reset token count (will be recalculated on next message)
-    manager.reset_tokens()
-    
-    return new_history
-
-def summarize_conversation_sync(
-    messages: List[Dict[str, str]],
-    llm_call: Callable,
-    manager: SummarizationManager,
-) -> List[Dict[str, str]]:
-    """
-    Synchronous version of summarize_conversation.
-    """
-    prompt = manager.generate_summary_prompt(messages)
-    summary = llm_call(prompt)
-    new_history = manager.prepare_summarized_history(messages, summary)
-    manager.reset_tokens()
-    return new_history

src/praisonai-agents/praisonaiagents/agent/tool_execution.py

@@ -14,6 +14,7 @@ import asyncio
 import inspect
 import contextvars
 import concurrent.futures
+import random
 from typing import List, Optional, Any, Dict, Union, TYPE_CHECKING
 from ..errors import ToolExecutionError
 from ..tools.trust import wrap_if_external
@@ -24,6 +25,30 @@ if TYPE_CHECKING:
     pass
 
 
+class BackoffPolicy:
+    """Exponential backoff policy for tool retries."""
+    
+    @staticmethod
+    def delay(attempt: int, initial_delay: float, backoff_factor: float, jitter: float, max_delay: float = 60.0) -> float:
+        """Calculate delay for a retry attempt.
+        
+        Args:
+            attempt: Attempt number (1-based)
+            initial_delay: Initial delay in seconds
+            backoff_factor: Exponential backoff multiplier
+            jitter: Fraction of base delay to add as random jitter
+            max_delay: Maximum delay to cap exponential growth
+            
+        Returns:
+            Delay in seconds
+        """
+        base = initial_delay * (backoff_factor ** (attempt - 1))
+        # Cap the base delay to prevent excessively long waits
+        base = min(base, max_delay)
+        jitter_amount = random.uniform(0, jitter * base)
+        return base + jitter_amount
+
+
 class ToolExecutionMixin:
     """Mixin providing toolexecution methods for the Agent class."""
 
@@ -231,9 +256,10 @@ class ToolExecutionMixin:
                     arguments.update(res.output.modified_data)
 
             # Loop guard check - prevent tool execution loops with graduated response
-            if hasattr(self, '_loop_guard') and self._loop_guard:
+            if hasattr(self, '_ensure_loop_guard'):
+                loop_guard = self._ensure_loop_guard()
                 from ..escalation.loop_guard import GuardAction
-                decision = self._loop_guard.check(function_name, arguments, is_pre_execution=True)
+                decision = loop_guard.check(function_name, arguments, is_pre_execution=True)
                 
                 if decision.action == GuardAction.WARN:
                     # Inject warning into tool result so LLM sees guidance
@@ -245,7 +271,6 @@ class ToolExecutionMixin:
                     blocked_result = {"error": f"[loop-guard] {decision.message}", "loop_blocked": True}
                 elif decision.action == GuardAction.HALT:
                     # Halt execution with exception
-                    from ..errors import ToolExecutionError
                     raise ToolExecutionError(
                         f"[loop-guard] {decision.message}",
                         tool_name=function_name,
@@ -273,32 +298,159 @@ class ToolExecutionMixin:
             if blocked_result is not None:
                 result = blocked_result
             else:
-                # P8/G11: Apply tool timeout if configured
-                tool_timeout = getattr(self, '_tool_timeout', None)
-                if tool_timeout and tool_timeout > 0:
-                    # Use copy_context to preserve injection context in executor thread
-                    ctx = contextvars.copy_context()
-                    
-                    def execute_with_context():
-                        with with_injection_context(state):
-                            return self._execute_tool_with_circuit_breaker(function_name, arguments)
-                    
-                    # Use reusable executor to prevent resource leaks
-                    if not hasattr(self, '_tool_executor'):
-                        self._tool_executor = concurrent.futures.ThreadPoolExecutor(
-                            max_workers=2, thread_name_prefix=f"tool-{self.name}"
-                        )
-                    
-                    future = self._tool_executor.submit(ctx.run, execute_with_context)
-                    try:
-                        result = future.result(timeout=tool_timeout)
-                    except concurrent.futures.TimeoutError:
-                        future.cancel()
-                        logging.warning(f"Tool {function_name} timed out after {tool_timeout}s")
-                        result = {"error": f"Tool timed out after {tool_timeout}s", "timeout": True}
+                # Apply tool retry logic with exponential backoff
+                execution_config = getattr(self, '_execution_config', None)
+                if execution_config is None:
+                    # Fall back to reading individual config attributes for backward compatibility
+                    max_retry_limit = getattr(self, 'max_retry_limit', 2)
+                    retry_initial_delay = 1.0
+                    retry_backoff_factor = 2.0
+                    retry_jitter = 0.1
                 else:
-                    with with_injection_context(state):
-                        result = self._execute_tool_with_circuit_breaker(function_name, arguments)
+                    max_retry_limit = execution_config.max_retry_limit
+                    retry_initial_delay = execution_config.retry_initial_delay
+                    retry_backoff_factor = execution_config.retry_backoff_factor
+                    retry_jitter = execution_config.retry_jitter
+                
+                result = None
+                last_exception = None
+                
+                # max_retry_limit is the number of retries (not total attempts)
+                # So total attempts = 1 (initial) + max_retry_limit (retries)
+                for attempt in range(1, max_retry_limit + 2):
+                    try:
+                        # P8/G11: Apply tool timeout if configured
+                        tool_timeout = getattr(self, '_tool_timeout', None)
+                        if tool_timeout and tool_timeout > 0:
+                            # Use copy_context to preserve injection context in executor thread
+                            ctx = contextvars.copy_context()
+                            
+                            def execute_with_context():
+                                with with_injection_context(state):
+                                    return self._execute_tool_with_circuit_breaker(function_name, arguments)
+                            
+                            # Use reusable executor to prevent resource leaks
+                            if not hasattr(self, '_tool_executor'):
+                                self._tool_executor = concurrent.futures.ThreadPoolExecutor(
+                                    max_workers=2, thread_name_prefix=f"tool-{self.name}"
+                                )
+                            
+                            future = self._tool_executor.submit(ctx.run, execute_with_context)
+                            try:
+                                result = future.result(timeout=tool_timeout)
+                            except concurrent.futures.TimeoutError:
+                                future.cancel()
+                                logging.warning(f"Tool {function_name} timed out after {tool_timeout}s")
+                                result = {"error": f"Tool timed out after {tool_timeout}s", "timeout": True}
+                        else:
+                            with with_injection_context(state):
+                                result = self._execute_tool_with_circuit_breaker(function_name, arguments)
+                        
+                        # Check if the result indicates a retryable error
+                        if isinstance(result, dict) and result.get("error"):
+                            # Check if this is a circuit breaker error (always retryable)
+                            if result.get("circuit_open"):
+                                raise ToolExecutionError(
+                                    result["error"],
+                                    tool_name=function_name,
+                                    agent_id=self.name,
+                                    is_retryable=True,
+                                )
+                            # Check if this is a timeout error (retryable)
+                            elif result.get("timeout"):
+                                raise ToolExecutionError(
+                                    result["error"],
+                                    tool_name=function_name,
+                                    agent_id=self.name,
+                                    is_retryable=True,
+                                )
+                            # For other error dicts, treat as non-retryable unless specified
+                            else:
+                                # Success path - return the result
+                                break
+                        else:
+                            # Success path - return the result
+                            break
+                            
+                    except ToolExecutionError as e:
+                        last_exception = e
+                        # Only retry if the error is marked as retryable and we have retries left
+                        # attempt starts at 1, so (attempt - 1) gives us the retry count
+                        if not e.is_retryable or (attempt - 1) >= max_retry_limit:
+                            raise e
+                        
+                        # Calculate delay for exponential backoff
+                        delay = BackoffPolicy.delay(attempt, retry_initial_delay, retry_backoff_factor, retry_jitter)
+                        logging.warning(
+                            f"Tool {function_name} failed (attempt {attempt}/{max_retry_limit + 1}): {e}. "
+                            f"Retrying in {delay:.2f}s..."
+                        )
+                        time.sleep(delay)
+                        
+                    except Exception as e:
+                        # Wrap unexpected exceptions in ToolExecutionError
+                        # Most tool errors are considered retryable unless they're programming errors
+                        is_retryable = not isinstance(e, (ValueError, TypeError, AttributeError))
+                        tool_error = ToolExecutionError(
+                            f"Tool '{function_name}' failed: {e}",
+                            tool_name=function_name,
+                            agent_id=self.name,
+                            is_retryable=is_retryable,
+                        )
+                        last_exception = tool_error
+                        
+                        # attempt starts at 1, so (attempt - 1) gives us the retry count
+                        if not is_retryable or (attempt - 1) >= max_retry_limit:
+                            raise tool_error from e
+                        
+                        # Calculate delay for exponential backoff
+                        delay = BackoffPolicy.delay(attempt, retry_initial_delay, retry_backoff_factor, retry_jitter)
+                        logging.warning(
+                            f"Tool {function_name} failed (attempt {attempt}/{max_retry_limit + 1}): {e}. "
+                            f"Retrying in {delay:.2f}s..."
+                        )
+                        time.sleep(delay)
+            
+            # Apply runtime-scoped middleware normalization BEFORE hooks fire
+            # Plugin harnesses can register middleware to normalize vendor-specific results
+            runtime_id = getattr(self, '_runtime_id', 'praisonai')  # Default to native runtime
+            normalized_result = None  # Track normalized result for hooks
+            if runtime_id != 'praisonai':  # Skip for native runtime to avoid allocation
+                try:
+                    from ..runtime import get_middleware, MiddlewareContext
+                    middleware = get_middleware(runtime_id)
+                    
+                    # Only allocate context and normalize if not using default pass-through
+                    if middleware.runtime_id != 'praisonai':
+                        mw_ctx = MiddlewareContext(
+                            tool_name=function_name,
+                            runtime_id=runtime_id,
+                            agent_id=self.name,
+                            session_id=getattr(self, '_session_id', None),
+                            execution_time_ms=(_time.time() - _tool_start_time) * 1000,
+                            metadata={'original_result_type': type(result).__name__}
+                        )
+                        
+                        normalized_result = middleware.normalize(result, function_name, mw_ctx)
+                        
+                        # Handle error cases by propagating error message as result
+                        if not normalized_result.success and normalized_result.error_message:
+                            # For failed tools, include error context in the result
+                            result = f"Tool Error: {normalized_result.error_message}"
+                        else:
+                            # For successful tools, use the normalized content
+                            result = normalized_result.content
+                        
+                        # Store normalized result for hooks to access full context
+                        self._last_normalized_result = normalized_result
+                        
+                        logger.debug(f"Applied runtime middleware for {runtime_id}: {function_name}")
+                except ImportError:
+                    # Runtime middleware not available - continue without normalization
+                    logger.debug("Runtime middleware not available, skipping normalization")
+                except Exception as e:
+                    # Don't let middleware failures break tool execution
+                    logger.warning(f"Runtime middleware failed for {runtime_id}: {e}")
             
             # Apply prompt injection protection for external tools
             # Zero-cost for trusted tools, wraps external content in security markers
@@ -312,7 +464,7 @@ class ToolExecutionMixin:
                     
                     if self.context_manager:
                         # Use context-aware truncation with configured budget
-                        truncated = self._truncate_tool_output(function_name, result_str)
+                        truncated = self._truncate_tool_output(function_name, result_str, tool_call_id)
                     else:
                         # Apply default limit even without context management
                         # This prevents runaway tool outputs from causing overflow
@@ -323,6 +475,21 @@ class ToolExecutionMixin:
                             head = result_str[:limit - tail_size]
                             tail = result_str[-tail_size:] if tail_size > 0 else ""
                             truncated = f"{head}\n...[{len(result_str):,} chars, showing first/last portions]...\n{tail}"
+                            
+                            # Store full output for later retrieval
+                            try:
+                                from ..runtime.tool_output_store import get_tool_output_store
+                                store = get_tool_output_store(getattr(self, '_run_id', None))
+                                metadata = store.store(function_name, result_str, call_id=tool_call_id)
+                                if metadata:
+                                    # Add reference to stored output in the truncated preview
+                                    truncated = store.format_reference(metadata, truncated)
+                                    logging.debug(f"Stored full {function_name} output ({len(result_str)} bytes) at {metadata.get('path')}")
+                            except ImportError:
+                                # Fallback if store not available
+                                pass
+                            except Exception as e:
+                                logging.debug(f"Failed to store tool output: {e}")
                         else:
                             truncated = result_str
                     
@@ -331,7 +498,7 @@ class ToolExecutionMixin:
                         # For dicts, truncate large string fields (e.g., raw_content from search)
                         if isinstance(result, dict):
                             max_field_chars = getattr(self, 'tool_output_limit', 16000) if not self.context_manager else None
-                            result = self._truncate_dict_fields(result, function_name, max_field_chars)
+                            result = self._truncate_dict_fields(result, function_name, max_field_chars, tool_call_id)
                         else:
                             result = truncated
                 except Exception as e:
@@ -373,6 +540,15 @@ class ToolExecutionMixin:
             
             # Trigger AFTER_TOOL hook
             from ..hooks import HookEvent, AfterToolInput
+            
+            # Extract error information from normalized result if available
+            tool_error = None
+            if hasattr(self, '_last_normalized_result') and self._last_normalized_result:
+                if not self._last_normalized_result.success and self._last_normalized_result.error_message:
+                    tool_error = self._last_normalized_result.error_message
+                # Clean up temporary attribute
+                delattr(self, '_last_normalized_result')
+            
             after_tool_input = AfterToolInput(
                 session_id=getattr(self, '_session_id', 'default'),
                 cwd=os.getcwd(),
@@ -382,6 +558,7 @@ class ToolExecutionMixin:
                 tool_name=function_name,
                 tool_input=arguments,
                 tool_output=result,
+                tool_error=tool_error,
                 execution_time_ms=(_time.time() - _tool_start_time) * 1000
             )
             self._hook_runner.execute_sync(HookEvent.AFTER_TOOL, after_tool_input, target=function_name)
@@ -394,11 +571,12 @@ class ToolExecutionMixin:
                     self._doom_loop_tracker.mark_progress(f"tool:{function_name}")
             
             # Record tool execution in loop guard
-            if hasattr(self, '_loop_guard') and self._loop_guard:
+            if hasattr(self, '_ensure_loop_guard'):
+                loop_guard = self._ensure_loop_guard()
                 is_success = result is not None and not (isinstance(result, dict) and result.get('error'))
-                self._loop_guard.record(function_name, arguments, is_success)
+                loop_guard.record(function_name, arguments, is_success)
                 # Handle warning injection for WARN decisions
-                decision = self._loop_guard.check(function_name, arguments, is_pre_execution=False) 
+                decision = loop_guard.check(function_name, arguments, is_pre_execution=False) 
                 if decision.action.value == "warn":
                     if isinstance(result, str):
                         result = f"{result}\n\n[loop-guard] {decision.message}"
@@ -417,22 +595,20 @@ class ToolExecutionMixin:
             
             return result
         except Exception as e:
-            # Emit tool call end with error
+            # Emit tool call end with error for exceptions that escape the retry loop
             _duration_ms = (_time.time() - _tool_start_time) * 1000
             _trace_emitter.tool_call_end(self.name, function_name, None, _duration_ms, str(e))
             
-            # Preserve existing ToolExecutionError unchanged to maintain loop guard HALT behavior
-            if isinstance(e, ToolExecutionError):
-                raise
-            
-            # Gap 3a fix: Wrap exceptions in ToolExecutionError for better observability
-            is_retryable = not isinstance(e, (ValueError, TypeError, AttributeError))
-            raise ToolExecutionError(
-                f"Tool '{function_name}' failed: {e}",
-                tool_name=function_name,
-                agent_id=self.name,
-                is_retryable=is_retryable,
-            ) from e
+            # Wrap the exception if it's not already a ToolExecutionError
+            if not isinstance(e, ToolExecutionError):
+                is_retryable = not isinstance(e, (ValueError, TypeError, AttributeError))
+                raise ToolExecutionError(
+                    f"Tool '{function_name}' failed: {e}",
+                    tool_name=function_name,
+                    agent_id=self.name,
+                    is_retryable=is_retryable,
+                ) from e
+            raise  # Re-raise if already wrapped
 
     def _trigger_after_agent_hook(self, prompt, response, start_time, tools_used=None):
         """Trigger AFTER_AGENT hook and return response."""
@@ -531,7 +707,7 @@ class ToolExecutionMixin:
             response=response,
         )
 
-    def _truncate_dict_fields(self, data: dict, tool_name: str, max_field_chars: int = None) -> dict:
+    def _truncate_dict_fields(self, data: dict, tool_name: str, max_field_chars: int = None, tool_call_id: str = None) -> dict:
         """Truncate large string fields in a dict to prevent context overflow."""
         if max_field_chars is None:
             # Use tool budget from context manager (default 5000 tokens * 4 chars/token = 20000 chars)
@@ -546,13 +722,30 @@ class ToolExecutionMixin:
                 tail_limit = int(max_field_chars * 0.15)
                 head = value[:head_limit]
                 tail = value[-tail_limit:] if tail_limit > 0 else ""
-                result[key] = f"{head}\n...[{len(value):,} chars, showing first/last portions]...\n{tail}"
+                truncated = f"{head}\n...[{len(value):,} chars, showing first/last portions]...\n{tail}"
+                
+                # Store full field value for later retrieval
+                try:
+                    from ..runtime.tool_output_store import get_tool_output_store
+                    from uuid import uuid4
+                    store = get_tool_output_store(getattr(self, '_run_id', None))
+                    # Add unique suffix to prevent collisions with repeated keys
+                    unique_suffix = uuid4().hex[:8]
+                    field_call_id = f"{tool_call_id}_{key}_{unique_suffix}" if tool_call_id else f"{tool_name}_{key}_{unique_suffix}"
+                    metadata = store.store(f"{tool_name}.{key}", value, call_id=field_call_id)
+                    if metadata:
+                        truncated = store.format_reference(metadata, truncated)
+                        logging.debug(f"Stored full {tool_name}.{key} field ({len(value)} bytes) at {metadata.get('path')}")
+                except Exception as e:
+                    logging.debug(f"Failed to store dict field: {e}")
+                
+                result[key] = truncated
                 logging.debug(f"Smart truncated field '{key}' from {len(value)} to ~{max_field_chars} chars")
             elif isinstance(value, dict):
-                result[key] = self._truncate_dict_fields(value, tool_name, max_field_chars)
+                result[key] = self._truncate_dict_fields(value, tool_name, max_field_chars, tool_call_id)
             elif isinstance(value, list):
                 result[key] = [
-                    self._truncate_dict_fields(item, tool_name, max_field_chars) if isinstance(item, dict)
+                    self._truncate_dict_fields(item, tool_name, max_field_chars, tool_call_id) if isinstance(item, dict)
                     else (self._smart_truncate_str(item, max_field_chars) if isinstance(item, str) and len(item) > max_field_chars else item)
                     for item in value
                 ]
@@ -584,12 +777,22 @@ class ToolExecutionMixin:
         from ..approval import get_approval_registry
         from ..approval.protocols import ApprovalRequest, ApprovalDecision
         from ..approval.registry import DEFAULT_DANGEROUS_TOOLS
+        from ..tools import get_registry as get_tool_registry
         
         backend = getattr(self, '_approval_backend', None)
         approve_all = getattr(self, '_approve_all_tools', False)
         
+        # Check tool trust level if available
+        tool_registry = get_tool_registry()
+        trust_level = tool_registry.get_trust_level(tool_name)
+        
         if backend is not None:
-            needs_approval = approve_all or tool_name in DEFAULT_DANGEROUS_TOOLS
+            # Check if tool needs approval based on multiple criteria
+            needs_approval = (
+                approve_all 
+                or tool_name in DEFAULT_DANGEROUS_TOOLS
+                or (trust_level == "external")  # External tools need approval
+            )
             if needs_approval:
                 request = ApprovalRequest(
                     tool_name=tool_name,

src/praisonai-agents/praisonaiagents/agent/unified_chat_mixin.py

@@ -1,152 +0,0 @@
-"""
-Unified Chat Mixin for Agent class.
-
-This module consolidates the dual execution paths (custom LLM vs OpenAI)
-into a single async-first protocol-driven implementation, eliminating
-the DRY violation and maintenance burden described in Issue #1304.
-"""
-
-import asyncio
-import logging
-import time
-from typing import Any, Dict, List, Optional, Union
-
-from ..llm import create_llm_dispatcher, UnifiedLLMDispatcher
-
-
-class UnifiedChatMixin:
-    """
-    Unified chat implementation that replaces dual execution paths.
-    
-    This mixin consolidates:
-    - Path A: LLM.get_response() / LLM.get_response_async() 
-    - Path B: OpenAIClient chat completion
-    
-    Into a single async-first protocol-driven dispatch using UnifiedLLMDispatcher.
-    """
-    
-    def _get_unified_dispatcher(self) -> UnifiedLLMDispatcher:
-        """
-        Get or create the unified LLM dispatcher for this agent.
-        
-        This method implements the provider selection logic that was
-        previously scattered in _chat_completion() method.
-        """
-        if hasattr(self, '_unified_dispatcher') and self._unified_dispatcher is not None:
-            return self._unified_dispatcher
-        
-        # Determine which adapter to use based on existing agent configuration
-        if self._using_custom_llm and hasattr(self, 'llm_instance'):
-            # Use LiteLLM adapter for custom LLM instances
-            dispatcher = create_llm_dispatcher(llm_instance=self.llm_instance)
-        else:
-            # Use OpenAI adapter for direct OpenAI access
-            if not hasattr(self, '_openai_client') or self._openai_client is None:
-                # Initialize OpenAI client if not present
-                from ..llm import get_openai_client
-                self._openai_client = get_openai_client(api_key=getattr(self, 'api_key', None))
-            dispatcher = create_llm_dispatcher(openai_client=self._openai_client, model=self.llm)
-        
-        # Cache the dispatcher
-        self._unified_dispatcher = dispatcher
-        return dispatcher
-    
-    def _unified_chat_completion(
-        self, 
-        messages: List[Dict[str, Any]], 
-        temperature: float = 1.0, 
-        tools: Optional[List[Dict[str, Any]]] = None, 
-        stream: bool = True,
-        reasoning_steps: bool = False,
-        task_name: Optional[str] = None,
-        task_description: Optional[str] = None,
-        task_id: Optional[str] = None,
-        response_format: Optional[Dict[str, Any]] = None
-    ) -> Any:
-        """
-        Unified sync chat completion method.
-        
-        This replaces the existing _chat_completion() method's dual dispatch logic
-        with a single path through UnifiedLLMDispatcher.
-        """
-        start_time = time.time()
-        
-        # Get unified dispatcher (handles provider selection)
-        dispatcher = self._get_unified_dispatcher()
-        
-        # Use the same message processing logic but route through unified dispatcher
-        formatted_tools = self._format_tools_for_completion(tools)
-        
-        try:
-            # Single dispatch path for all providers
-            final_response = dispatcher.chat_completion(
-                messages=messages,
-                tools=formatted_tools,
-                temperature=temperature,
-                max_tokens=getattr(self, 'max_tokens', None),
-                stream=stream,
-                response_format=response_format,
-                **{
-                    'reasoning_steps': reasoning_steps,
-                    'task_name': task_name,
-                    'task_description': task_description,
-                    'task_id': task_id,
-                }
-            )
-            
-            return final_response
-            
-        except Exception as e:
-            logging.error(f"Unified chat completion failed: {e}")
-            # Apply the same error handling as the original implementation
-            raise
-    
-    async def _unified_achat_completion(
-        self, 
-        messages: List[Dict[str, Any]], 
-        temperature: float = 1.0, 
-        tools: Optional[List[Dict[str, Any]]] = None, 
-        stream: bool = True,
-        reasoning_steps: bool = False,
-        task_name: Optional[str] = None,
-        task_description: Optional[str] = None,
-        task_id: Optional[str] = None,
-        response_format: Optional[Dict[str, Any]] = None
-    ) -> Any:
-        """
-        Unified async chat completion method.
-        
-        This provides native async implementation, eliminating the need
-        for separate async/sync implementations in the LLM layer.
-        """
-        start_time = time.time()
-        
-        # Get unified dispatcher (handles provider selection)
-        dispatcher = self._get_unified_dispatcher()
-        
-        # Use the same message processing logic but route through unified dispatcher
-        formatted_tools = self._format_tools_for_completion(tools)
-        
-        try:
-            # Single async dispatch path for all providers
-            final_response = await dispatcher.achat_completion(
-                messages=messages,
-                tools=formatted_tools,
-                temperature=temperature,
-                max_tokens=getattr(self, 'max_tokens', None),
-                stream=stream,
-                response_format=response_format,
-                **{
-                    'reasoning_steps': reasoning_steps,
-                    'task_name': task_name,
-                    'task_description': task_description,
-                    'task_id': task_id,
-                }
-            )
-            
-            return final_response
-            
-        except Exception as e:
-            logging.error(f"Unified async chat completion failed: {e}")
-            # Apply the same error handling as the original implementation
-            raise

src/praisonai-agents/praisonaiagents/agent/unified_execution_mixin.py

@@ -78,12 +78,37 @@ class UnifiedExecutionMixin:
         _trace_emitter.agent_start(self.name, {"role": self.role, "goal": self.goal})
         
         # Reset loop guard for new chat turn (always-on protection)
-        if hasattr(self, '_loop_guard') and self._loop_guard:
-            self._loop_guard.reset_turn()
+        if hasattr(self, '_ensure_loop_guard'):
+            self._ensure_loop_guard().reset_turn()
         
         try:
-            # CLI Backend routing - delegate entire turn if configured
-            if hasattr(self, '_cli_backend') and self._cli_backend is not None:
+            # Runtime resolution - check for model-scoped runtime before legacy CLI backend
+            runtime_instance = None
+            if getattr(self, '_runtime_config', None) is not None:
+                runtime_instance = await self._resolve_turn_runtime()
+            
+            # Model-scoped runtime routing - delegate entire turn to resolved runtime
+            if runtime_instance is not None:
+                return await self._chat_via_runtime(
+                    runtime_instance=runtime_instance,
+                    prompt=prompt,
+                    temperature=temperature,
+                    tools=tools,
+                    output_json=output_json,
+                    output_pydantic=output_pydantic,
+                    reasoning_steps=reasoning_steps,
+                    stream=stream,
+                    task_name=task_name,
+                    task_description=task_description,
+                    task_id=task_id,
+                    config=config,
+                    force_retrieval=force_retrieval,
+                    skip_retrieval=skip_retrieval,
+                    attachments=attachments,
+                    tool_choice=tool_choice
+                )
+            elif hasattr(self, '_cli_backend') and self._cli_backend is not None:
+                # Legacy CLI backend with deprecation warning (emitted in Agent.__init__)
                 return await self._chat_via_cli_backend(
                     prompt=prompt,
                     temperature=temperature,

src/praisonai-agents/praisonaiagents/agents/agents.py

@@ -40,6 +40,21 @@ class TaskStatus(Enum):
 # Set up logger
 logger = get_logger(__name__)
 
+
+def _launch_auth_token() -> Optional[str]:
+    return os.environ.get("PRAISONAI_LAUNCH_AUTH_TOKEN")
+
+
+def _authorise_launch_request(request) -> bool:
+    token = _launch_auth_token()
+    if not token:
+        return True
+    auth = request.headers.get("Authorization", "")
+    if auth.startswith("Bearer ") and auth[7:] == token:
+        return True
+    return request.headers.get("X-Auth-Token") == token
+
+
 # Agent server registry for thread-safe server management
 
 
@@ -245,7 +260,20 @@ Expected Output: {task.expected_output}."""
     # Add memory context if available
     if task.memory:
         try:
-            memory_context = task.memory.build_context_for_task(task.description)
+            # Use cache-optimized context if available for better prompt caching
+            if hasattr(task.memory, 'build_cache_optimized_context'):
+                try:
+                    cache_result = task.memory.build_cache_optimized_context(
+                        task_descr=task.description,
+                        include_cache_boundary=False  # Don't include boundary for agent task context
+                    )
+                    memory_context = cache_result.get('stable_prefix', '')
+                except Exception as e:
+                    # Fall back to standard context building
+                    logger.debug(f"Cache-optimized context failed for task '{task.description or task.id}', falling back to standard: {e}")
+                    memory_context = task.memory.build_context_for_task(task.description)
+            else:
+                memory_context = task.memory.build_context_for_task(task.description)
             if memory_context:
                 # Log detailed memory context for debugging
                 logger.debug(f"Memory context for task '{task.description}': {memory_context}")
@@ -652,7 +680,7 @@ class AgentTeam(SpawnAnnounceProtocol):
             _stream = _output_config.stream
         else:
             _verbose = 0
-            _stream = True
+            _stream = False
         
         # ─────────────────────────────────────────────────────────────────────
         # Resolve EXECUTION param using canonical resolver
@@ -1028,6 +1056,65 @@ class AgentTeam(SpawnAnnounceProtocol):
         task_result = _process_task_result(self, context, agent_output)
         return task_result.task_output
 
