fix: sync automation.meta.json hubRef to v1.1.15

Co-Authored-By: Claude Opus 5.5 <[email protected]>
chore(ci): bump github-automation-template to v1.1.15
2026-11-07 05:56:17 -10:00 · 2026-10-16 16:22:53 +00:00 · 2026-10-16 16:51:32 +01:00 · 2026-10-15 08:58:59 +01:00 · 2026-10-15 08:30:36 +01:00 · 2026-09-30 08:38:05 +01:00
105 changed files with 7989 additions and 388 deletions
@@ -3,7 +3,7 @@
  "stack": "claude",
  "profile": "sdk",
  "mode": "hub",
-  "hubRef": "v1.1.6",
-  "templateVersion": "1.1.6",
+  "hubRef": "v1.1.15",
+  "templateVersion": "1.1.7",
  "installedAt": "2026-09-07T10:02:11Z"
 }
@@ -4,6 +4,10 @@

 module.exports = {
  repoFullName: 'MervinPraison/PraisonAIUI',
+  gitUser: 'MervinPraison',
+  gitEmail: '[email protected]',
+  triggerLogins: ['MervinPraison', 'github-actions[bot]'],
+  allowedTriageBots: ['praisonai-triage-agent[bot]'],
  productPathPrefixes: ['src/praisonaiui/', 'src/frontend/', 'tests/'],
  sensitivePathPatterns: [
    /^\.github\/workflows\//,
@@ -10,7 +10,7 @@ on:
  pull_request_target:
    types: [opened, synchronize]
  schedule:
-    - cron: '30 3,9,15,21 * * *'
+    - cron: '30 * * * *'
  workflow_dispatch:

 env:
@@ -27,7 +27,7 @@ jobs:
    permissions:
      issues: write
      pull-requests: write
-    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].6
+    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].15
    secrets: inherit
    with:
-      hub_ref: v1.1.6
+      hub_ref: v1.1.15
@@ -16,8 +16,8 @@ permissions:

 jobs:
  recover:
-    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].6
+    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].15
    secrets: inherit
    with:
-      hub_ref: v1.1.6
+      hub_ref: v1.1.15
      pr_number: ${{ github.event.inputs.pr_number || '' }}
@@ -25,8 +25,8 @@ permissions:

 jobs:
  detect-and-trigger:
-    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].6
+    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].15
    secrets: inherit
    with:
-      hub_ref: v1.1.6
+      hub_ref: v1.1.15
      pr_number: ${{ github.event.inputs.pr_number || '' }}
@@ -36,6 +36,30 @@ jobs:
      - name: Integration smoke tests
        run: pytest tests/integration/test_agentic_roundtrip.py tests/test_feature_sdk_backends.py -q --no-cov --timeout=120

+  frontend:
+    name: frontend
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+          cache: npm
+          cache-dependency-path: src/frontend/package-lock.json
+
+      - name: Install dependencies
+        run: npm ci
+        working-directory: src/frontend
+
+      # `npm run build` is the dashboard ship path: dist/ is synced to
+      # src/praisonaiui/templates/frontend/ (see src/frontend/README.md).
+      - name: Build frontend
+        run: npm run build
+        working-directory: src/frontend
+
  docs:
    runs-on: ubuntu-latest
    if: github.ref == 'refs/heads/main'
@@ -43,9 +43,9 @@ jobs:
      id-token: write
    steps:
      - uses: actions/checkout@v4
-      - uses: MervinPraison/github-automation-template/.github/actions/[email protected].6
+      - uses: MervinPraison/github-automation-template/.github/actions/[email protected].15
        with:
-          hub_ref: v1.1.6
+          hub_ref: v1.1.15
      - name: Generate GitHub App token
        id: app-token
        uses: actions/create-github-app-token@v1
@@ -87,9 +87,9 @@ jobs:
      issues: write
      actions: read
      id-token: write
-    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].6
+    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].15
    secrets: inherit
    with:
-      hub_ref: v1.1.6
+      hub_ref: v1.1.15
      pr_number: ${{ github.event.inputs.pr_number || github.event.client_payload.pr_number || '' }}
      merge_method: ${{ github.event.inputs.merge_method || 'merge' }}
@@ -18,7 +18,7 @@ jobs:
      issues: write
      actions: read
      id-token: write
-    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].6
+    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].15
    secrets: inherit
    with:
-      hub_ref: v1.1.6
+      hub_ref: v1.1.15
@@ -25,9 +25,9 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
-      - uses: MervinPraison/github-automation-template/.github/actions/[email protected].6
+      - uses: MervinPraison/github-automation-template/.github/actions/[email protected].15
        with:
-          hub_ref: v1.1.6
+          hub_ref: v1.1.15
      - name: Detect conflicting PRs and trigger Claude
        uses: actions/github-script@v7
        with:
@@ -13,7 +13,7 @@ permissions:

 jobs:
  sync:
-    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].6
+    uses: MervinPraison/github-automation-template/.github/workflows/[email protected].15
    secrets: inherit
    with:
-      hub_ref: v1.1.6
+      hub_ref: v1.1.15
@@ -366,3 +366,7 @@ pytest tests -v
 ## License

 MIT © [Praison Limited](https://praison.ai)
+
+Third-party vendored source is documented in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
+See the [BeautifulUI upgrade playbook](docs/frontend/beautifului.md) for how BeautifulUI
+primitives are vendored and pinned.
@@ -0,0 +1,37 @@
+# Security Policy
+
+We take the security of PraisonAIUI seriously and appreciate responsible
+disclosure of vulnerabilities.
+
+## Supported Versions
+
+Security fixes are applied to the latest released version of PraisonAIUI.
+Please make sure you are on the most recent release before reporting.
+
+## Reporting a Vulnerability
+
+**Please do not open a public issue for security vulnerabilities.**
+
+Report privately through GitHub's **Private Vulnerability Reporting**:
+
+1. Go to the [Security tab](https://github.com/MervinPraison/PraisonAIUI/security)
+   of this repository.
+2. Click **"Report a vulnerability"** to open a private advisory.
+3. Include as much detail as you can:
+   - Affected component(s) and version.
+   - Steps to reproduce.
+   - Impact assessment.
+   - Any suggested fix or mitigation.
+
+If you are unable to use private reporting, you may open a public issue that
+contains **no technical specifics** and requests a private channel, and a
+maintainer will follow up.
+
+## What to Expect
+
+- We aim to acknowledge new reports within a few business days.
+- We will keep you informed as we investigate and work on a fix.
+- Once a fix is released, we are happy to credit you in the advisory (unless
+  you prefer to remain anonymous).
+
+Thank you for helping keep PraisonAIUI and its users safe.
@@ -0,0 +1,43 @@
+# Third-Party Notices
+
+PraisonAIUI incorporates or vendors source from the third-party projects listed below.
+Their licences and copyright notices are retained here as required.
+
+---
+
+## BeautifulUI
+
+- Project: BeautifulUI — <https://www.beautifului.dev/>
+- Source: <https://github.com/slev12397/beautiful-ui>
+- Licence: MIT
+
+BeautifulUI is distributed as a shadcn registry / copy-paste library (no npm package).
+When its React primitives are vendored into `src/frontend`, the following MIT licence
+applies to that copied source.
+
+```
+MIT License
+
+Copyright (c) Shane Levine
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+```
+
+See the upgrade playbook at [`docs/frontend/beautifului.md`](docs/frontend/beautifului.md)
+for how BeautifulUI is vendored and pinned.
@@ -59,6 +59,11 @@ navigation:
            - typescript/runtime
            - typescript/nextjs
            - typescript/components
+        - group: "Frontend"
+          icon: "layout"
+          prefix: "frontend"
+          pages:
+            - frontend/beautifului
        - group: "Examples"
          icon: "folder-open"
          prefix: "examples"
@@ -0,0 +1,56 @@
+# BeautifulUI Integration Strategy
+
+> **Epic:** [#292](https://github.com/MervinPraison/PraisonAIUI/issues/292) — BeautifulUI integration strategy for AI-native dashboard UX.
+> This is the durable, agreed strategy that the linked child issues execute against.
+
+[BeautifulUI](https://www.beautifului.dev/) (MIT, Shane Levine / Turbo) is a **copy-paste / shadcn-registry** library of React primitives for AI-native interfaces: thinking traces, streaming answers with inline sources, HITL approval cards, tool chips, task rows, context chunks, diff/records tables, prompt bars, and a reference chat harness.
+
+There is **no published npm package** named `beautiful-ui`. It ships only as source consumed via the shadcn registry or manual copy.
+
+## 1. Strategic decision
+
+BeautifulUI is adopted as **vendored source**, not a runtime dependency, and is exposed to dashboard plugins **only through the existing `@praisonaiui/react` package** — never by importing upstream file paths directly. Rollout is **phased behind a feature flag**, keeping the current vanilla dashboard as the fallback until parity is proven.
+
+Upstream stack (Next.js 15 · React 19 · Tailwind v4 · TypeScript) is closely aligned with PraisonAIUI's `src/frontend` (Vite · React 19 · Tailwind v4 · shadcn/Radix), so primitives port with adapter work rather than a runtime bridge.
+
+## 2. Pillars
+
+| # | Pillar | Decision |
+|---|--------|----------|
+| 1 | **Vendoring model** | Add primitives into `src/frontend` via `npx shadcn add https://www.beautifului.dev/r/<component>.json`, or manual copy from [github.com/slev12397/beautiful-ui](https://github.com/slev12397/beautiful-ui). Pin via lockfile plus a recorded registry URL / git SHA. No `beautiful-ui` npm dependency in production builds. |
+| 2 | **Adapter layer** | Map PraisonAIUI chat / approval / trace **protocol events** to stable internal props. Primitives sit behind `@praisonaiui/react` exports; dashboard plugins consume those exports, not upstream paths. See [#294](https://github.com/MervinPraison/PraisonAIUI/issues/294). |
+| 3 | **Hybrid rollout** | Mount React **islands** (or expand `@praisonaiui/react`) for high-value surfaces first — chat streaming/thinking/tools, then approvals. Vanilla chat (`templates/frontend/plugins/views/chat.js`) stays as the flag-gated fallback until parity is proven. |
+| 4 | **Token strategy** | Merge BeautifulUI foundation CSS **selectively** into the existing shadcn theme generation (`praisonaiui/themes.py`). One design system per page — avoid two competing token sets. |
+| 5 | **Icon licence** | Upstream `SidebarNav` uses the paid `@central-icons-react`. Port nav patterns with **lucide-react** (already in PraisonAIUI) instead. |
+
+## 3. Phasing
+
+1. **Spike** — evaluate primitives in the dev environment, confirm React 19 / Tailwind v4 alignment. [#293](https://github.com/MervinPraison/PraisonAIUI/issues/293)
+2. **Architecture** — adapter layer + upstream pin policy. [#294](https://github.com/MervinPraison/PraisonAIUI/issues/294)
+3. **CI gate** — frontend build + BeautifulUI compatibility check (no frontend build gate exists today). [#295](https://github.com/MervinPraison/PraisonAIUI/issues/295)
+4. **Migrations** — chat streaming / thinking / tool chips ([#296](https://github.com/MervinPraison/PraisonAIUI/issues/296)), then the approval card ([#297](https://github.com/MervinPraison/PraisonAIUI/issues/297)).
+5. **Docs** — upgrade playbook + MIT attribution. [#298](https://github.com/MervinPraison/PraisonAIUI/issues/298)
+
+## 4. Risks
+
+| Risk | Mitigation |
+|------|------------|
+| No semver npm package | Vendor the source; pin to a recorded registry URL / git SHA in the lockfile. |
+| Upstream breaking registry JSON | Vendoring decouples builds from live registry; re-sync is a deliberate, reviewed step. |
+| Icon licence (`@central-icons-react` is paid) | Use lucide-react for ported nav patterns; never vendor the paid package. |
+| Two competing design systems on one page | Selective CSS token merge into `themes.py`; adapter-only prop surface. |
+| React / Tailwind drift | Upstream (React 19 / Tailwind v4) already matches `src/frontend`; the CI gate catches drift. |
+
+## 5. Out of scope
+
+- Replacing the entire dashboard shell in one PR.
+- PyPI packaging of BeautifulUI.
+- iframe embedding of beautifului.dev.
+- Depending on any unpublished npm package for production builds.
+
+## 6. Acceptance criteria status
+
+- [x] Written integration strategy agreed (vendoring + adapter + phased rollout) — this document.
+- [x] Risks documented — section 4.
+- [x] Child issues filed and prioritised — [#293](https://github.com/MervinPraison/PraisonAIUI/issues/293)–[#298](https://github.com/MervinPraison/PraisonAIUI/issues/298).
+- [x] No requirement to depend on unpublished npm packages for production builds — pillar 1, out of scope.
@@ -0,0 +1,35 @@
+# BeautifulUI upstream pin record
+
+BeautifulUI is vendored (shadcn-style registry copy), **not** an npm dependency.
+This file is the single source of truth for what upstream revision is vendored.
+Update it in the **same PR** that adds or bumps any vendored component, and make
+sure the frontend CI gate (issue #295) passes.
+
+See [beautifului-adapter-architecture.md](./beautifului-adapter-architecture.md)
+for the adapter boundary and folder layout.
+
+## Pin
+
+| Field | Value |
+|-------|-------|
+| Upstream repo | _TBD — record on first vendor_ |
+| Git SHA | _TBD_ |
+| Vendored on (date) | _TBD_ |
+| Registry base URL | _TBD_ |
+| License | MIT (kept in-tree) |
+
+## Vendored components
+
+Record one row per component the first time it is copied in.
+
+| Component | Registry URL | SHA | Vendored path | Date |
+|-----------|--------------|-----|---------------|------|
+| _(none yet)_ | | | | |
+
+## Upgrade checklist
+
+1. Update the SHA / date / registry rows above.
+2. Re-copy changed components into `src/frontend/src/agent-ui/vendors/beautifului/`.
+3. Keep all upstream imports inside `src/agent-ui/` (adapter boundary rule).
+4. Run the frontend build + adapter unit tests (frontend CI gate, #295).
+5. Do **not** add a `beautiful-ui` entry to `package.json` dependencies.
@@ -0,0 +1,102 @@
+# ADR: BeautifulUI adapter layer & upstream pin policy
+
+**Status:** Accepted · **Issue:** [#294](https://github.com/MervinPraison/PraisonAIUI/issues/294)
+
+## Context
+
+BeautifulUI is **not** a semver npm library — it is distributed as shadcn-style
+registry components that are copied into the consuming repo. PraisonAIUI must
+integrate it without leaking upstream types or import paths across the codebase,
+and without breaking when either side updates.
+
+## Decision
+
+Introduce a single **adapter boundary** between transport payloads and the UI,
+and vendor BeautifulUI components behind that boundary.
+
+```
+Agent APIs (Python) ──► event bus / WS messages
+                              │
+                              ▼
+                    Agent-UI adapter (src/agent-ui/adapter.ts)
+                    (stable PraisonAIUI contracts)
+                              │
+                              ▼
+              @praisonaiui/react re-exports
+              (BeautifulUI vendored components)
+                              │
+                              ▼
+         Dashboard / chat mount points (React islands, then full shell)
+```
+
+### 1. Folder layout
+
+```
+src/frontend/src/agent-ui/
+├── contracts.ts        # Stable PraisonAIUI view models (the ONLY types consumers import)
+├── adapter.ts          # Pure mappers: wire payload -> contracts
+├── adapter.test.ts     # Unit tests (mock events -> props), node:test
+├── index.ts            # Barrel: contracts + adapters + (later) vendored re-exports
+└── vendors/
+    └── beautifului/    # Vendored BeautifulUI components (added during migration)
+```
+
+**Rule:** BeautifulUI / upstream prop types must **never** be imported outside
+`src/agent-ui/`. Consumers depend on `contracts.ts` only, so the vendored
+implementation can be swapped later without touching call sites.
+
+### 2. Adapter contracts
+
+Internal, transport-agnostic types live in `contracts.ts`:
+
+- `AgentStreamEvent` — normalized form of the WebSocket frames handled in
+  `chat.js` (`run_started`, `run_content`, `tool_call_started`, …). Unknown
+  frame `type`s degrade to `kind: 'unknown'` with `rawType` preserved instead
+  of throwing, so new upstream events never break the UI.
+- `ToolCallChipModel` — tool-call chip view model.
+- `ApprovalPromptModel` — approval card view model, mapped from the
+  `/api/approvals/pending` REST payload.
+
+Mappers (`adapter.ts`): `mapChatEventToUi(frame)` and
+`mapApprovalToCardProps(item)`. Both are **pure** and unit tested.
+
+### 3. Vendoring & pin policy
+
+| Mechanism | Purpose |
+|-----------|---------|
+| [`BEAUTIFULUI_UPSTREAM.md`](./BEAUTIFULUI_UPSTREAM.md) | Record git SHA, date, registry base URL for every vendored component |
+| shadcn registry URLs | Per-component install, pinned in the upstream doc |
+| **No** `dependencies.beautiful-ui` in `package.json` | Avoid a phantom npm package |
+
+**Upgrade rule:** bump upstream **only** via a PR that updates
+`BEAUTIFULUI_UPSTREAM.md` and passes the frontend CI gate (issue #295). Icons
+stay on `lucide-react`; do not vendor paid Central icons.
+
+### 4. Feature flag (default **off**)
+
+Integration is gated so the existing vanilla-JS chat/approvals plugins remain
+the fallback until migration milestones (#296, #297) complete.
+
+- **Flag:** `AIUI_BEAUTIFULUI` (build-time env / dashboard config JSON).
+- **Default:** off — vanilla JS renders; the adapter is imported but not mounted.
+- **On:** React islands mount using `src/agent-ui` contracts.
+- **Fallback:** if a vendored component fails to load, the flag is treated as
+  off and the vanilla path renders. No user-facing regression when off.
+
+### 5. Design tokens
+
+Scope agent-UI primitives under an `.aiui-agent-ui { … }` CSS layer (or extend
+the existing shadcn presets in `themes.py`) so vendored component tokens cannot
+collide with the global dashboard theme.
+
+### 6. Deprecation / upstream abandonment
+
+Vendored code stays MIT-licensed in-tree. Because consumers depend only on the
+`contracts.ts` view models, a future first-party implementation can satisfy the
+same contracts and replace the vendored components with no call-site churn.
+
+## Consequences
+
+- One place (`src/agent-ui/`) owns the upstream coupling.
+- Adapters are testable without a browser or the transport layer.
+- The default-off flag keeps the migration reversible at every step.
@@ -0,0 +1,107 @@
+# BeautifulUI Upgrade Playbook
+
+How PraisonAIUI vendors [BeautifulUI](https://www.beautifului.dev/) React primitives so
+upgrades are **repeatable** and licence obligations are **explicit**.
+
+!!! info "Related planning"
+    Epic [#292](https://github.com/MervinPraison/PraisonAIUI/issues/292) (integration
+    strategy) and the adapter/pin policy in
+    [#294](https://github.com/MervinPraison/PraisonAIUI/issues/294).
+
+## 1. What BeautifulUI is
+
+BeautifulUI (MIT, © Shane Levine / Turbo) is a library of React primitives aimed at
+**AI-native interfaces** — thinking traces, streaming answers with inline sources,
+human-in-the-loop approval cards, tool chips, task rows, context chunks, and a reference
+chat harness.
+
+**It is a shadcn registry + copy-paste library — not an npm package.** There is no
+published `beautiful-ui` package. You install components by resolving a registry URL, and
+the source is copied into your tree (like any other shadcn/ui primitive).
+
+- Website & registry: <https://www.beautifului.dev/>
+- Upstream source: <https://github.com/slev12397/beautiful-ui>
+
+Upstream stack (Next.js 15 · React 19 · Tailwind CSS v4 · TypeScript) is closely aligned
+with PraisonAIUI's `src/frontend` (Vite · React 19 · Tailwind v4 · shadcn/Radix), which is
+why copy-paste vendoring works without a framework shim.
+
+## 2. One-time foundation setup
+
+PraisonAIUI already carries the shadcn/ui foundation that BeautifulUI primitives depend on:
+
+- **shadcn config** — `src/frontend/components.json` (`style: new-york`,
+  `baseColor: neutral`, `iconLibrary: lucide`, `cssVariables: true`). Its `registries`
+  map is where pinned BeautifulUI registry URLs are added (see §3).
+- **Theme tokens** — `src/praisonaiui/themes.py` generates the shadcn CSS variables
+  (`--background`, `--foreground`, `--primary`, `--radius`, …) at build time for all 22
+  official presets, offline. BeautifulUI's foundation CSS (upstream `app/globals.css`)
+  must only contribute **new** tokens/utilities; it must **not** redefine variables that
+  `themes.py` already owns, or theme switching and dark mode break.
+
+**Rule:** when merging upstream `app/globals.css`, diff it against the variables emitted by
+`themes.py` and keep only the additive slices (new keyframes, component-scoped utilities).
+Record what was merged in the pin doc (§3).
+
+## 3. Adding or updating a primitive
+
+1. **Install with a pinned registry URL** from `src/frontend/`:
+   ```bash
+   npx shadcn add https://www.beautifului.dev/r/<component>.json
+   ```
+   The registry resolves dependencies and CSS slices. Commit the resulting
+   `package-lock.json` change in the same PR.
+2. **Re-export through the adapter** — add the component to
+   `src/frontend/src/components/index.ts` (`@praisonaiui/react`). Consumers never import
+   the raw vendored path (see §4).
+3. **Update the pin record** — create/update `docs/frontend/BEAUTIFULUI_UPSTREAM.md` with:
+   the upstream **git SHA**, the **date**, the **registry base URL**, and the **component
+   list** vendored so far. This file is the single source of truth for "what version are
+   we on". (It is created the first time a real primitive is vendored — do not add it
+   empty.)
+4. **Rebuild & sync the bundle** — the committed frontend bundle must be regenerated:
+   ```bash
+   cd src/frontend && npm run build
+   # then sync dist/{index.html,assets/,icon.svg} → src/praisonaiui/templates/frontend/
+   ```
+   (See `AGENTS.md` → "Frontend Build".)
+5. **Run frontend CI** — `.github/workflows/ci.yml` (ruff + pytest) must pass, plus the
+   frontend build step.
+
+## 4. Adapter boundary
+
+Consumers import from **`@praisonaiui/react` only** — never from raw vendored paths such as
+`@/components/ui/...` or any upstream module.
+
+```tsx
+// ✅ stable PraisonAIUI contract
+import { Button, ThemeToggle } from '@praisonaiui/react'
+
+// ❌ leaks upstream/shadcn paths across the codebase
+import { Button } from '@/components/ui/button'
+```
+
+This keeps the upstream surface behind one file (`components/index.ts`), so a BeautifulUI
+bump changes vendored source but not consumer imports.
+
+## 5. Rollback
+
+- BeautifulUI-backed views mount behind a **feature flag**; the **vanilla dashboard views**
+  (`src/praisonaiui/templates/frontend/plugins/*`) remain the fallback and are kept working.
+- To roll back a bad bump: revert the vendoring PR (source + `BEAUTIFULUI_UPSTREAM.md` +
+  rebuilt bundle move together in one commit), and disable the flag. No consumer import
+  changes are needed because everything routes through the adapter (§4).
+
+## 6. Licence
+
+BeautifulUI is **MIT, © Shane Levine**. The full licence text is retained in
+[`THIRD_PARTY_NOTICES.md`](https://github.com/MervinPraison/PraisonAIUI/blob/main/THIRD_PARTY_NOTICES.md)
+at the repository root. Any PR that vendors or bumps BeautifulUI source must keep that
+notice present and current.
+
+## 7. Commercial icons — do not vendor
+
+Do **not** use `@central-icons-react` without a licence — it is commercial. PraisonAIUI
+standardises on **`lucide-react`** (`iconLibrary: lucide` in `components.json`). If an
+upstream BeautifulUI primitive imports commercial icons, swap them for lucide equivalents
+during vendoring.
@@ -13,7 +13,9 @@ MEETING_AGENT_PORT=8000
 # ── Recall.ai (EU workspace: Praison) ───────────────────────────────────────
 # Region must match dashboard: eu-central-1 (Frankfurt).
 RECALL_REGION=eu-central-1
-RECALL_WORKSPACE_ID=33a035c7-03c1-4ec9-b752-18cdfbbefc42
+# Optional: your Recall workspace ID (shown in the config-status API only; the
+# API key already scopes requests to a workspace).
+RECALL_WORKSPACE_ID=