+    def _apply_task_guardrail(self, task, task_id, task_output):
+        """Apply guardrail validation to task output.
+        
+        Returns:
+            tuple: (task_output, should_retry) where should_retry is True if task should be retried
+            
+        Raises:
+            Exception: If guardrail validation fails after max retries
+        """
+        if not task._guardrail_fn:
+            return task_output, False
+            
+        try:
+            guardrail_result = task._process_guardrail(task_output)
+            if not guardrail_result.success:
+                if task.retry_count >= task.max_retries:
+                    raise Exception(
+                        f"Task failed guardrail validation after {task.max_retries} retries. "
+                        f"Last error: {guardrail_result.error}"
+                    )
+                
+                task.retry_count += 1
+                task.status = "in progress"  # Keep task in progress for retry
+                logger.warning(f"Task {task_id}: Guardrail validation failed (retry {task.retry_count}/{task.max_retries}): {guardrail_result.error}")
+                return task_output, True  # Signal retry needed
+            
+            # If guardrail passed and returned a modified result
+            if guardrail_result.result is not None:
+                if isinstance(guardrail_result.result, str):
+                    # Update the task output with the modified result
+                    task_output.raw = guardrail_result.result
+                    # Clear structured fields to avoid stale cache
+                    if hasattr(task_output, 'json_dict'):
+                        task_output.json_dict = None
+                    if hasattr(task_output, 'pydantic'):
+                        task_output.pydantic = None
+                    task.result = task_output
+                elif hasattr(guardrail_result.result, 'raw'):
+                    # Replace with the new task output
+                    task_output = guardrail_result.result
+                    task.result = task_output
+            
+            logger.info(f"Task {task_id}: Guardrail validation passed")
+            return task_output, False  # No retry needed
+            
+        except Exception as e:
+            logger.error(f"Task {task_id}: Error in guardrail processing: {e}")
+            # Handle guardrail failure with retry logic
+            if task.retry_count >= task.max_retries:
+                raise Exception(
+                    f"Task failed due to guardrail processing error after {task.max_retries} retries. "
+                    f"Last error: {e}"
+                ) from e
+            task.retry_count += 1
+            task.status = "in progress"
+            logger.warning(f"Task {task_id}: Guardrail processing error (retry {task.retry_count}/{task.max_retries}): {e}")
+            return task_output, True  # Signal retry needed
+
+
     async def arun_task(self, task_id):
         """Async version of run_task method"""
         if task_id not in self.tasks:
@@ -1046,53 +1133,11 @@ class AgentTeam(SpawnAnnounceProtocol):
             if task.status in ["not started", "in progress"]:
                 task_output = await self.aexecute_task(task_id)
                 if task_output and self.completion_checker(task, task_output.raw):
-                    # Run guardrail validation BEFORE marking task complete
-                    if task._guardrail_fn:
-                        try:
-                            guardrail_result = task._process_guardrail(task_output)
-                            if not guardrail_result.success:
-                                if task.retry_count >= task.max_retries:
-                                    raise Exception(
-                                        f"Task failed guardrail validation after {task.max_retries} retries. "
-                                        f"Last error: {guardrail_result.error}"
-                                    )
-                                
-                                task.retry_count += 1
-                                task.status = "in progress"  # Keep task in progress for retry
-                                logger.warning(f"Task {task_id}: Guardrail validation failed (retry {task.retry_count}/{task.max_retries}): {guardrail_result.error}")
-                                retries += 1
-                                continue  # Actually retry the task
-                            
-                            # If guardrail passed and returned a modified result
-                            if guardrail_result.result is not None:
-                                if isinstance(guardrail_result.result, str):
-                                    # Update the task output with the modified result
-                                    task_output.raw = guardrail_result.result
-                                    # Clear structured fields to avoid stale cache
-                                    if hasattr(task_output, 'json_dict'):
-                                        task_output.json_dict = None
-                                    if hasattr(task_output, 'pydantic'):
-                                        task_output.pydantic = None
-                                    task.result = task_output
-                                elif hasattr(guardrail_result.result, 'raw'):
-                                    # Replace with the new task output
-                                    task_output = guardrail_result.result
-                                    task.result = task_output
-                            
-                            logger.info(f"Task {task_id}: Guardrail validation passed")
-                        except Exception as e:
-                            logger.error(f"Task {task_id}: Error in guardrail processing: {e}")
-                            # Handle guardrail failure with retry logic
-                            if task.retry_count >= task.max_retries:
-                                raise Exception(
-                                    f"Task failed due to guardrail processing error after {task.max_retries} retries. "
-                                    f"Last error: {e}"
-                                ) from e
-                            task.retry_count += 1
-                            task.status = "in progress"
-                            logger.warning(f"Task {task_id}: Guardrail processing error (retry {task.retry_count}/{task.max_retries}): {e}")
-                            retries += 1
-                            continue  # Retry the task
+                    # Apply guardrail validation using shared helper
+                    task_output, should_retry = self._apply_task_guardrail(task, task_id, task_output)
+                    if should_retry:
+                        retries += 1
+                        continue
                     
                     task.status = "completed"
                     # Run execute_callback for memory operations
@@ -1106,20 +1151,6 @@ class AgentTeam(SpawnAnnounceProtocol):
                             raise
                         if hasattr(task, 'fail_on_memory_error') and task.fail_on_memory_error:
                             raise
-                    
-                    # Run task callback if exists
-                    if task.callback:
-                        try:
-                            if asyncio.iscoroutinefunction(task.callback):
-                                if run_coroutine_safely:
-                                    run_coroutine_safely(task.callback(task_output))
-                                else:
-                                    logger.warning("run_coroutine_safely not available, skipping async callback")
-                            else:
-                                task.callback(task_output)
-                        except Exception as e:
-                            logger.error(f"Error executing task callback for task {task_id}: {e}")
-                            logger.exception(e)
                             
                     self.save_output_to_file(task, task_output)
                     if self.verbose >= 1:
@@ -1334,52 +1365,11 @@ class AgentTeam(SpawnAnnounceProtocol):
             if task.status in ["not started", "in progress"]:
                 task_output = self.execute_task(task_id)
                 if task_output and self.completion_checker(task, task_output.raw):
-                    # Add guardrail validation (matches arun_task logic)
-                    if task._guardrail_fn:
-                        try:
-                            guardrail_result = task._process_guardrail(task_output)
-                            if not guardrail_result.success:
-                                if task.retry_count >= task.max_retries:
-                                    raise Exception(
-                                        f"Task failed guardrail validation after {task.max_retries} retries. "
-                                        f"Last error: {guardrail_result.error}"
-                                    )
-                                task.retry_count += 1
-                                task.status = "in progress"  # Keep task in progress for retry
-                                logger.warning(f"Task {task_id}: Guardrail validation failed (retry {task.retry_count}/{task.max_retries}): {guardrail_result.error}")
-                                retries += 1
-                                continue  # Retry the task
-                            
-                            # If guardrail passed and returned a modified result
-                            if guardrail_result.result is not None:
-                                if isinstance(guardrail_result.result, str):
-                                    # Update the task output with the modified result
-                                    task_output.raw = guardrail_result.result
-                                    # Clear structured fields to avoid stale cache
-                                    if hasattr(task_output, 'json_dict'):
-                                        task_output.json_dict = None
-                                    if hasattr(task_output, 'pydantic'):
-                                        task_output.pydantic = None
-                                    task.result = task_output
-                                elif hasattr(guardrail_result.result, 'raw'):
-                                    # Replace with the new task output
-                                    task_output = guardrail_result.result
-                                    task.result = task_output
-                            
-                            logger.info(f"Task {task_id}: Guardrail validation passed")
-                        except Exception as e:
-                            logger.error(f"Task {task_id}: Error in guardrail processing: {e}")
-                            # Handle guardrail failure with retry logic
-                            if task.retry_count >= task.max_retries:
-                                raise Exception(
-                                    f"Task failed due to guardrail processing error after {task.max_retries} retries. "
-                                    f"Last error: {e}"
-                                ) from e
-                            task.retry_count += 1
-                            task.status = "in progress"
-                            logger.warning(f"Task {task_id}: Guardrail processing error (retry {task.retry_count}/{task.max_retries}): {e}")
-                            retries += 1
-                            continue  # Retry the task
+                    # Apply guardrail validation using shared helper
+                    task_output, should_retry = self._apply_task_guardrail(task, task_id, task_output)
+                    if should_retry:
+                        retries += 1
+                        continue
                     
                     task.status = "completed"
                     # Run execute_callback for memory operations
@@ -1389,20 +1379,6 @@ class AgentTeam(SpawnAnnounceProtocol):
                     except Exception as e:
                         logger.error(f"Error executing memory callback for task {task_id}: {e}")
                         logger.exception(e)
-                    
-                    # Run task callback if exists
-                    if task.callback:
-                        try:
-                            if asyncio.iscoroutinefunction(task.callback):
-                                if run_coroutine_safely:
-                                    run_coroutine_safely(task.callback(task_output))
-                                else:
-                                    logger.warning("run_coroutine_safely not available, skipping async callback")
-                            else:
-                                task.callback(task_output)
-                        except Exception as e:
-                            logger.error(f"Error executing task callback for task {task_id}: {e}")
-                            logger.exception(e)
                             
                     self.save_output_to_file(task, task_output)
                     
@@ -1993,6 +1969,8 @@ class AgentTeam(SpawnAnnounceProtocol):
             # Define the endpoint handler
             @app.post(path)
             async def handle_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:
@@ -2071,7 +2049,9 @@ class AgentTeam(SpawnAnnounceProtocol):
             
             # Add GET endpoint to list available agents
             @app.get(f"{path}/list")
-            async def list_agents():
+            async def list_agents(request: Request):
+                if not _authorise_launch_request(request):
+                    raise HTTPException(status_code=401, detail="Unauthorized")
                 return {
                     "agents": [
                         {"name": agent.display_name, "id": agent.display_name.lower().replace(' ', '_')}
@@ -2086,6 +2066,8 @@ class AgentTeam(SpawnAnnounceProtocol):
                 # Create a closure to capture the agent instance
                 def create_agent_handler(agent):
                     async def handle_single_agent(request: Request):
+                        if not _authorise_launch_request(request):
+                            raise HTTPException(status_code=401, detail="Unauthorized")
                         try:
                             request_data = await request.json()
                             query = request_data.get("query", "")

src/praisonai-agents/praisonaiagents/app/config.py

@@ -53,3 +53,4 @@ class AgentAppConfig:
     workers: int = 1
     timeout: int = 60
     metadata: Dict[str, Any] = field(default_factory=dict)
+    api_key: Optional[str] = None

src/praisonai-agents/praisonaiagents/approval/protocols.py

@@ -74,11 +74,15 @@ class ApprovalConfig:
                     - *positive float* — wait up to that many seconds.
                     - ``None`` — wait **indefinitely** (no timeout).
                     - *unset / 0* — use the backend's own default timeout.
+        permissions: Declarative permission rules mapping patterns to actions
+                    (e.g., {"read:*": "allow", "bash:rm *": "deny"}).
+                    Used for CI-safe, non-interactive permission policies.
     """
 
     backend: Any = None
     all_tools: bool = False
     timeout: Optional[float] = 0
+    permissions: Optional[Dict[str, Any]] = None
 
 
 @runtime_checkable

src/praisonai-agents/praisonaiagents/approval/registry.py

@@ -194,9 +194,13 @@ class ApprovalRegistry:
 
     def is_yaml_approved(self, tool_name: str) -> bool:
         try:
-            return tool_name in self._yaml_approved_tools.get()
+            if tool_name not in self._yaml_approved_tools.get():
+                return False
         except LookupError:
             return False
+        if self.get_risk_level(tool_name) == "critical":
+            return False
+        return True
 
     # ── Env-var check ────────────────────────────────────────────────────
 

src/praisonai-agents/praisonaiagents/bots/__init__.py

@@ -24,6 +24,19 @@ if TYPE_CHECKING:
         HealthResult,
         EmailProtocol,
         EmailInbox,
+        SupportsPresentation,
+        PlatformCapabilities,
+    )
+    from .presentation import (
+        MessagePresentation,
+        PresentationBlock,
+        PresentationButton,
+        PresentationAction,
+        SelectOption,
+        PresentationLimits,
+        ActionType,
+        ButtonStyle,
+        BlockType,
     )
 from .protocols import (
     BotProtocol,
@@ -41,6 +54,19 @@ from .protocols import (
     HealthResult,
     EmailProtocol,
     EmailInbox,
+    SupportsPresentation,
+    PlatformCapabilities,
+)
+from .presentation import (
+    MessagePresentation,
+    PresentationBlock,
+    PresentationButton,
+    PresentationAction,
+    SelectOption,
+    PresentationLimits,
+    ActionType,
+    ButtonStyle,
+    BlockType,
 )
 from .config import BotConfig, BotOSConfig
 
@@ -62,4 +88,15 @@ __all__ = [
     "HealthResult",
     "EmailProtocol",
     "EmailInbox",
+    "MessagePresentation",
+    "PresentationBlock",
+    "PresentationButton",
+    "PresentationAction",
+    "SelectOption",
+    "PresentationLimits",
+    "SupportsPresentation",
+    "ActionType",
+    "ButtonStyle",
+    "BlockType",
+    "PlatformCapabilities",
 ]

src/praisonai-agents/praisonaiagents/bots/config.py

@@ -91,6 +91,16 @@ class BotConfig:
     # When set, stale sessions older than this are auto-reaped.
     session_ttl: int = 0
     
+    # In-flight run control settings for mid-run message handling
+    # busy_mode: Policy for handling messages during active runs
+    #   - "queue": Queue message for next turn (default)
+    #   - "interrupt": Cancel current run and start new one  
+    #   - "steer": Inject message into current run via steering
+    busy_mode: str = "queue"
+    
+    # Template for busy acknowledgment messages (use {action} placeholder)
+    busy_ack: str = "⏳ {action} — will be considered next"
+    
     # Workspace settings for file operation containment and security
     workspace_dir: Optional[str] = None  # default: ~/.praisonai/workspaces/<scope>/<session_key>
     workspace_access: str = "rw"  # "rw" (read-write) | "ro" (read-only) | "none" (copy-on-write sandbox)
@@ -101,6 +111,16 @@ class BotConfig:
     
     # Owner user ID for pairing approvals (platform-specific format)
     owner_user_id: Optional[str] = None
+    
+    # Progressive streaming for channel bots (default: False)
+    streaming: bool = False
+    
+    # Edit interval for streaming responses in milliseconds (default: 700ms)
+    stream_edit_interval_ms: int = 700
+    
+    # Intentional silence support
+    allow_silence: bool = False
+    silence_token: Optional[str] = None
 
     def __post_init__(self) -> None:
         if self.unknown_user_policy not in {"deny", "pair", "allow"}:
@@ -136,11 +156,17 @@ class BotConfig:
             "ack_emoji": self.ack_emoji,
             "done_emoji": self.done_emoji,
             "session_ttl": self.session_ttl,
+            "busy_mode": self.busy_mode,
+            "busy_ack": self.busy_ack,
             "workspace_dir": self.workspace_dir,
             "workspace_access": self.workspace_access,
             "workspace_scope": self.workspace_scope,
             "unknown_user_policy": self.unknown_user_policy,
             "owner_user_id": "***" if self.owner_user_id else None,
+            "streaming": self.streaming,
+            "stream_edit_interval_ms": self.stream_edit_interval_ms,
+            "allow_silence": self.allow_silence,
+            "silence_token": self.silence_token,
             "metadata": self.metadata,
         }
     

src/praisonai-agents/praisonaiagents/bots/presentation.py

@@ -0,0 +1,456 @@
+"""
+Presentation primitives for interactive UI in messaging bots.
+
+Defines typed, portable presentation blocks (buttons, selects, text)
+that channel adapters render as native widgets. Enables structured
+interactive UI across Telegram, Slack, Discord, and other platforms.
+
+This is a core protocol with no heavy implementations - channel-specific
+rendering belongs in the wrapper (praisonai).
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import Any, Dict, List, Literal, Optional, Union
+
+
+class ActionType(str, Enum):
+    """Types of actions that can be triggered by interactive elements."""
+    
+    COMMAND = "command"      # Execute a slash command
+    CALLBACK = "callback"     # Opaque callback data for the plugin
+    URL = "url"              # Open a URL
+    WEB_APP = "web_app"      # Open a web app (Telegram mini apps, etc)
+
+
+class ButtonStyle(str, Enum):
+    """Visual styles for buttons."""
+    
+    PRIMARY = "primary"      # Primary action (blue/green)
+    DANGER = "danger"        # Destructive action (red)
+    SECONDARY = "secondary"  # Secondary action (gray)
+    SUCCESS = "success"      # Success action (green)
+    WARNING = "warning"      # Warning action (yellow)
+
+
+class BlockType(str, Enum):
+    """Types of presentation blocks."""
+    
+    TEXT = "text"            # Text content
+    BUTTONS = "buttons"      # Button row/grid
+    SELECT = "select"        # Dropdown/select menu
+    DIVIDER = "divider"      # Visual separator
+    CONTEXT = "context"      # Contextual info (smaller text)
+
+
+@dataclass
+class PresentationAction:
+    """An action triggered by an interactive element.
+    
+    Attributes:
+        type: The action type (command, callback, url, web_app)
+        command: Slash command to execute (for type="command")
+        value: Callback data (for type="callback")
+        url: URL to open (for type="url")
+        web_app_url: Web app URL (for type="web_app")
+    """
+    
+    type: Union[ActionType, str]
+    command: Optional[str] = None
+    value: Optional[str] = None
+    url: Optional[str] = None
+    web_app_url: Optional[str] = None
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        data = {"type": self.type.value if isinstance(self.type, ActionType) else self.type}
+        if self.command is not None:
+            data["command"] = self.command
+        if self.value is not None:
+            data["value"] = self.value
+        if self.url is not None:
+            data["url"] = self.url
+        if self.web_app_url is not None:
+            data["web_app_url"] = self.web_app_url
+        return data
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "PresentationAction":
+        """Create from dictionary."""
+        return cls(
+            type=data.get("type", "callback"),
+            command=data.get("command"),
+            value=data.get("value"),
+            url=data.get("url"),
+            web_app_url=data.get("web_app_url"),
+        )
+
+
+@dataclass
+class PresentationButton:
+    """A button in an interactive presentation.
+    
+    Attributes:
+        label: Button text label
+        action: Action to trigger when clicked
+        url: Direct URL (alternative to action)
+        priority: Truncation priority (higher survives)
+        style: Visual style
+        disabled: Whether button is disabled
+    """
+    
+    label: str
+    action: Optional[PresentationAction] = None
+    url: Optional[str] = None
+    priority: int = 0
+    style: Optional[Union[ButtonStyle, str]] = None
+    disabled: bool = False
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        data = {"label": self.label, "priority": self.priority, "disabled": self.disabled}
+        if self.action is not None:
+            data["action"] = self.action.to_dict()
+        if self.url is not None:
+            data["url"] = self.url
+        if self.style is not None:
+            data["style"] = self.style.value if isinstance(self.style, ButtonStyle) else self.style
+        return data
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "PresentationButton":
+        """Create from dictionary."""
+        return cls(
+            label=data.get("label", ""),
+            action=PresentationAction.from_dict(data["action"]) if "action" in data else None,
+            url=data.get("url"),
+            priority=data.get("priority", 0),
+            style=data.get("style"),
+            disabled=data.get("disabled", False),
+        )
+
+
+@dataclass
+class SelectOption:
+    """An option in a select menu.
+    
+    Attributes:
+        label: Option text label
+        value: Option value
+        description: Optional description
+        emoji: Optional emoji icon
+        default: Whether this is the default option
+    """
+    
+    label: str
+    value: str
+    description: Optional[str] = None
+    emoji: Optional[str] = None
+    default: bool = False
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        data = {"label": self.label, "value": self.value, "default": self.default}
+        if self.description is not None:
+            data["description"] = self.description
+        if self.emoji is not None:
+            data["emoji"] = self.emoji
+        return data
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "SelectOption":
+        """Create from dictionary."""
+        return cls(
+            label=data.get("label", ""),
+            value=data.get("value", ""),
+            description=data.get("description"),
+            emoji=data.get("emoji"),
+            default=data.get("default", False),
+        )
+
+
+@dataclass
+class PresentationBlock:
+    """A block in a message presentation.
+    
+    Blocks are the building blocks of interactive messages.
+    Each block type has specific properties and behavior.
+    """
+    
+    type: Union[BlockType, str]
+    text: Optional[str] = None
+    buttons: Optional[List[PresentationButton]] = None
+    options: Optional[List[SelectOption]] = None
+    placeholder: Optional[str] = None
+    action_id: Optional[str] = None
+    
+    @staticmethod
+    def make_text(content: str, markdown: bool = True) -> "PresentationBlock":
+        """Create a text block."""
+        return PresentationBlock(type=BlockType.TEXT, text=content)
+    
+    @staticmethod
+    def make_buttons(items: List[PresentationButton]) -> "PresentationBlock":
+        """Create a buttons block."""
+        return PresentationBlock(type=BlockType.BUTTONS, buttons=items)
+    
+    @staticmethod
+    def make_select(
+        options: List[SelectOption],
+        placeholder: Optional[str] = None,
+        action_id: Optional[str] = None,
+    ) -> "PresentationBlock":
+        """Create a select menu block."""
+        return PresentationBlock(
+            type=BlockType.SELECT,
+            options=options,
+            placeholder=placeholder,
+            action_id=action_id,
+        )
+    
+    @staticmethod
+    def make_divider() -> "PresentationBlock":
+        """Create a divider block."""
+        return PresentationBlock(type=BlockType.DIVIDER)
+    
+    @staticmethod
+    def make_context(content: str) -> "PresentationBlock":
+        """Create a context block (smaller text)."""
+        return PresentationBlock(type=BlockType.CONTEXT, text=content)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        data = {"type": self.type.value if isinstance(self.type, BlockType) else self.type}
+        if self.text is not None:
+            data["text"] = self.text
+        if self.buttons is not None:
+            data["buttons"] = [b.to_dict() for b in self.buttons]
+        if self.options is not None:
+            data["options"] = [o.to_dict() for o in self.options]
+        if self.placeholder is not None:
+            data["placeholder"] = self.placeholder
+        if self.action_id is not None:
+            data["action_id"] = self.action_id
+        return data
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "PresentationBlock":
+        """Create from dictionary."""
+        return cls(
+            type=data.get("type", "text"),
+            text=data.get("text"),
+            buttons=(
+                [PresentationButton.from_dict(b) for b in data["buttons"]]
+                if "buttons" in data else None
+            ),
+            options=(
+                [SelectOption.from_dict(o) for o in data["options"]]
+                if "options" in data else None
+            ),
+            placeholder=data.get("placeholder"),
+            action_id=data.get("action_id"),
+        )
+
+
+@dataclass
+class MessagePresentation:
+    """A complete interactive message presentation.
+    
+    Attributes:
+        blocks: List of presentation blocks
+        tone: Optional tone/style hint for the whole message
+        ephemeral: Whether message should be ephemeral/temporary
+        replace_message_id: ID of message to replace (for updates)
+    """
+    
+    blocks: List[PresentationBlock] = field(default_factory=list)
+    tone: Optional[str] = None
+    ephemeral: bool = False
+    replace_message_id: Optional[str] = None
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        data = {"blocks": [b.to_dict() for b in self.blocks], "ephemeral": self.ephemeral}
+        if self.tone is not None:
+            data["tone"] = self.tone
+        if self.replace_message_id is not None:
+            data["replace_message_id"] = self.replace_message_id
+        return data
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "MessagePresentation":
+        """Create from dictionary."""
+        return cls(
+            blocks=[PresentationBlock.from_dict(b) for b in data.get("blocks", [])],
+            tone=data.get("tone"),
+            ephemeral=data.get("ephemeral", False),
+            replace_message_id=data.get("replace_message_id"),
+        )
+    
+    @staticmethod
+    def approval(
+        prompt: str,
+        approval_id: str,
+        allow_always: bool = False,
+        context: Optional[str] = None,
+    ) -> "MessagePresentation":
+        """Create a standard approval presentation.
+        
+        Args:
+            prompt: The approval prompt text
+            approval_id: Unique ID for this approval
+            allow_always: Whether to include "Allow Always" option
+            context: Optional context information
+            
+        Returns:
+            A presentation with standard approval buttons
+        """
+        blocks = []
+        
+        # Add prompt
+        blocks.append(PresentationBlock.make_text(prompt))
+        
+        # Add context if provided
+        if context:
+            blocks.append(PresentationBlock.make_context(context))
+        
+        # Create buttons
+        buttons = [
+            PresentationButton(
+                label="Allow Once",
+                action=PresentationAction(
+                    type=ActionType.COMMAND,
+                    command=f"/approve {approval_id} allow-once"
+                ),
+                style=ButtonStyle.PRIMARY,
+                priority=10,
+            ),
+            PresentationButton(
+                label="Deny",
+                action=PresentationAction(
+                    type=ActionType.COMMAND,
+                    command=f"/approve {approval_id} deny"
+                ),
+                style=ButtonStyle.DANGER,
+                priority=9,
+            ),
+        ]
+        
+        if allow_always:
+            buttons.insert(1, PresentationButton(
+                label="Allow Always",
+                action=PresentationAction(
+                    type=ActionType.COMMAND,
+                    command=f"/approve {approval_id} allow-always"
+                ),
+                style=ButtonStyle.SUCCESS,
+                priority=8,
+            ))
+        
+        blocks.append(PresentationBlock.make_buttons(buttons))
+        
+        return MessagePresentation(blocks=blocks)
+
+
+@dataclass
+class PresentationLimits:
+    """Channel-specific limits for presentations.
+    
+    Attributes:
+        max_buttons: Maximum buttons per row/message
+        max_button_rows: Maximum button rows
+        max_button_label: Maximum characters in button label
+        max_options: Maximum options in select menu
+        max_option_label: Maximum characters in option label
+        max_text_length: Maximum text block length
+        supports_markdown: Whether channel supports markdown
+        supports_select: Whether channel supports select menus
+        supports_web_apps: Whether channel supports web apps
+    """
+    
+    max_buttons: int = 10
+    max_button_rows: int = 5
+    max_button_label: int = 50
+    max_options: int = 25
+    max_option_label: int = 50
+    max_text_length: int = 4096
+    supports_markdown: bool = True
+    supports_select: bool = True
+    supports_web_apps: bool = False
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        return {
+            "max_buttons": self.max_buttons,
+            "max_button_rows": self.max_button_rows,
+            "max_button_label": self.max_button_label,
+            "max_options": self.max_options,
+            "max_option_label": self.max_option_label,
+            "max_text_length": self.max_text_length,
+            "supports_markdown": self.supports_markdown,
+            "supports_select": self.supports_select,
+            "supports_web_apps": self.supports_web_apps,
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "PresentationLimits":
+        """Create from dictionary."""
+        return cls(
+            max_buttons=data.get("max_buttons", 10),
+            max_button_rows=data.get("max_button_rows", 5),
+            max_button_label=data.get("max_button_label", 50),
+            max_options=data.get("max_options", 25),
+            max_option_label=data.get("max_option_label", 50),
+            max_text_length=data.get("max_text_length", 4096),
+            supports_markdown=data.get("supports_markdown", True),
+            supports_select=data.get("supports_select", True),
+            supports_web_apps=data.get("supports_web_apps", False),
+        )
+    
+    @staticmethod
+    def telegram() -> "PresentationLimits":
+        """Get Telegram-specific limits."""
+        return PresentationLimits(
+            max_buttons=8,
+            max_button_rows=100,
+            max_button_label=64,
+            max_options=0,  # No native select
+            max_text_length=4096,
+            supports_markdown=True,
+            supports_select=False,
+            supports_web_apps=True,
+        )
+    
+    @staticmethod
+    def slack() -> "PresentationLimits":
+        """Get Slack-specific limits."""
+        return PresentationLimits(
+            max_buttons=5,
+            max_button_rows=1,
+            max_button_label=75,
+            max_options=100,
+            max_option_label=75,
+            max_text_length=3000,
+            supports_markdown=True,
+            supports_select=True,
+            supports_web_apps=False,
+        )
+    
+    @staticmethod
+    def discord() -> "PresentationLimits":
+        """Get Discord-specific limits."""
+        return PresentationLimits(
+            max_buttons=25,
+            max_button_rows=5,
+            max_button_label=80,
+            max_options=25,
+            max_option_label=100,
+            max_text_length=2000,
+            supports_markdown=True,
+            supports_select=True,
+            supports_web_apps=False,
+        )
+
+

src/praisonai-agents/praisonaiagents/bots/protocols.py

@@ -22,6 +22,7 @@ from typing import (
     List,
     Optional,
     Protocol,
+    TypedDict,
     Union,
     runtime_checkable,
 )
@@ -30,6 +31,105 @@ if TYPE_CHECKING:
     from ..agent import Agent
 
 
+class ChannelCapabilities(TypedDict, total=False):
+    """Declares what features a bot channel supports.
+    
+    This allows shared engines to adapt behavior based on platform capabilities,
+    enabling graceful degradation when features aren't available.
+    
+    Attributes:
+        live_edit: Whether the channel supports editing messages in place
+        reactions: Whether the channel supports adding reactions to messages
+        typing: Whether the channel supports typing indicators
+        text_limit: Maximum message length (0 = unlimited)
+        edit_rate_limit: Minimum seconds between edits (for throttling)
+        reaction_rate_limit: Minimum seconds between reactions
+    """
+    live_edit: bool
+    reactions: bool
+    typing: bool
+    text_limit: int
+    edit_rate_limit: float
+    reaction_rate_limit: float
+
+
+class RunStatus(Enum):
+    """Run status states for progress feedback.
+    
+    Used by status engines to show agent execution state through
+    reactions or status lines.
+    """
+    QUEUED = "queued"
+    THINKING = "thinking"
+    TOOL = "tool"
+    DONE = "done"
+    ERROR = "error"
+
+
+@dataclass(frozen=True)
+class PlatformCapabilities:
+    """Platform-specific capabilities descriptor for bot adapters.
+    
+    Declares what a messaging platform can do, enabling the shared delivery
+    machinery to apply features like streaming, rate limiting, and chunking
+    uniformly across all platforms.
+    
+    Attributes:
+        max_message_length: Maximum message length in the platform's unit
+        length_unit: Unit for message length ("codepoints" or "utf16")
+        supports_edit: Whether the platform supports in-place message edits (for streaming)
+        supports_typing: Whether the platform supports typing indicators
+        markdown_dialect: Markdown flavor the platform uses (e.g., "markdown", "telegram_markdown_v2")
+        needs_rate_limit: Whether the platform needs rate limiting
+        edit_interval_ms: Minimum milliseconds between message edits (for streaming)
+        max_files_per_message: Maximum number of file attachments per message
+        max_file_size_mb: Maximum file size in megabytes
+        supported_file_types: List of supported file extensions/mime types
+    """
+    
+    max_message_length: int = 4096
+    length_unit: str = "codepoints"  # "codepoints" or "utf16"
+    supports_edit: bool = False
+    supports_typing: bool = True
+    markdown_dialect: str = "markdown"
+    needs_rate_limit: bool = True
+    edit_interval_ms: int = 1000  # Minimum ms between edits
+    max_files_per_message: int = 1
+    max_file_size_mb: int = 10
+    supported_file_types: List[str] = field(default_factory=lambda: ["*"])
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "max_message_length": self.max_message_length,
+            "length_unit": self.length_unit,
+            "supports_edit": self.supports_edit,
+            "supports_typing": self.supports_typing,
+            "markdown_dialect": self.markdown_dialect,
+            "needs_rate_limit": self.needs_rate_limit,
+            "edit_interval_ms": self.edit_interval_ms,
+            "max_files_per_message": self.max_files_per_message,
+            "max_file_size_mb": self.max_file_size_mb,
+            "supported_file_types": self.supported_file_types,
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "PlatformCapabilities":
+        """Create from dictionary."""
+        return cls(
+            max_message_length=data.get("max_message_length", 4096),
+            length_unit=data.get("length_unit", "codepoints"),
+            supports_edit=data.get("supports_edit", False),
+            supports_typing=data.get("supports_typing", True),
+            markdown_dialect=data.get("markdown_dialect", "markdown"),
+            needs_rate_limit=data.get("needs_rate_limit", True),
+            edit_interval_ms=data.get("edit_interval_ms", 1000),
+            max_files_per_message=data.get("max_files_per_message", 1),
+            max_file_size_mb=data.get("max_file_size_mb", 10),
+            supported_file_types=data.get("supported_file_types", ["*"]),
+        )
+
+
 class MessageType(str, Enum):
     """Types of bot messages."""
     
@@ -318,6 +418,80 @@ class ProbeResult:
         }
 
 
+class HealthReason(Enum):
+    """Reasons for channel health status."""
+    HEALTHY = "healthy"
+    NOT_RUNNING = "not-running"
+    DISCONNECTED = "disconnected"
+    STALE_SOCKET = "stale-socket"
+    STUCK = "stuck"
+    BUSY = "busy"
+    STARTUP_GRACE = "startup-grace"
+    ERROR = "error"
+    
+    @property
+    def is_recoverable(self) -> bool:
+        """Whether this reason indicates a recoverable state."""
+        return self in {
+            HealthReason.DISCONNECTED,
+            HealthReason.STALE_SOCKET,
+            HealthReason.STUCK,
+            HealthReason.ERROR,
+        }
+
+
+def evaluate_channel_health(
+    health: HealthResult,
+    startup_grace_seconds: float = 60.0,
+    stale_after_seconds: float = 120.0,
+    current_time: Optional[float] = None,
+) -> HealthReason:
+    """Evaluate channel health and return a reason.
+    
+    Pure function that evaluates a HealthResult and determines
+    the health reason based on various criteria.
+    
+    Args:
+        health: The health result to evaluate
+        startup_grace_seconds: Grace period for startup
+        stale_after_seconds: Time after which no activity is considered stale
+        current_time: Current timestamp (for testing)
+        
+    Returns:
+        HealthReason indicating the channel's health status
+    """
+    if current_time is None:
+        current_time = time.time()
+    
+    # Not running
+    if not health.is_running:
+        return HealthReason.NOT_RUNNING
+    
+    # Startup grace period
+    if health.uptime_seconds is not None and health.uptime_seconds < startup_grace_seconds:
+        return HealthReason.STARTUP_GRACE
+    
+    # Check for errors
+    if health.error:
+        return HealthReason.ERROR
+    
+    # Check probe result
+    if health.probe and not health.probe.ok:
+        return HealthReason.DISCONNECTED
+    
+    # Check for stale socket (no transport activity)
+    if health.last_activity is not None:
+        time_since_activity = current_time - health.last_activity
+        if time_since_activity > stale_after_seconds:
+            return HealthReason.STALE_SOCKET
+    
+    # Overall health status
+    if not health.ok:
+        return HealthReason.ERROR
+    
+    return HealthReason.HEALTHY
+
+
 @dataclass
 class HealthResult:
     """Detailed health status of a bot.
@@ -331,6 +505,8 @@ class HealthResult:
         sessions: Number of active sessions
         error: Error message (if unhealthy)
         details: Additional platform-specific health details
+        reason: Health status reason (optional)
+        last_activity: Last transport activity timestamp (optional)
     """
     
     ok: bool
@@ -341,6 +517,8 @@ class HealthResult:
     sessions: int = 0
     error: Optional[str] = None
     details: Dict[str, Any] = field(default_factory=dict)
+    reason: Optional[HealthReason] = None
+    last_activity: Optional[float] = None
     
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary."""
@@ -353,6 +531,8 @@ class HealthResult:
             "sessions": self.sessions,
             "error": self.error,
             "details": self.details,
+            "reason": self.reason.value if self.reason else None,
+            "last_activity": self.last_activity,
         }
 
 
@@ -388,6 +568,24 @@ class BotProtocol(Protocol):
         """The bot's user information."""
         ...
     
+    @property
+    def capabilities(self) -> ChannelCapabilities:
+        """Channel capabilities for feature discovery.
+        
+        Returns capabilities that shared engines use to adapt behavior.
+        Channels that don't support a feature should return False for it.
+        """
+        ...
+    
+    @property
+    def platform_capabilities(self) -> PlatformCapabilities:
+        """Platform capabilities descriptor.
+        
+        Returns the platform's capabilities for use by shared delivery code.
+        Adapters should override this to declare their specific capabilities.
+        """
+        ...
+    
     # Lifecycle methods
     async def start(self) -> None:
         """Start the bot (begin receiving messages)."""
@@ -499,6 +697,33 @@ class BotProtocol(Protocol):
         """
         ...
     
+    # Reactions
+    async def add_reaction(self, channel_id: str, message_id: str, emoji: str) -> bool:
+        """Add a reaction emoji to a message.
+        
+        Args:
+            channel_id: Channel containing the message
+            message_id: Message to react to
+            emoji: Emoji to add (Unicode or custom emoji ID)
+            
+        Returns:
+            True if reaction was added successfully
+        """
+        ...
+    
+    async def remove_reaction(self, channel_id: str, message_id: str, emoji: str) -> bool:
+        """Remove a reaction emoji from a message.
+        
+        Args:
+            channel_id: Channel containing the message
+            message_id: Message to remove reaction from
+            emoji: Emoji to remove
+            
+        Returns:
+            True if reaction was removed successfully
+        """
+        ...
+    
     # User/channel info
     async def get_user(self, user_id: str) -> Optional[BotUser]:
         """Get user information.
@@ -613,6 +838,55 @@ class ChatCommandProtocol(Protocol):
         ...
 
 
+# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+# SupportsPresentation — interactive UI presentation protocol
+# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+
+@runtime_checkable
+class SupportsPresentation(Protocol):
+    """Protocol for channel adapters that support presentations.
+    
+    Channel adapters implement this protocol to render
+    portable presentations as native widgets.
+    """
+    
+    @property
+    def presentation_limits(self) -> "PresentationLimits":
+        """Get channel-specific presentation limits."""
+        ...
+    
+    async def render_presentation(
+        self,
+        target: str,
+        presentation: "MessagePresentation",
+    ) -> Optional[str]:
+        """Render a presentation to a target (chat/channel).
+        
+        Args:
+            target: Target identifier (chat_id, channel_id, etc.)
+            presentation: The presentation to render
+            
+        Returns:
+            Message ID if sent successfully, None otherwise
+        """
+        ...
+    
+    def truncate_presentation(
+        self,
+        presentation: "MessagePresentation",
+    ) -> "MessagePresentation":
+        """Truncate a presentation to fit channel limits.
+        
+        Args:
+            presentation: The presentation to truncate
+            
+        Returns:
+            Truncated presentation that fits channel limits
+        """
+        ...
+
+
 # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 # EmailProtocol — email-specific bot capabilities
 # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

src/praisonai-agents/praisonaiagents/bots/silence.py

@@ -0,0 +1,52 @@
+"""
+Silence protocol for bot gateways.
+
+Provides a canonical contract for agents to signal intentional silence
+(no reply) in group/ambient channels. Zero dependencies.
+"""
+
+SILENT_REPLY_TOKEN = "NO_REPLY"
+
+_MARKERS = frozenset({"NO_REPLY", "[SILENT]", "SILENT"})
+
+
+def is_intentional_silence_response(text: str | None) -> bool:
+    """Check if a response is an intentional silence signal.
+    
+    True only when the reply is *exactly* a silence marker (not prose mentioning it).
+    Blank/empty responses are NOT treated as deliberate silence.
+    
+    Args:
+        text: The response text to check
+        
+    Returns:
+        True if the response is exactly a silence token, False otherwise
+        
+    Examples:
+        >>> is_intentional_silence_response("NO_REPLY")
+        True
+        >>> is_intentional_silence_response("[SILENT]")
+        True
+        >>> is_intentional_silence_response("I think NO_REPLY is good")
+        False
+        >>> is_intentional_silence_response("")
+        False
+        >>> is_intentional_silence_response(None)
+        False
+    """
+    if not text:
+        return False  # blank != deliberate silence
+    
+    normalized = text.strip().upper()
+    
+    # Check for exact match
+    if normalized in _MARKERS:
+        return True
+    
+    # Check for bracket-wrapped version (only for tokens that aren't already bracketed)
+    if normalized.startswith("[") and normalized.endswith("]"):
+        inner = normalized[1:-1]
+        if inner == "SILENT":  # Only [SILENT] is a valid bracketed marker
+            return True
+    
+    return False

src/praisonai-agents/praisonaiagents/bus/bus.py

@@ -170,6 +170,17 @@ class EventBus:
         Returns:
             The published Event object
         """
+        # Fast path: if no subscribers, return a minimal event without expensive operations
+        if not self._subscribers:
+            # 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 {},
+            )
+        
         # Convert EventType enum to string
         type_str = event_type.value if isinstance(event_type, EventType) else event_type
         
@@ -194,6 +205,10 @@ class EventBus:
         """
         logger.debug(f"Publishing event: {event.type}")
         
+        # Fast path: if no subscribers, skip expensive work
+        if not self._subscribers:
+            return event
+        
         # Store in history
         with self._lock:
             self._event_history.append(event)
@@ -245,6 +260,15 @@ class EventBus:
         """
         type_str = event_type.value if isinstance(event_type, EventType) else event_type
         
+        # Fast path: if no subscribers, return a minimal event without expensive operations
+        if not self._subscribers:
+            return Event(
+                type=type_str,
+                data=data or {},
+                source=source,
+                metadata=metadata or {},
+            )
+        
         event = Event(
             type=type_str,
             data=data or {},
@@ -321,6 +345,11 @@ class EventBus:
         with self._lock:
             return len(self._subscribers)
     
+    @property 
+    def has_subscribers(self) -> bool:
+        """Check if there are any subscribers (lock-free for performance)."""
+        return bool(self._subscribers)
+    
     def __repr__(self) -> str:
         return f"EventBus(subscribers={self.subscriber_count})"
 

src/praisonai-agents/praisonaiagents/compaction/__init__.py

@@ -33,6 +33,10 @@ __all__ = [
     "CompactionStrategy",
     # Results
     "CompactionResult",
+    # Protocols
+    "ToolResultPrunerProtocol",
+    "MessageFormatterProtocol", 
+    "SummaryBuilderProtocol",
 ]
 
 
@@ -54,4 +58,16 @@ def __getattr__(name: str):
         from .result import CompactionResult
         return CompactionResult
     
+    if name == "ToolResultPrunerProtocol":
+        from .protocols import ToolResultPrunerProtocol
+        return ToolResultPrunerProtocol
+    
+    if name == "MessageFormatterProtocol":
+        from .protocols import MessageFormatterProtocol
+        return MessageFormatterProtocol
+    
+    if name == "SummaryBuilderProtocol":
+        from .protocols import SummaryBuilderProtocol
+        return SummaryBuilderProtocol
+    
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

src/praisonai-agents/praisonaiagents/compaction/compactor.py

@@ -5,12 +5,13 @@ Manages context window by compacting messages when needed.
 """
 
 import re
-from typing import List, Dict, Any, Optional, Callable, Awaitable
+from typing import List, Dict, Any, Optional, Callable, Awaitable, Tuple
 import asyncio
 
 from .config import CompactionConfig, COMPACTION_PREFIX, SUMMARY_TEMPLATE
 from .strategy import CompactionStrategy
 from .result import CompactionResult
+from .protocols import ToolResultPrunerProtocol, MessageFormatterProtocol, SummaryBuilderProtocol
 
 
 class ContextCompactor:
@@ -34,7 +35,10 @@ class ContextCompactor:
         preserve_system: bool = True,
         preserve_recent: int = 5,
         config: Optional[CompactionConfig] = None,
-        llm_summarize_fn: Optional[Callable[[str], Awaitable[str]]] = None
+        llm_summarize_fn: Optional[Callable[[str], Awaitable[str]]] = None,
+        tool_pruner: Optional[ToolResultPrunerProtocol] = None,
+        message_formatter: Optional[MessageFormatterProtocol] = None,
+        summary_builder: Optional[SummaryBuilderProtocol] = None
     ):
         """
         Initialize the compactor.
@@ -45,30 +49,43 @@ class ContextCompactor:
             strategy: Compaction strategy to use
             preserve_system: Keep system messages
             preserve_recent: Number of recent messages to preserve
-<<<<<<< HEAD
-            config: Optional CompactionConfig to override defaults
+            config: Optional CompactionConfig for advanced settings
             llm_summarize_fn: Async function to call LLM for summarization
+            tool_pruner: Protocol implementation for tool result pruning
+            message_formatter: Protocol implementation for message formatting
+            summary_builder: Protocol implementation for summary building
         """
-        # Initialize config first
-        self.config = config or CompactionConfig()
+        # Use provided config or create default
+        self.config = config or CompactionConfig(
+            max_tokens=max_tokens,
+            target_tokens=target_tokens or int(max_tokens * 0.75),
+            preserve_system=preserve_system,
+            preserve_recent=preserve_recent
+        )
         
-        # Use config values if config provided, otherwise use constructor args
-        if config is not None:
-            self.max_tokens = config.max_tokens
-            self.target_tokens = config.target_tokens
-            self.preserve_system = config.preserve_system
-            self.preserve_recent = config.preserve_recent
-        else:
-            self.max_tokens = max_tokens
-            self.target_tokens = target_tokens or int(max_tokens * 0.75)
-            self.preserve_system = preserve_system
-            self.preserve_recent = preserve_recent
-            
+        # Set instance attributes from config to ensure consistency
+        self.max_tokens = self.config.max_tokens
+        self.target_tokens = self.config.target_tokens
         self.strategy = strategy
+        self.preserve_system = self.config.preserve_system
+        self.preserve_recent = self.config.preserve_recent
         self.llm_summarize_fn = llm_summarize_fn
         
-        # Track previous summaries for iterative update
+        # Protocol implementations (defaults to None - no heavy implementations in core)
+        self.tool_pruner = tool_pruner
+        self.message_formatter = message_formatter  
+        self.summary_builder = summary_builder
+        
+        # Anti-thrashing state tracking
+        self._last_savings_pct: float = 100.0  # Start high to allow first compaction
+        self._low_savings_streak: int = 0
+        
+        # Iterative summary state
         self._previous_summary: Optional[str] = None
+        self._previous_summary_global_idx: int = 0
+        
+        # Reset state tracking for each compact() call
+        self._used_previous_summary: bool = False
     
     def estimate_tokens(self, text: str) -> int:
         """Estimate token count for text."""
@@ -96,12 +113,26 @@ class ContextCompactor:
         return sum(self.count_message_tokens(m) for m in messages)
     
     def needs_compaction(self, messages: List[Dict[str, Any]]) -> bool:
-        """Check if messages need compaction."""
-        return self.count_total_tokens(messages) > self.max_tokens
+        """
+        Check if messages need compaction with anti-thrashing protection.
+        
+        Returns False if:
+        - Token count is below threshold
+        - We've had too many consecutive low-savings attempts
+        """
+        if self.count_total_tokens(messages) <= self.max_tokens:
+            return False
+            
+        # Anti-thrashing: skip if we've had too many low-savings attempts
+        if self._low_savings_streak >= self.config.max_consecutive_low_savings:
+            return False
+            
+        return True
     
     def compact(
         self,
-        messages: List[Dict[str, Any]]
+        messages: List[Dict[str, Any]],
+        focus_topic: str = ""
     ) -> tuple[List[Dict[str, Any]], CompactionResult]:
         """
         Compact messages to fit within token limit (synchronous version).
@@ -111,31 +142,49 @@ class ContextCompactor:
         
         Args:
             messages: List of messages to compact
+            focus_topic: Optional topic to focus on during summarization
             
         Returns:
             Tuple of (compacted messages, result)
         """
+        # Reset state for this compaction call
+        self._used_previous_summary = False
+        
         original_tokens = self.count_total_tokens(messages)
         
         if original_tokens <= self.max_tokens:
-            return messages, CompactionResult(
+            result = CompactionResult(
                 original_tokens=original_tokens,
                 compacted_tokens=original_tokens,
                 messages_removed=0,
                 messages_kept=len(messages),
                 strategy_used=self.strategy
             )
+            result.calculate_savings_pct()
+            return messages, result
+        
+        # Apply tool result deduplication pre-pass if enabled
+        processed_messages = messages
+        tool_results_pruned = 0
+        if self.config.tool_prune_before_summarise and self.tool_pruner:
+            processed_messages, tool_results_pruned = self.tool_pruner.prune(
+                messages, 
+                self.config.max_tool_result_size
+            )
+        
+        # Skip early exit - proceed with full compaction strategy
+        # Anti-thrashing check will be applied after strategy runs
         
         if self.strategy == CompactionStrategy.TRUNCATE:
-            compacted = self._truncate(messages)
+            compacted = self._truncate(processed_messages)
         elif self.strategy == CompactionStrategy.SLIDING:
-            compacted = self._sliding_window(messages)
+            compacted = self._sliding_window(processed_messages)
         elif self.strategy == CompactionStrategy.SUMMARIZE:
-            compacted = self._summarize(messages)
+            compacted = self._summarize(processed_messages)
         elif self.strategy == CompactionStrategy.SMART:
-            compacted = self._smart_compact(messages)
+            compacted = self._smart_compact(processed_messages)
         elif self.strategy == CompactionStrategy.PRUNE:
-            compacted = self._prune(messages)
+            compacted = self._prune(processed_messages)
         elif self.strategy == CompactionStrategy.LLM_SUMMARIZE:
             if self.llm_summarize_fn:
                 # For sync calls with LLM function, we need to run async
@@ -144,17 +193,17 @@ class ContextCompactor:
                     try:
                         loop = asyncio.get_running_loop()
                         # If in async context, fallback to naive summarization
-                        compacted = self._summarize(messages)
+                        compacted = self._summarize(processed_messages)
                     except RuntimeError:
                         # No running loop, safe to create one
-                        compacted = asyncio.run(self._llm_summarize_async(messages))
+                        compacted = asyncio.run(self._llm_summarize_async(processed_messages, focus_topic))
                 except Exception:
                     # Fallback to naive summarization if async fails
-                    compacted = self._summarize(messages)
+                    compacted = self._summarize(processed_messages)
             else:
-                compacted = self._llm_summarize(messages)
+                compacted = self._llm_summarize(processed_messages, focus_topic)
         else:
-            compacted = self._truncate(messages)
+            compacted = self._truncate(processed_messages)
         
         compacted_tokens = self.count_total_tokens(compacted)
         
@@ -163,49 +212,75 @@ class ContextCompactor:
             compacted_tokens=compacted_tokens,
             messages_removed=len(messages) - len(compacted),
             messages_kept=len(compacted),
-            strategy_used=self.strategy
+            strategy_used=self.strategy,
+            tool_results_pruned=tool_results_pruned,
+            previous_summary_reused=getattr(self, '_used_previous_summary', False)
         )
+        result.calculate_savings_pct()
+        
+        # Update anti-thrashing tracking based on actual results
+        self._last_savings_pct = result.savings_pct
+        if result.savings_pct >= self.config.min_savings_pct:
+            self._low_savings_streak = 0
+        else:
+            self._low_savings_streak += 1
         
         return compacted, result
 
     async def compact_async(
         self,
-        messages: List[Dict[str, Any]]
+        messages: List[Dict[str, Any]],
+        focus_topic: str = ""
     ) -> tuple[List[Dict[str, Any]], CompactionResult]:
         """
         Compact messages to fit within token limit (asynchronous version).
         
         Args:
             messages: List of messages to compact
+            focus_topic: Optional topic to focus on during summarization
             
         Returns:
             Tuple of (compacted messages, result)
         """
+        # Reset state for this compaction call
+        self._used_previous_summary = False
+        
         original_tokens = self.count_total_tokens(messages)
         
         if original_tokens <= self.max_tokens:
-            return messages, CompactionResult(
+            result = CompactionResult(
                 original_tokens=original_tokens,
                 compacted_tokens=original_tokens,
                 messages_removed=0,
                 messages_kept=len(messages),
                 strategy_used=self.strategy
             )
+            result.calculate_savings_pct()
+            return messages, result
+        
+        # Apply tool result deduplication pre-pass if enabled
+        processed_messages = messages
+        tool_results_pruned = 0
+        if self.config.tool_prune_before_summarise and self.tool_pruner:
+            processed_messages, tool_results_pruned = self.tool_pruner.prune(
+                messages, 
+                self.config.max_tool_result_size
+            )
         
         if self.strategy == CompactionStrategy.TRUNCATE:
-            compacted = self._truncate(messages)
+            compacted = self._truncate(processed_messages)
         elif self.strategy == CompactionStrategy.SLIDING:
-            compacted = self._sliding_window(messages)
+            compacted = self._sliding_window(processed_messages)
         elif self.strategy == CompactionStrategy.SUMMARIZE:
-            compacted = self._summarize(messages)
+            compacted = self._summarize(processed_messages)
         elif self.strategy == CompactionStrategy.SMART:
-            compacted = self._smart_compact(messages)
+            compacted = self._smart_compact(processed_messages)
         elif self.strategy == CompactionStrategy.PRUNE:
-            compacted = self._prune(messages)
+            compacted = self._prune(processed_messages)
         elif self.strategy == CompactionStrategy.LLM_SUMMARIZE:
-            compacted = await self._llm_summarize_async(messages)
+            compacted = await self._llm_summarize_async(processed_messages, focus_topic)
         else:
-            compacted = self._truncate(messages)
+            compacted = self._truncate(processed_messages)
         
         compacted_tokens = self.count_total_tokens(compacted)
         
@@ -214,11 +289,35 @@ class ContextCompactor:
             compacted_tokens=compacted_tokens,
             messages_removed=len(messages) - len(compacted),
             messages_kept=len(compacted),
-            strategy_used=self.strategy
+            strategy_used=self.strategy,
+            tool_results_pruned=tool_results_pruned,
+            previous_summary_reused=getattr(self, '_used_previous_summary', False)
         )
+        result.calculate_savings_pct()
+        
+        # Update anti-thrashing tracking based on actual results
+        self._last_savings_pct = result.savings_pct
+        if result.savings_pct >= self.config.min_savings_pct:
+            self._low_savings_streak = 0
+        else:
+            self._low_savings_streak += 1
         
         return compacted, result
     
+    def _prune_tool_results(self, messages: List[Dict[str, Any]]) -> Tuple[List[Dict[str, Any]], int]:
+        """
+        Delegate tool result pruning to injected protocol implementation.
+        
+        If no tool pruner is provided, returns messages unchanged.
+        This maintains backward compatibility while enforcing protocol-driven design.
+        
+        Returns:
+            Tuple of (processed messages, number of tool results pruned)
+        """
+        if self.tool_pruner:
+            return self.tool_pruner.prune(messages, self.config.max_tool_result_size)
+        return messages, 0
+    
     def _truncate(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         """Truncate oldest messages."""
         result = []
@@ -244,11 +343,17 @@ class ContextCompactor:
         
         # If still over limit, truncate more
         while self.count_total_tokens(result) > self.target_tokens and len(result) > 1:
-            # Remove oldest non-system message
+            # Remove oldest message (respecting preserve_system setting)
+            removed = False
             for i, msg in enumerate(result):
-                if msg.get("role") != "system":
+                # Skip system messages only if preserve_system is True
+                if not (self.preserve_system and msg.get("role") == "system"):
                     result.pop(i)
+                    removed = True
                     break
+            if not removed:
+                # Only system messages remain but still over budget — stop to avoid infinite loop
+                break
         
         return result
     
@@ -360,72 +465,19 @@ class ContextCompactor:
         result.extend(recent)
         return result
     
-    def _llm_summarize(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    def _llm_summarize(self, messages: List[Dict[str, Any]], focus_topic: str = "") -> List[Dict[str, Any]]:
         """
-        Use LLM to summarize older messages with anti-injection framing.
+        Use LLM to summarize older messages with iterative support.
         
-        Note: This implementation uses structured templates and anti-injection prefixes
-        to prevent the model from treating summarized content as active instructions.
-        """
-        result = []
+        Supports iterative summarization - if we have a previous summary,
+        we only summarize the new messages since that summary.
+        Also includes anti-injection framing when configured.
         
-        # Keep system messages (excluding previous compacted summaries)
-        system_msgs = [
-            m for m in messages 
-            if m.get("role") == "system" and not m.get("_compacted")
-        ]
-        other_msgs = [m for m in messages if m.get("role") != "system"]
-        
-        result.extend(system_msgs)
-        
-        # Keep recent messages
-        recent = other_msgs[-self.preserve_recent:]
-        older = other_msgs[:-self.preserve_recent] if len(other_msgs) > self.preserve_recent else []
-        
-        if older:
-            if self.config.structured_template:
-                structured = self._build_structured_summary(older)
-            else:
-                # Fallback to simple summary
-                summary_parts = []
-                for msg in older:
-                    role = msg.get("role", "unknown")
-                    content = msg.get("content", "")
-                    if isinstance(content, str) and content:
-                        summary_parts.append(f"[{role}]: {content[:150]}...")
-                structured = "\n".join(summary_parts[:10])
-            
-            if structured:
-                # Apply anti-injection prefix
-                prefixed = f"{self.config.compaction_prefix}\n\n{structured}"
-                
-                # Handle iterative update if enabled
-                if self.config.iterative_update and self._previous_summary:
-                    structured = self._merge_summaries(self._previous_summary, structured)
-                    prefixed = f"{self.config.compaction_prefix}\n\n{structured}"
-                
-                # Store for next iteration
-                if self.config.iterative_update:
-                    self._previous_summary = structured
-                
-                result.append({
-                    "role": "system",
-                    "content": prefixed,
-                    "_compacted": True,
-                    "_original_count": len(older),
-                    "_anti_injection": True,
-                })
-        
-        result.extend(recent)
-        return result
-
-    async def _llm_summarize_async(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-        """
-        Use LLM to intelligently summarize older messages.
-        
-        This method invokes the agent's LLM to create a meaningful summary
-        that preserves key facts, identifiers, and the user's intent.
+        Args:
+            messages: List of messages to compact
+            focus_topic: Optional topic to focus on during summarization
         """
+        self._used_previous_summary = False
         result = []
         
         # Keep system messages
@@ -436,33 +488,169 @@ class ContextCompactor:
         
         # Keep recent messages
         recent = other_msgs[-self.preserve_recent:]
-        older = other_msgs[:-self.preserve_recent] if len(other_msgs) > self.preserve_recent else []
+        
+        # Determine what to summarize
+        total_original_messages = len(messages)
+        if (self.config.enable_iterative_summary and 
+            self._previous_summary and 
+            total_original_messages > self._previous_summary_global_idx):
+            # Iterative: summarize only new messages since previous summary
+            # Calculate how many old messages to summarize based on global position
+            messages_since_summary = total_original_messages - self._previous_summary_global_idx
+            new_older_messages = max(0, messages_since_summary - self.preserve_recent)
+            if new_older_messages > 0:
+                to_summarize = other_msgs[-messages_since_summary:-self.preserve_recent]
+            else:
+                to_summarize = []
+            self._used_previous_summary = True
+        else:
+            # Fresh summary: summarize all older messages
+            to_summarize = other_msgs[:-self.preserve_recent] if len(other_msgs) > self.preserve_recent else []
+            self._used_previous_summary = False
+        
+        if to_summarize or self._previous_summary:
+            # Build summary content
+            summary_parts = []
+            
+            # Include previous summary if doing iterative summarization
+            if self._used_previous_summary and self._previous_summary:
+                summary_parts.append(f"[Previous Summary]: {self._previous_summary}")
+                summary_parts.append("[New Activity]:")
+            
+            # Add new messages to summarize
+            for msg in to_summarize:
+                role = msg.get("role", "unknown")
+                content = msg.get("content", "")
+                if isinstance(content, str) and content:
+                    # Extract key information with focus consideration
+                    if focus_topic and focus_topic.lower() in content.lower():
+                        # Prioritize content related to focus topic
+                        summary_parts.append(f"[{role}] *FOCUS*: {content[:200]}...")
+                    else:
+                        summary_parts.append(f"[{role}]: {content[:150]}...")
+            
+            if summary_parts:
+                focus_hint = f"\nFocus on: {focus_topic}" if focus_topic else ""
+                
+                if self._used_previous_summary:
+                    summary = (
+                        f"[Iterative conversation summary - update with new activity{focus_hint}]\n"
+                        + "\n".join(summary_parts[:15])
+                    )
+                else:
+                    summary = (
+                        f"[Compacted conversation history - summarize key points{focus_hint}]\n"
+                        + "\n".join(summary_parts[:10])
+                    )
+                
+                # Apply anti-injection prefix if configured
+                if hasattr(self.config, 'compaction_prefix') and self.config.compaction_prefix:
+                    summary = f"{self.config.compaction_prefix}\n\n{summary}"
+                
+                summary_msg = {
+                    "role": "system",
+                    "content": summary,
+                    "_compacted": True,
+                    "_original_count": len(to_summarize),
+                    "_iterative": self._used_previous_summary,
+                    "_focus_topic": focus_topic
+                }
+                result.append(summary_msg)
+                
+                # Update tracking for future iterative summarization
+                self._previous_summary = summary
+                self._previous_summary_global_idx = len(messages)
+        
+        result.extend(recent)
+        return result
+
+    async def _llm_summarize_async(self, messages: List[Dict[str, Any]], focus_topic: str = "") -> List[Dict[str, Any]]:
+        """
+        Use LLM to intelligently summarize older messages.
+        
+        This method invokes the agent's LLM to create a meaningful summary
+        that preserves key facts, identifiers, and the user's intent.
+        
+        Args:
+            messages: List of messages to compact
+            focus_topic: Optional topic to focus on during summarization
+        """
+        self._used_previous_summary = False
+        result = []
+        
+        # Keep system messages
+        system_msgs = [m for m in messages if m.get("role") == "system"]
+        other_msgs = [m for m in messages if m.get("role") != "system"]
+        
+        result.extend(system_msgs)
+        
+        # Keep recent messages
+        recent = other_msgs[-self.preserve_recent:]
+        
+        # Determine what to summarize based on iterative settings
+        total_original_messages = len(messages)
+        if (self.config.enable_iterative_summary and 
+            self._previous_summary and 
+            total_original_messages > self._previous_summary_global_idx):
+            # Iterative: summarize only new messages since previous summary
+            messages_since_summary = total_original_messages - self._previous_summary_global_idx
+            new_older_messages = max(0, messages_since_summary - self.preserve_recent)
+            if new_older_messages > 0:
+                older = other_msgs[-messages_since_summary:-self.preserve_recent]
+            else:
+                older = []
+            self._used_previous_summary = True
+        else:
+            # Fresh summary: summarize all older messages
+            older = other_msgs[:-self.preserve_recent] if len(other_msgs) > self.preserve_recent else []
+            self._used_previous_summary = False
         
         if older and self.llm_summarize_fn:
             try:
-                # Format messages for summarization
-                history_text = self._format_messages_for_summary(older)
+                # Format messages for summarization using protocol
+                if self.message_formatter:
+                    history_text = self.message_formatter.format_for_summary(older)
+                else:
+                    # Fallback to basic formatting if no protocol implementation
+                    history_text = str(older)
                 
                 # Create summarization prompt that preserves important information
+                focus_hint = f" Focus especially on: {focus_topic}." if focus_topic else ""
+                iterative_hint = ""
+                if self._used_previous_summary and self._previous_summary:
+                    iterative_hint = f"\n\n[Previous Summary]: {self._previous_summary}\n[New Activity to Add]:"
+                
                 prompt = (
                     "Summarise the following conversation history. Preserve verbatim: "
                     "all file paths, IDs, hashes, URLs, task references, error messages, "
                     "tool outputs, and the user's requests. Be concise but complete. "
-                    "Focus on facts and actions taken, not general conversation.\n\n"
-                    f"{history_text}"
+                    f"Focus on facts and actions taken, not general conversation.{focus_hint}"
+                    f"{iterative_hint}\n\n{history_text}"
                 )
                 
                 # Call the LLM for summarization
                 summary = await self.llm_summarize_fn(prompt)
                 
                 # Add the LLM-generated summary as a system message
+                summary_content = summary
+                if self._used_previous_summary:
+                    summary_content = f"[Iterative conversation summary]\n{summary}"
+                else:
+                    summary_content = f"[Previous conversation summary]\n{summary}"
+                
                 result.append({
                     "role": "system",
-                    "content": f"[Previous conversation summary]\n{summary}",
+                    "content": summary_content,
                     "_compacted": True,
                     "_original_count": len(older),
                     "_llm_generated": True,
+                    "_iterative": self._used_previous_summary,
+                    "_focus_topic": focus_topic
                 })
+                
+                # Update tracking for future iterative summarization
+                self._previous_summary = summary
+                self._previous_summary_global_idx = len(messages)
             except Exception as e:
                 # Fallback to naive summarization if LLM call fails
                 import logging
@@ -506,133 +694,41 @@ class ContextCompactor:
                     "content": summary,
                     "_compacted": True,
                     "_original_count": len(older),
+                    "_fallback": True,
                 })
         
         result.extend(recent)
         return result
 
     def _format_messages_for_summary(self, messages: List[Dict[str, Any]]) -> str:
-        """Format messages for LLM summarization."""
-        formatted = []
-        for i, msg in enumerate(messages):
-            role = msg.get("role", "unknown")
-            content = msg.get("content", "")
-            
-            # Handle tool calls
-            if msg.get("tool_calls"):
-                tool_calls = msg.get("tool_calls", [])
-                tool_names = [tc.get("function", {}).get("name", "unknown") for tc in tool_calls]
-                formatted.append(f"{i+1}. {role}: Called tools: {', '.join(tool_names)}")
-            
-            # Handle content
-            if isinstance(content, str) and content.strip():
-                # Truncate very long content but preserve structure
-                if len(content) > 1000:
-                    content = content[:800] + "...[truncated]..." + content[-200:]
-                formatted.append(f"{i+1}. {role}: {content}")
-            elif isinstance(content, list):
-                # Handle multi-part content
-                parts = []
-                for part in content[:3]:  # Limit to first 3 parts
-                    if isinstance(part, dict):
-                        text = part.get("text", "")
-                        if text:
-                            parts.append(text[:200] + "..." if len(text) > 200 else text)
-                if parts:
-                    formatted.append(f"{i+1}. {role}: {' | '.join(parts)}")
-            
-            # Handle tool results with IDs
-            if msg.get("tool_call_id"):
-                tool_id = msg.get("tool_call_id", "")
-                formatted.append(f"{i+1}. {role} (tool {tool_id}): {str(content)[:500]}...")
+        """
+        Delegate message formatting to protocol implementation.
         