 # REST API key from Recall dashboard (Developers → API Keys). Server-side only.
 RECALL_API_KEY=
@@ -17,9 +17,10 @@ Meeting Bots for live Zoom / Meet / Teams capture.

 ## Recall.ai workspace

-This integration targets the **Praison** workspace in **EU (Frankfurt)**:
+This integration targets your Recall workspace in **EU (Frankfurt)** by default:

- Workspace ID: `33a035c7-03c1-4ec9-b752-18cdfbbefc42`
+- Workspace ID: set `RECALL_WORKSPACE_ID` to your own (optional — the API key
+  already scopes requests to a workspace)
 - API region: `eu-central-1`

 Follow the [Recall onboarding guide](https://docs.recall.ai/docs/agent-quickstarts) via Recall MCP (`recall://guides/onboarding`) for credential and webhook setup.
@@ -102,7 +102,7 @@ _agent: Any | None = None
 def _recall_schedule_tool():
    from praisonai_tools.tools.decorator import tool

-    from integrations.recall.service import schedule_recall_bot as _schedule
+    from meeting_integrations.recall.service import schedule_recall_bot as _schedule

    @tool
    def schedule_recall_bot(meeting_url: str, title: str = "") -> dict:
@@ -115,7 +115,7 @@ def _recall_schedule_tool():
 def _upcoming_meetings_tool():
    from praisonai_tools.tools.decorator import tool

-    from integrations.recall.calendar_service import list_upcoming_calendar_meetings
+    from meeting_integrations.recall.calendar_service import list_upcoming_calendar_meetings

    @tool
    def get_upcoming_meetings(hours: int = 24) -> list[dict]:
@@ -194,8 +194,8 @@ def _sync_stale_recall_status(record: dict[str, Any]) -> dict[str, Any]:
    if not os.getenv("RECALL_API_KEY"):
        return record
    try:
-        from integrations.recall.client import RecallClient
-        from integrations.recall.config import load_recall_settings
+        from meeting_integrations.recall.client import RecallClient
+        from meeting_integrations.recall.config import load_recall_settings
        from pipeline import merge_meeting_metadata

        bot = RecallClient(load_recall_settings()).get_bot(str(bot_id))
@@ -464,7 +464,7 @@ async def calendar_page():
    status_note = "Set RECALL_API_KEY to load calendar status."
    if os.getenv("RECALL_API_KEY"):
        try:
-            from integrations.recall.calendar_service import list_upcoming_calendar_meetings
+            from meeting_integrations.recall.calendar_service import list_upcoming_calendar_meetings

            upcoming = list_upcoming_calendar_meetings(hours=48)
            status_note = f"{len(upcoming)} upcoming event(s) with video links (48h window)."
@@ -1,13 +0,0 @@
-"""Recall.ai Meeting Bot integration for the meeting-agent example."""
-
-from integrations.recall.client import RecallClient
-from integrations.recall.config import RecallSettings, load_recall_settings
-from integrations.recall.verify import VerificationError, verify_request_from_recall
-
-__all__ = [
-    "RecallClient",
-    "RecallSettings",
-    "VerificationError",
-    "load_recall_settings",
-    "verify_request_from_recall",
-]
@@ -0,0 +1,13 @@
+"""Recall.ai Meeting Bot integration for the meeting-agent example."""
+
+from meeting_integrations.recall.client import RecallClient
+from meeting_integrations.recall.config import RecallSettings, load_recall_settings
+from meeting_integrations.recall.verify import VerificationError, verify_request_from_recall
+
+__all__ = [
+    "RecallClient",
+    "RecallSettings",
+    "VerificationError",
+    "load_recall_settings",
+    "verify_request_from_recall",
+]
@@ -6,9 +6,9 @@ import logging
 from datetime import datetime, timedelta, timezone
 from typing import Any

-from integrations.recall.client import RecallAPIError, RecallClient
-from integrations.recall.config import RecallSettings
-from integrations.recall.store import RecallStore
+from meeting_integrations.recall.client import RecallAPIError, RecallClient
+from meeting_integrations.recall.config import RecallSettings
+from meeting_integrations.recall.store import RecallStore

 logger = logging.getLogger(__name__)

@@ -64,7 +64,7 @@ def list_upcoming_calendar_meetings(
    client: RecallClient | None = None,
 ) -> list[dict[str, Any]]:
    """List upcoming Recall calendar events with video links (for UI + agent tool)."""
-    from integrations.recall.config import load_recall_settings
+    from meeting_integrations.recall.config import load_recall_settings

    settings = settings or load_recall_settings()
    client = client or RecallClient(settings)
@@ -10,7 +10,7 @@ from typing import Any

 import httpx

-from integrations.recall.config import RecallSettings
+from meeting_integrations.recall.config import RecallSettings

 logger = logging.getLogger(__name__)

@@ -93,9 +93,7 @@ def load_recall_settings(*, require_public_url: bool = False) -> RecallSettings:
        api_key=_require("RECALL_API_KEY"),
        webhook_verification_secret=_require("RECALL_WEBHOOK_VERIFICATION_SECRET"),
        public_api_base_url=public_url,
-        workspace_id=os.getenv(
-            "RECALL_WORKSPACE_ID", "33a035c7-03c1-4ec9-b752-18cdfbbefc42"
-        ).strip(),
+        workspace_id=os.getenv("RECALL_WORKSPACE_ID", "").strip(),
        bot_name=os.getenv("RECALL_BOT_NAME", "Praison Meeting Agent").strip()
        or "Praison Meeting Agent",
        calendar_auto_record=auto_record not in {"0", "false", "no", "off"},
@@ -5,10 +5,10 @@ from __future__ import annotations
 import logging
 from typing import Any

-from integrations.recall.client import RecallAPIError, RecallClient
-from integrations.recall.config import RecallSettings
-from integrations.recall.store import RecallStore, webhook_event_key
-from integrations.recall.transcript import transcript_data_to_line, transcript_download_to_text
+from meeting_integrations.recall.client import RecallAPIError, RecallClient
+from meeting_integrations.recall.config import RecallSettings
+from meeting_integrations.recall.store import RecallStore, webhook_event_key
+from meeting_integrations.recall.transcript import transcript_data_to_line, transcript_download_to_text

 logger = logging.getLogger(__name__)

@@ -22,7 +22,7 @@ def _speaker_prefix(line: str) -> str:
 def _upsert_live_transcript(meeting_id: str, line: str, *, partial: bool = False) -> str:
    """Merge one live caption line; partial updates replace the last same-speaker line."""
    from praisonai_tools.tools.meeting_tools import get_meeting
-    from integrations.recall.live import publish_live_chunk
+    from meeting_integrations.recall.live import publish_live_chunk
    from pipeline import _metadata_lock