-        return "\n".join(formatted)
+        If no formatter is provided, returns basic string representation.
+        """
+        if self.message_formatter:
+            return self.message_formatter.format_for_summary(messages)
+        return str(messages)
     
     def _build_structured_summary(self, messages: List[Dict[str, Any]]) -> str:
         """
-        Build a structured summary using the configured template.
+        Delegate structured summary building to protocol implementation.
         
-        Args:
-            messages: Messages to summarize
-            
-        Returns:
-            Structured summary string
+        If no summary builder is provided, returns basic summary.
         """
-        # Extract information from messages
-        active_task = "No specific task identified"
-        completed = []
-        in_progress = []
-        pending = []
-        files = set()
-        remaining = []
-        
-        for msg in messages:
-            role = msg.get("role", "")
-            content = msg.get("content", "")
-            
-            if not isinstance(content, str):
-                continue
-                
-            content_lower = content.lower()
-            
-            # Extract file paths
-            file_matches = re.findall(r'[\w/\.\-]+\.[a-zA-Z]{1,4}', content)
-            files.update(file_matches[:5])  # Limit to 5 files
-            
-            # Categorize content based on keywords and role
-            content_snippet = content[:200] + ("..." if len(content) > 200 else "")
-            
-            if role == "user":
-                if any(word in content_lower for word in ["please", "can you", "help"]) or re.search(r'\bdo\b', content_lower):
-                    active_task = content_snippet
-                elif "?" in content:
-                    pending.append(content_snippet)
-                    
-            elif role == "assistant":
-                if any(word in content_lower for word in ["completed", "done", "finished"]):
-                    completed.append(content_snippet)
-                elif any(word in content_lower for word in ["working", "processing", "analyzing"]):
-                    in_progress.append(content_snippet)
-                elif any(word in content_lower for word in ["will", "plan to", "next"]):
-                    remaining.append(content_snippet)
-                    
-            elif role == "tool":
-                # Tool results are generally completed actions
-                tool_name = msg.get("name", "unknown")
-                completed.append(f"Tool {tool_name}: {content_snippet}")
-        
-        # Format using template
-        return SUMMARY_TEMPLATE.format(
-            active_task=active_task,
-            completed="\n".join(f"- {item}" for item in completed[:3]) or "None identified",
-            in_progress="\n".join(f"- {item}" for item in in_progress[:3]) or "None identified",
-            pending="\n".join(f"- {item}" for item in pending[:3]) or "None identified",
-            files=", ".join(list(files)[:5]) or "None mentioned",
-            remaining="\n".join(f"- {item}" for item in remaining[:3]) or "None identified"
-        )
+        if self.summary_builder:
+            return self.summary_builder.build_structured_summary(messages)
+        return f"Summary of {len(messages)} messages"
     
     def _merge_summaries(self, previous: str, current: str) -> str:
         """
-        Merge previous summary with current one for iterative updates.
+        Delegate summary merging to protocol implementation.
         
-        Args:
-            previous: Previous summary content
-            current: Current summary content
-            
-        Returns:
-            Merged summary
+        If no summary builder is provided, returns current summary only.
         """
-        # Simple merge strategy: prioritize current but preserve unique info from previous
-        if not previous:
-            return current
-            
-        # Preserve context but avoid excessive growth
-        MAX_PREVIOUS_LENGTH = 500
-        if len(previous) > MAX_PREVIOUS_LENGTH:
-            truncated_previous = previous[:MAX_PREVIOUS_LENGTH]
-            return f"{current}\n\n[Previous context]: {truncated_previous}..."
-        else:
-            return f"{current}\n\n[Previous context]: {previous}"
+        if self.summary_builder:
+            return self.summary_builder.merge_summaries(previous, current)
+        return current
     
     def get_stats(self, messages: List[Dict[str, Any]]) -> Dict[str, Any]:
         """Get statistics about messages."""

src/praisonai-agents/praisonaiagents/compaction/config.py

@@ -50,3 +50,28 @@ class CompactionConfig:
     compaction_prefix: str = COMPACTION_PREFIX  # Override for custom framing
     structured_template: bool = True  # Emit Active Task / Remaining Work sections
     iterative_update: bool = True  # Merge previous summary on re-compaction
+    
+    # Anti-thrashing protection
+    min_savings_pct: float = 10.0  # Skip if projected saving < 10% (0-100 scale)
+    max_consecutive_low_savings: int = 2  # Abort after N low-savings attempts
+    
+    # Tool result optimization
+    tool_prune_before_summarise: bool = True  # Deduplicate tool results before summarization
+    max_tool_result_size: int = 500  # Maximum size for tool results before pruning
+    
+    # Iterative summarization
+    enable_iterative_summary: bool = True  # Build on previous summaries instead of starting from scratch
+    
+    def __post_init__(self):
+        """Validate and normalize configuration values."""
+        # Normalize min_savings_pct to 0-100 scale if provided as ratio
+        if self.min_savings_pct < 1.0:
+            self.min_savings_pct *= 100.0
+        
+        # Validate ranges
+        if not 0.0 <= self.min_savings_pct <= 100.0:
+            raise ValueError("min_savings_pct must be between 0 and 100")
+        if self.max_consecutive_low_savings < 0:
+            raise ValueError("max_consecutive_low_savings must be >= 0")
+        if self.max_tool_result_size <= 0:
+            raise ValueError("max_tool_result_size must be > 0")

src/praisonai-agents/praisonaiagents/compaction/protocols.py

@@ -0,0 +1,73 @@
+"""
+Compaction Protocols for PraisonAI Agents.
+
+Protocol-driven interfaces for context compaction functionality.
+"""
+
+from typing import Protocol, List, Dict, Any, Tuple, Optional
+
+
+class ToolResultPrunerProtocol(Protocol):
+    """Protocol for pruning and deduplicating tool results."""
+    
+    def prune(
+        self,
+        messages: List[Dict[str, Any]],
+        max_tool_result_size: int = 500
+    ) -> Tuple[List[Dict[str, Any]], int]:
+        """
+        Prune and deduplicate tool results to reduce token usage.
+        
+        Args:
+            messages: List of messages to process
+            max_tool_result_size: Maximum size for tool results before truncation
+            
+        Returns:
+            Tuple of (processed messages, number of tool results pruned)
+        """
+        ...
+
+
+class MessageFormatterProtocol(Protocol):
+    """Protocol for formatting messages for summarization."""
+    
+    def format_for_summary(self, messages: List[Dict[str, Any]]) -> str:
+        """
+        Format messages for LLM summarization.
+        
+        Args:
+            messages: Messages to format
+            
+        Returns:
+            Formatted string suitable for summarization
+        """
+        ...
+
+
+class SummaryBuilderProtocol(Protocol):
+    """Protocol for building structured summaries."""
+    
+    def build_structured_summary(self, messages: List[Dict[str, Any]]) -> str:
+        """
+        Build a structured summary from messages.
+        
+        Args:
+            messages: Messages to summarize
+            
+        Returns:
+            Structured summary string
+        """
+        ...
+    
+    def merge_summaries(self, previous: str, current: str) -> str:
+        """
+        Merge previous summary with current one for iterative updates.
+        
+        Args:
+            previous: Previous summary content
+            current: Current summary content
+            
+        Returns:
+            Merged summary
+        """
+        ...

src/praisonai-agents/praisonaiagents/compaction/result.py

@@ -18,6 +18,16 @@ class CompactionResult:
     strategy_used: CompactionStrategy
     summary: str = ""
     
+    # Anti-thrashing and optimization info
+    savings_pct: float = field(default=0.0, init=False)  # Percentage of tokens saved (computed)
+    tool_results_pruned: int = 0  # Number of tool results deduplicated/pruned
+    previous_summary_reused: bool = False  # Whether iterative summary was used
+    was_skipped_due_to_low_savings: bool = False  # Whether compaction was skipped due to anti-thrashing
+    
+    def __post_init__(self):
+        """Initialize computed fields after construction."""
+        self.calculate_savings_pct()
+    
     @property
     def tokens_saved(self) -> int:
         """Get number of tokens saved."""
@@ -30,6 +40,13 @@ class CompactionResult:
             return 0.0
         return self.compacted_tokens / self.original_tokens
     
+    def calculate_savings_pct(self) -> None:
+        """Calculate and set the savings percentage."""
+        if self.original_tokens > 0:
+            self.savings_pct = (self.tokens_saved / self.original_tokens) * 100.0
+        else:
+            self.savings_pct = 0.0
+    
     @property
     def was_compacted(self) -> bool:
         """Check if compaction occurred."""
@@ -45,5 +62,9 @@ class CompactionResult:
             "messages_kept": self.messages_kept,
             "strategy_used": self.strategy_used.value,
             "compression_ratio": self.compression_ratio,
+            "savings_pct": self.savings_pct,
+            "tool_results_pruned": self.tool_results_pruned,
+            "previous_summary_reused": self.previous_summary_reused,
+            "was_skipped_due_to_low_savings": self.was_skipped_due_to_low_savings,
             "summary": self.summary
         }

src/praisonai-agents/praisonaiagents/config/__init__.py

@@ -26,6 +26,7 @@ __all__ = [
     "OutputPreset",
     "ExecutionPreset",
     # Config classes
+    "LLMConfig",
     "MemoryConfig",
     "KnowledgeConfig",
     "PlanningConfig",
@@ -110,6 +111,7 @@ _MODULE_MAP = {
     "WebSearchProvider": "feature_configs",
     "OutputPreset": "feature_configs",
     "ExecutionPreset": "feature_configs",
+    "LLMConfig": "feature_configs",
     "MemoryConfig": "feature_configs",
     "KnowledgeConfig": "feature_configs",
     "PlanningConfig": "feature_configs",

src/praisonai-agents/praisonaiagents/config/feature_configs.py

@@ -724,6 +724,9 @@ class ExecutionConfig:
     
     # Retry settings
     max_retry_limit: int = 2
+    retry_initial_delay: float = 1.0  # seconds
+    retry_backoff_factor: float = 2.0
+    retry_jitter: float = 0.1  # fraction of computed delay
     
     # Tool call limits (loop protection)
     max_tool_calls_per_turn: int = 10
@@ -766,7 +769,7 @@ class ExecutionConfig:
     parallel_tool_calls: bool = False
 
     def __post_init__(self) -> None:
-        """Post-initialization processing with deprecation warnings."""
+        """Post-initialization processing with deprecation warnings and validation."""
         # Handle context_compaction serialization round-trip
         if isinstance(self.context_compaction, dict):
             from ..context.policy import ContextCompactionPolicy
@@ -774,29 +777,47 @@ class ExecutionConfig:
                 self.context_compaction
             )
         
-        # Emit deprecation warning for default behavior change
-        # Only warn if context_compaction was not explicitly set (i.e., using the default False)
-        import warnings
-        import inspect
-        frame = inspect.currentframe()
-        try:
-            # Check if this is being called from user code (not internal)
-            caller_frame = frame.f_back
-            if (caller_frame and 
-                not any(path_part in caller_frame.f_code.co_filename 
-                       for path_part in ['praisonaiagents', 'test_', '__pycache__'])):
-                # This is being called from user code
-                if self.context_compaction is False:
-                    warnings.warn(
-                        "ExecutionConfig.context_compaction will default to True in the next "
-                        "release for proactive context overflow protection. To disable, explicitly "
-                        "set context_compaction=False. To use the new default early, set "
-                        "context_compaction=True.",
-                        DeprecationWarning,
-                        stacklevel=3
-                    )
-        finally:
-            del frame
+        # Emit deprecation warning once per process for default behavior change.
+        # Walk the stack once — dataclass __init__ uses caller filename "<string>".
+        if self.context_compaction is False:
+            if getattr(ExecutionConfig, '_context_compaction_internal', False):
+                return
+            if getattr(ExecutionConfig, '_context_compaction_warned', False):
+                return
+            import warnings
+            import inspect
+            _INTERNAL = ('praisonaiagents', 'test_', '__pycache__')
+            frame = inspect.currentframe()
+            try:
+                caller = frame.f_back if frame else None
+                is_internal = False
+                while caller is not None:
+                    if any(p in caller.f_code.co_filename for p in _INTERNAL):
+                        is_internal = True
+                        break
+                    caller = caller.f_back
+                if is_internal:
+                    ExecutionConfig._context_compaction_internal = True
+                    return
+                warnings.warn(
+                    "ExecutionConfig.context_compaction will default to True in the next "
+                    "release for proactive context overflow protection. To disable, explicitly "
+                    "set context_compaction=False. To use the new default early, set "
+                    "context_compaction=True.",
+                    DeprecationWarning,
+                    stacklevel=3,
+                )
+                ExecutionConfig._context_compaction_warned = True
+            finally:
+                del frame
+        
+        # Validate retry configuration parameters
+        if self.retry_initial_delay <= 0:
+            raise ValueError("ExecutionConfig.retry_initial_delay must be positive.")
+        if self.retry_backoff_factor < 1.0:
+            raise ValueError("ExecutionConfig.retry_backoff_factor must be >= 1.0.")
+        if self.retry_jitter < 0:
+            raise ValueError("ExecutionConfig.retry_jitter must be non-negative.")
 
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary."""
@@ -805,6 +826,9 @@ class ExecutionConfig:
             "max_rpm": self.max_rpm,
             "max_execution_time": self.max_execution_time,
             "max_retry_limit": self.max_retry_limit,
+            "retry_initial_delay": self.retry_initial_delay,
+            "retry_backoff_factor": self.retry_backoff_factor,
+            "retry_jitter": self.retry_jitter,
             "code_execution": self.code_execution,
             "code_mode": self.code_mode,
             "code_sandbox_mode": self.code_sandbox_mode,
@@ -843,6 +867,69 @@ class ExecutionConfig:
         )
 
 
+@dataclass
+class LLMConfig:
+    """
+    Configuration for LLM model settings.
+    
+    Groups all LLM-related parameters including model selection, 
+    API settings, and fallback configuration.
+    
+    Usage:
+        # With primary model only
+        Agent(llm_config=LLMConfig(model="gpt-4o"))
+        
+        # With fallback chain
+        Agent(llm_config=LLMConfig(
+            model="gpt-4o",
+            fallback_models=["claude-3-5-sonnet", "gpt-4o-mini"],
+            base_url="https://api.example.com",
+            api_key="sk-...",
+        ))
+        
+        # With LiteLLM-style provider prefix
+        Agent(llm_config=LLMConfig(
+            model="anthropic/claude-3-5-sonnet",
+            fallback_models=["openai/gpt-4o", "openai/gpt-4o-mini"],
+        ))
+    """
+    # Primary model to use (required)
+    model: str
+    
+    # Ordered fallback models for resilience (optional)
+    fallback_models: Optional[List[str]] = None
+    
+    # API endpoint override (optional)
+    base_url: Optional[str] = None
+    
+    # API key (optional, defaults to env vars)
+    api_key: Optional[str] = None
+    
+    # Additional auth headers (optional)
+    auth: Optional[Dict[str, str]] = None
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "model": self.model,
+            "fallback_models": list(self.fallback_models) if self.fallback_models else None,
+            "base_url": self.base_url,
+            "api_key": self.api_key,
+            "auth": dict(self.auth) if self.auth else None,
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "LLMConfig":
+        """Create LLMConfig from dictionary."""
+        return cls(
+            model=data["model"],
+            fallback_models=list(data["fallback_models"]) if data.get("fallback_models") else None,
+            base_url=data.get("base_url"),
+            api_key=data.get("api_key"),
+            auth=dict(data["auth"]) if data.get("auth") else None,
+        )
+
+
 @dataclass  
 class ToolConfig:
     """
@@ -1089,8 +1176,8 @@ class MultiAgentOutputConfig:
     # Verbosity level (0=silent, 1=minimal, 2+=verbose)
     verbose: int = 0
     
-    # Streaming output
-    stream: bool = True
+    # Streaming output - False by default for sync multi-agent compatibility
+    stream: bool = False
     
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary."""
@@ -1265,79 +1352,49 @@ def resolve_memory(value: MemoryParam) -> Optional[MemoryConfig]:
     """
     Resolve memory= parameter following precedence ladder.
     
-    Precedence: Instance > Config > Dict > String > Bool > Default
+    Delegates to the canonical resolver in param_resolver.py with
+    special handling for backward compatibility.
     
     Args:
         value: Memory parameter in any supported form
         
     Returns:
         MemoryConfig if enabled, None if disabled
-        
-    Examples:
-        >>> resolve_memory(None)  # Default: disabled
-        None
-        >>> resolve_memory(False)  # Explicit disable
-        None
-        >>> resolve_memory(True)  # Enable with defaults
-        MemoryConfig(backend='file')
-        >>> resolve_memory("redis")  # String shorthand
-        MemoryConfig(backend='redis')
-        >>> resolve_memory({"backend": "sqlite", "user_id": "alice"})  # Dict
-        MemoryConfig(backend='sqlite', user_id='alice')
-        >>> resolve_memory(MemoryConfig(backend="postgres"))  # Config passthrough
-        MemoryConfig(backend='postgres')
     """
-    # Default: disabled
-    if value is None:
-        return None
+    from .param_resolver import resolve_memory as _resolve
     
-    # Bool: False = disabled, True = defaults
-    if value is False:
-        return None
-    if value is True:
-        return MemoryConfig()
-    
-    # String: backend shorthand
+    # Special case: handle string backends not in MEMORY_PRESETS
     if isinstance(value, str):
         try:
-            backend = MemoryBackend(value.lower())
+            # Try to resolve with canonical resolver first
+            return _resolve(value, MemoryConfig)
         except ValueError:
-            backend = value  # Allow custom backend strings
-        return MemoryConfig(backend=backend)
-    
-    # Dict: expand to config
-    if isinstance(value, dict):
-        # Handle backend enum conversion
-        backend = value.get("backend", MemoryBackend.FILE)
-        if isinstance(backend, str):
+            # Fall back to old behavior for custom/unknown backends
             try:
-                backend = MemoryBackend(backend.lower())
+                backend = MemoryBackend(value.lower())
             except ValueError:
-                pass  # Keep as string for custom backends
-        return MemoryConfig(
-            backend=backend,
-            user_id=value.get("user_id"),
-            session_id=value.get("session_id"),
-            auto_memory=value.get("auto_memory", False),
-            claude_memory=value.get("claude_memory", False),
-            db=value.get("db"),
-            config=value.get("config"),
-        )
+                backend = value  # Allow custom backend strings
+            return MemoryConfig(backend=backend)
     
-    # Config: passthrough
-    if isinstance(value, MemoryConfig):
-        return value
+    # Special case: handle arbitrary instances with passthrough
+    if not isinstance(value, (type(None), bool, dict, list, tuple, MemoryConfig)):
+        # Try canonical resolver first
+        result = _resolve(value, MemoryConfig)
+        # If canonical resolver returns None for an instance, pass it through
+        if result is None and hasattr(value, '__class__'):
+            return value
+        return result
     
-    # Instance (MemoryManager): return as-is (caller handles)
-    # This is the highest precedence - user provided a pre-configured instance
-    return value
+    # Default: use canonical resolver
+    return _resolve(value, MemoryConfig)
 
 
 def resolve_knowledge(value: KnowledgeParam) -> Optional[KnowledgeConfig]:
     """
     Resolve knowledge= parameter following precedence ladder.
     
-    Precedence: Instance > Config > Array > Dict > String > Bool > Default
+    Delegates to the canonical resolver in param_resolver.py.
+    Kept for backward compatibility with tests.
     
     Args:
         value: Knowledge parameter in any supported form
@@ -1345,164 +1402,91 @@ def resolve_knowledge(value: KnowledgeParam) -> Optional[KnowledgeConfig]:
     Returns:
         KnowledgeConfig if enabled, None if disabled
     """
-    # Default: disabled
-    if value is None:
-        return None
-    
-    # Bool: False = disabled, True = defaults
-    if value is False:
-        return None
-    if value is True:
-        return KnowledgeConfig()
-    
-    # String: single source
-    if isinstance(value, str):
-        return KnowledgeConfig(sources=[value])
-    
-    # Array: list of sources
-    if isinstance(value, list):
-        return KnowledgeConfig(sources=value)
-    
-    # Dict: expand to config
-    if isinstance(value, dict):
-        return KnowledgeConfig(
-            sources=value.get("sources", []),
-            embedder=value.get("embedder", "openai"),
-            embedder_config=value.get("embedder_config"),
-            chunking_strategy=value.get("chunking_strategy", ChunkingStrategy.SEMANTIC),
-            chunk_size=value.get("chunk_size", 1000),
-            chunk_overlap=value.get("chunk_overlap", 200),
-            retrieval_k=value.get("retrieval_k", 5),
-            retrieval_threshold=value.get("retrieval_threshold", 0.0),
-            rerank=value.get("rerank", False),
-            rerank_model=value.get("rerank_model"),
-            auto_retrieve=value.get("auto_retrieve", True),
-        )
-    
-    # Config: passthrough
-    if isinstance(value, KnowledgeConfig):
-        return value
-    
-    # Instance: return as-is
-    return value
+    from .param_resolver import resolve_knowledge as _resolve
+    return _resolve(value, KnowledgeConfig)
 
 
 def resolve_planning(value: PlanningParam) -> Optional[PlanningConfig]:
-    """Resolve planning= parameter following precedence ladder."""
-    if value is None or value is False:
-        return None
-    if value is True:
-        return PlanningConfig()
-    if isinstance(value, dict):
-        return PlanningConfig(**value)
-    if isinstance(value, PlanningConfig):
-        return value
-    return value
+    """
+    Resolve planning= parameter following precedence ladder.
+    
+    Delegates to the canonical resolver in param_resolver.py.
+    Kept for backward compatibility with tests.
+    """
+    from .param_resolver import resolve_planning as _resolve
+    return _resolve(value, PlanningConfig)
 
 
 def resolve_reflection(value: ReflectionParam) -> Optional[ReflectionConfig]:
-    """Resolve reflection= parameter following precedence ladder."""
-    if value is None or value is False:
-        return None
-    if value is True:
-        return ReflectionConfig()
-    if isinstance(value, dict):
-        return ReflectionConfig(**value)
-    if isinstance(value, ReflectionConfig):
-        return value
-    return value
+    """
+    Resolve reflection= parameter following precedence ladder.
+    
+    Delegates to the canonical resolver in param_resolver.py.
+    Kept for backward compatibility with tests.
+    """
+    from .param_resolver import resolve_reflection as _resolve
+    return _resolve(value, ReflectionConfig)
 
 
 def resolve_guardrails(value: GuardrailParam) -> Optional[GuardrailConfig]:
-    """Resolve guardrails= parameter following precedence ladder."""
-    if value is None or value is False:
-        return None
-    if value is True:
-        return GuardrailConfig()
-    if callable(value):
+    """
+    Resolve guardrails= parameter following precedence ladder.
+    
+    Delegates to the canonical resolver in param_resolver.py with
+    special handling for callable validators.
+    """
+    from .param_resolver import resolve_guardrails as _resolve
+    
+    # Special case: wrap callable in GuardrailConfig for backward compatibility
+    if callable(value) and not isinstance(value, type):
         return GuardrailConfig(validator=value)
-    if isinstance(value, dict):
-        return GuardrailConfig(**value)
-    if isinstance(value, GuardrailConfig):
-        return value
-    return value
+    
+    # Default: use canonical resolver
+    return _resolve(value, GuardrailConfig)
 
 
 def resolve_web(value: WebParam) -> Optional[WebConfig]:
-    """Resolve web= parameter following precedence ladder."""
-    if value is None or value is False:
-        return None
-    if value is True:
-        return WebConfig()
-    if isinstance(value, dict):
-        return WebConfig(**value)
-    if isinstance(value, WebConfig):
-        return value
-    return value
+    """
+    Resolve web= parameter following precedence ladder.
+    
+    Delegates to the canonical resolver in param_resolver.py.
+    Kept for backward compatibility with tests.
+    """
+    from .param_resolver import resolve_web as _resolve
+    return _resolve(value, WebConfig)
 
 
-def resolve_output(value: OutputParam) -> Optional[OutputConfig]:
-    """Resolve output= parameter following precedence ladder."""
-    if value is None:
-        return None
-    if isinstance(value, str):
-        try:
-            preset = OutputPreset(value.lower())
-            return OutputConfig.from_preset(preset)
-        except ValueError:
-            return None
-    if isinstance(value, dict):
-        return OutputConfig(**value)
-    if isinstance(value, OutputConfig):
-        return value
-    return value
-
-
-def resolve_execution(value: ExecutionParam) -> Optional[ExecutionConfig]:
-    """Resolve execution= parameter following precedence ladder."""
-    if value is None:
-        return None
-    if isinstance(value, str):
-        try:
-            preset = ExecutionPreset(value.lower())
-            return ExecutionConfig.from_preset(preset)
-        except ValueError:
-            return None
-    if isinstance(value, dict):
-        return ExecutionConfig(**value)
-    if isinstance(value, ExecutionConfig):
-        return value
-    return value
-
 
 def resolve_caching(value: CachingParam) -> Optional[CachingConfig]:
-    """Resolve caching= parameter following precedence ladder."""
-    if value is None or value is False:
-        return None
-    if value is True:
-        return CachingConfig()
-    if isinstance(value, dict):
-        return CachingConfig(**value)
-    if isinstance(value, CachingConfig):
-        return value
-    return value
+    """
+    Resolve caching= parameter following precedence ladder.
+    
+    Delegates to the canonical resolver in param_resolver.py.
+    Kept for backward compatibility with tests.
+    """
+    from .param_resolver import resolve_caching as _resolve
+    return _resolve(value, CachingConfig)
 
 
 def resolve_autonomy(value: AutonomyParam) -> Optional[AutonomyConfig]:
-    """Resolve autonomy= parameter following precedence ladder."""
-    if value is None or value is False:
-        return None
-    if value is True:
-        return AutonomyConfig()
-    if isinstance(value, dict):
-        return AutonomyConfig(**value)
-    if isinstance(value, AutonomyConfig):
-        return value
-    return value
+    """
+    Resolve autonomy= parameter following precedence ladder.
+    
+    Delegates to the canonical resolver in param_resolver.py.
+    Kept for backward compatibility with tests.
+    """
+    from .param_resolver import resolve_autonomy as _resolve
+    return _resolve(value, AutonomyConfig)
 
 
 def resolve_tool_search(value: ToolSearchParam) -> Optional[ToolSearchConfig]:
-    """Resolve tool_search= parameter following precedence ladder."""
+    """
+    Resolve tool_search= parameter following precedence ladder.
+    
+    NOTE: This resolver has zero references in the codebase but is kept
+    for backward compatibility. Consider removing in a future version.
+    """
+    # Simple implementation since it's unused
     if value is None or value is False:
         return None
     if value is True:
@@ -1585,8 +1569,6 @@ __all__ = [
     "resolve_reflection",
     "resolve_guardrails",
     "resolve_web",
-    "resolve_output",
-    "resolve_execution",
     "resolve_caching",
     "resolve_autonomy",
     "resolve_tool_search",

src/praisonai-agents/praisonaiagents/config/loader.py

@@ -83,6 +83,7 @@ class ConfigValidationError(Exception):
 # Global config cache (lazy loaded)
 _config_cache: Optional[Dict[str, Any]] = None
 _config_loaded: bool = False
+_defaults_has_values: Optional[bool] = None
 
 
 def _default_project_plugins_dir():
@@ -418,11 +419,27 @@ def get_enabled_plugins() -> Optional[List[str]]:
     return None  # All plugins enabled
 
 
+def _defaults_has_any_values() -> bool:
+    """Return True if config file defines any defaults (cached after first check)."""
+    global _defaults_has_values
+    if _defaults_has_values is None:
+        defaults = get_defaults_config()
+        _defaults_has_values = any(
+            getattr(defaults, key, None) is not None
+            for key in (
+                "model", "base_url", "memory", "knowledge", "planning",
+                "reflection", "web", "output", "execution", "caching", "autonomy",
+            )
+        )
+    return _defaults_has_values
+
+
 def clear_config_cache() -> None:
     """Clear the config cache (for testing)."""
-    global _config_cache, _config_loaded
+    global _config_cache, _config_loaded, _defaults_has_values
     _config_cache = None
     _config_loaded = False
+    _defaults_has_values = None
 
 
 def validate_config(config: Dict[str, Any], raise_on_error: bool = False) -> List[str]:
@@ -603,6 +620,8 @@ def apply_config_defaults(
     if explicit_value is not None:
         # If True and we have config defaults, merge them
         if explicit_value is True:
+            if not _defaults_has_any_values():
+                return explicit_value
             config_defaults = get_default(param_name)
             if config_defaults and isinstance(config_defaults, dict):
                 if config_class:
@@ -612,6 +631,10 @@ def apply_config_defaults(
                 return config_defaults
         return explicit_value
     
+    # Fast path: no config defaults file values — skip lookups
+    if not _defaults_has_any_values():
+        return None
+    
     # Check if config has defaults for this param
     config_defaults = get_default(param_name)
     if config_defaults is None:

src/praisonai-agents/praisonaiagents/config/presets.py

@@ -239,12 +239,12 @@ EXECUTION_PRESETS: Dict[str, Dict[str, Any]] = {
 # =============================================================================
 
 WEB_PRESETS: Dict[str, Dict[str, Any]] = {
-    # Providers
-    "duckduckgo": {"search": True, "fetch": True, "search_provider": "duckduckgo"},
-    "tavily": {"search": True, "fetch": True, "search_provider": "tavily"},
-    "google": {"search": True, "fetch": True, "search_provider": "google"},
-    "bing": {"search": True, "fetch": True, "search_provider": "bing"},
-    "serper": {"search": True, "fetch": True, "search_provider": "serper"},
+    # Providers - fetch disabled by default for OpenAI model compatibility
+    "duckduckgo": {"search": True, "fetch": False, "search_provider": "duckduckgo"},
+    "tavily": {"search": True, "fetch": False, "search_provider": "tavily"},
+    "google": {"search": True, "fetch": False, "search_provider": "google"},
+    "bing": {"search": True, "fetch": False, "search_provider": "bing"},
+    "serper": {"search": True, "fetch": False, "search_provider": "serper"},
     # Modes
     "search_only": {"search": True, "fetch": False},
     "fetch_only": {"search": False, "fetch": True},
@@ -322,8 +322,8 @@ CACHING_PRESETS: Dict[str, Dict[str, Any]] = {
 # =============================================================================
 
 MULTI_AGENT_OUTPUT_PRESETS: Dict[str, Dict[str, Any]] = {
-    "verbose": {"verbose": 2, "stream": True},
-    "minimal": {"verbose": 1, "stream": True},
+    "verbose": {"verbose": 2, "stream": False},
+    "minimal": {"verbose": 1, "stream": False},
     "silent": {"verbose": 0, "stream": False},
 }
 

src/praisonai-agents/praisonaiagents/context/manager.py

@@ -94,65 +94,6 @@ class SessionDeduplicationCache:
             self._cache.clear()
             self._stats = {"duplicates_prevented": 0, "tokens_saved": 0}
 
-def deduplicate_topics(topics: list, key: str = "title", similarity_threshold: float = 0.8) -> list:
-    """
-    Programmatic deduplication of topics/items before agent processing.
-    
-    This helps prevent duplicate content from being passed to downstream agents,
-    reducing token waste and improving quality.
-    
-    Args:
-        topics: List of topic dicts or strings
-        key: Key to use for comparison if topics are dicts (default: "title")
-        similarity_threshold: Similarity threshold for fuzzy matching (0.0-1.0)
-        
-    Returns:
-        Deduplicated list of topics
-    """
-    if not topics:
-        return topics
-    
-    seen_hashes = set()
-    seen_normalized = set()
-    unique_topics = []
-    
-    for topic in topics:
-        # Get the content to compare
-        if isinstance(topic, dict):
-            content = str(topic.get(key, topic.get("content", str(topic))))
-        else:
-            content = str(topic)
-        
-        # Normalize for comparison
-        normalized = content.lower().strip()
-        # Remove common words for better matching
-        normalized = " ".join(w for w in normalized.split() if len(w) > 3)
-        
-        # Check exact hash match
-        content_hash = hashlib.sha256(normalized.encode()).hexdigest()
-        if content_hash in seen_hashes:
-            continue
-        
-        # Check fuzzy match using simple word overlap
-        is_duplicate = False
-        for seen in seen_normalized:
-            # Calculate Jaccard similarity
-            words1 = set(normalized.split())
-            words2 = set(seen.split())
-            if words1 and words2:
-                intersection = len(words1 & words2)
-                union = len(words1 | words2)
-                similarity = intersection / union if union > 0 else 0
-                if similarity >= similarity_threshold:
-                    is_duplicate = True
-                    break
-        
-        if not is_duplicate:
-            seen_hashes.add(content_hash)
-            seen_normalized.add(normalized)
-            unique_topics.append(topic)
-    
-    return unique_topics
 
 class EstimationMode(str, Enum):
     """Token estimation modes."""
@@ -785,7 +726,7 @@ class ContextManager:
         if protected and tool_name not in self.config.protected_tools:
             self.config.protected_tools.append(tool_name)
     
-    def truncate_tool_output(self, tool_name: str, output: str) -> str:
+    def truncate_tool_output(self, tool_name: str, output: str, tool_call_id: str = None, run_id: str = None) -> str:
         """Truncate tool output according to its budget."""
         max_tokens = self.get_tool_budget(tool_name)
         
@@ -805,6 +746,17 @@ class ContextManager:
         tail = output[-tail_size:] if tail_size > 0 else ""
         truncated = f"{head}\n...[{len(output):,} chars, showing first/last portions]...\n{tail}"
         
+        # Store full output for later retrieval
+        try:
+            from ..runtime.tool_output_store import get_tool_output_store
+            store = get_tool_output_store(run_id)
+            metadata = store.store(tool_name, output, call_id=tool_call_id)
+            if metadata:
+                truncated = store.format_reference(metadata, truncated)
+                logger.debug(f"Stored full {tool_name} output ({len(output)} bytes) at {metadata.get('path')}")
+        except Exception as e:
+            logger.debug(f"Failed to store tool output: {e}")
+        
         self._add_history_event(
             OptimizationEventType.CAP_OUTPUTS,
             tokens_before=current_tokens,

src/praisonai-agents/praisonaiagents/gateway/__init__.py

@@ -22,6 +22,7 @@ from .protocols import (
     PushChannelProtocol,
     PresenceProtocol,
     DeliveryGuaranteeProtocol,
+    OutboundDeliveryProtocol,
     ChannelInfo,
     PresenceInfo,
 )
@@ -89,6 +90,7 @@ __all__ = [
     "PushChannelProtocol",
     "PresenceProtocol",
     "DeliveryGuaranteeProtocol",
+    "OutboundDeliveryProtocol",
     "ChannelInfo",
     "PresenceInfo",
     # Config (always available)

src/praisonai-agents/praisonaiagents/gateway/config.py

@@ -17,14 +17,33 @@ class SessionConfig:
         max_messages: Maximum messages to keep in history (0 = unlimited)
         persist: Whether to persist session state
         persist_path: Path for session persistence
+        resume_window: How long (seconds) a session stays resumable after disconnect
+        max_inbox: Maximum queued messages per session (0 = unlimited, default 256)
         metadata: Additional session metadata
+        mirror_runtime_state: Enable runtime state mirroring for native transcript replay (Issue #1943)
     """
     
     timeout: int = 3600  # 1 hour default
     max_messages: int = 1000
     persist: bool = False
     persist_path: Optional[str] = None
+    resume_window: int = 86400  # 24 hours default
+    max_inbox: int = 256  # Default bounded queue size
     metadata: Dict[str, Any] = field(default_factory=dict)
+    mirror_runtime_state: bool = False  # Opt-in to avoid storage bloat
+    
+    def __post_init__(self) -> None:
+        """Validate configuration values."""
+        if self.max_inbox < 0:
+            raise ValueError(
+                "max_inbox must be >= 0 (use 0 for unlimited queue size)"
+            )
+        if self.timeout < 0:
+            raise ValueError("timeout must be >= 0")
+        if self.max_messages < 0:
+            raise ValueError("max_messages must be >= 0")
+        if self.resume_window < 0:
+            raise ValueError("resume_window must be >= 0")
     
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary."""
@@ -33,7 +52,10 @@ class SessionConfig:
             "max_messages": self.max_messages,
             "persist": self.persist,
             "persist_path": self.persist_path,
+            "resume_window": self.resume_window,
+            "max_inbox": self.max_inbox,
             "metadata": self.metadata,
+            "mirror_runtime_state": self.mirror_runtime_state,
         }
 
 
@@ -206,6 +228,7 @@ class GatewayConfig:
         reconnect_timeout: Time to wait for reconnection before closing session
         ssl_cert: Path to SSL certificate (for HTTPS/WSS)
         ssl_key: Path to SSL key
+        max_buffered_bytes: Maximum buffered bytes before slow consumer disconnect (default 1MB)
         push: Push notification service configuration
     """
     
@@ -222,12 +245,23 @@ class GatewayConfig:
     reconnect_timeout: int = 60
     ssl_cert: Optional[str] = None
     ssl_key: Optional[str] = None
+    max_buffered_bytes: int = 1024 * 1024  # 1MB default
     push: PushConfig = field(default_factory=PushConfig)
 
     def __post_init__(self) -> None:
-        """Post-initialization to set bind_host from host if not specified."""
+        """Post-initialization to set bind_host from host if not specified and validate values."""
         if self.bind_host is None:
             self.bind_host = self.host
+        if self.max_buffered_bytes < 0:
+            raise ValueError(
+                "max_buffered_bytes must be >= 0 (use 0 to disable slow-consumer checks)"
+            )
+        if self.max_connections < 0:
+            raise ValueError("max_connections must be >= 0")
+        if self.heartbeat_interval < 0:
+            raise ValueError("heartbeat_interval must be >= 0")
+        if self.reconnect_timeout < 0:
+            raise ValueError("reconnect_timeout must be >= 0")
     
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary (hides sensitive data)."""
@@ -243,6 +277,7 @@ class GatewayConfig:
             "heartbeat_interval": self.heartbeat_interval,
             "reconnect_timeout": self.reconnect_timeout,
             "ssl_enabled": bool(self.ssl_cert and self.ssl_key),
+            "max_buffered_bytes": self.max_buffered_bytes,
             "push": self.push.to_dict(),
         }
     
@@ -365,6 +400,22 @@ class MultiChannelGatewayConfig:
         """
         # Parse gateway section
         gw_data = data.get("gateway", {})
+        
+        # Parse session config if provided
+        session_config = SessionConfig()
+        if "session_config" in gw_data:
+            sc_data = gw_data["session_config"]
+            if isinstance(sc_data, dict):
+                session_config = SessionConfig(
+                    timeout=sc_data.get("timeout", 3600),
+                    max_messages=sc_data.get("max_messages", 1000),
+                    persist=sc_data.get("persist", False),
+                    persist_path=sc_data.get("persist_path"),
+                    resume_window=sc_data.get("resume_window", 86400),
+                    max_inbox=sc_data.get("max_inbox", 256),
+                    metadata=sc_data.get("metadata", {}),
+                )
+        
         gateway_config = GatewayConfig(
             host=gw_data.get("host", "127.0.0.1"),
             port=gw_data.get("port", 8765),
@@ -372,6 +423,13 @@ class MultiChannelGatewayConfig:
             allowed_origins=gw_data.get("allowed_origins", []),
             auth_token=gw_data.get("auth_token"),
             max_connections=gw_data.get("max_connections", 1000),
+            max_sessions_per_agent=gw_data.get("max_sessions_per_agent", 0),
+            session_config=session_config,
+            heartbeat_interval=gw_data.get("heartbeat_interval", 30),
+            reconnect_timeout=gw_data.get("reconnect_timeout", 60),
+            ssl_cert=gw_data.get("ssl_cert"),
+            ssl_key=gw_data.get("ssl_key"),
+            max_buffered_bytes=gw_data.get("max_buffered_bytes", 1024 * 1024),
         )
         
         # Parse agents section (pass through as dicts)

src/praisonai-agents/praisonaiagents/gateway/protocols.py

@@ -17,6 +17,7 @@ from enum import Enum
 from typing import (
     TYPE_CHECKING,
     Any,
+    Awaitable,
     Callable,
     Dict,
     List,
@@ -27,9 +28,24 @@ from typing import (
     runtime_checkable,
 )
 
+# Gateway protocol versioning constants
+GATEWAY_PROTOCOL_VERSION = 1
+MIN_CLIENT_PROTOCOL_VERSION = 1
+
 if TYPE_CHECKING:
     from praisonai.gateway.pairing import PairedChannel
     from ..agent import Agent
+    from ..bots.presentation import MessagePresentation
+    from ..scheduler.models import DeliveryTarget
+
+
+class ConnectErrorCode(str, Enum):
+    """Structured error codes for connection failures."""
+    AUTH_REQUIRED = "auth_required"
+    AUTH_UNAUTHORIZED = "auth_unauthorized"
+    PROTOCOL_UNSUPPORTED = "protocol_unsupported"
+    PAIRING_REQUIRED = "pairing_required"
+    AGENT_NOT_FOUND = "agent_not_found"
 
 
 class EventType(str, Enum):
@@ -53,6 +69,7 @@ class EventType(str, Enum):
     # Message events
     MESSAGE = "message"
     MESSAGE_ACK = "message_ack"
+    MESSAGE_ABORT = "message_abort"
     TYPING = "typing"
     
     # Streaming events (relayed from agent's StreamEventEmitter)
@@ -84,6 +101,65 @@ class EventType(str, Enum):
     # Polling events
     POLL_REQUEST = "poll_request"
     POLL_RESPONSE = "poll_response"
+    
+    # Handshake events
+    HELLO = "hello"
+    HELLO_OK = "hello_ok"
+    HELLO_ERROR = "hello_error"
+
+
+@dataclass
+class HelloParams:
+    """Parameters for initiating a versioned handshake.
+    
+    Attributes:
+        agent_id: The agent to connect to
+        protocol_min: Minimum protocol version the client supports
+        protocol_max: Maximum protocol version the client supports
+        capabilities: Optional list of capability tokens the client supports
+        session_id: Optional session to resume
+        since: Optional cursor for event replay
+    """
+    agent_id: str
+    protocol_min: int
+    protocol_max: int
+    capabilities: List[str] = field(default_factory=list)
+    session_id: Optional[str] = None
+    since: Optional[int] = None
+
+
+@dataclass
+class HelloResult:
+    """Result of a successful handshake negotiation.
+    
+    Attributes:
+        protocol: The negotiated protocol version
+        features: Supported methods and events
+        policy: Gateway policy limits (max_payload, heartbeat_ms, etc.)
+        session_id: The session ID for this connection
+        resumed: Whether an existing session was resumed
+        cursor: Current event cursor position
+    """
+    protocol: int
+    features: Dict[str, List[str]]  # {"methods": [...], "events": [...]}
+    policy: Dict[str, int]  # {"max_payload": ..., "heartbeat_ms": ...}
+    session_id: str
+    resumed: bool
+    cursor: int
+
+
+@dataclass
+class HelloError:
+    """Error response for failed handshake.
+    
+    Attributes:
+        code: Structured error code
+        message: Human-readable error message
+        next_action: Suggested next action (e.g., "upgrade_client", "pair_device")
+    """
+    code: ConnectErrorCode
+    message: str
+    next_action: Optional[str] = None
 
 
 @dataclass
@@ -148,6 +224,7 @@ class GatewayMessage:
         timestamp: Message creation time
         metadata: Additional message metadata
         reply_to: ID of message being replied to (optional)
+        presentation: Optional interactive presentation (buttons, menus, etc.)
     """
     
     content: Union[str, Dict[str, Any]]
@@ -157,10 +234,11 @@ class GatewayMessage:
     timestamp: float = field(default_factory=time.time)
     metadata: Dict[str, Any] = field(default_factory=dict)
     reply_to: Optional[str] = None
+    presentation: Optional["MessagePresentation"] = None
     
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary for serialization."""
-        return {
+        data = {
             "content": self.content,
             "sender_id": self.sender_id,
             "session_id": self.session_id,
@@ -169,10 +247,20 @@ class GatewayMessage:
             "metadata": self.metadata,
             "reply_to": self.reply_to,
         }
+        if self.presentation is not None:
+            data["presentation"] = self.presentation.to_dict()
+        return data
     
     @classmethod
     def from_dict(cls, data: Dict[str, Any]) -> "GatewayMessage":
         """Create from dictionary."""
+        from ..bots.presentation import MessagePresentation
+        
+        presentation = None
+        raw_presentation = data.get("presentation")
+        if isinstance(raw_presentation, dict):
+            presentation = MessagePresentation.from_dict(raw_presentation)
+        
         return cls(
             content=data.get("content", ""),
             sender_id=data.get("sender_id", "unknown"),
@@ -181,6 +269,7 @@ class GatewayMessage:
             timestamp=data.get("timestamp", time.time()),
             metadata=data.get("metadata", {}),
             reply_to=data.get("reply_to"),
+            presentation=presentation,
         )
 
 
@@ -688,6 +777,126 @@ class DeliveryGuaranteeProtocol(Protocol):
         ...
 
 
+@runtime_checkable
+class OutboundDeliveryProtocol(Protocol):
+    """Protocol for durable outbound message delivery.
+    
+    Ensures messages sent to external channels (Telegram, Slack, Discord, etc.)
+    are persisted before sending and can be retried on failure. This provides
+    crash-safe at-least-once delivery for channel replies.
+    
+    Example usage (implementation in praisonai wrapper):
+        from praisonai.bots import OutboundQueue
+        
+        outbox = OutboundQueue(path="~/.praisonai/state/outbox.sqlite")
+        
+        # Enqueue before sending
+        key = await outbox.enqueue(
+            idempotency_key="msg-123",
+            target_channel="telegram:12345",
+            payload={"text": "Hello", "metadata": {...}}
+        )
+        
+        # Attempt delivery
+        success = await deliver_with_retry(adapter, channel_id, payload)
+        
+        # Mark as sent only if successful
+        if success:
+            await outbox.mark_sent(key)
+        
+        # On restart, drain pending messages
+        await outbox.drain(delivery_handler)
+    """
+    
+    async def enqueue(
+        self,
+        idempotency_key: str,
+        target: str,
+        payload: Dict[str, Any],
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> str:
+        """Persist an outbound message for delivery.
+        
+        Args:
+            idempotency_key: Unique key to prevent duplicate sends
+            target: Target channel identifier (e.g., "telegram:12345")
+            payload: Message payload to deliver
+            metadata: Optional metadata for tracking/routing
+            
+        Returns:
+            Unique entry key for tracking this message
+        """
+        ...
+    
+    async def mark_sent(self, key: str) -> bool:
+        """Mark a message as successfully sent.
+        
+        Args:
+            key: The entry key returned by enqueue()
+            
+        Returns:
+            True if marked successfully, False if not found
+        """
+        ...
+    
+    async def mark_failed(
+        self,
+        key: str,
+        error: str,
+        permanent: bool = False,
+    ) -> bool:
+        """Mark a message as failed.
+        
+        Args:
+            key: The entry key returned by enqueue()
+            error: Error description
+            permanent: If True, won't retry this message
+            
+        Returns:
+            True if marked successfully, False if not found
+        """
+        ...
+    
+    async def drain(
+        self,
+        sender: Callable[[str, Dict[str, Any]], Awaitable[bool]],
+        limit: Optional[int] = None,
+    ) -> tuple[int, int]:
+        """Process pending messages.
+        
+        Called on startup to retry unsent messages. Messages are processed
+        oldest-first to maintain order.
+        
+        Args:
+            sender: Async function that attempts delivery. Should return
+                    True on success, False to retry later.
+            limit: Optional max messages to process
+            
+        Returns:
+            Tuple of (succeeded, failed) counts
+        """
+        ...
+    
+    def pending_count(self) -> int:
+        """Get count of pending messages awaiting delivery."""
+        ...
+    
+    def size(self) -> int:
+        """Get total number of messages in queue."""
+        ...
+    
+    async def purge_old(self, max_age_seconds: int = 86400 * 7) -> int:
+        """Remove old sent messages.
+        
+        Args:
+            max_age_seconds: Age threshold for removal
+            
+        Returns:
+            Number of messages purged
+        """
+        ...
+
+
 # ---------------------------------------------------------------------------
 # Auth Mode protocols and helpers (bind-aware authentication posture)
 # ---------------------------------------------------------------------------
@@ -904,3 +1113,83 @@ class SessionBindingProtocol(Protocol):
             Principal information if found, None otherwise
         """
         ...
+
+
+# ---------------------------------------------------------------------------
+# Home Channel and Delivery Routing Protocols
+# ---------------------------------------------------------------------------
+
+@runtime_checkable
+class HomeChannelRegistryProtocol(Protocol):
+    """Protocol for managing default delivery targets per platform.
+    
+    Home channels provide a per-platform default delivery target that can be
+    set once from inside a chat and persisted, so scheduled jobs can deliver
+    results without requiring explicit channel IDs.
+    """
+    
+    def set_home(
+        self, 
+        platform: str, 
+        chat_id: str, 
+        thread_id: Optional[str] = None
+    ) -> None:
+        """Set the home channel for a platform.
+        
+        Args:
+            platform: Platform name (e.g., "telegram", "slack", "discord")
+            chat_id: Platform-specific chat/channel ID
+            thread_id: Optional thread ID for threaded platforms
+        """
+        ...
+    
+    def get_home(self, platform: str) -> Optional[tuple[str, Optional[str]]]:
+        """Get the home channel for a platform.
+        
+        Args:
+            platform: Platform name to look up
+            
+        Returns:
+            Tuple of (chat_id, thread_id) if set, None otherwise
+        """
+        ...
+    
+    def platforms_with_home(self) -> List[str]:
+        """List all platforms that have a home channel configured.
+        
+        Returns:
+            List of platform names with home channels
+        """
+        ...
+
+
+@runtime_checkable
+class DeliveryResolverProtocol(Protocol):
+    """Protocol for resolving delivery routing tokens.
+    
+    Resolves tokens like "origin", "telegram", "all" to concrete delivery
+    targets at fire time, enabling ergonomic routing without hard-coded IDs.
+    """
+    
+    def resolve(
+        self, 
+        token: str, 
+        *, 
+        origin: Optional["DeliveryTarget"] = None
+    ) -> List["DeliveryTarget"]:
+        """Resolve a routing token to concrete delivery targets.
+        
+        Token formats:
+        - "origin": Reply to the chat where the job was created (requires origin)
+        - "<platform>": That platform's home channel
+        - "<platform>:<chat_id>[:<thread_id>]": Explicit target
+        - "all": Fan-out to every connected platform with a home channel
+        
+        Args:
+            token: Routing token to resolve
+            origin: Original delivery target (for "origin" token)
+            
+        Returns:
+            List of concrete delivery targets
+        """
+        ...

src/praisonai-agents/praisonaiagents/hooks/runner.py

@@ -83,7 +83,8 @@ class HookRunner:
         self,
         event: HookEvent,
         input_data: HookInput,
-        target: Optional[str] = None
+        target: Optional[str] = None,
+        _hooks: Optional[List["HookDefinition"]] = None
     ) -> List[HookExecutionResult]:
         """
         Execute all hooks for an event.
@@ -92,11 +93,12 @@ class HookRunner:
             event: The event to execute hooks for
             input_data: Input data for the hooks
             target: Optional target to filter hooks (e.g., tool name)
+            _hooks: Internal parameter to avoid duplicate get_hooks() calls
             
         Returns:
             List of execution results
         """
-        hooks = self._registry.get_hooks(event, target)
+        hooks = _hooks if _hooks is not None else self._registry.get_hooks(event, target)
         
         if not hooks:
             return []
@@ -153,8 +155,14 @@ class HookRunner:
                 "Use 'await runner.execute(event, input_data, target)' instead in async contexts."
             )
 
-        # No running loop — safe to create one
-        return asyncio.run(self.execute(event, input_data, target))
+        # Fast path: no hooks → no loop creation needed
+        # This mirrors the same check in execute() but avoids event loop creation overhead
+        hooks = self._registry.get_hooks(event, target)
+        if not hooks:
+            return []
+
+        # No running loop and hooks exist — safe to create one
+        return asyncio.run(self.execute(event, input_data, target, _hooks=hooks))
     
     async def _execute_parallel(
         self,

src/praisonai-agents/praisonaiagents/llm/llm.py

@@ -24,6 +24,7 @@ import xml.etree.ElementTree as ET
 from ..errors import AgentErrorKind, FailoverDecision, IdleTimeoutBreaker
 # Gap 2: Tool call execution imports
 from ..tools.call_executor import ToolCall, create_tool_call_executor
+from ..tools.schema import annotation_to_json_schema, get_parameter_requirements
 # Display functions - lazy loaded to avoid importing rich at startup
 # These are only needed when output=verbose
 _display_module = None
@@ -5749,8 +5750,6 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         if content:
                             response_text += content
                             live.update(_get_display_functions()['display_generating'](response_text, start_time))
-                        if content:
-                            response_text += content
             else:
                 # Use retry wrapper for non-streaming calls
                 response = self._completion_with_retry(**completion_params)
@@ -5846,8 +5845,6 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         if content:
                             response_text += content
                             live.update(_get_display_functions()['display_generating'](response_text, start_time))
-                        if content:
-                            response_text += content
             else:
                 # Use retry wrapper for non-streaming async calls
                 response = await self._acompletion_with_retry(**completion_params)
@@ -5886,34 +5883,74 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
             function_name = function_or_name
             logging.debug(f"Attempting to generate tool definition for: {function_name}")
             
-            # First try to get the tool definition if it exists
-            tool_def_name = f"{function_name}_definition"
-            tool_def = globals().get(tool_def_name)
-            logging.debug(f"Looking for {tool_def_name} in globals: {tool_def is not None}")
-            
-            if not tool_def:
-                import __main__
-                tool_def = getattr(__main__, tool_def_name, None)
-                logging.debug(f"Looking for {tool_def_name} in __main__: {tool_def is not None}")
-            
-            if tool_def:
-                logging.debug(f"Found tool definition: {tool_def}")
-                return tool_def
-
-            # Try to find the function
-            func = globals().get(function_name)
-            logging.debug(f"Looking for {function_name} in globals: {func is not None}")
-            
-            if not func:
-                import __main__
-                func = getattr(__main__, function_name, None)
-                logging.debug(f"Looking for {function_name} in __main__: {func is not None}")
-            
-            if not func or not callable(func):
-                logging.debug(f"Function {function_name} not found or not callable")
-                return None
+            # First try to get from the tool registry (preferred method)
+            try:
+                from ..tools.registry import get_registry
+                registry = get_registry()
+                tool = registry.get(function_name)
+                if tool:
+                    logging.debug(f"Found tool in registry: {function_name}")
+                    # If it's a BaseTool, get its schema
+                    if hasattr(tool, 'get_schema'):
+                        tool_def = tool.get_schema()
+                        # Apply same normalization as the callable path
+                        if (
+                            isinstance(tool_def, dict)
+                            and isinstance(tool_def.get("function"), dict)
+                            and isinstance(tool_def["function"].get("parameters"), dict)
+                        ):
+                            tool_def = tool_def.copy()
+                            tool_def["function"] = tool_def["function"].copy()
+                            tool_def["function"]["parameters"] = self._fix_array_schemas(
+                                tool_def["function"]["parameters"]
+                            )
+                        return tool_def
+                    # If it's a callable, use it as func
+                    if callable(tool):
+                        func = tool
+                    else:
+                        logging.debug(f"Tool {function_name} in registry is not callable")
+                        return None
+                else:
+                    logging.debug(f"Tool {function_name} not found in registry, falling back to globals/__main__")
+                    # Fall back to globals and __main__ for backward compatibility
+                    tool_def_name = f"{function_name}_definition"
+                    tool_def = globals().get(tool_def_name)
+                    if not tool_def:
+                        import __main__
+                        tool_def = getattr(__main__, tool_def_name, None)
+                    if tool_def:
+                        return tool_def
+                    
+                    func = globals().get(function_name)
+                    if not func:
+                        import __main__
+                        func = getattr(__main__, function_name, None)
+                    if not func or not callable(func):
+                        logging.debug(f"Function {function_name} not found or not callable")
+                        return None
+            except ImportError:
+                logging.debug("Tool registry not available, falling back to globals/__main__")
+                # Fall back to globals and __main__ when registry unavailable
+                tool_def_name = f"{function_name}_definition"
+                tool_def = globals().get(tool_def_name)
+                if not tool_def:
+                    import __main__
+                    tool_def = getattr(__main__, tool_def_name, None)
+                if tool_def:
+                    return tool_def
+                
+                func = globals().get(function_name)
+                if not func:
+                    import __main__
+                    func = getattr(__main__, function_name, None)
+                if not func or not callable(func):
+                    logging.debug(f"Function {function_name} not found or not callable")
+                    return None
 
         import inspect
+        from typing import get_type_hints
+        
         # Handle Langchain and CrewAI tools
         if inspect.isclass(func) and hasattr(func, 'run') and not hasattr(func, '_run'):
             original_func = func
@@ -5961,26 +5998,26 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
         
         logging.debug(f"Parameter descriptions: {param_descriptions}")
 
+        # Get type hints for proper schema generation
+        try:
+            hints = get_type_hints(func) if getattr(func, "__annotations__", None) else {}
+        except (NameError, TypeError, AttributeError):
+            hints = getattr(func, "__annotations__", {}) or {}
+
         for name, param in parameters_list:
-            param_type = "string"  # Default type
-            if param.annotation != inspect.Parameter.empty:
-                if param.annotation == int:
-                    param_type = "integer"
-                elif param.annotation == float:
-                    param_type = "number"
-                elif param.annotation == bool:
-                    param_type = "boolean"
-                elif param.annotation == list:
-                    param_type = "array"
-                elif param.annotation == dict:
-                    param_type = "object"
+            # Get type annotation
+            param_type = hints.get(name, param.annotation if param.annotation != inspect.Parameter.empty else str)
             
-            parameters["properties"][name] = {
-                "type": param_type,
-                "description": param_descriptions.get(name, "Parameter description not available")
-            }
+            # Use new schema utility for proper type handling
+            prop_schema = annotation_to_json_schema(param_type)
             
-            if param.default == inspect.Parameter.empty:
+            # Add description from docstring
+            prop_schema["description"] = param_descriptions.get(name, "Parameter description not available")
+            
+            parameters["properties"][name] = prop_schema
+            
+            # Check if required using improved logic
+            if get_parameter_requirements(sig, name):
                 parameters["required"].append(name)
         
         logging.debug(f"Generated parameters: {parameters}")

src/praisonai-agents/praisonaiagents/llm/unified_adapters.py

@@ -141,22 +141,6 @@ class LiteLLMAdapter:
         
         return self._convert_response_to_standard_format(response)
     
-    def _convert_messages_to_prompt(self, messages: List[Dict[str, Any]]) -> str:
-        """Convert OpenAI-style messages to prompt string."""
-        if not messages:
-            return ""
-        
-        # If single message, return its content
-        if len(messages) == 1:
-            return messages[0].get('content', '')
-        
-        # For multiple messages, use the last user message as prompt
-        for msg in reversed(messages):
-            if msg.get('role') == 'user':
-                return msg.get('content', '')
-        
-        return messages[-1].get('content', '') if messages else ''
-    
     async def _convert_stream_response(self, response_stream) -> AsyncIterator[Dict[str, Any]]:
         """Convert LLM stream response to standard format."""
         # Implementation depends on the actual stream format from LLM.get_response_async

src/praisonai-agents/praisonaiagents/mcp/mcp_security.py

@@ -246,3 +246,74 @@ class SecurityConfig:
     def get_bind_address(self) -> str:
         """Get recommended bind address based on config."""
         return "127.0.0.1" if self.bind_localhost_only else "0.0.0.0"
+
+
+def load_sse_security_from_env():
+    """Load optional SSE security settings from environment."""
+    import os
+
+    token = os.environ.get("MCP_SSE_AUTH_TOKEN") or os.environ.get("MCP_AUTH_TOKEN")
+    origins_raw = os.environ.get("MCP_SSE_ALLOWED_ORIGINS")
+    if not token and not origins_raw:
+        return None, None
+
+    origins = (
+        [origin.strip() for origin in origins_raw.split(",") if origin.strip()]
+        if origins_raw
+        else None
+    )
+    config = SecurityConfig(
+        require_auth=bool(token),
+        validate_origin=origins is not None,
+        allowed_origins=origins or SecurityConfig().allowed_origins,
+    )
+    return config, token
+
+
+def build_sse_security_app(app, security: Optional["SecurityConfig"] = None):
+    """Wrap a Starlette app with SSE security middleware when configured."""
+    config = security
+    token = None
+    if config is None:
+        config, token = load_sse_security_from_env()
+    elif config.require_auth:
+        import os
+
+        token = os.environ.get("MCP_SSE_AUTH_TOKEN") or os.environ.get("MCP_AUTH_TOKEN")
+    if config is None or (not config.require_auth and not config.validate_origin):
+        return app
+
+    class _SSESecurityMiddleware:
+        def __init__(self, inner, sec: SecurityConfig, expected_token: Optional[str]):
+            self._inner = inner
+            self._sec = sec
+            self._token = expected_token
+
+        async def __call__(self, scope, receive, send):
+            if scope.get("type") != "http":
+                await self._inner(scope, receive, send)
+                return
+
+            headers = {k.decode().lower(): v.decode() for k, v in scope.get("headers", [])}
+
+            if self._sec.require_auth and self._token:
+                auth = headers.get("authorization", "")
+                if not auth.startswith("Bearer ") or auth[7:] != self._token:
+                    from starlette.responses import JSONResponse
+
+                    response = JSONResponse({"error": "Unauthorized"}, status_code=401)
+                    await response(scope, receive, send)
+                    return
+
+            if self._sec.validate_origin:
+                origin = headers.get("origin") or None
+                if not self._sec.is_origin_allowed(origin):
+                    from starlette.responses import JSONResponse
+
+                    response = JSONResponse({"error": "Origin not allowed"}, status_code=403)
+                    await response(scope, receive, send)
+                    return
+
+            await self._inner(scope, receive, send)
+
+    return _SSESecurityMiddleware(app, config, token)

src/praisonai-agents/praisonaiagents/mcp/mcp_server.py

@@ -242,12 +242,18 @@ class ToolsMCPServer:
         
         mcp.run()
     
-    def run_sse(self, host: str = "0.0.0.0", port: int = 8080) -> None:
+    def run_sse(
+        self,
+        host: str = "0.0.0.0",
+        port: int = 8080,
+        security: "Optional[SecurityConfig]" = None,
+    ) -> None:
         """Run the MCP server using SSE transport.
         