    # Hold the merge lock across the read too: streaming transcript events are
@@ -120,7 +120,7 @@ def process_recall_webhook(

    if event_type == "bot.in_call_recording" and meeting_id:
        _merge_metadata(meeting_id, {"live_status": "live", "status": "live", "error": ""})
-        from integrations.recall.live import set_live_status
+        from meeting_integrations.recall.live import set_live_status

        set_live_status(meeting_id, "live")
        return
@@ -139,7 +139,7 @@ def process_recall_webhook(

    if event_type == "bot.call_ended" and meeting_id:
        _merge_metadata(meeting_id, {"live_status": "ended"})
-        from integrations.recall.live import set_live_status
+        from meeting_integrations.recall.live import set_live_status

        set_live_status(meeting_id, "ended")
        return
@@ -6,8 +6,8 @@ import logging
 import re
 from typing import Any

-from integrations.recall.client import RecallAPIError, RecallClient
-from integrations.recall.config import RecallConfigError, RecallSettings, load_recall_settings
+from meeting_integrations.recall.client import RecallAPIError, RecallClient
+from meeting_integrations.recall.config import RecallConfigError, RecallSettings, load_recall_settings

 logger = logging.getLogger(__name__)

@@ -9,7 +9,7 @@ from contextlib import closing
 from pathlib import Path
 from typing import Any

-from integrations.recall.client import RecallClient
+from meeting_integrations.recall.client import RecallClient


 def _state_db_path() -> str:
@@ -100,42 +100,42 @@ class RecallStore:

    @staticmethod
    def extract_bot_id(payload: dict[str, Any]) -> str | None:
-        from integrations.recall.client import RecallClient
+        from meeting_integrations.recall.client import RecallClient

        return RecallClient.extract_bot_id(payload)

    @staticmethod
    def extract_recording_id(payload: dict[str, Any]) -> str | None:
-        from integrations.recall.client import RecallClient
+        from meeting_integrations.recall.client import RecallClient

        return RecallClient.extract_recording_id(payload)

    @staticmethod
    def extract_transcript_id(payload: dict[str, Any]) -> str | None:
-        from integrations.recall.client import RecallClient
+        from meeting_integrations.recall.client import RecallClient

        return RecallClient.extract_transcript_id(payload)

    @staticmethod
    def mark_event_processed(event_key: str, event_type: str) -> bool:
-        """Return False if this event was already processed."""
+        """Claim an event for processing; return False if already processed.
+
+        Uses an atomic ``INSERT OR IGNORE`` so concurrent duplicate webhook
+        deliveries (Svix retries/fan-out) cannot both pass the check — the
+        second insert is ignored and reports zero changed rows.
+        """
        from datetime import datetime, timezone

        with closing(_connect()) as conn, conn:
-            row = conn.execute(
-                "SELECT 1 FROM recall_webhook_events WHERE event_key = ?",
-                (event_key,),
-            ).fetchone()
-            if row:
-                return False
-            conn.execute(
+            cursor = conn.execute(
                """
-                INSERT INTO recall_webhook_events (event_key, event_type, processed_at)
+                INSERT OR IGNORE INTO recall_webhook_events
+                    (event_key, event_type, processed_at)
                VALUES (?, ?, ?)
                """,
                (event_key, event_type, datetime.now(timezone.utc).isoformat()),
            )
-        return True
+            return cursor.rowcount > 0

    @staticmethod
    def save_calendar_setup(setup_id: str, platform: str, state: dict[str, Any]) -> None:
@@ -11,18 +11,18 @@ from typing import Any
 from starlette.requests import Request
 from starlette.responses import JSONResponse, PlainTextResponse, Response, StreamingResponse

-from integrations.recall.calendar import forward_calendar_callback
-from integrations.recall.calendar_service import (
+from meeting_integrations.recall.calendar import forward_calendar_callback
+from meeting_integrations.recall.calendar_service import (
    list_upcoming_calendar_meetings,
    process_calendar_webhook,
    sync_calendar_events,
 )
-from integrations.recall.config import RecallConfigError, load_recall_settings
-from integrations.recall.live import snapshot_for_api, wait_for_update
-from integrations.recall.processor import process_recall_webhook
-from integrations.recall.service import cancel_recall_bot, schedule_recall_bot
-from integrations.recall.store import RecallStore
-from integrations.recall.verify import VerificationError, verify_request_from_recall
+from meeting_integrations.recall.config import RecallConfigError, load_recall_settings
+from meeting_integrations.recall.live import snapshot_for_api, wait_for_update
+from meeting_integrations.recall.processor import process_recall_webhook
+from meeting_integrations.recall.service import cancel_recall_bot, schedule_recall_bot
+from meeting_integrations.recall.store import RecallStore
+from meeting_integrations.recall.verify import VerificationError, verify_request_from_recall

 logger = logging.getLogger(__name__)

@@ -208,7 +208,7 @@ async def api_recall_calendar_status(_request: Request) -> JSONResponse:
    public_url = os.getenv("PUBLIC_API_BASE_URL", "").strip()
    calendars: list[dict[str, Any]] = []
    try:
-        from integrations.recall.client import RecallClient
+        from meeting_integrations.recall.client import RecallClient

        settings = load_recall_settings()
        calendars = RecallClient(settings).list_calendars(status="connected")
@@ -45,9 +45,9 @@ def _duration_seconds(started_at: str | None, completed_at: str | None) -> int |


 def recover(meeting_id: str, bot_id: str | None = None) -> dict:
-    from integrations.recall.client import RecallClient
-    from integrations.recall.config import load_recall_settings
-    from integrations.recall.transcript import transcript_download_to_text
+    from meeting_integrations.recall.client import RecallClient
+    from meeting_integrations.recall.config import load_recall_settings
+    from meeting_integrations.recall.transcript import transcript_download_to_text
    from pipeline import merge_meeting_metadata, run_post_transcript_pipeline
    from praisonai_tools.tools.meeting_tools import get_meeting

@@ -39,12 +39,12 @@ def run_smoke() -> None:
    os.environ.setdefault("RECALL_REGION", "eu-central-1")
    os.environ.setdefault("RECALL_API_KEY", "smoke-test-key")
    os.environ.setdefault("RECALL_WEBHOOK_VERIFICATION_SECRET", "whsec_" + base64.b64encode(b"smoke-secret-key-32-bytes-long!!").decode())
-    os.environ.setdefault("RECALL_WORKSPACE_ID", "33a035c7-03c1-4ec9-b752-18cdfbbefc42")
+    os.environ.setdefault("RECALL_WORKSPACE_ID", "smoke-workspace")
    os.environ.setdefault("PUBLIC_API_BASE_URL", "https://smoke.example.test")

    import app as meeting_app  # noqa: WPS433
-    from integrations.recall.config import load_recall_settings
-    from integrations.recall.verify import verify_request_from_recall
+    from meeting_integrations.recall.config import load_recall_settings
+    from meeting_integrations.recall.verify import verify_request_from_recall

    settings = load_recall_settings()
    assert settings.api_key == "smoke-test-key"
@@ -4,13 +4,48 @@ OPENAI_API_KEY=

 VOICE_AGENT_PORT=8001

+# LLM for voice assistant (synced to provider on setup / start_dev)
+VOICE_AGENT_MODEL=gpt-4o-mini
+
 # Voice telephony provider
 VOICE_API_BASE=
 VOICE_PRIVATE_API_KEY=
+VOICE_PUBLIC_API_KEY=
 VOICE_SERVER_URL_SECRET=
 VOICE_ASSISTANT_ID=
 VOICE_PHONE_NUMBER_ID=

+# Phase 2 — browser web talk (provider web SDK)
+# VOICE_WEB_SDK_URL=https://esm.sh/@vapi-ai/[email protected]
+# VOICE_WEB_SDK_GLOBAL=VoiceWebClient
+
+# Dynamic assistant (assistant-request returns inline config + agent tools)
+# VOICE_DYNAMIC_ASSISTANT=true
+# VOICE_AGENT_EXTRA_TOOLS=meeting_search
+
+# ElevenLabs Speech Engine — OpenAI tool calling on the sidecar LLM path
+# VOICE_SPEECH_USE_TOOLS=true
+
+# Cross-call memory — stable browser user id → PraisonAI session reload
+# VOICE_MEMORY_ENABLED=true
+# VOICE_MEMORY_HISTORY_LIMIT=20
+# VOICE_MEMORY_USE_KNOWLEDGE=true
+
+# OpenAI Realtime (browser speech-to-speech — higher audio token cost; keep demos short)
+# REALTIME_MODEL=gpt-realtime-2.1-mini
+# REALTIME_VOICE=marin
+# REALTIME_REASONING_EFFORT=low
+# REALTIME_INSTRUCTIONS=You are a helpful voice assistant. Keep replies to one or two short sentences.
+# REALTIME_FIRST_MESSAGE=
+
+# ElevenLabs Speech Engine (Phase 3 — browser voice, BYO LLM)
+# ELEVENLABS_API_KEY=
+# SPEECH_ENGINE_ID=          # auto-written by setup_speech_engine.py
+# SPEECH_ENGINE_PUBLIC_URL=  # auto-written — dedicated tunnel for sidecar :8002
+# SPEECH_ENGINE_PORT=8002
+# SPEECH_ENGINE_FIRST_MESSAGE=Hello! How can I help you today?
+# ELEVENLABS_VOICE_ID=          # Praison/custom voice from ElevenLabs dashboard
+
 # Public HTTPS URL for provider webhook (Cloudflare tunnel)
 # cloudflared tunnel --url http://127.0.0.1:8001
 PUBLIC_API_BASE_URL=
@@ -34,16 +34,80 @@ Or run `py -3.13 setup_voice_provider.py` after `start_dev.ps1`.
 | `/api/voice/calls` | POST | Outbound call `{ "customer_number": "+1..." }` |
 | `/api/voice/calls` | GET | List stored calls |
 | `/api/voice/calls/{id}` | GET | Call detail + transcript |
+| `/api/voice/calls/{id}/live-transcript` | GET | SSE live transcript stream |
+| `/api/voice/web-config` | GET | Browser talk widget config |
 | `/webhooks/voice` | POST | Provider webhooks (tool-calls, transcript, end-of-call) |
 | `/api/voice/config` | GET | Config status |
+| `/api/voice/doctor` | GET | Voice stack health checks (.env, ports, APIs, tunnel) |
+
+## Phase 2 features
+
+- **Live transcript** — `GET /api/voice/calls/{id}/live-transcript` (SSE); **Call detail** page uses `plugin.js` + EventSource
+- **Dynamic assistant** — set `VOICE_DYNAMIC_ASSISTANT=true` so `assistant-request` returns inline config with agent tool names
+- **Agent tools** — `tool-calls` run through `integrations/voice/agent_runner.py` (same registry as dashboard chat)
+- **Web talk** — **Web talk** page loads provider web SDK from `VOICE_WEB_SDK_URL` / `VOICE_WEB_SDK_GLOBAL`
+- **Chat bridge** — final transcript lines and end-of-call summaries appear in **Chat** sidebar as `📞 Call …` sessions
+
+Optional: `VOICE_AGENT_EXTRA_TOOLS=meeting_search` adds meeting search when `praisonai_tools` is installed.
+
+## Cross-call memory (PraisonAI sessions)
+
+Stable browser `voice-user-id` (localStorage) maps to `voice-user-{id}` PraisonAI session folder. Prior turns reload on new calls/restarts when `VOICE_MEMORY_ENABLED=true` (default).
+
+| Route | Purpose |
+|-------|---------|
+| `/api/voice/memory/attach-user` | ElevenLabs — attach user before conversation starts |
+| `/api/voice/memory/bind` | Bind `call_id` + `user_id` |
+| `/api/voice/memory/resolve?user_id=` | Resolve Praison session id |
+
+Knowledge base entries are injected into voice prompts when available.
+
+## Call detail UI (PraisonAIUI components)
+
+Call detail uses server-built components (`code_block`, `badge`, `key_value_list`, `tabs`) via `GET /api/voice/calls/{id}/ui`. Live SSE updates the transcript code block. Mic/WebRTC talk pages still use minimal `plugin.js`.
+
+## End-of-call summary & analytics
+
+When a session ends (ElevenLabs, GPT Realtime, or telephony webhook), `integrations/voice/call_finalize.py`:
+
+- Generates a one-sentence summary (OpenAI) when the provider did not send one
+- Stores **duration** and **turn count** in call metadata
+- Posts the summary to the **Chat** sidebar via the chat bridge
+
+**Voice calls** table shows Duration and Turns. **Config** page runs the voice doctor (`GET /api/voice/doctor` or `py -3.13 doctor.py`).
+
+## MCP manager
+
+The dashboard **MCP** page lists configured Model Context Protocol servers (`GET /api/mcp/servers`) and supports connect/disconnect from the UI.
+
+## Phase 3 — ElevenLabs Speech Engine
+
+Bring-your-own-LLM browser voice ([Speech Engine docs](https://elevenlabs.io/docs/overview/capabilities/speech-engine)):
+
+```powershell
+pip install -r requirements-voice.txt
+# Add ELEVENLABS_API_KEY to .env
+.\start_dev.ps1 -Restart
+```
+
+- **ElevenLabs talk** page — WebRTC client, token from `/api/voice/speech-engine/token`
+- **Speech Engine sidecar** (`speech_engine_sidecar.py` on port **8002**) — official SDK WebSocket server
+- **Dedicated tunnel** — `SPEECH_ENGINE_PUBLIC_URL` → `wss://…/ws` (separate from dashboard tunnel)
+- **`setup_speech_engine.py`** — registers the public WebSocket URL with ElevenLabs
+
+| Route | Purpose |
+|-------|---------|
+| `/api/voice/speech-engine/config` | Client config |
+| `/api/voice/speech-engine/token` | Browser conversation token |
+| `/ws` | Speech Engine WebSocket (ElevenLabs → your server) |

 ## Architecture

 External platform handles voice pipeline (phone, STT, TTS, turn-taking). This app handles:

- **tool-calls** → PraisonAI-registered tools (`integrations/voice/tools.py`)
- **transcript** / **end-of-call-report** → SQLite store + dashboard pages
- **assistant-request** → returns `VOICE_ASSISTANT_ID`
+- **tool-calls** → PraisonAI agent tools (`integrations/voice/agent_runner.py`)
+- **transcript** / **end-of-call-report** → SQLite store, SSE hub, chat bridge
+- **assistant-request** → default `VOICE_ASSISTANT_ID` or dynamic inline assistant

 Same pattern as `examples/python/meeting-agent` (Recall webhooks).