         Args:
             host: Host to bind to
             port: Port to listen on
+            security: Optional security config; when unset, reads MCP_SSE_* env vars
         """
         try:
             import uvicorn
@@ -260,6 +266,8 @@ class ToolsMCPServer:
                 f"SSE dependencies not available: {e}. "
                 "Install with: pip install uvicorn starlette"
             )
+
+        from .mcp_security import SecurityConfig, build_sse_security_app
         
         mcp = self.get_fastmcp()
         
@@ -286,7 +294,8 @@ class ToolsMCPServer:
                 Mount(messages_path, app=sse_transport.handle_post_message),
             ]
         )
-        
+
+        app = build_sse_security_app(app, security)
         print(f"🚀 Starting MCP server '{self.name}' with SSE transport")
         print(f"📡 SSE endpoint: http://{host}:{port}{sse_path}")
         print(f"🛠️  Available tools: {', '.join(self.get_tool_names())}")

src/praisonai-agents/praisonaiagents/memory/core.py

@@ -378,27 +378,34 @@ class MemoryCoreMixin:
     def _emit_memory_event(self, event_type: str, memory_type: str, content: str, metadata: Dict):
         """Emit a memory-related event for monitoring/hooks."""
         try:
-            # Integrate with the existing event system
-            event_data = {
-                'type': event_type,
-                'memory_type': memory_type,
-                'content_preview': content[:100] + ('...' if len(content) > 100 else ''),
-                'metadata': metadata,
-                'timestamp': datetime.now().isoformat(),
-                'provider': getattr(self, 'provider', 'unknown')
-            }
+            # Helper to build event data lazily
+            def _build_event_data():
+                return {
+                    'type': event_type,
+                    'memory_type': memory_type,
+                    'content_preview': content[:100] + ('...' if len(content) > 100 else ''),
+                    'metadata': metadata,
+                    'timestamp': datetime.now().isoformat(),
+                    'provider': getattr(self, 'provider', 'unknown')
+                }
             
             # Try to emit to the EventBus if available
             try:
                 from ..bus import get_default_bus
                 bus = get_default_bus()
-                bus.publish(f"memory.{event_type}.{memory_type}", event_data)
+                
+                # Fast path: only build expensive payload if someone is listening
+                if not bus.has_subscribers:
+                    return
+                    
+                # Integrate with the existing event system
+                bus.publish(f"memory.{event_type}.{memory_type}", _build_event_data())
             except ImportError:
                 # EventBus not available, fall back to logging
-                logging.debug(f"Memory event: {event_type}.{memory_type} - {event_data}")
+                logging.debug(f"Memory event: {event_type}.{memory_type} - {_build_event_data()}")
             except Exception as bus_error:
                 # EventBus failed, log the event instead
-                logging.debug(f"Memory event (bus failed): {event_type}.{memory_type} - {event_data}")
+                logging.debug(f"Memory event (bus failed): {event_type}.{memory_type} - {_build_event_data()}")
                 logging.debug(f"EventBus error: {bus_error}")
         except Exception as e:
             logging.debug(f"Failed to emit memory event: {e}")

src/praisonai-agents/praisonaiagents/memory/file_memory.py

@@ -115,6 +115,16 @@ class FileMemory:
         "importance_threshold": 0.7,   # Min importance for long-term
         "auto_promote": True,          # Auto-promote important short-term to long-term
     }
+
+    @staticmethod
+    def _sanitise_user_id(user_id: str) -> str:
+        """Reject path traversal in user_id before using it as a directory name."""
+        if not user_id or not isinstance(user_id, str):
+            return "default"
+        if ".." in user_id or "/" in user_id or "\\" in user_id:
+            raise ValueError("user_id must not contain path separators or parent references")
+        safe = user_id.strip()
+        return safe or "default"
     
     def __init__(
         self,
@@ -132,7 +142,7 @@ class FileMemory:
             config: Configuration overrides
             verbose: Verbosity level (0=quiet, 1=info, 2+=debug)
         """
-        self.user_id = user_id
+        self.user_id = self._sanitise_user_id(user_id)
         self.verbose = verbose
         
         # Set up paths
@@ -142,7 +152,7 @@ class FileMemory:
             from ..paths import get_project_data_dir
             self.base_path = get_project_data_dir() / "memory"
         
-        self.user_path = self.base_path / user_id
+        self.user_path = self.base_path / self.user_id
         self.episodic_path = self.user_path / "episodic"
         
         # Create directories

src/praisonai-agents/praisonaiagents/memory/memory.py

@@ -10,7 +10,6 @@ from praisonaiagents._logging import get_logger
 from datetime import datetime, timezone
 
 # Decomposed memory functionality - imported as mixins for backward compatibility
-from .storage import StorageMixin
 from .search import SearchMixin
 from .core import MemoryCoreMixin
 
@@ -29,9 +28,14 @@ logger = get_logger(__name__, extra_data={"subsystem": "memory"})
 TRACE_LEVEL = 5
 logging.addLevelName(TRACE_LEVEL, 'TRACE')
 
+# Cache boundary constants for prompt prefix caching optimization
+# NOTE: This is a placeholder - current implementation does not provide
+# functional prompt caching as it requires structured API metadata
+CACHE_BOUNDARY = ""
 
 
-class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
+
+class Memory(SearchMixin, MemoryCoreMixin):
     """
     A single-file memory manager covering:
     - Short-term memory (STM) for ephemeral context
@@ -230,6 +234,34 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
         
         # Determine embedding dimensions for legacy compatibility
         self.embedding_dimensions = self._get_embedding_dimensions(self.embedding_model)
+        
+        # Copy adapter attributes to self for backward compatibility
+        # This ensures legacy code that checks hasattr(self, 'attribute') still works
+        if self.use_mem0 and hasattr(adapter, 'mem0_client'):
+            self.mem0_client = adapter.mem0_client
+            
+        if self.use_rag and hasattr(adapter, 'collection'):
+            self.chroma_col = adapter.collection
+            self.chroma_client = adapter.client
+            
+        if self.use_mongodb:
+            if hasattr(adapter, 'client'):
+                self.mongo_client = adapter.client
+            if hasattr(adapter, 'db'):
+                self.mongo_db = adapter.db
+            if hasattr(adapter, 'short_collection'):
+                self.mongo_short_term = adapter.short_collection
+            if hasattr(adapter, 'long_collection'):
+                self.mongo_long_term = adapter.long_collection
+            # CRITICAL FIX: Copy use_vector_search from adapter to self
+            # Without this, lines 579, 758, 857 will crash with AttributeError
+            if hasattr(adapter, 'use_vector_search'):
+                self.use_vector_search = adapter.use_vector_search
+            else:
+                self.use_vector_search = False
+        else:
+            # Initialize use_vector_search to False for non-MongoDB providers
+            self.use_vector_search = False
 
     def _get_adapter_config(self) -> Dict[str, Any]:
         """Get configuration for adapter initialization."""
@@ -381,80 +413,6 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
     # -------------------------------------------------------------------------
     #                          Initialization
     # -------------------------------------------------------------------------
-    def _init_stm(self):
-        """Creates or verifies short-term memory table."""
-        os.makedirs(os.path.dirname(self.short_db) or ".", exist_ok=True)
-        conn = self._get_stm_conn()
-        c = conn.cursor()
-        c.execute("""
-        CREATE TABLE IF NOT EXISTS short_mem (
-            id TEXT PRIMARY KEY,
-            content TEXT,
-            meta TEXT,
-            created_at REAL
-        )
-        """)
-        conn.commit()
-
-    def _init_ltm(self):
-        """Creates or verifies long-term memory table."""
-        os.makedirs(os.path.dirname(self.long_db) or ".", exist_ok=True)
-        conn = self._get_ltm_conn()
-        c = conn.cursor()
-        c.execute("""
-        CREATE TABLE IF NOT EXISTS long_mem (
-            id TEXT PRIMARY KEY,
-            content TEXT,
-            meta TEXT,
-            created_at REAL
-        )
-        """)
-        conn.commit()
-
-    def _init_mem0(self):
-        """Initialize Mem0 client for agent or user memory with optional graph support."""
-        mem_cfg = self.cfg.get("config", {})
-        api_key = mem_cfg.get("api_key", os.getenv("MEM0_API_KEY"))
-        org_id = mem_cfg.get("org_id")
-        proj_id = mem_cfg.get("project_id")
-        
-        # Check if graph memory is enabled
-        graph_config = mem_cfg.get("graph_store")
-        use_graph = graph_config is not None
-        
-        if use_graph:
-            # Initialize with graph memory support
-            from mem0 import Memory
-            self._log_verbose("Initializing Mem0 with graph memory support")
-            
-            # Build Mem0 config with graph store
-            mem0_config = {}
-            
-            # Add graph store configuration
-            mem0_config["graph_store"] = graph_config
-            
-            # Add other configurations if provided
-            if "vector_store" in mem_cfg:
-                mem0_config["vector_store"] = mem_cfg["vector_store"]
-            if "llm" in mem_cfg:
-                mem0_config["llm"] = mem_cfg["llm"]
-            if "embedder" in mem_cfg:
-                mem0_config["embedder"] = mem_cfg["embedder"]
-            
-            # Initialize Memory with graph support
-            self.mem0_client = Memory.from_config(config_dict=mem0_config)
-            self.graph_enabled = True
-            self._log_verbose("Graph memory initialized successfully")
-        else:
-            # Use traditional MemoryClient
-            from mem0 import MemoryClient
-            self._log_verbose("Initializing Mem0 with traditional memory client")
-            
-            if org_id and proj_id:
-                self.mem0_client = MemoryClient(api_key=api_key, org_id=org_id, project_id=proj_id)
-            else:
-                self.mem0_client = MemoryClient(api_key=api_key)
-            self.graph_enabled = False
 
     def _init_chroma(self):
         """Initialize a local Chroma client for embedding-based search."""
@@ -491,105 +449,6 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
             self._log_verbose(f"Failed to initialize ChromaDB: {e}", logging.ERROR)
             self.use_rag = False
 
-    def _init_mongodb(self):
-        """Initialize MongoDB client for memory storage."""
-        try:
-            # Import MongoClient locally to handle optional dependency
-            from pymongo import MongoClient
-            
-            mongo_cfg = self.cfg.get("config", {})
-            self.connection_string = mongo_cfg.get("connection_string", "mongodb://localhost:27017/")
-            self.database_name = mongo_cfg.get("database", "praisonai")
-            self.use_vector_search = mongo_cfg.get("use_vector_search", False)
-            
-            # Initialize MongoDB client
-            self.mongo_client = MongoClient(
-                self.connection_string,
-                maxPoolSize=mongo_cfg.get("max_pool_size", 50),
-                minPoolSize=mongo_cfg.get("min_pool_size", 10),
-                maxIdleTimeMS=mongo_cfg.get("max_idle_time", 30000),
-                serverSelectionTimeoutMS=mongo_cfg.get("server_selection_timeout", 5000),
-                retryWrites=True,
-                retryReads=True
-            )
-            
-            # Test connection
-            self.mongo_client.admin.command('ping')
-            
-            # Setup database and collections
-            self.mongo_db = self.mongo_client[self.database_name]
-            self.mongo_short_term = self.mongo_db.short_term_memory
-            self.mongo_long_term = self.mongo_db.long_term_memory
-            self.mongo_entities = self.mongo_db.entity_memory
-            self.mongo_users = self.mongo_db.user_memory
-            
-            # Create indexes for better performance
-            self._create_mongodb_indexes()
-            
-            self._log_verbose("MongoDB initialized successfully")
-            
-        except Exception as e:
-            self._log_verbose(f"Failed to initialize MongoDB: {e}", logging.ERROR)
-            self.use_mongodb = False
-
-    def _create_mongodb_indexes(self):
-        """Create MongoDB indexes for better performance."""
-        try:
-            # Text search indexes
-            self.mongo_short_term.create_index([("content", "text")])
-            self.mongo_long_term.create_index([("content", "text")])
-            
-            # Compound indexes for filtering
-            self.mongo_short_term.create_index([("created_at", -1), ("metadata.quality", -1)])
-            self.mongo_long_term.create_index([("created_at", -1), ("metadata.quality", -1)])
-            
-            # User-specific indexes
-            self.mongo_users.create_index([("user_id", 1), ("created_at", -1)])
-            
-            # Entity indexes
-            self.mongo_entities.create_index([("entity_name", 1), ("entity_type", 1)])
-            
-            # Vector search indexes for Atlas (if enabled)
-            if self.use_vector_search:
-                self._create_vector_search_indexes()
-                
-        except Exception as e:
-            self._log_verbose(f"Warning: Could not create MongoDB indexes: {e}", logging.WARNING)
-
-    def _create_vector_search_indexes(self):
-        """Create vector search indexes for Atlas."""
-        try:
-            vector_index_def = {
-                "mappings": {
-                    "dynamic": True,
-                    "fields": {
-                        "embedding": {
-                            "type": "knnVector",
-                            "dimensions": self.embedding_dimensions,
-                            "similarity": "cosine"
-                        }
-                    }
-                }
-            }
-            
-            # Create vector indexes for both short and long term collections
-            try:
-                # Use SearchIndexModel for PyMongo 4.6+ compatibility
-                try:
-                    from pymongo.operations import SearchIndexModel
-                    search_index_model = SearchIndexModel(definition=vector_index_def, name="vector_index")
-                    self.mongo_short_term.create_search_index(search_index_model)
-                    self.mongo_long_term.create_search_index(search_index_model)
-                except ImportError:
-                    # Fallback for older PyMongo versions
-                    self.mongo_short_term.create_search_index(vector_index_def, "vector_index")
-                    self.mongo_long_term.create_search_index(vector_index_def, "vector_index")
-                self._log_verbose("Vector search indexes created successfully")
-            except Exception as e:
-                self._log_verbose(f"Could not create vector search indexes: {e}", logging.WARNING)
-                
-        except Exception as e:
-            self._log_verbose(f"Error creating vector search indexes: {e}", logging.WARNING)
 
     def _get_embedding(self, text: str) -> List[float]:
         """Get embedding for text using the unified embedding module."""
@@ -1704,6 +1563,41 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
         entities = self.search_entity(q, limit=max_items)
         user_mem = self.search_user_memory(user_id, q, limit=max_items) if user_id else []
 
+        # Apply stable sorting to ensure deterministic order for prompt caching
+        # Sort by timestamp descending, then by content hash for stable ordering
+        def _sort_memory_results(results: List[Any]) -> List[Any]:
+            """Stable sort so identical query sets produce identical context strings."""
+            import hashlib
+            if not results:
+                return results
+            
+            def sort_key(r):
+                if not isinstance(r, dict):
+                    return (0.0, hashlib.sha256(str(r).encode()).hexdigest())
+                
+                # Prefer created_at over timestamp for better determinism
+                timestamp = r.get("created_at") or r.get("timestamp") or 0
+                if isinstance(timestamp, str):
+                    try:
+                        from datetime import datetime
+                        timestamp = datetime.fromisoformat(timestamp.replace("Z", "+00:00")).timestamp()
+                    except ValueError:
+                        timestamp = 0.0
+                timestamp = -(float(timestamp))  # Negative for descending order
+                
+                # Use full content hash for true deterministic ordering
+                content = r.get("text", "") if isinstance(r, dict) else str(r)
+                content_hash = hashlib.sha256(content.encode()).hexdigest()
+                return (timestamp, content_hash)
+            
+            return sorted(results, key=sort_key)
+
+        # Apply stable sorting to all memory results
+        short_term = _sort_memory_results(short_term)
+        long_term = _sort_memory_results(long_term)
+        entities = _sort_memory_results(entities)
+        user_mem = _sort_memory_results(user_mem)
+
         # Add sections in order of priority
         add_section("Short-term Memory Context", short_term)
         add_section("Long-term Memory Context", long_term)
@@ -1713,6 +1607,51 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
 
         return "\n".join(lines) if lines else ""
 
+    def build_cache_optimized_context(
+        self,
+        task_descr: str,
+        user_id: Optional[str] = None,
+        additional: str = "",
+        max_items: int = 3,
+        include_cache_boundary: bool = True,
+        include_in_output: Optional[bool] = None
+    ) -> Dict[str, str]:
+        """
+        Build context with cache boundary markers for prompt prefix caching optimization.
+        
+        Returns a dictionary with 'stable_prefix' and 'cache_boundary' keys.
+        The stable prefix contains deterministically ordered content that should
+        remain constant between turns for effective prompt caching.
+        
+        Args:
+            task_descr: Task description for memory search
+            user_id: Optional user ID for personalized memory
+            additional: Additional context to include in search
+            max_items: Maximum items per memory category
+            include_cache_boundary: Whether to include cache boundary marker
+            include_in_output: Whether to include memory content in output
+            
+        Returns:
+            Dict with 'stable_prefix' and 'cache_boundary' keys
+        """
+        # Preserve original behavior: include_in_output=None means debug-level-only
+        # Don't change the default behavior for backward compatibility
+            
+        # Build the stable context using existing method
+        stable_context = self.build_context_for_task(
+            task_descr=task_descr,
+            user_id=user_id,
+            additional=additional,
+            max_items=max_items,
+            include_in_output=include_in_output
+        )
+        
+        result = {"stable_prefix": stable_context}
+        
+        result["cache_boundary"] = CACHE_BOUNDARY if include_cache_boundary else ""
+            
+        return result
+
     # -------------------------------------------------------------------------
     #                      Master Reset (Everything)
     # -------------------------------------------------------------------------

src/praisonai-agents/praisonaiagents/memory/search.py

@@ -14,6 +14,9 @@ from datetime import datetime
 class SearchMixin:
     """Mixin class containing search and retrieval methods for the Memory class."""
     
+    # Allowed table names for SQLite queries to prevent SQL injection
+    _ALLOWED_SQLITE_TABLES = frozenset({"short_term", "long_term"})
+    
     def search_short_term(self, query: str, limit: int = 5, metadata_filter: Optional[Dict] = None, 
                          min_quality: Optional[float] = None, user_id: Optional[str] = None, **kwargs) -> List[Dict[str, Any]]:
         """
@@ -170,10 +173,10 @@ class SearchMixin:
         
         return self._format_vector_results(results)
     
-    def _search_mongodb_stm(self, query: str, limit: int, metadata_filter: Optional[Dict],
-                           min_quality: Optional[float], user_id: Optional[str]) -> List[Dict[str, Any]]:
-        """Search short-term memory in MongoDB."""
-        if not hasattr(self, 'stm_collection'):
+    def _search_mongodb(self, collection, query: str, limit: int, metadata_filter: Optional[Dict],
+                        min_quality: Optional[float], user_id: Optional[str]) -> List[Dict[str, Any]]:
+        """Search MongoDB collection with given parameters."""
+        if not collection:
             return []
         
         # Build search filter
@@ -191,120 +194,82 @@ class SearchMixin:
             search_filter["metadata.user_id"] = user_id
         
         # Perform search with text score
-        cursor = self.stm_collection.find(
+        cursor = collection.find(
             search_filter,
             {"score": {"$meta": "textScore"}}
         ).sort([("score", {"$meta": "textScore"})]).limit(limit)
         
         return [self._format_mongodb_result(doc) for doc in cursor]
     
+    def _search_mongodb_stm(self, query: str, limit: int, metadata_filter: Optional[Dict],
+                           min_quality: Optional[float], user_id: Optional[str]) -> List[Dict[str, Any]]:
+        """Search short-term memory in MongoDB."""
+        collection = getattr(self, 'stm_collection', None)
+        return self._search_mongodb(collection, query, limit, metadata_filter, min_quality, user_id)
+    
     def _search_mongodb_ltm(self, query: str, limit: int, metadata_filter: Optional[Dict],
                            min_quality: Optional[float], user_id: Optional[str]) -> List[Dict[str, Any]]:
         """Search long-term memory in MongoDB."""
-        if not hasattr(self, 'ltm_collection'):
+        collection = getattr(self, 'ltm_collection', None)
+        return self._search_mongodb(collection, query, limit, metadata_filter, min_quality, user_id)
+    
+    def _search_sqlite(self, conn, table: str, query: str, limit: int, metadata_filter: Optional[Dict],
+                       min_quality: Optional[float], user_id: Optional[str]) -> List[Dict[str, Any]]:
+        """Search SQLite table with given parameters."""
+        if not conn:
             return []
         
-        # Build search filter
-        search_filter = {"$text": {"$search": query}}
+        # Validate table name to prevent SQL injection
+        if table not in self._ALLOWED_SQLITE_TABLES:
+            allowed = ", ".join(sorted(self._ALLOWED_SQLITE_TABLES))
+            raise ValueError(f"Invalid table name: {table}. Allowed tables: {allowed}")
         
-        # Add additional filters
-        if metadata_filter:
-            for key, value in metadata_filter.items():
-                search_filter[f"metadata.{key}"] = value
+        # Build WHERE clause
+        where_clauses = ["content LIKE ?"]
+        params = [f"%{query}%"]
         
         if min_quality is not None:
-            search_filter["quality_score"] = {"$gte": min_quality}
+            where_clauses.append("quality_score >= ?")
+            params.append(min_quality)
+        
+        # Simple metadata filtering (JSON contains check)
+        if metadata_filter:
+            for key, value in metadata_filter.items():
+                where_clauses.append("metadata LIKE ?")
+                params.append(f'%"{key}": "{value}"%')
         
         if user_id:
-            search_filter["metadata.user_id"] = user_id
+            where_clauses.append("metadata LIKE ?")
+            params.append(f'%"user_id": "{user_id}"%')
         
-        # Perform search with text score
-        cursor = self.ltm_collection.find(
-            search_filter,
-            {"score": {"$meta": "textScore"}}
-        ).sort([("score", {"$meta": "textScore"})]).limit(limit)
+        where_clause = " AND ".join(where_clauses)
         
-        return [self._format_mongodb_result(doc) for doc in cursor]
+        query_sql = f"""
+            SELECT id, content, metadata, timestamp, quality_score
+            FROM {table}
+            WHERE {where_clause}
+            ORDER BY quality_score DESC, timestamp DESC
+            LIMIT ?
+        """
+        
+        params.append(limit)
+        
+        cursor = conn.execute(query_sql, params)
+        rows = cursor.fetchall()
+        
+        return [self._format_sqlite_result(dict(row)) for row in rows]
     
     def _search_sqlite_stm(self, query: str, limit: int, metadata_filter: Optional[Dict],
                           min_quality: Optional[float], user_id: Optional[str]) -> List[Dict[str, Any]]:
         """Search short-term memory in SQLite."""
-        conn = self._get_stm_conn()
-        
-        # Build WHERE clause
-        where_clauses = ["content LIKE ?"]
-        params = [f"%{query}%"]
-        
-        if min_quality is not None:
-            where_clauses.append("quality_score >= ?")
-            params.append(min_quality)
-        
-        # Simple metadata filtering (JSON contains check)
-        if metadata_filter:
-            for key, value in metadata_filter.items():
-                where_clauses.append("metadata LIKE ?")
-                params.append(f'%"{key}": "{value}"%')
-        
-        if user_id:
-            where_clauses.append("metadata LIKE ?")
-            params.append(f'%"user_id": "{user_id}"%')
-        
-        where_clause = " AND ".join(where_clauses)
-        
-        query_sql = f"""
-            SELECT id, content, metadata, timestamp, quality_score
-            FROM short_term
-            WHERE {where_clause}
-            ORDER BY quality_score DESC, timestamp DESC
-            LIMIT ?
-        """
-        
-        params.append(limit)
-        
-        cursor = conn.execute(query_sql, params)
-        rows = cursor.fetchall()
-        
-        return [self._format_sqlite_result(dict(row)) for row in rows]
+        conn = self._get_stm_conn()  # Let AttributeError propagate naturally
+        return self._search_sqlite(conn, "short_term", query, limit, metadata_filter, min_quality, user_id)
     
     def _search_sqlite_ltm(self, query: str, limit: int, metadata_filter: Optional[Dict],
                           min_quality: Optional[float], user_id: Optional[str]) -> List[Dict[str, Any]]:
         """Search long-term memory in SQLite."""
-        conn = self._get_ltm_conn()
-        
-        # Build WHERE clause
-        where_clauses = ["content LIKE ?"]
-        params = [f"%{query}%"]
-        
-        if min_quality is not None:
-            where_clauses.append("quality_score >= ?")
-            params.append(min_quality)
-        
-        # Simple metadata filtering (JSON contains check)
-        if metadata_filter:
-            for key, value in metadata_filter.items():
-                where_clauses.append("metadata LIKE ?")
-                params.append(f'%"{key}": "{value}"%')
-        
-        if user_id:
-            where_clauses.append("metadata LIKE ?")
-            params.append(f'%"user_id": "{user_id}"%')
-        
-        where_clause = " AND ".join(where_clauses)
-        
-        query_sql = f"""
-            SELECT id, content, metadata, timestamp, quality_score
-            FROM long_term
-            WHERE {where_clause}
-            ORDER BY quality_score DESC, timestamp DESC
-            LIMIT ?
-        """
-        
-        params.append(limit)
-        
-        cursor = conn.execute(query_sql, params)
-        rows = cursor.fetchall()
-        
-        return [self._format_sqlite_result(dict(row)) for row in rows]
+        conn = self._get_ltm_conn()  # Let AttributeError propagate naturally
+        return self._search_sqlite(conn, "long_term", query, limit, metadata_filter, min_quality, user_id)
     
     def _format_vector_results(self, results: Dict) -> List[Dict[str, Any]]:
         """Format ChromaDB vector search results."""

src/praisonai-agents/praisonaiagents/memory/storage.py

@@ -1,233 +0,0 @@
-"""
-Storage and persistence functionality for Memory class.
-
-This module contains methods related to data storage, database connections, 
-and persistence operations. Split from the main memory.py file for better maintainability.
-"""
-
-import os
-import sqlite3
-import threading
-import logging
-from typing import Any, Dict, List, Optional, Set
-
-# Import shared utility functions
-from .utils import _check_chromadb, _get_chromadb, _check_pymongo, _get_pymongo
-
-
-class StorageMixin:
-    """Mixin class containing storage and persistence methods for the Memory class."""
-    
-    def _get_stm_conn(self):
-        """Get thread-local short-term memory SQLite connection."""
-        if not hasattr(self._local, 'stm_conn') or self._local.stm_conn is None:
-            stm_db_path = self.cfg.get("short_db", "short_term_memory.db")
-            self._local.stm_conn = sqlite3.connect(stm_db_path, check_same_thread=False)
-            self._local.stm_conn.row_factory = sqlite3.Row
-            
-            # Register for cleanup
-            with self._connection_lock:
-                self._all_connections.add(self._local.stm_conn)
-                
-        return self._local.stm_conn
-
-    def _get_ltm_conn(self):
-        """Get thread-local long-term memory SQLite connection.""" 
-        if not hasattr(self._local, 'ltm_conn') or self._local.ltm_conn is None:
-            ltm_db_path = self.cfg.get("long_db", "long_term_memory.db")
-            self._local.ltm_conn = sqlite3.connect(ltm_db_path, check_same_thread=False)
-            self._local.ltm_conn.row_factory = sqlite3.Row
-            
-            # Register for cleanup
-            with self._connection_lock:
-                self._all_connections.add(self._local.ltm_conn)
-                
-        return self._local.ltm_conn
-    
-    def _init_stm(self):
-        """Initialize short-term memory table."""
-        conn = self._get_stm_conn()
-        with self._write_lock:
-            conn.execute('''
-                CREATE TABLE IF NOT EXISTS short_term (
-                    id INTEGER PRIMARY KEY AUTOINCREMENT,
-                    content TEXT NOT NULL,
-                    metadata TEXT,
-                    timestamp TEXT,
-                    quality_score REAL DEFAULT 0.0
-                )
-            ''')
-            conn.commit()
-
-    def _init_ltm(self):
-        """Initialize long-term memory table.""" 
-        conn = self._get_ltm_conn()
-        with self._write_lock:
-            conn.execute('''
-                CREATE TABLE IF NOT EXISTS long_term (
-                    id INTEGER PRIMARY KEY AUTOINCREMENT,
-                    content TEXT NOT NULL,
-                    metadata TEXT,
-                    timestamp TEXT,
-                    quality_score REAL DEFAULT 0.0
-                )
-            ''')
-            conn.commit()
-    
-    def _init_chroma(self):
-        """Initialize ChromaDB for vector storage."""
-        if not _check_chromadb():
-            return
-            
-        chromadb = _get_chromadb()
-        chroma_settings = chromadb["settings"]
-        
-        rag_db_path = self.cfg.get("rag_db_path", "rag_memory_db")
-        
-        # Create directory if it doesn't exist
-        os.makedirs(rag_db_path, exist_ok=True)
-        
-        self.chroma_client = chromadb["module"].PersistentClient(
-            path=rag_db_path,
-            settings=chroma_settings(anonymized_telemetry=False)
-        )
-        
-        # Initialize collections
-        try:
-            self.stm_collection = self.chroma_client.get_collection("short_term_memory")
-        except ValueError:  # ChromaDB raises ValueError for missing collections
-            self.stm_collection = self.chroma_client.create_collection("short_term_memory")
-        except Exception as e:
-            from .._logging import get_logger
-            logger = get_logger(__name__)
-            logger.error(f"Failed to initialize short_term_memory collection: {e}")
-            raise
-            
-        try:
-            self.ltm_collection = self.chroma_client.get_collection("long_term_memory")
-        except ValueError:  # ChromaDB raises ValueError for missing collections
-            self.ltm_collection = self.chroma_client.create_collection("long_term_memory")
-        except Exception as e:
-            from .._logging import get_logger
-            logger = get_logger(__name__)
-            logger.error(f"Failed to initialize long_term_memory collection: {e}")
-            raise
-    
-    def _init_mongodb(self):
-        """Initialize MongoDB for storage."""
-        if not _check_pymongo():
-            return
-            
-        pymongo = _get_pymongo()
-        
-        config = self.cfg.get("config", {})
-        connection_string = config.get("connection_string", "mongodb://localhost:27017/")
-        database_name = config.get("database", "praisonai")
-        
-        # Connection pool settings
-        max_pool_size = config.get("max_pool_size", 50)
-        min_pool_size = config.get("min_pool_size", 10)
-        max_idle_time = config.get("max_idle_time", 30000)
-        server_selection_timeout = config.get("server_selection_timeout", 5000)
-        
-        self.mongo_client = pymongo["MongoClient"](
-            connection_string,
-            maxPoolSize=max_pool_size,
-            minPoolSize=min_pool_size,
-            maxIdleTimeMS=max_idle_time,
-            serverSelectionTimeoutMS=server_selection_timeout
-        )
-        
-        self.mongo_db = self.mongo_client[database_name]
-        self.stm_collection = self.mongo_db.short_term_memory
-        self.ltm_collection = self.mongo_db.long_term_memory
-        
-        # Create indexes for better performance
-        self._create_mongodb_indexes()
-        
-        # Initialize vector search if enabled
-        use_vector_search = config.get("use_vector_search", False)
-        if use_vector_search:
-            self._create_vector_search_indexes()
-    
-    def _create_mongodb_indexes(self):
-        """Create MongoDB indexes for better query performance."""
-        try:
-            # Text indexes for search
-            self.stm_collection.create_index([("content", "text")])
-            self.ltm_collection.create_index([("content", "text")])
-            
-            # Compound indexes for common queries
-            self.stm_collection.create_index([("timestamp", -1), ("quality_score", -1)])
-            self.ltm_collection.create_index([("timestamp", -1), ("quality_score", -1)])
-            
-            logging.info("MongoDB indexes created successfully")
-        except Exception as e:
-            logging.warning(f"Failed to create MongoDB indexes: {e}")
-    
-    def _create_vector_search_indexes(self):
-        """Create MongoDB Atlas Vector Search indexes.""" 
-        try:
-            # Vector search indexes would be created via MongoDB Atlas UI
-            # or using the Atlas Admin API
-            logging.info("Vector search indexes configuration completed")
-        except Exception as e:
-            logging.warning(f"Failed to configure vector search indexes: {e}")
-    
-    def reset_short_term(self):
-        """Clear all short-term memory."""
-        if self.use_rag and hasattr(self, 'stm_collection'):
-            try:
-                # Clear ChromaDB collection
-                self.stm_collection.delete()
-                # Recreate collection
-                self.stm_collection = self.chroma_client.create_collection("short_term_memory") 
-                self._log_verbose("Short-term RAG memory reset")
-            except Exception as e:
-                logging.warning(f"Failed to reset ChromaDB STM: {e}")
-        
-        if self.use_mongodb and hasattr(self, 'stm_collection'):
-            try:
-                self.stm_collection.delete_many({})
-                self._log_verbose("Short-term MongoDB memory reset")
-            except Exception as e:
-                logging.warning(f"Failed to reset MongoDB STM: {e}")
-        
-        # Clear SQLite STM
-        try:
-            conn = self._get_stm_conn()
-            with self._write_lock:
-                conn.execute("DELETE FROM short_term")
-                conn.commit()
-            self._log_verbose("Short-term SQLite memory reset")
-        except Exception as e:
-            logging.warning(f"Failed to reset SQLite STM: {e}")
-
-    def reset_long_term(self):
-        """Clear all long-term memory."""
-        if self.use_rag and hasattr(self, 'ltm_collection'):
-            try:
-                # Clear ChromaDB collection
-                self.ltm_collection.delete()
-                # Recreate collection  
-                self.ltm_collection = self.chroma_client.create_collection("long_term_memory")
-                self._log_verbose("Long-term RAG memory reset")
-            except Exception as e:
-                logging.warning(f"Failed to reset ChromaDB LTM: {e}")
-        
-        if self.use_mongodb and hasattr(self, 'ltm_collection'):
-            try:
-                self.ltm_collection.delete_many({})
-                self._log_verbose("Long-term MongoDB memory reset")
-            except Exception as e:
-                logging.warning(f"Failed to reset MongoDB LTM: {e}")
-        
-        # Clear SQLite LTM
-        try:
-            conn = self._get_ltm_conn()
-            with self._write_lock:
-                conn.execute("DELETE FROM long_term")
-                conn.commit()
-            self._log_verbose("Long-term SQLite memory reset")
-        except Exception as e:
-            logging.warning(f"Failed to reset SQLite LTM: {e}")

src/praisonai-agents/praisonaiagents/memory/workflows.py

@@ -27,7 +27,7 @@ from pathlib import Path
 from typing import Any, Dict, List, Optional, Callable
 from dataclasses import dataclass, field
 from ..task.task import Task
-from ..workflows.results import StepResult, StepError, WorkflowResult, StepStatus, ErrorStrategy
+from ..workflows.results import StepError, WorkflowResult, StepStatus, ErrorStrategy
 
 logger = get_logger(__name__)
 

src/praisonai-agents/praisonaiagents/permissions/__init__.py

@@ -34,6 +34,7 @@ __all__ = [
     "PermissionAction",
     "PermissionMode",
     "PermissionResult",
+    "PersistentApproval",
     "DoomLoopDetector",
 ]
 
@@ -60,6 +61,10 @@ def __getattr__(name: str):
         from .rules import PermissionResult
         return PermissionResult
     
+    if name == "PersistentApproval":
+        from .rules import PersistentApproval
+        return PersistentApproval
+    
     if name == "DoomLoopDetector":
         from .doom_loop import DoomLoopDetector
         return DoomLoopDetector

src/praisonai-agents/praisonaiagents/permissions/manager.py

@@ -111,6 +111,11 @@ class PermissionManager:
         except IOError as e:
             logger.error(f"Failed to save rules: {e}")
     
+    def save_rules(self):
+        """Public method to save rules to disk."""
+        with self._lock:
+            self._save_rules()
+    
     def _load_approvals(self):
         """Load approvals from disk."""
         path = self._get_approvals_path()
@@ -359,3 +364,71 @@ class PermissionManager:
             self._approvals = [PersistentApproval.from_dict(a) for a in data.get("approvals", [])]
             if "agent_name" in data:
                 self.agent_name = data["agent_name"]
+    
+    def load_rules_from_config(self, permissions_config: Dict[str, Any], priority_base: int = 50, persist: bool = False):
+        """
+        Load permission rules from configuration (YAML/CLI/Python).
+        
+        Args:
+            permissions_config: Dictionary mapping patterns to actions or detailed configs
+                               e.g., {"read:*": "allow", "bash:rm *": {"action": "deny", "description": "..."}}
+            priority_base: Base priority for these rules (default: 50, between default and user rules)
+            persist: Whether to persist these rules to disk (default: False for ephemeral CI/YAML rules)
+        """
+        if not permissions_config:
+            return
+        
+        with self._lock:
+            incoming_rules = []
+            
+            for pattern, config in permissions_config.items():
+                try:
+                    if isinstance(config, str):
+                        # Simple format: pattern -> action
+                        rule = PermissionRule.from_config(pattern, config, priority=priority_base)
+                    elif isinstance(config, dict):
+                        # Detailed format: pattern -> {action, description, ...}
+                        action = config.get("action", "ask")
+                        rule = PermissionRule.from_config(
+                            pattern,
+                            action,
+                            description=config.get("description"),
+                            is_regex=config.get("is_regex", False),
+                            priority=config.get("priority", priority_base),
+                            agent_name=config.get("agent_name"),
+                            enabled=config.get("enabled", True),
+                        )
+                    else:
+                        logger.warning(
+                            "Invalid permission config for pattern '%s': %r. "
+                            "Use 'allow|deny|ask' or {action, description, is_regex, priority, agent_name, enabled}.",
+                            pattern,
+                            config,
+                        )
+                        continue
+                except (TypeError, ValueError) as e:
+                    logger.warning(
+                        "Skipping invalid permission rule for pattern '%s': %s. "
+                        "Valid actions are: allow, deny, ask.",
+                        pattern,
+                        e,
+                    )
+                    continue
+                    
+                incoming_rules.append(rule)
+            
+            # Remove existing rules with same pattern to avoid duplicates
+            # (pattern, agent_name, is_regex) uniquely identifies a rule for matching
+            incoming_keys = {(r.pattern, r.agent_name, r.is_regex) for r in incoming_rules}
+            self._rules = [
+                r for r in self._rules
+                if (r.pattern, r.agent_name, r.is_regex) not in incoming_keys
+            ]
+            self._rules.extend(incoming_rules)
+            
+            # Re-sort by priority
+            self._rules.sort(key=lambda r: r.priority, reverse=True)
+            
+            # Only persist if explicitly requested (not for ephemeral CI/YAML rules)
+            if persist:
+                self._save_rules()

src/praisonai-agents/praisonaiagents/permissions/rules.py

@@ -114,6 +114,29 @@ class PermissionRule:
             enabled=data.get("enabled", True),
             created_at=data.get("created_at", time.time()),
         )
+    
+    @classmethod
+    def from_config(cls, pattern: str, action: str, **kwargs) -> "PermissionRule":
+        """
+        Create rule from configuration (YAML/CLI/Python).
+        
+        Args:
+            pattern: The permission pattern (e.g., "read:*", "bash:git *")
+            action: The action string (allow/deny/ask)
+            **kwargs: Optional attributes like description, priority, is_regex
+            
+        Returns:
+            PermissionRule instance
+        """
+        return cls(
+            pattern=pattern,
+            action=PermissionAction(action),
+            description=kwargs.get("description", f"Rule for {pattern}"),
+            is_regex=kwargs.get("is_regex", False),
+            priority=kwargs.get("priority", 0),
+            agent_name=kwargs.get("agent_name"),
+            enabled=kwargs.get("enabled", True),
+        )
 
 
 @dataclass

src/praisonai-agents/praisonaiagents/planning/approval.py

@@ -16,6 +16,7 @@ Features:
 import asyncio
 import inspect
 import logging
+import sys
 from praisonaiagents._logging import get_logger
 from typing import Callable, Optional, Union, TYPE_CHECKING
 
@@ -93,6 +94,12 @@ class ApprovalCallback:
                 return False
                 
         # Default: require explicit approval (return False)
+        # However, auto-approve when not in an interactive environment (e.g., CI/scripts)
+        if sys.stdin is None or not sys.stdin.isatty():
+            logger.debug(f"Auto-approving plan in non-interactive environment: {plan.name}")
+            plan.approve()
+            return True
+        
         logger.debug(f"Plan requires explicit approval: {plan.name}")
         return False
     
@@ -127,6 +134,12 @@ class ApprovalCallback:
                 return False
                 
         # Default: require explicit approval (return False)
+        # However, auto-approve when not in an interactive environment (e.g., CI/scripts)
+        if sys.stdin is None or not sys.stdin.isatty():
+            logger.debug(f"Auto-approving plan in non-interactive environment: {plan.name}")
+            plan.approve()
+            return True
+        
         logger.debug(f"Plan requires explicit approval: {plan.name}")
         return False
     

src/praisonai-agents/praisonaiagents/process/process.py

@@ -196,8 +196,13 @@ class Process:
         except (json.JSONDecodeError, ValueError, TypeError) as e:
             raise Exception(f"Failed to parse response: {response}") from e
 
-    def _create_loop_subtasks(self, loop_task: Task):
-        """Create subtasks for a loop task from input file."""
+    def _create_loop_subtasks(self, loop_task: Task, decision_mode: bool = False):
+        """Create subtasks for a loop task from input file.
+        
+        Args:
+            loop_task: The loop task to create subtasks for
+            decision_mode: If True, create decision-type subtasks with conditions (for start tasks)
+        """
         if not loop_task.input_file:
             logging.warning(f"_create_loop_subtasks called for {loop_task.name} but no input_file specified")
             return
@@ -229,20 +234,45 @@ class Process:
                         task_count += 1
                         logging.debug(f"Creating subtask from CSV row {i+1}: {task_desc}")
 
-                        row_task = Task(
-                            description=f"{loop_task.description}\n{task_desc}" if loop_task.description else task_desc,
-                            agent=loop_task.agent,
-                            name=f"{loop_task.name}_{task_count}" if loop_task.name else task_desc,
-                            expected_output=getattr(loop_task, 'expected_output', None),
-                            on_task_complete=loop_task.callback,  # Inherit callback from parent loop task
-                            is_start=(task_count == 1),
-                            task_type="task"
-                        )
+                        # Configure task based on decision_mode
+                        if decision_mode:
+                            # For start tasks: create decision-type tasks with conditions and inherited next_tasks
+                            inherited_next_tasks = loop_task.next_tasks if loop_task.next_tasks else []
+                            row_name = f"{loop_task.name}_{task_count}" if loop_task.name else task_desc
+                            row_task = Task(
+                                description=f"{loop_task.description}\n{task_desc}" if loop_task.description else task_desc,
+                                agent=loop_task.agent,
+                                name=row_name,
+                                expected_output=getattr(loop_task, 'expected_output', None),
+                                on_task_complete=loop_task.callback,  # Inherit callback from parent loop task
+                                is_start=(task_count == 1),
+                                task_type="decision",  # Change to decision type for start tasks
+                                next_tasks=inherited_next_tasks,  # Inherit parent's next tasks
+                                condition={
+                                    "done": inherited_next_tasks if inherited_next_tasks else [],
+                                    "retry": [row_name],
+                                    "exit": []  # Empty list for exit condition
+                                }
+                            )
+                        else:
+                            # For regular loop tasks: create basic task-type subtasks
+                            row_task = Task(
+                                description=f"{loop_task.description}\n{task_desc}" if loop_task.description else task_desc,
+                                agent=loop_task.agent,
+                                name=f"{loop_task.name}_{task_count}" if loop_task.name else task_desc,
+                                expected_output=getattr(loop_task, 'expected_output', None),
+                                on_task_complete=loop_task.callback,  # Inherit callback from parent loop task
+                                is_start=(task_count == 1),
+                                task_type="task"
+                            )
+                        
                         self.tasks[row_task.id] = row_task
                         new_tasks.append(row_task)
 
                         if previous_task:
                             previous_task.next_tasks = [row_task.name]
+                            if decision_mode:
+                                previous_task.condition["done"] = [row_task.name]  # Use "done" consistently for decision tasks
                         previous_task = row_task
 
                     logging.info(f"Created {task_count} subtasks from CSV file for {loop_task.name}")
@@ -254,21 +284,42 @@ class Process:
                     for i, line in enumerate(lines):
                         if not line.strip():  # Skip empty lines
                             continue
-                            
-                        row_task = Task(
-                            description=f"{loop_task.description}\n{line.strip()}" if loop_task.description else line.strip(),
-                            agent=loop_task.agent,
-                            name=f"{loop_task.name}_{i+1}" if loop_task.name else line.strip(),
-                            expected_output=getattr(loop_task, 'expected_output', None),
-                            on_task_complete=loop_task.callback,  # Inherit callback from parent loop task
-                            is_start=(i == 0),
-                            task_type="task"
-                        )
+                        
+                        # Configure task based on decision_mode
+                        if decision_mode:
+                            # For start tasks: create tasks with conditions 
+                            row_task = Task(
+                                description=f"{loop_task.description}\n{line.strip()}" if loop_task.description else line.strip(),
+                                agent=loop_task.agent,
+                                name=f"{loop_task.name}_{i+1}" if loop_task.name else line.strip(),
+                                expected_output=getattr(loop_task, 'expected_output', None),
+                                on_task_complete=loop_task.callback,  # Inherit callback from parent loop task
+                                is_start=(i == 0),
+                                task_type="task",
+                                condition={
+                                    "complete": ["next"],
+                                    "retry": ["current"]
+                                }
+                            )
+                        else:
+                            # For regular loop tasks: create basic tasks
+                            row_task = Task(
+                                description=f"{loop_task.description}\n{line.strip()}" if loop_task.description else line.strip(),
+                                agent=loop_task.agent,
+                                name=f"{loop_task.name}_{i+1}" if loop_task.name else line.strip(),
+                                expected_output=getattr(loop_task, 'expected_output', None),
+                                on_task_complete=loop_task.callback,  # Inherit callback from parent loop task
+                                is_start=(i == 0),
+                                task_type="task"
+                            )
+                        
                         self.tasks[row_task.id] = row_task
                         new_tasks.append(row_task)
 
                         if previous_task:
                             previous_task.next_tasks = [row_task.name]
+                            if decision_mode:
+                                previous_task.condition["complete"] = [row_task.name]
                         previous_task = row_task
 
                     logging.info(f"Created {len(new_tasks)} subtasks from text file for {loop_task.name}")
@@ -276,7 +327,12 @@ class Process:
             if new_tasks and loop_task.next_tasks:
                 # Connect last subtask to loop task's next tasks
                 last_task = new_tasks[-1]
-                last_task.next_tasks = loop_task.next_tasks
+                if decision_mode:
+                    # For decision mode, ensure the last task points to parent's next tasks
+                    if not last_task.next_tasks:
+                        last_task.next_tasks = loop_task.next_tasks
+                else:
+                    last_task.next_tasks = loop_task.next_tasks
 
         except Exception as e:
             logging.error(f"Failed to create subtasks for loop task {loop_task.name}: {e}")
@@ -562,7 +618,6 @@ class Process:
 
         current_task = start_task
         visited_tasks = set()
-        loop_data = {}  # Store loop-specific data
 
         # TODO: start task with loop feature is not available in aworkflow method
 
@@ -732,9 +787,6 @@ Subtask: {st.name}
                 logging.debug(f"Task next_tasks: {current_task.next_tasks}")
                 yield task_id
                 visited_tasks.add(task_id)
-                # Reset description to original after execution to prevent context accumulation
-                if hasattr(current_task, '_original_description'):
-                    current_task.description = current_task._original_description
 
                 # Only end workflow if no next_tasks AND no conditions
                 if not current_task.next_tasks and not current_task.condition and not any(
@@ -772,22 +824,6 @@ Subtask: {st.name}
                         logging.debug(f"=== Skipping reset for loop/decision/subtask/parallel or rerun=False: {subtask_name} ===")
                         logging.debug(f"Keeping status as: {self.tasks[task_id].status}")
 
-            # Handle loop progression
-            if current_task.task_type == "loop":
-                loop_key = f"loop_{current_task.name}"
-                if loop_key in loop_data:
-                    loop_info = loop_data[loop_key]
-                    loop_info["index"] += 1
-                    has_more = loop_info["remaining"] > 0
-
-                    # Update result to trigger correct condition
-                    if current_task.result:
-                        result = current_task.result.raw
-                        if has_more:
-                            result += "\nmore"
-                        else:
-                            result += "\ndone"
-                        current_task.result.raw = result
 
             # Determine next task based on result
             next_task = None
@@ -1112,104 +1148,29 @@ Provide a JSON with the structure:
         if start_task and start_task.task_type == "loop" and not start_task.input_file:
             start_task.input_file = "tasks.csv"
 
-        # --- If loop + input_file, read file & create tasks
+        # --- If loop + input_file, read file & create tasks using consolidated helper
         if start_task and start_task.task_type == "loop" and getattr(start_task, "input_file", None):
             try:
-                file_ext = os.path.splitext(start_task.input_file)[1].lower()
-                new_tasks = []
-
-                if file_ext == ".csv":
-                    with open(start_task.input_file, "r", encoding="utf-8") as f:
-                        reader = csv.reader(f, quotechar='"', escapechar='\\')  # Handle quoted/escaped fields
-                        previous_task = None
-                        task_count = 0
-
-                        for i, row in enumerate(reader):
-                            if not row:  # Skip truly empty rows
-                                continue
-
-                            # Properly handle Q&A pairs with potential commas
-                            task_desc = row[0].strip() if row else ""
-                            if len(row) > 1:
-                                # Preserve all fields in case of multiple commas
-                                question = row[0].strip()
-                                answer = ",".join(field.strip() for field in row[1:])
-                                task_desc = f"Question: {question}\nAnswer: {answer}"
-
-                            if not task_desc:  # Skip rows with empty content
-                                continue
-
-                            task_count += 1
-                            logging.debug(f"Processing CSV row {i+1}: {task_desc}")
-
-                            # Inherit next_tasks from parent loop task
-                            inherited_next_tasks = start_task.next_tasks if start_task.next_tasks else []
-
-                            row_task = Task(
-                                description=f"{start_task.description}\n{task_desc}" if start_task.description else task_desc,
-                                agent=start_task.agent,
-                                name=f"{start_task.name}_{task_count}" if start_task.name else task_desc,
-                                expected_output=getattr(start_task, 'expected_output', None),
-                                on_task_complete=start_task.callback,  # Inherit callback from parent loop task
-                                is_start=(task_count == 1),
-                                task_type="decision",  # Change to decision type
-                                next_tasks=inherited_next_tasks,  # Inherit parent's next tasks
-                                condition={
-                                    "done": inherited_next_tasks if inherited_next_tasks else ["next"],  # Use full inherited_next_tasks
-                                    "retry": ["current"],
-                                    "exit": []  # Empty list for exit condition
-                                }
-                            )
-                            self.tasks[row_task.id] = row_task
-                            new_tasks.append(row_task)
-
-                            if previous_task:
-                                previous_task.next_tasks = [row_task.name]
-                                previous_task.condition["done"] = [row_task.name]  # Use "done" consistently
-                            previous_task = row_task
-
-                            # For the last task in the loop, ensure it points to parent's next tasks
-                            if task_count > 0 and not row_task.next_tasks:
-                                row_task.next_tasks = inherited_next_tasks
-
-                        logging.info(f"Processed {task_count} rows from CSV file")
-                else:
-                    # If not CSV, read lines
-                    with open(start_task.input_file, "r", encoding="utf-8") as f:
-                        lines = f.read().splitlines()
-                        previous_task = None
-                        for i, line in enumerate(lines):
-                            row_task = Task(
-                                description=f"{start_task.description}\n{line.strip()}" if start_task.description else line.strip(),
-                                agent=start_task.agent,
-                                name=f"{start_task.name}_{i+1}" if start_task.name else line.strip(),
-                                expected_output=getattr(start_task, 'expected_output', None),
-                                on_task_complete=start_task.callback,  # Inherit callback from parent loop task
-                                is_start=(i == 0),
-                                task_type="task",
-                                condition={
-                                    "complete": ["next"],
-                                    "retry": ["current"]
-                                }
-                            )
-                            self.tasks[row_task.id] = row_task
-                            new_tasks.append(row_task)
-
-                            if previous_task:
-                                previous_task.next_tasks = [row_task.name]
-                                previous_task.condition["complete"] = [row_task.name]
-                            previous_task = row_task
-
-                if new_tasks:
-                    start_task = new_tasks[0]
-                    logging.info(f"Created {len(new_tasks)} tasks from: {start_task.input_file}")
+                parent_loop_task = start_task
+                parent_input_file = parent_loop_task.input_file
+                self._create_loop_subtasks(parent_loop_task, decision_mode=True)
+                # Get the first created subtask as the new start task
+                subtasks = [
+                    t for t in self.tasks.values()
+                    if t.name.startswith(parent_loop_task.name + "_")
+                ]
+                if subtasks:
+                    # Mark parent loop task as completed and find start subtask
+                    parent_loop_task.status = "completed"
+                    parent_loop_task._subtasks_created = True
+                    start_task = next((t for t in subtasks if t.is_start), subtasks[0])
+                    logging.info(f"Created {len(subtasks)} tasks from: {parent_input_file}")
             except Exception as e:
                 logging.error(f"Failed to read file tasks: {e}")
 
         # end of start task handling
         current_task = start_task
         visited_tasks = set()
-        loop_data = {}  # Store loop-specific data
 
         while current_task:
             current_iter += 1
@@ -1256,7 +1217,7 @@ Tasks by type:
                 break  # Exit immediately to prevent task reset
 
 
-            # Handle loop task file reading at runtime
+            # Handle loop task file reading at runtime using consolidated helper
             if (current_task.task_type == "loop" and
                 current_task is not start_task and
                 getattr(current_task, "_subtasks_created", False) is not True):
@@ -1266,66 +1227,18 @@ Tasks by type:
 
                 if getattr(current_task, "input_file", None):
                     try:
-                        file_ext = os.path.splitext(current_task.input_file)[1].lower()
-                        new_tasks = []
-
-                        if file_ext == ".csv":
-                            with open(current_task.input_file, "r", encoding="utf-8") as f:
-                                reader = csv.reader(f)
-                                previous_task = None
-                                for i, row in enumerate(reader):
-                                    if row:  # Skip empty rows
-                                        task_desc = row[0]  # Take first column
-                                        row_task = Task(
-                                            description=f"{current_task.description}\n{task_desc}" if current_task.description else task_desc,
-                                            agent=current_task.agent,
-                                            name=f"{current_task.name}_{i+1}" if current_task.name else task_desc,
-                                            expected_output=getattr(current_task, 'expected_output', None),
-                                            on_task_complete=current_task.callback,  # Inherit callback from parent loop task
-                                            is_start=(i == 0),
-                                            task_type="task",
-                                            condition={
-                                                "complete": ["next"],
-                                                "retry": ["current"]
-                                            }
-                                        )
-                                        self.tasks[row_task.id] = row_task
-                                        new_tasks.append(row_task)
-
-                                        if previous_task:
-                                            previous_task.next_tasks = [row_task.name]
-                                            previous_task.condition["complete"] = [row_task.name]
-                                        previous_task = row_task
-                        else:
-                            with open(current_task.input_file, "r", encoding="utf-8") as f:
-                                lines = f.read().splitlines()
-                                previous_task = None
-                                for i, line in enumerate(lines):
-                                    row_task = Task(
-                                        description=f"{current_task.description}\n{line.strip()}" if current_task.description else line.strip(),
-                                        agent=current_task.agent,
-                                        name=f"{current_task.name}_{i+1}" if current_task.name else line.strip(),
-                                        expected_output=getattr(current_task, 'expected_output', None),
-                                        on_task_complete=current_task.callback,  # Inherit callback from parent loop task
-                                        is_start=(i == 0),
-                                        task_type="task",
-                                        condition={
-                                            "complete": ["next"],
-                                            "retry": ["current"]
-                                        }
-                                    )
-                                    self.tasks[row_task.id] = row_task
-                                    new_tasks.append(row_task)
-
-                                    if previous_task:
-                                        previous_task.next_tasks = [row_task.name]
-                                        previous_task.condition["complete"] = [row_task.name]
-                                    previous_task = row_task
-
-                        if new_tasks:
-                            current_task.next_tasks = [new_tasks[0].name]
+                        self._create_loop_subtasks(current_task, decision_mode=False)
+                        # Update current task to point to first subtask
+                        subtasks = [
+                            t for t in self.tasks.values()
+                            if t.name.startswith(current_task.name + "_")
+                        ]
+                        if subtasks:
+                            # Find first subtask by is_start flag or first in list
+                            first_subtask = next((t for t in subtasks if t.is_start), subtasks[0])
+                            current_task.next_tasks = [first_subtask.name]
                             current_task._subtasks_created = True
-                            logging.info(f"Created {len(new_tasks)} tasks from: {current_task.input_file} for loop task {current_task.name}")
+                            logging.info(f"Created {len(subtasks)} tasks from: {current_task.input_file} for loop task {current_task.name}")
                     except Exception as e:
                         logging.error(f"Failed to read file tasks for loop task {current_task.name}: {e}")
 
@@ -1451,9 +1364,6 @@ Subtask: {st.name}
                 logging.debug(f"Task next_tasks: {current_task.next_tasks}")
                 yield task_id
                 visited_tasks.add(task_id)
-                # Reset description to original after execution to prevent context accumulation
-                if hasattr(current_task, '_original_description'):
-                    current_task.description = current_task._original_description
 
                 # Only end workflow if no next_tasks AND no conditions
                 if not current_task.next_tasks and not current_task.condition and not any(
@@ -1489,22 +1399,6 @@ Subtask: {st.name}
                     logging.debug(f"Keeping status as: {self.tasks[task_id].status}")
 
 
-            # Handle loop progression
-            if current_task.task_type == "loop":
-                loop_key = f"loop_{current_task.name}"
-                if loop_key in loop_data:
-                    loop_info = loop_data[loop_key]
-                    loop_info["index"] += 1
-                    has_more = loop_info["remaining"] > 0
-
-                    # Update result to trigger correct condition
-                    if current_task.result:
-                        result = current_task.result.raw
-                        if has_more:
-                            result += "\nmore"
-                        else:
-                            result += "\ndone"
-                        current_task.result.raw = result
 