@@ -2,7 +2,7 @@

  - Outbound calls     → POST /api/voice/calls
  - Provider webhooks  → POST /webhooks/voice (tool-calls, transcript, end-of-call)
-  - Dashboard          → Calls list + call detail + chat
+  - Dashboard          → Calls list + live call detail + web talk + chat bridge

 Run:
    cd examples/python/voice-agent
@@ -20,17 +20,40 @@ import sys
 from pathlib import Path
 from typing import Any

-from starlette.routing import Route
-from voice_routes import (
-    api_voice_call_detail,
-    api_voice_config_status,
-    api_voice_create_call,
-    api_voice_list_calls,
-    webhook_voice,
-)
-
 import praisonaiui as aiui
 from praisonaiui.server import create_app
+from starlette.routing import Route, WebSocketRoute
+
+from memory_routes import (
+    api_voice_memory_attach_user,
+    api_voice_memory_bind,
+    api_voice_memory_config,
+    api_voice_memory_for_call,
+    api_voice_memory_resolve,
+)
+from realtime_routes import (
+    api_realtime_config,
+    api_realtime_end,
+    api_realtime_new_call,
+    api_realtime_session,
+    api_realtime_transcript,
+)
+from speech_engine_routes import (
+    api_speech_engine_config,
+    api_speech_engine_token,
+    speech_engine_websocket,
+)
+from voice_routes import (
+    api_voice_call_detail,
+    api_voice_call_detail_ui,
+    api_voice_config_status,
+    api_voice_create_call,
+    api_voice_doctor,
+    api_voice_list_calls,
+    api_voice_live_transcript,
+    api_voice_web_config,
+    webhook_voice,
+)

 _EXAMPLE_DIR = Path(__file__).resolve().parent
 if str(_EXAMPLE_DIR) not in sys.path:
@@ -47,68 +70,54 @@ def _load_local_env() -> None:
            continue
        key, _, value = line.partition("=")
        key, value = key.strip(), value.strip()
-        if key and key not in os.environ:
+        if key:
            os.environ[key] = value


 _load_local_env()

-aiui.set_pages(["chat", "voice-calls", "call-detail", "config"])
+aiui.set_pages(
+    [
+        "chat",
+        "voice-calls",
+        "call-detail",
+        "web-talk",
+        "eleven-talk",
+        "realtime-talk",
+        "mcp",
+        "config",
+    ]
+)
 aiui.set_style("dashboard")
-
-_agent: Any | None = None
+aiui.set_custom_js(_EXAMPLE_DIR / "plugin.js")


 def get_agent():
-    global _agent
-    if _agent is not None:
-        return _agent
-    from integrations.voice.tools import echo_message, get_current_time, register_tool
-    from praisonaiagents import Agent
+    from integrations.voice.agent_runner import get_voice_agent

-    register_tool("get_current_time", get_current_time)
-    register_tool("echo_message", echo_message)
-
-    _agent = Agent(
-        name="Voice Assistant",
-        instructions=(
-            "You help with voice and phone agent demos. "
-            "Use tools when the caller asks for the time or wants a message echoed."
-        ),
-        model="gpt-4o-mini",
-        tools=[get_current_time, echo_message],
-    )
-    from praisonaiagents.escalation.loop_guard import LoopGuard, LoopGuardConfig
-
-    max_turn_sec = float(os.getenv("VOICE_AGENT_LOOP_GUARD_MAX_SEC", "600"))
-    _agent._loop_guard = LoopGuard(
-        LoopGuardConfig(enabled=True, max_time_per_turn=max_turn_sec)
-    )
-    return _agent
+    return get_voice_agent()


-def _demo_tool_reply(text: str) -> str | None:
-    """Fast path for demo tool queries — no LLM required."""
-    from integrations.voice.tools import echo_message, get_current_time
+aiui.register_agent("voice-assistant", get_agent())

-    lower = text.lower().strip()
-    if any(p in lower for p in ("what time", "current time", "tell me the time", "time is it")):
-        return str(get_current_time()["spoken"])
-    if lower.startswith("echo "):
-        return str(echo_message(text[5:])["spoken"])
-    if lower.startswith("repeat "):
-        return str(echo_message(text[7:])["spoken"])
-    return None
+
+@aiui.on_app_startup
+async def _start_voice_chat_bridge() -> None:
+    from integrations.voice.chat_bridge import start_chat_bridge_worker
+
+    start_chat_bridge_worker()


@aiui.reply
 async def on_message(message: str):
    """Chat with the voice demo agent."""
+    from integrations.voice.tools import demo_tool_reply
+
    text = str(message).strip()
    if not text:
        return

-    demo = _demo_tool_reply(text)
+    demo = demo_tool_reply(text)
    if demo is not None:
        await aiui.say(demo)
        return
@@ -126,19 +135,25 @@ async def on_message(message: str):


 def _list_calls() -> list[dict[str, Any]]:
+    from integrations.voice.call_finalize import analytics_for_record
    from integrations.voice.store import VoiceCallStore

    rows = VoiceCallStore.list_calls()
-    return [
-        {
-            "id": r["call_id"],
-            "status": r["status"],
-            "customer_number": r.get("customer_number") or "—",
-            "transcript_preview": (r.get("transcript") or "")[:120],
-            "updated_at": r.get("updated_at") or "",
-        }
-        for r in rows
-    ]
+    out: list[dict[str, Any]] = []
+    for r in rows:
+        analytics = analytics_for_record(r)
+        out.append(
+            {
+                "id": r["call_id"],
+                "status": r["status"],
+                "customer_number": r.get("customer_number") or "—",
+                "transcript_preview": (r.get("transcript") or "")[:120],
+                "duration": analytics.get("duration") or "—",
+                "turn_count": analytics.get("turn_count") or 0,
+                "updated_at": r.get("updated_at") or "",
+            }
+        )
+    return out


 def _pick_call(calls: list[dict[str, Any]]) -> dict[str, Any]:
@@ -149,19 +164,22 @@ def _pick_call(calls: list[dict[str, Any]]) -> dict[str, Any]:
@aiui.page("voice-calls", title="Voice calls", icon="📞", group="Voice", order=1)
 async def voice_calls_page():
    calls = _list_calls()
-    rows = [[c["id"], c["customer_number"], c["status"], c["updated_at"]] for c in calls]
+    rows = [
+        [c["id"], c["customer_number"], c["status"], c["duration"], str(c["turn_count"]), c["updated_at"]]
+        for c in calls
+    ]
    return aiui.layout(
        [
            aiui.text("Outbound and inbound voice calls via telephony provider"),
            aiui.table(
-                headers=["Call ID", "Customer", "Status", "Updated"],
-                rows=rows or [["—", "—", "—", "—"]],
+                headers=["Call ID", "Customer", "Status", "Duration", "Turns", "Updated"],
+                rows=rows or [["—", "—", "—", "—", "—", "—"]],
            ),
            aiui.alert(
-                "Start a call: POST /api/voice/calls with {\"customer_number\": \"+1...\"}. "
-                "Set provider webhook URL to {PUBLIC_API_BASE_URL}/webhooks/voice",
+                "Open **Call detail** for live transcript (SSE). "
+                "Voice sessions also appear in **Chat** sidebar as 📞 entries.",
                variant="info",
-                title="Setup",
+                title="Phase 2",
            ),
        ]
    )
@@ -169,23 +187,65 @@ async def voice_calls_page():

@aiui.page("call-detail", title="Call detail", icon="🎙️", group="Voice", order=2)
 async def call_detail_page():
+    """Server fallback — plugin.js registerView overrides with live SSE + UI components."""
+    from integrations.voice.call_detail_ui import build_call_detail_layout
    from integrations.voice.store import VoiceCallStore

    calls = _list_calls()
    picked = _pick_call(calls)
    record = VoiceCallStore.get_call(picked["id"]) if picked.get("id") and picked["id"] != "—" else None
-    transcript = (record or {}).get("transcript") or "(empty)"
-    summary = (record or {}).get("summary") or "(not available yet)"
-    status = (record or {}).get("status") or picked.get("status") or "unknown"
+    if record:
+        return build_call_detail_layout(record)
    return aiui.layout(
        [
-            aiui.text(f"Call {picked.get('id', '—')}"),
-            aiui.badge(status, variant="default"),
-            aiui.tabs(
-                [
-                    {"label": "Transcript", "children": [aiui.code_block(transcript, language="text")]},
-                    {"label": "Summary", "children": [aiui.text(summary)]},
-                ]
+            aiui.alert(
+                "No calls yet. Start from ElevenLabs talk, GPT Realtime, or Web talk.",
+                variant="info",
+                title="Call detail",
+            ),
+        ]
+    )
+
+
+@aiui.page("web-talk", title="Web talk", icon="🎤", group="Voice", order=3)
+async def web_talk_page():
+    return aiui.layout(
+        [
+            aiui.text("Browser voice widget — client view loads from plugin.js"),
+            aiui.alert(
+                "Set VOICE_PUBLIC_API_KEY, VOICE_ASSISTANT_ID, VOICE_WEB_SDK_URL, and "
+                "VOICE_WEB_SDK_GLOBAL in .env.",
+                variant="info",
+                title="Web SDK",
+            ),
+        ]
+    )
+
+
+@aiui.page("eleven-talk", title="ElevenLabs talk", icon="🗣️", group="Voice", order=4)
+async def eleven_talk_page():
+    return aiui.layout(
+        [
+            aiui.text("Browser voice via ElevenLabs Speech Engine — loads from plugin.js"),
+            aiui.alert(
+                "Set ELEVENLABS_API_KEY, run setup_speech_engine.py after start_dev.ps1, "
+                "then click Start conversation.",
+                variant="info",
+                title="Speech Engine",
+            ),
+        ]
+    )
+
+
+@aiui.page("realtime-talk", title="GPT Realtime", icon="⚡", group="Voice", order=5)
+async def realtime_talk_page():
+    return aiui.layout(
+        [
+            aiui.text("Browser voice via OpenAI gpt-realtime-2.1-mini — loads from plugin.js"),
+            aiui.alert(
+                "Set OPENAI_API_KEY and REALTIME_MODEL=gpt-realtime-2.1-mini.",
+                variant="warning",
+                title="OpenAI Realtime",
            ),
        ]
    )
@@ -196,8 +256,25 @@ app.routes[0:0] = [
    Route("/api/voice/calls", api_voice_create_call, methods=["POST"]),
    Route("/api/voice/calls", api_voice_list_calls, methods=["GET"]),
    Route("/api/voice/calls/{call_id}", api_voice_call_detail, methods=["GET"]),
+    Route("/api/voice/calls/{call_id}/ui", api_voice_call_detail_ui, methods=["GET"]),
+    Route("/api/voice/calls/{call_id}/live-transcript", api_voice_live_transcript, methods=["GET"]),
    Route("/api/voice/config", api_voice_config_status, methods=["GET"]),
+    Route("/api/voice/doctor", api_voice_doctor, methods=["GET"]),
+    Route("/api/voice/web-config", api_voice_web_config, methods=["GET"]),
+    Route("/api/voice/speech-engine/config", api_speech_engine_config, methods=["GET"]),
+    Route("/api/voice/speech-engine/token", api_speech_engine_token, methods=["GET"]),
+    Route("/api/voice/memory/config", api_voice_memory_config, methods=["GET"]),
+    Route("/api/voice/memory/bind", api_voice_memory_bind, methods=["POST"]),
+    Route("/api/voice/memory/attach-user", api_voice_memory_attach_user, methods=["POST"]),
+    Route("/api/voice/memory/resolve", api_voice_memory_resolve, methods=["GET"]),
+    Route("/api/voice/memory/for-call", api_voice_memory_for_call, methods=["GET"]),
+    Route("/api/voice/realtime/config", api_realtime_config, methods=["GET"]),
+    Route("/api/voice/realtime/new-call", api_realtime_new_call, methods=["POST"]),
+    Route("/api/voice/realtime/session", api_realtime_session, methods=["POST"]),
+    Route("/api/voice/realtime/transcript", api_realtime_transcript, methods=["POST"]),
+    Route("/api/voice/realtime/end", api_realtime_end, methods=["POST"]),
    Route("/webhooks/voice", webhook_voice, methods=["POST"]),
+    WebSocketRoute("/ws", speech_engine_websocket),
 ]

 if __name__ == "__main__":
@@ -0,0 +1,41 @@
+"""CLI for voice-agent stack diagnostics."""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+from pathlib import Path
+
+_EXAMPLE_DIR = Path(__file__).resolve().parent
+if str(_EXAMPLE_DIR) not in sys.path:
+    sys.path.insert(0, str(_EXAMPLE_DIR))
+
+
+def _load_local_env() -> None:
+    env_path = _EXAMPLE_DIR / ".env"
+    if not env_path.is_file():
+        return
+    for raw in env_path.read_text(encoding="utf-8").splitlines():
+        line = raw.strip()
+        if not line or line.startswith("#") or "=" not in line:
+            continue
+        key, _, value = line.partition("=")
+        key, value = key.strip(), value.strip()
+        if key:
+            os.environ[key] = value
+
+
+def main() -> int:
+    _load_local_env()
+    from integrations.voice.doctor import run_voice_doctor
+
+    app_port = int(os.getenv("VOICE_AGENT_PORT", "8001"))
+    sidecar_port = int(os.getenv("SPEECH_ENGINE_SIDECAR_PORT", "8002"))
+    report = run_voice_doctor(app_port=app_port, sidecar_port=sidecar_port)
+    print(json.dumps(report, indent=2))
+    return 1 if report["summary"]["failed"] else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
@@ -0,0 +1,171 @@
+"""Run voice tool-calls through the PraisonAI agent tool registry."""
+
+from __future__ import annotations
+
+import json
+import os
+from typing import Any, Callable
+
+from integrations.voice.tools import execute_tool, register_tool
+
+_agent: Any | None = None
+_session_agents: dict[str, Any] = {}
+
+
+def _tool_name(tool: Any) -> str | None:
+    for attr in ("name", "__name__"):
+        value = getattr(tool, attr, None)
+        if isinstance(value, str) and value.strip():
+            return value.strip()
+    fn = getattr(tool, "func", None) or getattr(tool, "function", None)
+    if fn is not None:
+        return _tool_name(fn)
+    return None
+
+
+def _build_voice_agent(*, session_id: str | None = None):
+    from praisonaiagents import Agent
+    from praisonaiagents.config.feature_configs import MemoryConfig
+    from praisonaiagents.escalation.loop_guard import LoopGuard, LoopGuardConfig
+
+    from integrations.voice.session_memory import memory_enabled
+    from integrations.voice.tools import echo_message, get_current_time
+
+    register_tool("get_current_time", get_current_time)
+    register_tool("echo_message", echo_message)
+
+    tools: list[Any] = [get_current_time, echo_message]
+    extra = os.getenv("VOICE_AGENT_EXTRA_TOOLS", "").strip()
+    if extra == "meeting_search":
+        try:
+            from praisonai_tools.tools.meeting_tools import search_meetings
+
+            tools.append(search_meetings)
+            register_tool("search_meetings", search_meetings)
+        except ImportError:
+            pass
+
+    memory_cfg = None
+    if memory_enabled() and session_id:
+        memory_cfg = MemoryConfig(
+            history=True,
+            session_id=session_id,
+            history_limit=int(os.getenv("VOICE_MEMORY_HISTORY_LIMIT", "20")),
+        )
+
+    agent = Agent(
+        name="Voice Assistant",
+        instructions=(
+            "You help with voice and phone agent demos. "
+            "Use tools when the caller asks for the time, wants a message echoed, "
+            "or asks about past meetings."
+        ),
+        model=os.getenv("VOICE_AGENT_MODEL", "gpt-4o-mini"),
+        tools=tools,
+        memory=memory_cfg or False,
+    )
+    max_turn_sec = float(os.getenv("VOICE_AGENT_LOOP_GUARD_MAX_SEC", "600"))
+    agent._loop_guard = LoopGuard(LoopGuardConfig(enabled=True, max_time_per_turn=max_turn_sec))
+
+    for tool in tools:
+        name = _tool_name(tool)
+        if name:
+            register_tool(name, _callable_tool(tool))
+    return agent
+
+
+def get_voice_agent(session_id: str | None = None):
+    """Lazy-load the voice assistant (optionally bound to a PraisonAI session)."""
+    global _agent
+    if session_id:
+        cached = _session_agents.get(session_id)
+        if cached is not None:
+            return cached
+        agent = _build_voice_agent(session_id=session_id)
+        _session_agents[session_id] = agent
+        return agent
+    if _agent is not None:
+        return _agent
+    _agent = _build_voice_agent()
+    return _agent
+
+
+def _callable_tool(tool: Any) -> Callable[..., Any]:
+    fn = getattr(tool, "func", None) or getattr(tool, "run", None) or tool
+    return fn  # type: ignore[return-value]
+
+
+def execute_voice_tool(
+    name: str,
+    parameters: dict[str, Any] | None,
+    *,
+    session_id: str | None = None,
+) -> str:
+    """Execute via agent tool registry, then fall back to static registry."""
+    params = parameters or {}
+    agent = get_voice_agent(session_id)
+    for tool in agent.tools or []:
+        tname = _tool_name(tool)
+        if tname != name:
+            continue
+        fn = _callable_tool(tool)
+        try:
+            result = fn(**params)
+        except TypeError:
+            try:
+                result = fn(params)
+            except Exception as exc:  # noqa: BLE001
+                return json.dumps({"error": str(exc)})
+        except Exception as exc:  # noqa: BLE001
+            return json.dumps({"error": str(exc)})
+        if isinstance(result, str):
+            return result
+        return json.dumps(result)
+    return execute_tool(name, params)
+
+
+def list_registered_tool_names() -> list[str]:
+    agent = get_voice_agent()
+    names: list[str] = []
+    for tool in agent.tools or []:
+        tname = _tool_name(tool)
+        if tname:
+            names.append(tname)
+    return names
+
+
+def build_assistant_request_response(settings: Any) -> dict[str, Any]:
+    """Return provider assistant config for dynamic assistant-request events."""
+    if settings.default_assistant_id and not os.getenv("VOICE_DYNAMIC_ASSISTANT", "").lower() in (
+        "1",
+        "true",
+        "yes",
+    ):
+        return {"assistantId": settings.default_assistant_id}
+
+    tool_names = list_registered_tool_names()
+    assistant: dict[str, Any] = {
+        "name": "Praison Voice Assistant",
+        "firstMessage": os.getenv("VOICE_FIRST_MESSAGE", "Hello! How can I help you today?"),
+        "model": {
+            "provider": os.getenv("VOICE_MODEL_PROVIDER", "openai"),
+            "model": os.getenv("VOICE_AGENT_MODEL", "gpt-4o-mini"),
+            "messages": [
+                {
+                    "role": "system",
+                    "content": (
+                        "You are a helpful voice assistant. When asked for the time, call get_current_time. "
+                        "When asked to repeat or echo text, call echo_message with the message parameter. "
+                        "Available tools: "
+                        + ", ".join(tool_names or ["get_current_time", "echo_message"])
+                    ),
+                }
+            ],
+        },
+    }
+    if settings.webhook_url and settings.server_url_secret:
+        assistant["server"] = {
+            "url": settings.webhook_url,
+            "headers": {"X-Voice-Webhook-Secret": settings.server_url_secret},
+        }
+    return {"assistant": assistant}
@@ -0,0 +1,57 @@
+"""Build PraisonAIUI components for the call detail page."""
+
+from __future__ import annotations
+
+from typing import Any
+
+import praisonaiui as aiui
+
+
+def build_call_detail_layout(record: dict[str, Any]) -> dict[str, Any]:
+    """Return dashboard component tree for call detail (code_block + metadata)."""
+    from integrations.voice.call_finalize import analytics_for_record
+
+    metadata = record.get("metadata") or {}
+    analytics = analytics_for_record(record)
+    transcript = (record.get("transcript") or "").strip() or "(empty — waiting for speech…)"
+    summary = (record.get("summary") or "").strip() or "(not available yet)"
+    status = str(record.get("status") or "unknown")
+
+    kv_items: list[dict[str, str]] = [
+        {"label": "Call ID", "value": str(record.get("call_id") or "—")},
+        {"label": "Status", "value": status},
+        {"label": "Customer", "value": str(record.get("customer_number") or "—")},
+    ]
+    if metadata.get("user_id"):
+        kv_items.append({"label": "User ID", "value": str(metadata["user_id"])})
+    if metadata.get("praison_session_id"):
+        kv_items.append({"label": "Memory session", "value": str(metadata["praison_session_id"])})
+    kv_items.append({"label": "Duration", "value": str(analytics.get("duration") or "—")})
+    kv_items.append({"label": "Turns", "value": str(analytics.get("turn_count") or 0)})
+
+    return aiui.layout(
+        [
+            aiui.text("Live voice call transcript and metadata"),
+            aiui.badge(status, variant="default"),
+            aiui.definition_list(kv_items, title="Call metadata"),
+            aiui.alert(
+                "Transcript updates live via SSE. Final lines use PraisonAIUI code blocks.",
+                variant="info",
+                title="Live feed",
+            ),
+            aiui.tabs(
+                [
+                    {
+                        "label": "Transcript",
+                        "children": [
+                            aiui.code_block(transcript, language="text"),
+                        ],
+                    },
+                    {
+                        "label": "Summary",
+                        "children": [aiui.text(summary)],
+                    },
+                ]
+            ),
+        ]
+    )
@@ -0,0 +1,126 @@
+"""End-of-call summary, analytics, and chat bridge on session close."""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import os
+from datetime import datetime, timezone
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+
+def count_turns(transcript: str) -> int:
+    """Count user/agent lines in stored transcript."""
+    turns = 0
+    for line in (transcript or "").splitlines():
+        text = line.strip()
+        if text.startswith("user:") or text.startswith("agent:") or text.startswith("assistant:"):
+            turns += 1
+    return turns
+
+
+def duration_seconds(created_at: str | None, ended_at: str | None) -> int | None:
+    if not created_at or not ended_at:
+        return None
+    try:
+        start = datetime.fromisoformat(created_at.replace("Z", "+00:00"))
+        end = datetime.fromisoformat(ended_at.replace("Z", "+00:00"))
+        return max(0, int((end - start).total_seconds()))
+    except ValueError:
+        return None
+
+
+def format_duration(seconds: int | None) -> str:
+    if seconds is None:
+        return "—"
+    if seconds < 60:
+        return f"{seconds}s"
+    minutes, rem = divmod(seconds, 60)
+    return f"{minutes}m {rem}s"
+
+
+async def generate_call_summary(transcript: str) -> str:
+    """One-sentence LLM summary for call detail + chat sidebar."""
+    text = (transcript or "").strip()
+    if not text:
+        return "Empty call — no conversation captured."
+    api_key = os.getenv("OPENAI_API_KEY", "").strip()
+    if not api_key:
+        lines = [ln.strip() for ln in text.splitlines() if ln.strip()]
+        preview = " · ".join(lines[:3])
+        return f"Call covered: {preview[:240]}"
+
+    from openai import AsyncOpenAI
+
+    client = AsyncOpenAI(api_key=api_key)
+    try:
+        response = await asyncio.wait_for(
+            client.chat.completions.create(
+                model=os.getenv("VOICE_AGENT_MODEL", "gpt-4o-mini"),
+                messages=[
+                    {
+                        "role": "system",
+                        "content": (
+                            "Summarize this voice call in one concise sentence for a dashboard. "
+                            "Mention main topic and outcome only."
+                        ),
+                    },
+                    {"role": "user", "content": text[:6000]},
+                ],
+                max_tokens=80,
+            ),
+            timeout=15.0,
+        )
+        summary = (response.choices[0].message.content or "").strip()
+        return summary or "Call completed."
+    except Exception:  # noqa: BLE001
+        logger.exception("Call summary generation failed")
+        return "Call completed — summary unavailable."
+
+
+async def finalize_call(call_id: str, *, status: str) -> None:
+    """Persist analytics, optional summary, and notify chat bridge."""
+    from integrations.voice.chat_bridge import enqueue_call_summary
+    from integrations.voice.store import VoiceCallStore
+
+    record = VoiceCallStore.get_call(call_id)
+    if not record:
+        return
+
+    transcript = (record.get("transcript") or "").strip()
+    metadata: dict[str, Any] = dict(record.get("metadata") or {})
+    ended_at = datetime.now(timezone.utc).isoformat()
+    metadata["ended_at"] = ended_at
+    metadata["turn_count"] = count_turns(transcript)
+    metadata["duration_sec"] = duration_seconds(record.get("created_at"), ended_at)
+
+    summary = (record.get("summary") or "").strip()
+    if not summary and transcript:
+        summary = await generate_call_summary(transcript)
+
+    VoiceCallStore.upsert_call(
+        call_id,
+        status=status,
+        summary=summary or None,
+        metadata=metadata,
+    )
+    enqueue_call_summary(call_id, summary or "(no summary)", status)
+
+
+def analytics_for_record(record: dict[str, Any]) -> dict[str, Any]:
+    """Derived analytics for API + dashboard tables."""
+    metadata = record.get("metadata") or {}
+    transcript = record.get("transcript") or ""
+    turns = metadata.get("turn_count")
+    if turns is None:
+        turns = count_turns(transcript)
+    duration_sec = metadata.get("duration_sec")
+    if duration_sec is None:
+        duration_sec = duration_seconds(record.get("created_at"), metadata.get("ended_at"))
+    return {
+        "turn_count": turns,
+        "duration_sec": duration_sec,
+        "duration": format_duration(duration_sec if isinstance(duration_sec, int) else None),
+    }
@@ -0,0 +1,150 @@
+"""Bridge voice call events into Chat UI sessions (sidebar + history)."""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import queue
+import threading
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+_bridge_queue: queue.Queue[dict[str, Any]] = queue.Queue(maxsize=512)
+_worker_lock = threading.Lock()
+_worker_task: asyncio.Task[None] | None = None
+
+
+def voice_session_id(call_id: str) -> str:
+    return f"voice-{call_id}"
+
+
+def enqueue_transcript(call_id: str, role: str, text: str, *, final: bool) -> None:
+    if not call_id or not text.strip():
+        return
+    _enqueue(
+        {
+            "kind": "transcript",
+            "call_id": call_id,
+            "role": role,
+            "text": text.strip(),
+            "final": final,
+        }
+    )
+
+
+def enqueue_call_summary(call_id: str, summary: str, status: str) -> None:
+    if not call_id:
+        return
+    _enqueue(
+        {
+            "kind": "summary",
+            "call_id": call_id,
+            "summary": summary.strip(),
+            "status": status,
+        }
+    )
+
+
+def _enqueue(item: dict[str, Any]) -> None:
+    try:
+        _bridge_queue.put_nowait(item)
+    except queue.Full:
+        logger.warning("Voice chat bridge queue full; dropping event")
+
+
+async def _ensure_session(call_id: str) -> str:
+    from praisonaiui.server import _datastore
+
+    session_id = voice_session_id(call_id)
+    existing = await _datastore.get_session(session_id)
+    if existing is None:
+        await _datastore.create_session(session_id)
+        await _datastore.update_session(
+            session_id,
+            platform="voice",
+            icon="📞",
+            title=f"📞 Call {call_id[:8]}",
+        )
+    return session_id
+
+
+async def _process_item(item: dict[str, Any]) -> None:
+    from praisonaiui.features.chat import get_chat_manager
+    from praisonaiui.server import _datastore
+
+    call_id = str(item.get("call_id") or "")
+    if not call_id:
+        return
+    session_id = await _ensure_session(call_id)
+    mgr = get_chat_manager()
+
+    if item.get("kind") == "transcript":
+        role = str(item.get("role") or "unknown")
+        text = str(item.get("text") or "")
+        final = bool(item.get("final"))
+        if not final:
+            await mgr.broadcast(
+                session_id,
+                {
+                    "type": "voice_transcript_partial",
+                    "session_id": session_id,
+                    "call_id": call_id,
+                    "role": role,
+                    "content": text,
+                },
+            )
+            return
+        content = f"**{role}:** {text}"
+        await _datastore.add_message(
+            session_id,
+            {"role": "user" if role == "user" else "assistant", "content": content, "platform": "voice"},
+        )
+        await mgr.broadcast(
+            session_id,
+            {
+                "type": "voice_transcript",
+                "session_id": session_id,
+                "call_id": call_id,
+                "role": role,
+                "content": text,
+            },
+        )
+        return
+
+    if item.get("kind") == "summary":
+        summary = str(item.get("summary") or "(call ended)")
+        status = str(item.get("status") or "ended")
+        content = f"**Call {status}**\n\n{summary}"
+        await _datastore.add_message(session_id, {"role": "assistant", "content": content, "platform": "voice"})
+        await mgr.broadcast(
+            session_id,
+            {
+                "type": "voice_call_ended",
+                "session_id": session_id,
+                "call_id": call_id,
+                "status": status,
+                "summary": summary,
+            },
+        )
+
+
+async def run_chat_bridge_worker() -> None:
+    while True:
+        try:
+            item = await asyncio.to_thread(_bridge_queue.get, True, 1.0)
+        except queue.Empty:
+            continue
+        try:
+            await _process_item(item)
+        except Exception as exc:  # noqa: BLE001
+            logger.debug("Voice chat bridge error: %s", exc)
+
+
+def start_chat_bridge_worker() -> None:
+    global _worker_task
+    with _worker_lock:
+        if _worker_task is not None and not _worker_task.done():
+            return
+        loop = asyncio.get_running_loop()
+        _worker_task = loop.create_task(run_chat_bridge_worker())
@@ -18,6 +18,7 @@ class VoiceSettings:
    public_api_base_url: str | None
    default_assistant_id: str | None
    default_phone_number_id: str | None
+    public_api_key: str | None
    api_base: str