             # Determine next task based on result
             next_task = None

src/praisonai-agents/praisonaiagents/runtime/README.md

@@ -0,0 +1,219 @@
+# Runtime Tool Result Middleware
+
+This module provides runtime-scoped tool result middleware for plugin harnesses. Plugin harnesses return tool results in vendor-specific shapes, and this middleware normalizes them before they reach hooks and memory adapters.
+
+## Overview
+
+The middleware system follows the protocol-driven design of PraisonAI Agents:
+
+- **Core Protocol**: `RuntimeToolResultMiddleware` - defines the normalization interface
+- **Standardized Format**: `NormalizedToolResult` - consistent result format for downstream consumers
+- **Registry**: `RuntimeRegistry` - manages per-runtime middleware instances
+- **Zero Overhead**: Native `praisonai` runtime bypasses middleware to avoid allocation
+
+## Architecture
+
+```
+Tool Execution Flow:
+┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
+│ Plugin Harness  │───▶│ Runtime         │───▶│ Global Hooks    │───▶│ Memory          │
+│ (vendor result) │    │ Middleware      │    │ (normalized)    │    │ Persistence     │
+└─────────────────┘    └─────────────────┘    └─────────────────┘    └─────────────────┘
+                              │
+                              ▼
+                    ┌─────────────────┐
+                    │ NormalizedTool  │
+                    │ Result          │
+                    └─────────────────┘
+```
+
+## For Plugin Authors
+
+### 1. Implement the Middleware Protocol
+
+```python
+from praisonaiagents.runtime import RuntimeToolResultMiddleware, NormalizedToolResult, MiddlewareContext
+
+class MyHarnessMiddleware:
+    @property
+    def runtime_id(self) -> str:
+        return "my_plugin_harness"
+    
+    def normalize(self, result: Any, tool_name: str, ctx: MiddlewareContext) -> NormalizedToolResult:
+        # Convert your vendor-specific result to standard format
+        if isinstance(result, MyVendorResult):
+            return NormalizedToolResult(
+                content=result.data,
+                success=result.status == "ok",
+                error_message=result.error_msg if result.status != "ok" else None,
+                metadata={
+                    "vendor": "my_vendor",
+                    "tool_version": result.version,
+                    "execution_id": result.exec_id
+                },
+                execution_time_ms=ctx.execution_time_ms,
+                raw_result=result  # For debugging/advanced use cases
+            )
+        
+        # Fallback for unknown result types
+        return NormalizedToolResult(content=result)
+```
+
+### 2. Register Your Middleware
+
+```python
+from praisonaiagents.runtime import register_middleware
+
+# During plugin initialization
+middleware = MyHarnessMiddleware()
+register_middleware("my_plugin_harness", middleware)
+```
+
+### 3. Set Runtime ID on Agent
+
+```python
+from praisonaiagents import Agent
+
+# Your plugin harness should set the runtime_id on agents it creates
+agent = Agent(name="MyAgent", tools=[...])
+agent._runtime_id = "my_plugin_harness"  # This triggers middleware lookup
+```
+
+## Standard Result Format
+
+The `NormalizedToolResult` provides a consistent interface for downstream consumers:
+
+```python
+@dataclass
+class NormalizedToolResult:
+    # Core result data
+    content: Any  # The actual tool result (can be any type)
+    
+    # Standardized metadata
+    success: bool = True
+    error_message: Optional[str] = None
+    
+    # Rich metadata for downstream consumers
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    
+    # Execution context
+    execution_time_ms: float = 0.0
+    timestamp: float = field(default_factory=time.time)
+    
+    # Original raw result for debugging/advanced use cases
+    raw_result: Optional[Any] = None
+```
+
+## Common Patterns
+
+### Error Handling
+
+```python
+def normalize(self, result, tool_name, ctx):
+    if isinstance(result, MyErrorResult):
+        return NormalizedToolResult(
+            content=None,
+            success=False,
+            error_message=result.error_description,
+            metadata={
+                "error_code": result.code,
+                "error_category": result.category,
+                "retryable": result.is_retryable
+            },
+            raw_result=result
+        )
+```
+
+### Rich Metadata
+
+```python
+def normalize(self, result, tool_name, ctx):
+    return NormalizedToolResult(
+        content=result.data,
+        success=True,
+        metadata={
+            # Vendor identification
+            "vendor": "my_vendor",
+            "vendor_version": "2.1.0",
+            
+            # Execution metadata
+            "tool_version": result.tool_version,
+            "execution_id": result.exec_id,
+            "cache_hit": result.from_cache,
+            
+            # Performance metadata
+            "network_time_ms": result.network_latency,
+            "processing_time_ms": result.processing_time,
+            
+            # Context metadata
+            "region": result.execution_region,
+            "model": result.model_used if hasattr(result, 'model_used') else None
+        },
+        raw_result=result
+    )
+```
+
+### Conditional Normalization
+
+```python
+def normalize(self, result, tool_name, ctx):
+    # Different handling based on tool type
+    if tool_name.startswith("search_"):
+        return self._normalize_search_result(result, ctx)
+    elif tool_name.startswith("db_"):
+        return self._normalize_db_result(result, ctx)
+    else:
+        return self._normalize_generic_result(result, ctx)
+```
+
+## Performance Considerations
+
+- **Native Runtime**: The `praisonai` runtime bypasses middleware entirely to avoid allocation overhead
+- **Lazy Loading**: Middleware modules use lazy loading to minimize startup time
+- **Error Isolation**: Middleware failures don't break tool execution - they log warnings and continue
+- **Thread Safety**: The registry is thread-safe for concurrent agent execution
+
+## Testing
+
+Use the provided test utilities to verify your middleware:
+
+```python
+from praisonaiagents.runtime.middleware import MiddlewareContext
+
+def test_my_middleware():
+    middleware = MyHarnessMiddleware()
+    
+    ctx = MiddlewareContext(
+        tool_name="test_tool",
+        runtime_id="my_plugin_harness",
+        agent_id="test_agent",
+        execution_time_ms=100.0
+    )
+    
+    result = MyVendorResult(status="ok", data="test")
+    normalized = middleware.normalize(result, "test_tool", ctx)
+    
+    assert normalized.success is True
+    assert normalized.content == "test"
+    assert normalized.metadata["vendor"] == "my_vendor"
+```
+
+## Registry Management
+
+The registry supports runtime management for testing and plugin lifecycle:
+
+```python
+from praisonaiagents.runtime import get_default_registry
+
+registry = get_default_registry()
+
+# List registered runtimes
+runtimes = registry.list_runtimes()
+
+# Check if middleware is registered
+if registry.has_middleware("my_runtime"):
+    middleware = registry.get_middleware("my_runtime")
+
+# Unregister for cleanup
+registry.unregister("my_runtime")
+```

src/praisonai-agents/praisonaiagents/runtime/__init__.py

@@ -0,0 +1,80 @@
+""" 
+Runtime execution components for PraisonAI agents.
+
+This module provides runtime execution abstractions and protocols for 
+standardizing agent execution across different harness types.
+
+Key Components:
+- Runtime execution contexts: PreparedTurnContext, TurnRuntimeProtocol  
+- Tool result middleware: RuntimeToolResultMiddleware, NormalizedToolResult
+- Registry patterns for runtime resolution and middleware management
+"""
+
+from .._lazy import create_lazy_getattr_with_groups
+
+__all__ = [
+    # Runtime execution contexts
+    'PreparedTurnContext',
+    'TurnRuntimeProtocol',
+    'TurnContextBuilderProtocol',
+    'ModelReference',
+    'ToolSchema', 
+    'TranscriptWindow',
+    'DeliveryChannels',
+    'SessionCorrelation',
+    'RuntimeMode',
+    'create_default_model_ref',
+    'create_empty_transcript',
+    'create_default_delivery',
+    'create_session_correlation',
+    'DefaultTurnContextBuilder',
+    'default_context_builder',
+    # Tool result middleware
+    "RuntimeToolResultMiddleware",
+    "NormalizedToolResult", 
+    "MiddlewareContext",
+    # Middleware registry management  
+    "MiddlewareRegistry",
+    "get_default_middleware_registry",
+    "register_middleware",
+    "get_middleware",
+]
+
+# Grouped lazy imports for efficient loading
+_LAZY_GROUPS = {
+    'turn_context': {
+        'PreparedTurnContext': ('praisonaiagents.runtime.turn_context', 'PreparedTurnContext'),
+        'ModelReference': ('praisonaiagents.runtime.turn_context', 'ModelReference'),
+        'ToolSchema': ('praisonaiagents.runtime.turn_context', 'ToolSchema'),
+        'TranscriptWindow': ('praisonaiagents.runtime.turn_context', 'TranscriptWindow'),
+        'DeliveryChannels': ('praisonaiagents.runtime.turn_context', 'DeliveryChannels'),
+        'SessionCorrelation': ('praisonaiagents.runtime.turn_context', 'SessionCorrelation'),
+        'RuntimeMode': ('praisonaiagents.runtime.turn_context', 'RuntimeMode'),
+        'create_default_model_ref': ('praisonaiagents.runtime.turn_context', 'create_default_model_ref'),
+        'create_empty_transcript': ('praisonaiagents.runtime.turn_context', 'create_empty_transcript'),
+        'create_default_delivery': ('praisonaiagents.runtime.turn_context', 'create_default_delivery'),
+        'create_session_correlation': ('praisonaiagents.runtime.turn_context', 'create_session_correlation'),
+    },
+    'protocols': {
+        'TurnRuntimeProtocol': ('praisonaiagents.runtime.protocols', 'TurnRuntimeProtocol'),
+        'TurnContextBuilderProtocol': ('praisonaiagents.runtime.protocols', 'TurnContextBuilderProtocol'),
+    },
+    'context_builder': {
+        'DefaultTurnContextBuilder': ('praisonaiagents.runtime.context_builder', 'DefaultTurnContextBuilder'),
+        'default_context_builder': ('praisonaiagents.runtime.context_builder', 'default_context_builder'),
+    },
+    'middleware': {
+        'RuntimeToolResultMiddleware': ('praisonaiagents.runtime.middleware', 'RuntimeToolResultMiddleware'),
+        'NormalizedToolResult': ('praisonaiagents.runtime.middleware', 'NormalizedToolResult'),
+        'MiddlewareContext': ('praisonaiagents.runtime.middleware', 'MiddlewareContext'),
+    },
+    'middleware_registry': {
+        'MiddlewareRegistry': ('praisonaiagents.runtime.middleware_registry', 'MiddlewareRegistry'),
+        'get_default_middleware_registry': ('praisonaiagents.runtime.middleware_registry', 'get_default_middleware_registry'),
+        'register_middleware': ('praisonaiagents.runtime.middleware_registry', 'register_middleware'),
+        'get_middleware': ('praisonaiagents.runtime.middleware_registry', 'get_middleware'),
+    },
+}
+
+# Create the __getattr__ function using centralized utility
+__getattr__ = create_lazy_getattr_with_groups(_LAZY_GROUPS, __name__)

src/praisonai-agents/praisonaiagents/runtime/config.py

@@ -0,0 +1,169 @@
+"""AgentRuntimeConfig dataclass for model-scoped runtime selection.
+
+Protocol-driven design following AGENTS.md:
+- Lightweight dataclass for configuration
+- No heavy implementations (those live in wrapper layer)
+- Support for per-model and per-provider runtime configuration
+"""
+
+from dataclasses import dataclass, field
+from typing import Optional, Dict, Any, Union
+
+
+@dataclass
+class AgentRuntimeConfig:
+    """Model-scoped runtime configuration for PraisonAI Agents.
+    
+    Supports the resolution order:
+    per-model runtime → per-provider default → auto → built-in default
+    
+    This replaces agent-level cli_backend with model-scoped runtime selection.
+    """
+    # Runtime identifier (e.g., "claude-code", "codex-cli", "praisonai")
+    runtime: Optional[str] = None
+    
+    # Runtime-specific configuration overrides
+    config_overrides: Dict[str, Any] = field(default_factory=dict)
+    
+    # Provider-level default runtime (for YAML providers.<name>.runtime_default)
+    provider_default: Optional[str] = None
+    
+    # Enable auto-selection if no explicit runtime is specified
+    enable_auto_selection: bool = True
+    
+    # Metadata for runtime resolution
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    
+    def __post_init__(self):
+        """Validate runtime configuration on initialization."""
+        # Ensure config_overrides is a dictionary
+        if not isinstance(self.config_overrides, dict):
+            raise TypeError("config_overrides must be a dictionary")
+        
+        # Ensure metadata is a dictionary  
+        if not isinstance(self.metadata, dict):
+            raise TypeError("metadata must be a dictionary")
+    
+    @classmethod
+    def from_runtime_id(cls, runtime_id: str, **kwargs) -> "AgentRuntimeConfig":
+        """Create AgentRuntimeConfig from a runtime ID string.
+        
+        Args:
+            runtime_id: Runtime identifier (e.g., "claude-code")
+            **kwargs: Additional configuration options
+            
+        Returns:
+            AgentRuntimeConfig instance
+        """
+        return cls(runtime=runtime_id, **kwargs)
+    
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any]) -> "AgentRuntimeConfig":
+        """Create AgentRuntimeConfig from a dictionary.
+        
+        Args:
+            config_dict: Configuration dictionary
+            
+        Returns:
+            AgentRuntimeConfig instance
+            
+        Raises:
+            TypeError: If config_dict is not a dictionary
+            ValueError: If required fields are missing or invalid
+        """
+        if not isinstance(config_dict, dict):
+            raise TypeError("config_dict must be a dictionary")
+        
+        # Extract known fields
+        runtime = config_dict.get("runtime")
+        config_overrides = config_dict.get("config_overrides", {})
+        provider_default = config_dict.get("provider_default")
+        enable_auto_selection = config_dict.get("enable_auto_selection", True)
+        metadata = config_dict.get("metadata", {})
+        
+        # Validate types
+        if config_overrides is not None and not isinstance(config_overrides, dict):
+            raise TypeError("config_overrides must be a dictionary")
+        
+        if metadata is not None and not isinstance(metadata, dict):
+            raise TypeError("metadata must be a dictionary")
+        
+        return cls(
+            runtime=runtime,
+            config_overrides=config_overrides,
+            provider_default=provider_default,
+            enable_auto_selection=enable_auto_selection,
+            metadata=metadata
+        )
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert AgentRuntimeConfig to a dictionary.
+        
+        Returns:
+            Dictionary representation of the configuration
+        """
+        return {
+            "runtime": self.runtime,
+            "config_overrides": self.config_overrides.copy(),
+            "provider_default": self.provider_default,
+            "enable_auto_selection": self.enable_auto_selection,
+            "metadata": self.metadata.copy()
+        }
+    
+    def merge_overrides(self, overrides: Dict[str, Any]) -> "AgentRuntimeConfig":
+        """Create a new AgentRuntimeConfig with merged configuration overrides.
+        
+        Args:
+            overrides: Configuration overrides to merge
+            
+        Returns:
+            New AgentRuntimeConfig instance with merged overrides
+        """
+        if not isinstance(overrides, dict):
+            raise TypeError("overrides must be a dictionary")
+        
+        # Create new config overrides by merging
+        new_config_overrides = {**self.config_overrides, **overrides}
+        
+        return AgentRuntimeConfig(
+            runtime=self.runtime,
+            config_overrides=new_config_overrides,
+            provider_default=self.provider_default,
+            enable_auto_selection=self.enable_auto_selection,
+            metadata=self.metadata.copy()
+        )
+    
+    def with_runtime(self, runtime_id: str) -> "AgentRuntimeConfig":
+        """Create a new AgentRuntimeConfig with a different runtime ID.
+        
+        Args:
+            runtime_id: New runtime identifier
+            
+        Returns:
+            New AgentRuntimeConfig instance with updated runtime
+        """
+        return AgentRuntimeConfig(
+            runtime=runtime_id,
+            config_overrides=self.config_overrides.copy(),
+            provider_default=self.provider_default,
+            enable_auto_selection=self.enable_auto_selection,
+            metadata=self.metadata.copy()
+        )
+    
+    def is_explicit(self) -> bool:
+        """Check if this configuration specifies an explicit runtime.
+        
+        Returns:
+            True if runtime is explicitly specified, False otherwise
+        """
+        return self.runtime is not None and self.runtime != ""
+    
+    def __repr__(self) -> str:
+        """String representation of AgentRuntimeConfig."""
+        return (
+            f"AgentRuntimeConfig("
+            f"runtime={self.runtime!r}, "
+            f"config_overrides={self.config_overrides!r}, "
+            f"provider_default={self.provider_default!r}, "
+            f"enable_auto_selection={self.enable_auto_selection!r})"
+        )

src/praisonai-agents/praisonaiagents/runtime/context_builder.py

@@ -0,0 +1,452 @@
+"""
+Context Builder for PreparedTurnContext.
+
+This module provides the DefaultTurnContextBuilder that prepares
+PreparedTurnContext instances from agent configuration and request
+parameters. It consolidates the scattered context preparation logic
+from chat_mixin.py, chat_handler.py, and other execution paths.
+"""
+
+from __future__ import annotations
+
+import copy
+import logging
+import time
+import uuid
+from typing import Any, Dict, List, Optional, TYPE_CHECKING
+
+from .turn_context import (
+    PreparedTurnContext,
+    ModelReference, 
+    ToolSchema,
+    TranscriptWindow,
+    DeliveryChannels,
+    SessionCorrelation,
+    RuntimeMode,
+)
+from .protocols import TurnContextBuilderProtocol
+
+if TYPE_CHECKING:
+    from ..agent.protocols import AgentProtocol
+    from ..streaming.events import StreamEventEmitter
+
+logger = logging.getLogger(__name__)
+
+
+class DefaultTurnContextBuilder:
+    """
+    Default implementation of turn context building.
+    
+    This builder extracts and normalizes configuration from Agent instances
+    to create PreparedTurnContext objects. It consolidates the logic that
+    was previously scattered across multiple mixins.
+    """
+    
+    def __init__(self):
+        self._cache: Dict[str, Any] = {}
+    
+    def build_context(
+        self,
+        agent: AgentProtocol,
+        prompt: str,
+        **kwargs: Any
+    ) -> PreparedTurnContext:
+        """
+        Build a prepared turn context from agent and request.
+        
+        Args:
+            agent: The agent instance
+            prompt: The user prompt for this turn
+            **kwargs: Additional request parameters including:
+                - model: Override model ID
+                - tools: Override tools list
+                - stream: Enable streaming
+                - temperature: Model temperature
+                - max_tokens: Token limit
+                - session_id: Session identifier
+                - turn_id: Turn identifier
+                
+        Returns:
+            A prepared turn context ready for execution
+        """
+        try:
+            # Extract runtime mode from kwargs
+            runtime_mode = self._determine_runtime_mode(kwargs)
+            
+            # Build model reference
+            model_ref = self._build_model_reference(agent, kwargs)
+            
+            # Normalize tools
+            tools = self._build_tool_schemas(agent, kwargs)
+            
+            # Build transcript window
+            transcript = self._build_transcript_window(agent, prompt, kwargs)
+            
+            # Setup delivery channels
+            delivery = self._build_delivery_channels(agent, kwargs)
+            
+            # Create correlation IDs
+            correlation = self._build_session_correlation(agent, kwargs)
+            
+            # Create the context
+            context = PreparedTurnContext(
+                model_ref=model_ref,
+                agent_runtime=agent,
+                tools=tools,
+                transcript=transcript,
+                delivery=delivery,
+                correlation=correlation,
+                runtime_mode=runtime_mode,
+                turn_metadata=self._extract_turn_metadata(agent, kwargs),
+                created_at=time.time()
+            )
+            
+            logger.debug(
+                f"Built turn context: {context.get_message_count()} messages, "
+                f"{len(context.tools)} tools, mode={context.runtime_mode.value}"
+            )
+            
+            return context
+            
+        except Exception as e:
+            logger.error(f"Failed to build turn context: {e}")
+            raise
+    
+    def _determine_runtime_mode(self, kwargs: Dict[str, Any]) -> RuntimeMode:
+        """Determine runtime mode from request parameters."""
+        is_stream = kwargs.get('stream', False)
+        is_async = kwargs.get('async_execution', False)
+        
+        if is_async and is_stream:
+            return RuntimeMode.ASYNC_STREAM
+        elif is_async:
+            return RuntimeMode.ASYNC
+        elif is_stream:
+            return RuntimeMode.STREAM
+        else:
+            return RuntimeMode.SYNC
+    
+    def _build_model_reference(
+        self, 
+        agent: AgentProtocol, 
+        kwargs: Dict[str, Any]
+    ) -> ModelReference:
+        """Build model reference from agent configuration and overrides."""
+        # Get model ID (kwargs override, then agent default)
+        model_id = kwargs.get('model') or getattr(agent, 'model', 'gpt-3.5-turbo')
+        
+        # Determine provider from model ID or agent config
+        provider = self._determine_provider(model_id, agent)
+        
+        # Get model capabilities
+        supports_streaming = kwargs.get('stream', False) or getattr(agent, '_supports_streaming', True)
+        supports_tools = len(getattr(agent, 'tools', [])) > 0
+        supports_system_prompts = getattr(agent, 'use_system_prompt', True)
+        
+        # Build model config - agent defaults first, kwargs override
+        model_config = {}
+        if hasattr(agent, 'model_config'):
+            model_config.update(agent.model_config)
+        if 'temperature' in kwargs:
+            model_config['temperature'] = kwargs['temperature']
+        if 'max_tokens' in kwargs:
+            model_config['max_tokens'] = kwargs['max_tokens']
+        
+        return ModelReference(
+            model_id=model_id,
+            provider=provider,
+            supports_streaming=supports_streaming,
+            supports_tools=supports_tools,
+            supports_system_prompts=supports_system_prompts,
+            max_tokens=kwargs.get('max_tokens'),
+            temperature=kwargs.get('temperature'),
+            model_config=model_config
+        )
+    
+    def _determine_provider(self, model_id: str, agent: AgentProtocol) -> str:
+        """Determine provider from model ID or agent configuration."""
+        # Check if agent has explicit provider
+        if hasattr(agent, 'provider'):
+            return agent.provider
+        
+        # Infer from model ID
+        if model_id.startswith('gpt-'):
+            return 'openai'
+        elif model_id.startswith('claude-'):
+            return 'anthropic'
+        elif model_id.startswith('gemini-'):
+            return 'google'
+        elif 'llama' in model_id.lower():
+            return 'meta'
+        else:
+            return 'openai'  # Default fallback
+    
+    def _build_tool_schemas(
+        self,
+        agent: AgentProtocol,
+        kwargs: Dict[str, Any]
+    ) -> List[ToolSchema]:
+        """Build normalized tool schemas from agent tools."""
+        # Get tools from kwargs override or agent - respect explicit empty list
+        if 'tools' in kwargs:
+            raw_tools = kwargs['tools']
+        else:
+            raw_tools = getattr(agent, 'tools', [])
+        
+        if not raw_tools:
+            return []
+        
+        normalized_tools = []
+        for tool in raw_tools:
+            try:
+                schema = self._normalize_tool_schema(tool)
+                if schema:
+                    normalized_tools.append(schema)
+            except Exception as e:
+                logger.warning(f"Failed to normalize tool {tool}: {e}")
+                continue
+        
+        return normalized_tools
+    
+    def _normalize_tool_schema(self, tool: Any) -> Optional[ToolSchema]:
+        """Normalize a single tool into ToolSchema format."""
+        if tool is None:
+            return None
+        
+        # Handle different tool formats
+        if callable(tool):
+            # Function tool
+            return self._normalize_function_tool(tool)
+        elif isinstance(tool, dict):
+            # Dictionary tool (OpenAI format, custom format, etc.)
+            return self._normalize_dict_tool(tool)
+        elif hasattr(tool, 'name') and hasattr(tool, 'description'):
+            # Object with name/description attributes
+            return self._normalize_object_tool(tool)
+        else:
+            logger.warning(f"Unknown tool format: {type(tool)}")
+            return None
+    
+    def _normalize_function_tool(self, func: callable) -> ToolSchema:
+        """Normalize a callable function into ToolSchema."""
+        import inspect
+        
+        name = getattr(func, '__name__', 'unknown_function')
+        description = getattr(func, '__doc__', '') or f"Function: {name}"
+        
+        # Try to extract parameters from function signature
+        try:
+            sig = inspect.signature(func)
+            parameters = {
+                "type": "object",
+                "properties": {},
+                "required": []
+            }
+            
+            for param_name, param in sig.parameters.items():
+                if param_name == 'self':
+                    continue
+                    
+                prop = {"type": "string"}  # Default type
+                if param.annotation != inspect.Parameter.empty:
+                    prop["type"] = self._python_type_to_json_type(param.annotation)
+                    
+                parameters["properties"][param_name] = prop
+                
+                if param.default == inspect.Parameter.empty:
+                    parameters["required"].append(param_name)
+            
+        except Exception:
+            # Fallback to empty parameters
+            parameters = {"type": "object", "properties": {}}
+        
+        return ToolSchema(
+            name=name,
+            description=description.strip(),
+            parameters=parameters,
+            callable=func,
+            source_type="function"
+        )
+    
+    def _normalize_dict_tool(self, tool_dict: Dict[str, Any]) -> ToolSchema:
+        """Normalize a dictionary tool into ToolSchema."""
+        if 'function' in tool_dict:
+            # OpenAI format: {"type": "function", "function": {...}}
+            func_def = tool_dict['function']
+            return ToolSchema(
+                name=func_def.get('name', 'unknown'),
+                description=func_def.get('description', ''),
+                parameters=func_def.get('parameters', {}),
+                source_type="openai"
+            )
+        else:
+            # Direct format: {"name": ..., "description": ..., "parameters": ...}
+            return ToolSchema(
+                name=tool_dict.get('name', 'unknown'),
+                description=tool_dict.get('description', ''),
+                parameters=tool_dict.get('parameters', {}),
+                source_type="custom"
+            )
+    
+    def _normalize_object_tool(self, tool_obj: Any) -> ToolSchema:
+        """Normalize a tool object into ToolSchema."""
+        return ToolSchema(
+            name=getattr(tool_obj, 'name', 'unknown'),
+            description=getattr(tool_obj, 'description', ''),
+            parameters=getattr(tool_obj, 'parameters', {}),
+            callable=getattr(tool_obj, '__call__', None),
+            source_type="object"
+        )
+    
+    def _python_type_to_json_type(self, py_type: type) -> str:
+        """Convert Python type annotation to JSON schema type."""
+        if py_type == str:
+            return "string"
+        elif py_type == int:
+            return "integer"
+        elif py_type == float:
+            return "number"
+        elif py_type == bool:
+            return "boolean"
+        elif py_type == list:
+            return "array"
+        elif py_type == dict:
+            return "object"
+        else:
+            return "string"  # Default fallback
+    
+    def _build_transcript_window(
+        self,
+        agent: AgentProtocol,
+        prompt: str,
+        kwargs: Dict[str, Any]
+    ) -> TranscriptWindow:
+        """Build transcript window from agent conversation history."""
+        # Get system prompt
+        system_prompt = None
+        if hasattr(agent, '_build_system_prompt'):
+            try:
+                system_prompt = agent._build_system_prompt()
+            except Exception as e:
+                logger.warning(f"Failed to build system prompt: {e}")
+        
+        # Get conversation history - deep copy to prevent cross-agent mutations
+        messages = []
+        if hasattr(agent, 'chat_history') and agent.chat_history:
+            messages = copy.deepcopy(agent.chat_history)
+        
+        # Add current prompt as user message
+        if prompt:
+            messages.append({"role": "user", "content": prompt})
+        
+        # Estimate token count (rough approximation)
+        total_tokens = self._estimate_tokens(messages, system_prompt)
+        
+        return TranscriptWindow(
+            messages=messages,
+            total_tokens=total_tokens,
+            system_prompt=system_prompt,
+            context_metadata={
+                "original_prompt": prompt,
+                "history_length": len(messages) - (1 if prompt else 0)
+            }
+        )
+    
+    def _estimate_tokens(
+        self, 
+        messages: List[Dict[str, Any]], 
+        system_prompt: Optional[str]
+    ) -> int:
+        """Rough token estimation for transcript."""
+        total_chars = 0
+        
+        if system_prompt:
+            total_chars += len(system_prompt)
+        
+        for msg in messages:
+            content = msg.get('content', '')
+            if isinstance(content, str):
+                total_chars += len(content)
+        
+        # Rough approximation: 1 token ≈ 4 characters
+        return total_chars // 4
+    
+    def _build_delivery_channels(
+        self,
+        agent: AgentProtocol,
+        kwargs: Dict[str, Any]
+    ) -> DeliveryChannels:
+        """Build delivery channels configuration."""
+        enable_streaming = kwargs.get('stream', False)
+        enable_metrics = kwargs.get('enable_metrics', False)
+        
+        # Get stream emitter from agent if available
+        stream_emitter = None
+        if enable_streaming:
+            if hasattr(agent, '_stream_emitter'):
+                stream_emitter = agent._stream_emitter
+            elif hasattr(agent, 'stream_emitter'):
+                stream_emitter = agent.stream_emitter
+        
+        # Get callbacks
+        callbacks = kwargs.get('callbacks', [])
+        async_callbacks = kwargs.get('async_callbacks', [])
+        
+        return DeliveryChannels(
+            stream_emitter=stream_emitter,
+            callbacks=callbacks,
+            async_callbacks=async_callbacks,
+            enable_streaming=enable_streaming,
+            enable_metrics=enable_metrics
+        )
+    
+    def _build_session_correlation(
+        self,
+        agent: AgentProtocol,
+        kwargs: Dict[str, Any]
+    ) -> SessionCorrelation:
+        """Build session correlation identifiers."""
+        # Use provided IDs or generate new ones
+        session_id = kwargs.get('session_id')
+        if not session_id and hasattr(agent, 'session_id'):
+            session_id = agent.session_id
+        if not session_id:
+            session_id = f"session-{uuid.uuid4().hex[:8]}"
+        
+        turn_id = kwargs.get('turn_id') or f"turn-{uuid.uuid4().hex[:8]}"
+        agent_id = kwargs.get('agent_id') or getattr(agent, 'name', 'agent')
+        run_id = kwargs.get('run_id') or f"run-{uuid.uuid4().hex[:8]}"
+        parent_id = kwargs.get('parent_id')
+        
+        return SessionCorrelation(
+            session_id=session_id,
+            turn_id=turn_id,
+            agent_id=agent_id,
+            run_id=run_id,
+            parent_id=parent_id,
+            trace_metadata=kwargs.get('trace_metadata', {})
+        )
+    
+    def _extract_turn_metadata(
+        self,
+        agent: AgentProtocol,
+        kwargs: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """Extract additional turn metadata."""
+        metadata = {}
+        
+        # Add agent metadata
+        if hasattr(agent, 'role'):
+            metadata['agent_role'] = agent.role
+        if hasattr(agent, 'goal'):
+            metadata['agent_goal'] = agent.goal
+        
+        # Add request metadata
+        if 'metadata' in kwargs:
+            metadata.update(kwargs['metadata'])
+        
+        return metadata
+
+
+# Global instance for convenience
+default_context_builder = DefaultTurnContextBuilder()