    @property
@@ -61,5 +62,13 @@ def load_voice_settings(*, require_key: bool = True) -> VoiceSettings:
        public_api_base_url=_env("PUBLIC_API_BASE_URL") or None,
        default_assistant_id=_env("VOICE_ASSISTANT_ID") or None,
        default_phone_number_id=_env("VOICE_PHONE_NUMBER_ID") or None,
+        public_api_key=_env("VOICE_PUBLIC_API_KEY") or None,
        api_base=api_base.rstrip("/"),
    )
+
+
+def get_web_sdk_settings() -> tuple[str, str]:
+    """Browser SDK URL + legacy global name (global unused for ESM imports)."""
+    url = _env("VOICE_WEB_SDK_URL") or "https://esm.sh/@vapi-ai/[email protected]"
+    global_name = _env("VOICE_WEB_SDK_GLOBAL") or "Vapi"
+    return url, global_name
@@ -0,0 +1,119 @@
+"""Voice agent stack diagnostics (OpenClaw-style doctor)."""
+
+from __future__ import annotations
+
+import os
+import socket
+from typing import Any
+from urllib.error import HTTPError, URLError
+from urllib.request import Request, urlopen
+
+
+def _tcp_open(host: str, port: int, timeout: float = 2.0) -> bool:
+    try:
+        with socket.create_connection((host, port), timeout=timeout):
+            return True
+    except OSError:
+        return False
+
+
+_MCP_PROBE_TIMEOUT_SEC = 2.0
+
+
+def _is_slow_mcp_probe(detail: str) -> bool:
+    lowered = detail.lower()
+    return any(token in lowered for token in ("timed out", "timeout", "temporarily unavailable"))
+
+
+def _mcp_servers_probe(app_port: int) -> tuple[str, str]:
+    """Probe MCP list endpoint; slow stdio connect must not fail the whole doctor run."""
+    status, detail = _http_status(
+        f"http://127.0.0.1:{app_port}/api/mcp/servers",
+        timeout=_MCP_PROBE_TIMEOUT_SEC,
+    )
+    if status == "fail" and _is_slow_mcp_probe(detail):
+        return "warn", detail
+    return status, detail
+
+
+def _http_status(url: str, timeout: float = 5.0) -> tuple[str, str]:
+    try:
+        req = Request(url, headers={"User-Agent": "praisonai-voice-doctor"})
+        with urlopen(req, timeout=timeout) as resp:
+            code = resp.getcode()
+            if 200 <= code < 500:
+                return "pass", f"HTTP {code}"
+            return "warn", f"HTTP {code}"
+    except HTTPError as exc:
+        if 200 <= exc.code < 500:
+            return "pass", f"HTTP {exc.code}"
+        return "fail", f"HTTP {exc.code}"
+    except URLError as exc:
+        return "fail", str(exc.reason)
+    except Exception as exc:  # noqa: BLE001
+        return "fail", str(exc)
+
+
+def _env_check(name: str, *, required: bool = False) -> dict[str, Any]:
+    value = os.getenv(name, "").strip()
+    if value:
+        masked = value[:4] + "…" if len(value) > 8 else "set"
+        return {"name": name, "status": "pass", "detail": masked}
+    if required:
+        return {"name": name, "status": "fail", "detail": "missing"}
+    return {"name": name, "status": "warn", "detail": "not set"}
+
+
+def run_voice_doctor(*, app_port: int = 8001, sidecar_port: int = 8002) -> dict[str, Any]:
+    """Run local voice-agent health checks."""
+    checks: list[dict[str, Any]] = []
+
+    checks.append(_env_check("OPENAI_API_KEY", required=True))
+    checks.append(_env_check("ELEVENLABS_API_KEY"))
+    checks.append(_env_check("SPEECH_ENGINE_ID"))
+    checks.append(_env_check("SPEECH_ENGINE_PUBLIC_URL"))
+    checks.append(_env_check("PUBLIC_API_BASE_URL"))
+    checks.append(_env_check("VOICE_PRIVATE_API_KEY"))
+
+    if _tcp_open("127.0.0.1", app_port):
+        checks.append({"name": f"App port {app_port}", "status": "pass", "detail": "listening"})
+        status, detail = _http_status(f"http://127.0.0.1:{app_port}/api/voice/config")
+        checks.append({"name": "Voice config API", "status": status, "detail": detail})
+        status, detail = _http_status(f"http://127.0.0.1:{app_port}/api/voice/speech-engine/config")
+        checks.append({"name": "Speech Engine config", "status": status, "detail": detail})
+        status, detail = _http_status(f"http://127.0.0.1:{app_port}/api/voice/realtime/config")
+        checks.append({"name": "GPT Realtime config", "status": status, "detail": detail})
+        status, detail = _mcp_servers_probe(app_port)
+        checks.append({"name": "MCP servers API", "status": status, "detail": detail})
+    else:
+        checks.append(
+            {
+                "name": f"App port {app_port}",
+                "status": "fail",
+                "detail": "not listening — run .\\start_dev.ps1",
+            }
+        )
+
+    if _tcp_open("127.0.0.1", sidecar_port):
+        checks.append({"name": f"Speech sidecar {sidecar_port}", "status": "pass", "detail": "listening"})
+    else:
+        checks.append(
+            {
+                "name": f"Speech sidecar {sidecar_port}",
+                "status": "warn",
+                "detail": "not listening (ElevenLabs path needs sidecar)",
+            }
+        )
+
+    speech_url = os.getenv("SPEECH_ENGINE_PUBLIC_URL", "").strip().rstrip("/")
+    if speech_url:
+        status, detail = _http_status(speech_url)
+        checks.append({"name": "Speech Engine tunnel", "status": status, "detail": detail})
+
+    passed = sum(1 for c in checks if c["status"] == "pass")
+    warnings = sum(1 for c in checks if c["status"] == "warn")
+    failed = sum(1 for c in checks if c["status"] == "fail")
+    return {
+        "checks": checks,
+        "summary": {"passed": passed, "warnings": warnings, "failed": failed},
+    }
@@ -0,0 +1,51 @@
+"""Live transcript pub/sub for voice calls (SSE)."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import queue
+import threading
+from collections import defaultdict
+from typing import Any, AsyncIterator
+
+
+class TranscriptHub:
+    """Thread-safe hub broadcasting transcript events per call_id."""
+
+    def __init__(self) -> None:
+        self._queues: dict[str, list[queue.Queue[dict[str, Any]]]] = defaultdict(list)
+        self._lock = threading.Lock()
+
+    def publish(self, call_id: str, event: dict[str, Any]) -> None:
+        if not call_id:
+            return
+        with self._lock:
+            queues = list(self._queues.get(call_id, []))
+        for sub in queues:
+            try:
+                sub.put_nowait(event)
+            except queue.Full:
+                pass
+
+    async def subscribe(self, call_id: str) -> AsyncIterator[dict[str, Any]]:
+        sub: queue.Queue[dict[str, Any]] = queue.Queue(maxsize=256)
+        with self._lock:
+            self._queues[call_id].append(sub)
+        try:
+            while True:
+                try:
+                    event = await asyncio.to_thread(sub.get, True, 1.0)
+                except queue.Empty:
+                    continue
+                yield event
+        finally:
+            with self._lock:
+                if call_id in self._queues and sub in self._queues[call_id]:
+                    self._queues[call_id].remove(sub)
+
+    def snapshot_sse(self, event: dict[str, Any]) -> str:
+        return f"data: {json.dumps(event, ensure_ascii=False)}\n\n"
+
+
+transcript_hub = TranscriptHub()
@@ -2,13 +2,15 @@

 from __future__ import annotations

+import asyncio
 import json
 import logging
 from typing import Any

+from integrations.voice.agent_runner import build_assistant_request_response, execute_voice_tool
 from integrations.voice.config import VoiceSettings
+from integrations.voice.live import transcript_hub
 from integrations.voice.store import VoiceCallStore, webhook_event_key
-from integrations.voice.tools import execute_tool

 logger = logging.getLogger(__name__)

@@ -35,6 +37,34 @@ def _customer_number(payload: dict[str, Any]) -> str | None:
    return None


+def _tool_parameters(item: dict[str, Any]) -> dict[str, Any]:
+    """Extract tool params from provider payloads (parameters or arguments)."""
+    for key in ("parameters", "arguments"):
+        value = item.get(key)
+        if isinstance(value, dict):
+            return value
+    fn = item.get("function")
+    if isinstance(fn, dict):
+        for key in ("parameters", "arguments"):
+            value = fn.get(key)
+            if isinstance(value, dict):
+                return value
+    return {}
+
+
+def _spoken_tool_result(raw: str) -> str:
+    """Return plain spoken text for the voice pipeline (no newlines)."""
+    try:
+        data = json.loads(raw)
+        if isinstance(data, dict):
+            spoken = data.get("spoken") or data.get("message")
+            if spoken:
+                return str(spoken).replace("\n", " ").strip()
+    except json.JSONDecodeError:
+        pass
+    return str(raw).replace("\n", " ").strip()
+
+
 def process_voice_webhook(
    event_type: str, payload: dict[str, Any], *, settings: VoiceSettings
 ) -> dict[str, Any] | None:
@@ -49,21 +79,36 @@ def process_voice_webhook(
    call_id = _call_id(payload)

    if event_type == "assistant-request":
-        assistant_id = settings.default_assistant_id
-        if not assistant_id:
-            return {"error": "No assistant configured on server"}
-        return {"assistantId": assistant_id}
+        return build_assistant_request_response(settings)

    if event_type == "tool-calls" and call_id:
        tool_calls = msg.get("toolCallList") or []
+        if not tool_calls and isinstance(msg.get("toolWithToolCallList"), list):
+            for entry in msg["toolWithToolCallList"]:
+                if not isinstance(entry, dict):
+                    continue
+                tool_call = entry.get("toolCall")
+                if isinstance(tool_call, dict):
+                    tool_calls.append(
+                        {
+                            "id": tool_call.get("id"),
+                            "name": entry.get("name") or (tool_call.get("function") or {}).get("name"),
+                            "parameters": _tool_parameters(tool_call),
+                            "arguments": _tool_parameters(tool_call),
+                        }
+                    )
        results = []
        for item in tool_calls:
            if not isinstance(item, dict):
                continue
            tool_call_id = str(item.get("id") or "")
            name = str(item.get("name") or "")
-            params = item.get("parameters") if isinstance(item.get("parameters"), dict) else {}
-            result = execute_tool(name, params)
+            if not name:
+                fn = item.get("function")
+                if isinstance(fn, dict):
+                    name = str(fn.get("name") or "")
+            params = _tool_parameters(item)
+            result = _spoken_tool_result(execute_voice_tool(name, params))
            results.append({"toolCallId": tool_call_id, "name": name, "result": result})
        VoiceCallStore.upsert_call(call_id, status="in-progress", customer_number=_customer_number(payload))
        return {"results": results}
@@ -76,8 +121,25 @@ def process_voice_webhook(
    if event_type == "transcript" and call_id:
        role = str(msg.get("role") or "unknown")
        text = str(msg.get("transcript") or "").strip()
-        if text and str(msg.get("transcriptType") or "final") == "final":
-            VoiceCallStore.append_transcript_line(call_id, f"{role}: {text}")
+        kind = str(msg.get("transcriptType") or "final")
+        if text:
+            line = f"{role}: {text}"
+            if kind == "final":
+                VoiceCallStore.append_transcript_line(call_id, line)
+            else:
+                VoiceCallStore.upsert_call(
+                    call_id,
+                    status="in-progress",
+                    customer_number=_customer_number(payload),
+                    metadata={"partial_transcript": line},
+                )
+            transcript_hub.publish(
+                call_id,
+                {"type": "transcript", "role": role, "text": text, "transcriptType": kind},
+            )
+            from integrations.voice.chat_bridge import enqueue_transcript
+
+            enqueue_transcript(call_id, role, text, final=kind == "final")
        return None

    if event_type == "end-of-call-report" and call_id:
@@ -93,6 +155,9 @@ def process_voice_webhook(
            summary=summary or None,
            metadata={"artifact": artifact},
        )
+        from integrations.voice.call_finalize import finalize_call
+
+        asyncio.run(finalize_call(call_id, status=f"ended ({ended})"))
        return None

    if event_type == "conversation-update" and call_id:
@@ -0,0 +1,180 @@
+"""OpenAI Realtime API — browser WebRTC voice (gpt-realtime-2.1-mini)."""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import uuid
+from dataclasses import dataclass
+from typing import Any
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+OPENAI_REALTIME_CALLS_URL = "https://api.openai.com/v1/realtime/calls"
+OPENAI_CLIENT_SECRETS_URL = "https://api.openai.com/v1/realtime/client_secrets"
+
+
+class RealtimeConfigError(Exception):
+    pass
+
+
+@dataclass(frozen=True)
+class RealtimeSettings:
+    api_key: str
+    model: str
+    voice: str
+    reasoning_effort: str
+    instructions: str
+    first_message: str | None
+
+
+def load_realtime_settings() -> RealtimeSettings:
+    api_key = os.getenv("OPENAI_API_KEY", "").strip()
+    if not api_key:
+        raise RealtimeConfigError("Set OPENAI_API_KEY in .env")
+    first_message = os.getenv("REALTIME_FIRST_MESSAGE", "").strip() or None
+    return RealtimeSettings(
+        api_key=api_key,
+        model=os.getenv("REALTIME_MODEL", "gpt-realtime-2.1-mini").strip(),
+        voice=os.getenv("REALTIME_VOICE", "marin").strip(),
+        reasoning_effort=os.getenv("REALTIME_REASONING_EFFORT", "low").strip(),
+        instructions=os.getenv(
+            "REALTIME_INSTRUCTIONS",
+            "You are a helpful voice assistant for PraisonAI demos. "
+            "Keep every reply to one or two short sentences.",
+        ).strip(),
+        first_message=first_message,
+    )
+
+
+def realtime_enabled() -> bool:
+    try:
+        load_realtime_settings()
+    except RealtimeConfigError:
+        return False
+    return True
+
+
+def build_session_config(
+    settings: RealtimeSettings | None = None,
+    *,
+    praison_session_id: str | None = None,
+) -> dict[str, Any]:
+    from integrations.voice.session_memory import instructions_with_memory
+
+    cfg = settings or load_realtime_settings()
+    session: dict[str, Any] = {
+        "type": "realtime",
+        "model": cfg.model,
+        "instructions": instructions_with_memory(cfg.instructions, session_id=praison_session_id),
+        "audio": {"output": {"voice": cfg.voice}},
+        # Required for user transcript events in the browser data channel.
+        "input_audio_transcription": {
+            "model": os.getenv("REALTIME_TRANSCRIBE_MODEL", "gpt-4o-mini-transcribe"),
+        },
+    }
+    if cfg.reasoning_effort:
+        session["reasoning"] = {"effort": cfg.reasoning_effort}
+    return session
+
+
+def new_call_id() -> str:
+    return f"rt-{uuid.uuid4().hex[:12]}"
+
+
+async def create_webrtc_session(sdp_offer: str, *, call_id: str | None = None) -> str:
+    """Relay browser SDP to OpenAI unified Realtime calls endpoint."""
+    from integrations.voice.session_memory import resolve_praison_session_id
+
+    settings = load_realtime_settings()
+    praison_session_id = resolve_praison_session_id(call_id)
+    session_json = json.dumps(build_session_config(settings, praison_session_id=praison_session_id))
+    files = {
+        "sdp": (None, sdp_offer, "application/sdp"),
+        "session": (None, session_json, "application/json"),
+    }
+    async with httpx.AsyncClient(timeout=30.0) as client:
+        response = await client.post(
+            OPENAI_REALTIME_CALLS_URL,
+            headers={"Authorization": f"Bearer {settings.api_key}"},
+            files=files,
+        )
+    if response.status_code >= 400:
+        detail = response.text[:500]
+        logger.error("OpenAI Realtime session failed (%s): %s", response.status_code, detail)
+        raise RealtimeConfigError(f"OpenAI Realtime session failed: HTTP {response.status_code}")
+    return response.text
+
+
+async def mint_client_secret() -> str:
+    """Mint ephemeral key (alternative to unified SDP relay)."""
+    settings = load_realtime_settings()
+    payload = {"session": build_session_config(settings)}
+    async with httpx.AsyncClient(timeout=30.0) as client:
+        response = await client.post(
+            OPENAI_CLIENT_SECRETS_URL,
+            headers={
+                "Authorization": f"Bearer {settings.api_key}",
+                "Content-Type": "application/json",
+            },
+            json=payload,
+        )
+    if response.status_code >= 400:
+        detail = response.text[:500]
+        raise RealtimeConfigError(f"OpenAI client secret failed: HTTP {response.status_code} — {detail}")
+    data = response.json()
+    token = data.get("value") or data.get("client_secret", {}).get("value")
+    if not token:
+        raise RealtimeConfigError("OpenAI client secret response missing value")
+    return str(token)
+
+
+def persist_transcript_line(call_id: str, role: str, text: str) -> None:
+    from integrations.voice.live import transcript_hub
+    from integrations.voice.store import VoiceCallStore
+
+    text = text.strip()
+    if not call_id or not text:
+        return
+    record = VoiceCallStore.get_call(call_id) or {}
+    prefix = f"{role}: {text}"
+    existing = (record.get("transcript") or "").strip()
+    merged = f"{existing}\n{prefix}".strip() if existing else prefix
+    VoiceCallStore.upsert_call(
+        call_id,
+        status=record.get("status") or "in-progress",
+        customer_number="web-openai-realtime",
+        transcript=merged,
+    )
+    transcript_hub.publish(
+        call_id,
+        {
+            "type": "transcript",
+            "role": role,
+            "transcriptType": "final",
+            "text": text,
+        },
+    )
+
+
+def mark_session_started(call_id: str) -> None:
+    from integrations.voice.store import VoiceCallStore
+
+    VoiceCallStore.upsert_call(
+        call_id,
+        status="in-progress",
+        customer_number="web-openai-realtime",
+    )
+
+
+def mark_session_ended(call_id: str) -> None:
+    from integrations.voice.store import VoiceCallStore
+
+    VoiceCallStore.upsert_call(
+        call_id,
+        status="ended (customer-ended-call)",
+        customer_number="web-openai-realtime",
+    )
@@ -0,0 +1,251 @@
+"""Cross-call voice memory via PraisonAI session reload by stable user ID."""
+
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+_CALL_USER: dict[str, str] = {}
+
+
+def memory_enabled() -> bool:
+    return os.getenv("VOICE_MEMORY_ENABLED", "true").lower() in ("1", "true", "yes")
+
+
+def voice_user_session_id(user_id: str) -> str:
+    clean = str(user_id or "").strip()
+    if not clean:
+        raise ValueError("user_id required")
+    return f"voice-user-{clean}"
+
+
+def attach_pending_user(user_id: str) -> None:
+    from integrations.voice.store import VoiceCallStore
+
+    clean = str(user_id or "").strip()
+    if clean:
+        VoiceCallStore.set_pending_user(clean)
+
+
+def consume_pending_user() -> str | None:
+    from integrations.voice.store import VoiceCallStore
+
+    return VoiceCallStore.consume_pending_user()
+
+
+def bind_call_user(call_id: str, user_id: str) -> str:
+    """Associate a provider call/conversation id with a stable browser user id."""
+    from integrations.voice.store import VoiceCallStore
+
+    call_id = str(call_id or "").strip()
+    user_id = str(user_id or "").strip()
+    if not call_id or not user_id:
+        raise ValueError("call_id and user_id required")
+    session_id = voice_user_session_id(user_id)
+    _CALL_USER[call_id] = user_id
+    VoiceCallStore.upsert_call(
+        call_id,
+        metadata={"user_id": user_id, "praison_session_id": session_id},
+    )
+    return session_id
+
+
+def resolve_call_user(call_id: str | None) -> str | None:
+    if not call_id:
+        return None
+    if call_id in _CALL_USER:
+        return _CALL_USER[call_id]
+    from integrations.voice.store import VoiceCallStore
+
+    record = VoiceCallStore.get_call(call_id)
+    if not record:
+        return None
+    user_id = (record.get("metadata") or {}).get("user_id")
+    if user_id:
+        _CALL_USER[call_id] = str(user_id)
+    return str(user_id) if user_id else None
+
+
+def resolve_praison_session_id(call_id: str | None) -> str | None:
+    user_id = resolve_call_user(call_id)
+    if not user_id:
+        return None
+    return voice_user_session_id(user_id)
+
+
+def _history_limit() -> int:
+    try:
+        return max(1, int(os.getenv("VOICE_MEMORY_HISTORY_LIMIT", "20")))
+    except ValueError:
+        return 20
+
+
+def load_session_history(session_id: str) -> list[dict[str, str]]:
+    """Load prior turns from PraisonAI DefaultSessionStore."""
+    try:
+        from praisonaiagents.session.store import DefaultSessionStore
+
+        store = DefaultSessionStore()
+        return store.get_chat_history(session_id, max_messages=_history_limit())
+    except Exception:  # noqa: BLE001
+        logger.debug("Failed to load session history for %s", session_id, exc_info=True)
+        return []
+
+
+def append_session_turn(session_id: str, role: str, content: str) -> None:
+    """Persist one voice turn into the PraisonAI session folder."""
+    text = str(content or "").strip()
+    if not session_id or not text:
+        return
+    try:
+        from praisonaiagents.session.store import DefaultSessionStore
+
+        store = DefaultSessionStore()
+        if role == "user":
+            store.add_user_message(session_id, text, metadata={"platform": "voice"})
+        else:
+            store.add_assistant_message(session_id, text, metadata={"platform": "voice"})
+    except Exception:  # noqa: BLE001
+        logger.debug("Failed to append session turn for %s", session_id, exc_info=True)
+
+
+async def ensure_ui_session(session_id: str, *, user_id: str | None = None, call_id: str | None = None) -> None:
+    """Ensure PraisonAIUI datastore session exists for sidebar reload."""
+    try:
+        from praisonaiui.server import _datastore
+    except Exception:  # noqa: BLE001
+        return
+    existing = await _datastore.get_session(session_id)
+    title = "Voice memory"
+    if call_id:
+        title = f"Voice user {user_id or call_id[:8]}"
+    if existing is None:
+        await _datastore.create_session(session_id)
+        await _datastore.update_session(
+            session_id,
+            platform="voice",
+            icon="🎙️",
+            title=title,
+        )
+
+
+async def append_ui_message(session_id: str, role: str, content: str) -> None:
+    try:
+        from praisonaiui.server import _datastore
+
+        await _datastore.add_message(
+            session_id,
+            {"role": "user" if role == "user" else "assistant", "content": content, "platform": "voice"},
+        )
+    except Exception:  # noqa: BLE001
+        logger.debug("Failed to append UI message for %s", session_id, exc_info=True)
+
+
+def _apply_knowledge_augmentation(prompt: str) -> str:
+    from praisonaiui.features.knowledge import get_knowledge_manager
+
+    k_mgr = get_knowledge_manager()
+    if not k_mgr.list_all():
+        return prompt
+    results = k_mgr.search(prompt, limit=5)
+    lines = [r.get("text", "") for r in results if r.get("text")]
+    if not lines:
+        return prompt
+    block = "\n".join(lines)
+    return f"[Knowledge Context]\n{block}\n[/Knowledge Context]\n\n{prompt}"
+
+
+def build_knowledge_augmented_prompt(prompt: str, *, timeout_s: float | None = 2.0) -> str:
+    """Inject knowledge search results like dashboard chat run_agent().
+
+    Knowledge indexing can take tens of seconds on cold start; Speech Engine
+    disconnects if the LLM path blocks too long, so callers use a short timeout.
+    """
+    if not os.getenv("VOICE_MEMORY_USE_KNOWLEDGE", "true").lower() in ("1", "true", "yes"):
+        return prompt
+    try:
+        if timeout_s is None:
+            return _apply_knowledge_augmentation(prompt)
+        from concurrent.futures import ThreadPoolExecutor, TimeoutError as FuturesTimeoutError
+
+        with ThreadPoolExecutor(max_workers=1) as pool:
+            future = pool.submit(_apply_knowledge_augmentation, prompt)
+            return future.result(timeout=timeout_s)
+    except FuturesTimeoutError:
+        logger.warning("Knowledge augmentation timed out after %.1fs; using plain prompt", timeout_s)
+        return prompt
+    except Exception:  # noqa: BLE001
+        logger.debug("Knowledge augmentation skipped", exc_info=True)
+        return prompt
+
+
+def history_context_block(session_id: str | None) -> str:
+    if not memory_enabled() or not session_id:
+        return ""
+    history = load_session_history(session_id)
+    if not history:
+        return ""
+    lines = []
+    for item in history[-_history_limit() :]:
+        role = item.get("role", "user")
+        content = str(item.get("content") or "").strip()
+        if content:
+            lines.append(f"{role}: {content}")
+    if not lines:
+        return ""
+    return "Previous conversation with this user:\n" + "\n".join(lines)
+
+
+def openai_messages_with_memory(
+    transcript_lines: list[str],
+    *,
+    session_id: str | None,
+    base_instructions: str,
+) -> list[dict[str, str]]:
+    messages: list[dict[str, str]] = [{"role": "system", "content": base_instructions}]
+    prior = history_context_block(session_id)
+    if prior:
+        messages.append({"role": "system", "content": prior})
+    for line in transcript_lines:
+        if line.startswith("user:"):
+            messages.append({"role": "user", "content": line[5:].strip()})
+        elif line.startswith("agent:"):
+            messages.append({"role": "assistant", "content": line[6:].strip()})
+    return messages
+
+
+def bind_call_from_pending(call_id: str) -> str | None:
+    user_id = consume_pending_user()
+    if not user_id:
+        return None
+    return bind_call_user(call_id, user_id)
+
+
+async def record_voice_turn(
+    call_id: str | None,
+    *,
+    user_text: str | None = None,
+    agent_text: str | None = None,
+) -> None:
+    """Save latest user/agent lines to Praison session + chat UI."""
+    session_id = resolve_praison_session_id(call_id)
+    if not session_id:
+        return
+    user_id = resolve_call_user(call_id)
+    await ensure_ui_session(session_id, user_id=user_id, call_id=call_id)
+    if user_text:
+        append_session_turn(session_id, "user", user_text)
+        await append_ui_message(session_id, "user", user_text)
+    if agent_text:
+        append_session_turn(session_id, "assistant", agent_text)
+        await append_ui_message(session_id, "assistant", agent_text)
+
+
+def instructions_with_memory(base: str, *, session_id: str | None) -> str:
+    prior = history_context_block(session_id)
+    if not prior:
+        return base
+    return f"{base}\n\n{prior}\nUse prior conversation when the user refers to earlier calls."
@@ -0,0 +1,346 @@
+"""ElevenLabs Speech Engine — WebSocket LLM bridge for browser voice."""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import os
+from dataclasses import dataclass
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+_engine_resource: Any | None = None
+_engine_lock = asyncio.Lock()
+
+
+class SpeechEngineConfigError(Exception):
+    pass
+
+
+@dataclass(frozen=True)
+class SpeechEngineSettings:
+    api_key: str
+    engine_id: str | None
+    public_ws_url: str | None
+    first_message: str
+
+
+def _public_ws_url() -> str | None:
+    speech_base = os.getenv("SPEECH_ENGINE_PUBLIC_URL", "").strip().rstrip("/")
+    base = speech_base or os.getenv("PUBLIC_API_BASE_URL", "").strip().rstrip("/")
+    if not base:
+        return None
+    if base.startswith("https://"):
+        return "wss://" + base[len("https://") :] + "/ws"
+    if base.startswith("http://"):
+        return "ws://" + base[len("http://") :] + "/ws"
+    return None
+
+
+def load_speech_engine_settings(*, require_engine_id: bool = False) -> SpeechEngineSettings:
+    api_key = os.getenv("ELEVENLABS_API_KEY", "").strip()
+    if not api_key:
+        raise SpeechEngineConfigError("Set ELEVENLABS_API_KEY in .env")
+    engine_id = os.getenv("SPEECH_ENGINE_ID", "").strip() or None
+    if require_engine_id and not engine_id:
+        raise SpeechEngineConfigError("Set SPEECH_ENGINE_ID in .env (run setup_speech_engine.py)")
+    return SpeechEngineSettings(
+        api_key=api_key,
+        engine_id=engine_id,
+        public_ws_url=_public_ws_url(),
+        first_message=os.getenv(
+            "SPEECH_ENGINE_FIRST_MESSAGE", "Hello! How can I help you today?"
+        ).strip(),
+    )
+
+
+def speech_engine_enabled() -> bool:
+    try:
+        settings = load_speech_engine_settings()
+    except SpeechEngineConfigError:
+        return False
+    return bool(settings.engine_id and settings.public_ws_url)
+
+
+async def get_engine_resource():
+    """Lazy-load the ElevenLabs SpeechEngineResource."""
+    global _engine_resource
+    if _engine_resource is not None:
+        return _engine_resource
+
+    async with _engine_lock:
+        if _engine_resource is not None:
+            return _engine_resource
+        settings = load_speech_engine_settings(require_engine_id=True)
+        try:
+            from elevenlabs import AsyncElevenLabs
+        except ImportError as exc:
+            raise SpeechEngineConfigError(
+                "Install Speech Engine SDK: pip install 'elevenlabs>=2.47.0'"
+            ) from exc
+
+        if not hasattr(AsyncElevenLabs(api_key=settings.api_key), "speech_engine"):
+            raise SpeechEngineConfigError(
+                "elevenlabs package is too old. Run: pip install -U 'elevenlabs>=2.47.0'"
+            )
+
+        client = AsyncElevenLabs(api_key=settings.api_key)
+        _engine_resource = await client.speech_engine.get(settings.engine_id)
+        return _engine_resource
+
+
+def _conversation_id(session: Any) -> str | None:
+    cid = getattr(session, "conversation_id", None)
+    return str(cid) if cid else None
+
+
+def _transcript_display_text(transcript: list[Any], agent_reply: str | None = None) -> str:
+    text = _transcript_to_prompt(transcript)
+    if agent_reply:
+        agent_line = f"agent: {agent_reply.strip()}"
+        text = f"{text}\n{agent_line}".strip() if text else agent_line
+    return text
+
+
+def _persist_conversation_turn(call_id: str, transcript: list[Any], agent_reply: str) -> None:
+    from integrations.voice.live import transcript_hub
+    from integrations.voice.store import VoiceCallStore
+
+    display = _transcript_display_text(transcript, agent_reply)
+    VoiceCallStore.upsert_call(
+        call_id,
+        status="in-progress",
+        customer_number="web-speech-engine",
+        transcript=display,
+    )
+    prompt = _transcript_to_prompt(transcript)
+    if prompt:
+        last_line = prompt.splitlines()[-1]
+        role, _, content = last_line.partition(": ")
+        transcript_hub.publish(
+            call_id,
+            {
+                "type": "transcript",
+                "role": role or "user",
+                "transcriptType": "final",
+                "text": content,
+            },
+        )
+    transcript_hub.publish(
+        call_id,
+        {
+            "type": "transcript",
+            "role": "assistant",
+            "transcriptType": "final",
+            "text": agent_reply,
+        },
+    )
+
+
+def _transcript_to_prompt(transcript: list[Any]) -> str:
+    lines: list[str] = []
+    for message in transcript:
+        role = getattr(message, "role", None) or (message.get("role") if isinstance(message, dict) else "user")
+        content = getattr(message, "content", None) or (
+            message.get("content") if isinstance(message, dict) else ""
+        )
+        text = str(content or "").strip()
+        if text:
+            lines.append(f"{role}: {text}")
+    if not lines:
+        return ""
+    return "\n".join(lines)
+
+
+def _try_demo_tool_reply(transcript: list[Any]) -> str | None:
+    from integrations.voice.tools import demo_tool_reply
+
+    user_text = _latest_user_text(transcript)
+    if not user_text:
+        return None
+    return demo_tool_reply(user_text)
+
+
+async def _run_openai_reply(transcript: list[Any], *, session_id: str | None = None) -> str:
+    """Voice LLM path with demo fast-path + OpenAI tool calling."""
+    import json
+
+    from openai import AsyncOpenAI
+
+    from integrations.voice.agent_runner import execute_voice_tool
+    from integrations.voice.session_memory import openai_messages_with_memory
+    from integrations.voice.tools import openai_tool_schemas, tool_result_for_voice
+
+    prompt = _transcript_to_prompt(transcript)
+    if not prompt:
+        return "I did not catch that. Could you repeat?"
+
+    fast = _try_demo_tool_reply(transcript)
+    if fast:
+        return fast
+
+    client = AsyncOpenAI(api_key=os.getenv("OPENAI_API_KEY", "").strip())
+    model = os.getenv("VOICE_AGENT_MODEL", "gpt-4o-mini")
+    base_instructions = (
+        "You are a helpful voice assistant for PraisonAI demos. "
+        "Keep answers concise (1-3 sentences) and conversational. "
+        "When the caller asks for the time, call get_current_time. "
+        "When they ask you to repeat or echo text, call echo_message."
+    )
+    messages = openai_messages_with_memory(
+        prompt.splitlines(),
+        session_id=session_id,
+        base_instructions=base_instructions,
+    )
+    use_tools = os.getenv("VOICE_SPEECH_USE_TOOLS", "true").lower() in ("1", "true", "yes")
+    create_kwargs: dict[str, Any] = {
+        "model": model,
+        "messages": messages,
+        "max_tokens": 200,
+    }
+    if use_tools:
+        create_kwargs["tools"] = openai_tool_schemas()
+        create_kwargs["tool_choice"] = "auto"
+
+    response = await client.chat.completions.create(**create_kwargs)
+    choice = response.choices[0].message
+
+    if use_tools and choice.tool_calls:
+        messages.append(choice.model_dump(exclude_none=True))
+        fallback = "Done."
+        for tool_call in choice.tool_calls:
+            fn = tool_call.function
+            name = fn.name if fn else ""
+            try:
+                args = json.loads(fn.arguments or "{}") if fn else {}
+            except json.JSONDecodeError:
+                args = {}
+            raw = execute_voice_tool(name, args, session_id=session_id)
+            fallback = tool_result_for_voice(raw)
+            messages.append(
+                {
+                    "role": "tool",
+                    "tool_call_id": tool_call.id,
+                    "content": raw,
+                }
+            )
+        final = await client.chat.completions.create(
+            model=model,
+            messages=messages,
+            max_tokens=200,
+        )
+        text = (final.choices[0].message.content or "").strip()
+        return text or fallback
+
+    text = (choice.content or "").strip()
+    return text or "Sorry, I could not generate a response."
+
+
+async def _run_agent_reply(transcript: list[Any], *, session_id: str | None = None) -> str:
+    prompt = _transcript_to_prompt(transcript)
+    if not prompt:
+        return "I did not catch that. Could you repeat?"
+
+    # Prefer fast OpenAI streaming for Speech Engine turn latency.
+    if os.getenv("OPENAI_API_KEY", "").strip():
+        try:
+            return await asyncio.wait_for(
+                _run_openai_reply(transcript, session_id=session_id),
+                timeout=20.0,
+            )
+        except TimeoutError:
+            return "Sorry, that took too long. Could you ask again with a shorter question?"
+        except Exception:  # noqa: BLE001
+            logger.exception("OpenAI Speech Engine reply failed; falling back to agent")
+
+    from integrations.voice.agent_runner import get_voice_agent
+
+    agent = get_voice_agent(session_id)
+    reply = await asyncio.wait_for(asyncio.to_thread(agent.chat, prompt), timeout=25.0)
+    return str(reply).strip() or "Sorry, I could not generate a response."
+
+
+def _latest_user_text(transcript: list[Any]) -> str:
+    prompt = _transcript_to_prompt(transcript)
+    if not prompt:
+        return ""
+    for line in reversed(prompt.splitlines()):
+        if line.startswith("user:"):
+            return line[5:].strip()
+    return ""
+
+
+async def handle_user_transcript(transcript: list[Any], session: Any) -> None:
+    """Run voice LLM on transcript and send text back to Speech Engine."""
+    from integrations.voice.session_memory import record_voice_turn, resolve_praison_session_id
+
+    call_id = _conversation_id(session)
+    praison_session_id = resolve_praison_session_id(call_id)
+    preview = _transcript_to_prompt(transcript)
+    print(f"[speech sidecar] transcript: {preview[:200]!r}", flush=True)
+    reply = ""
+    try:
+        reply = await _run_agent_reply(transcript, session_id=praison_session_id)
+        print(f"[speech sidecar] reply: {reply[:200]!r}", flush=True)
+        await session.send_response(reply)
+    except TimeoutError:
+        reply = "Sorry, that took too long. Please try a shorter question."
+        await session.send_response(reply)
+    except Exception:  # noqa: BLE001
+        logger.exception("Speech Engine agent reply failed")
+        reply = "Sorry, something went wrong on my side."
+        await session.send_response(reply)
+    if call_id and reply:
+        _persist_conversation_turn(call_id, transcript, reply)
+        user_text = _latest_user_text(transcript)
+        await record_voice_turn(call_id, user_text=user_text or None, agent_text=reply)
+
+
+async def handle_session_init(conversation_id: str, session: Any) -> None:
+    from integrations.voice.session_memory import bind_call_from_pending, ensure_ui_session, resolve_praison_session_id
+    from integrations.voice.store import VoiceCallStore
+
+    call_id = str(conversation_id)
+    bind_call_from_pending(call_id)
+    VoiceCallStore.upsert_call(
+        call_id,
+        status="in-progress",
+        customer_number="web-speech-engine",
+    )
+    session_id = resolve_praison_session_id(call_id)
+    if session_id:
+        await ensure_ui_session(session_id, call_id=call_id)
+    logger.info("Speech Engine session started: %s", conversation_id)
+
+
+async def handle_session_close(session: Any) -> None:
+    from integrations.voice.call_finalize import finalize_call
+
+    conversation_id = _conversation_id(session)
+    if conversation_id:
+        await finalize_call(conversation_id, status="ended (customer-ended-call)")
+        logger.info("Speech Engine session ended: %s", conversation_id)
+
+
+async def handle_session_disconnect(session: Any) -> None:
+    """Fallback when WebSocket drops without a formal close message."""
+    await handle_session_close(session)
+
+
+def get_webrtc_token() -> str:
+    """Mint a browser conversation token (keeps API key server-side)."""
+    settings = load_speech_engine_settings(require_engine_id=True)
+    try:
+        from elevenlabs import ElevenLabs
+    except ImportError as exc:
+        raise SpeechEngineConfigError("Install elevenlabs: pip install elevenlabs") from exc
+
+    client = ElevenLabs(api_key=settings.api_key)
+    response = client.conversational_ai.conversations.get_webrtc_token(
+        agent_id=settings.engine_id,
+    )
+    token = getattr(response, "token", None) or (response.get("token") if isinstance(response, dict) else None)
+    if not token:
+        raise SpeechEngineConfigError("Could not mint Speech Engine conversation token")
+    return str(token)
@@ -43,6 +43,15 @@ def _connect() -> sqlite3.Connection:
        )
        """
    )
+    conn.execute(
+        """
+        CREATE TABLE IF NOT EXISTS voice_memory_pending (
+            key TEXT PRIMARY KEY,
+            user_id TEXT NOT NULL,
+            updated_at TEXT NOT NULL
+        )
+        """
+    )
    return conn


@@ -146,14 +155,25 @@ class VoiceCallStore:

    @staticmethod
    def append_transcript_line(call_id: str, line: str) -> str:
-        with closing(_connect()) as conn:
+        now = datetime.now(timezone.utc).isoformat()
+        with closing(_connect()) as conn, conn:
            row = conn.execute(
                "SELECT transcript FROM voice_calls WHERE call_id = ?",
                (call_id,),
            ).fetchone()
-        existing = row[0] if row else ""
-        full = f"{existing}\n{line}".strip() if existing else line
-        VoiceCallStore.upsert_call(call_id, transcript=full)
+            existing = row[0] if row else ""
+            full = f"{existing}\n{line}".strip() if existing else line
+            conn.execute(
+                """
+                INSERT INTO voice_calls (
+                    call_id, status, transcript, summary, metadata_json, created_at, updated_at
+                ) VALUES (?, 'unknown', ?, '', '{}', ?, ?)
+                ON CONFLICT(call_id) DO UPDATE SET
+                    transcript = excluded.transcript,
+                    updated_at = excluded.updated_at
+                """,
+                (call_id, full, now, now),
+            )
        return full

    @staticmethod
@@ -161,7 +181,8 @@ class VoiceCallStore:
        with closing(_connect()) as conn:
            rows = conn.execute(
                """
-                SELECT call_id, status, customer_number, transcript, summary, metadata_json, updated_at
+                SELECT call_id, status, customer_number, transcript, summary, metadata_json,
+                       created_at, updated_at
                FROM voice_calls ORDER BY updated_at DESC LIMIT ?
                """,
                (max(1, limit),),
@@ -176,17 +197,43 @@ class VoiceCallStore:
                    "transcript": row[3],
                    "summary": row[4],
                    "metadata": json.loads(row[5] or "{}"),
-                    "updated_at": row[6],
+                    "created_at": row[6],
+                    "updated_at": row[7],
                }
            )
        return out

+    @staticmethod
+    def set_pending_user(user_id: str) -> None:
+        now = datetime.now(timezone.utc).isoformat()
+        with closing(_connect()) as conn, conn:
+            conn.execute(
+                """
+                INSERT INTO voice_memory_pending (key, user_id, updated_at)
+                VALUES ('browser', ?, ?)
+                ON CONFLICT(key) DO UPDATE SET user_id = excluded.user_id, updated_at = excluded.updated_at
+                """,
+                (user_id, now),
+            )
+
+    @staticmethod
+    def consume_pending_user() -> str | None:
+        with closing(_connect()) as conn, conn:
+            row = conn.execute(
+                "SELECT user_id FROM voice_memory_pending WHERE key = 'browser'",
+            ).fetchone()
+            if not row:
+                return None
+            conn.execute("DELETE FROM voice_memory_pending WHERE key = 'browser'")
+            return str(row[0])
+
    @staticmethod
    def get_call(call_id: str) -> dict[str, Any] | None:
        with closing(_connect()) as conn:
            row = conn.execute(
                """
-                SELECT call_id, status, customer_number, transcript, summary, metadata_json, updated_at
+                SELECT call_id, status, customer_number, transcript, summary, metadata_json,
+                       created_at, updated_at
                FROM voice_calls WHERE call_id = ?
                """,
                (call_id,),
@@ -200,5 +247,6 @@ class VoiceCallStore:
            "transcript": row[3],
            "summary": row[4],
            "metadata": json.loads(row[5] or "{}"),
-            "updated_at": row[6],
+            "created_at": row[6],
+            "updated_at": row[7],
        }
@@ -19,6 +19,65 @@ def echo_message(message: str = "") -> dict[str, Any]:
    return {"message": text, "spoken": f"You said {text}"}


+def demo_tool_reply(text: str) -> str | None:
+    """Fast path for common voice tool queries — no LLM round-trip."""
+    lower = text.lower().strip()
+    if any(p in lower for p in ("what time", "current time", "tell me the time", "time is it")):
+        return str(get_current_time()["spoken"])
+    if lower.startswith("echo "):
+        return str(echo_message(text[5:])["spoken"])
+    if lower.startswith("repeat "):
+        return str(echo_message(text[7:])["spoken"])
+    return None
+
+
+def openai_tool_schemas() -> list[dict[str, Any]]:
+    """OpenAI function schemas for Speech Engine / voice LLM paths."""
+    return [
+        {
+            "type": "function",
+            "function": {
+                "name": "get_current_time",
+                "description": "Return the current UTC time when the caller asks for the time.",
+                "parameters": {"type": "object", "properties": {}, "required": []},
+            },
+        },
+        {
+            "type": "function",
+            "function": {
+                "name": "echo_message",
+                "description": "Repeat or echo back a phrase the caller wants repeated.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "message": {
+                            "type": "string",
+                            "description": "The text to echo back to the caller",
+                        }
+                    },
+                    "required": ["message"],
+                },
+            },
+        },
+    ]
+
+
+def tool_result_for_voice(raw_result: str) -> str:
+    """Convert tool JSON output to speakable voice text."""
+    try:
+        data = json.loads(raw_result)
+    except json.JSONDecodeError:
+        return raw_result.strip() or "Done."
+    if isinstance(data, dict):
+        spoken = data.get("spoken")
+        if spoken:
+            return str(spoken)
+        err = data.get("error")
+        if err:
+            return f"Sorry, {err}"
+    return raw_result.strip() or "Done."
+
+
 _TOOL_REGISTRY: dict[str, Callable[..., Any]] = {
    "get_current_time": get_current_time,
    "echo_message": echo_message,
@@ -0,0 +1,76 @@
+"""HTTP routes for stable voice user sessions."""
+
+from __future__ import annotations
+
+from starlette.requests import Request
+from starlette.responses import JSONResponse
+
+from integrations.voice.session_memory import (
+    attach_pending_user,
+    bind_call_user,
+    memory_enabled,
+    resolve_praison_session_id,
+    voice_user_session_id,
+)
+
+
+async def api_voice_memory_config(_request: Request) -> JSONResponse:
+    return JSONResponse({"enabled": memory_enabled()})
+
+
+async def api_voice_memory_bind(request: Request) -> JSONResponse:
+    try:
+        payload = await request.json()
+    except Exception:  # noqa: BLE001
+        return JSONResponse({"error": "Invalid JSON"}, status_code=400)
+    call_id = str(payload.get("call_id") or "").strip()
+    user_id = str(payload.get("user_id") or "").strip()
+    if not call_id or not user_id:
+        return JSONResponse({"error": "call_id and user_id required"}, status_code=400)
+    try:
+        session_id = bind_call_user(call_id, user_id)
+    except ValueError as exc:
+        return JSONResponse({"error": str(exc)}, status_code=400)
+    return JSONResponse(
+        {
+            "ok": True,
+            "call_id": call_id,
+            "user_id": user_id,
+            "praison_session_id": session_id,
+        }
+    )
+
+
+async def api_voice_memory_resolve(request: Request) -> JSONResponse:
+    user_id = request.query_params.get("user_id", "").strip()
+    if not user_id:
+        return JSONResponse({"error": "user_id required"}, status_code=400)
+    session_id = voice_user_session_id(user_id)
+    return JSONResponse({"user_id": user_id, "praison_session_id": session_id})
+
+
+async def api_voice_memory_attach_user(request: Request) -> JSONResponse:
+    """ElevenLabs: browser attaches user before conversation id exists."""
+    try:
+        payload = await request.json()
+    except Exception:  # noqa: BLE001
+        return JSONResponse({"error": "Invalid JSON"}, status_code=400)
+    user_id = str(payload.get("user_id") or "").strip()
+    if not user_id:
+        return JSONResponse({"error": "user_id required"}, status_code=400)
+    attach_pending_user(user_id)
+    return JSONResponse(
+        {
+            "ok": True,
+            "user_id": user_id,
+            "praison_session_id": voice_user_session_id(user_id),
+        }
+    )
+
+
+async def api_voice_memory_for_call(request: Request) -> JSONResponse:
+    call_id = request.query_params.get("call_id", "").strip()
+    if not call_id:
+        return JSONResponse({"error": "call_id required"}, status_code=400)
+    session_id = resolve_praison_session_id(call_id)
+    return JSONResponse({"call_id": call_id, "praison_session_id": session_id})
@@ -0,0 +1,781 @@
+/**
+ * Voice Agent Phase 2 — live transcript + browser talk widget.
+ *
+ * custom.js can run before dashboard.js defines window.aiui, so we wait.
+ */
+(function () {
+  function whenReady(fn) {
+    if (window.aiui && typeof window.aiui.registerView === 'function') {
+      fn();
+      return;
+    }
+    var n = 0;
+    var t = setInterval(function () {
+      n += 1;
+      if (window.aiui && typeof window.aiui.registerView === 'function') {
+        clearInterval(t);
+        fn();
+      } else if (n > 150) {
+        clearInterval(t);
+        console.warn('[voice plugin] window.aiui never became ready');
+      }
+    }, 100);
+  }
+
+  function el(tag, style, text) {
+    var node = document.createElement(tag);
+    if (style) node.style.cssText = style;
+    if (text != null) node.textContent = text;
+    return node;
+  }
+
+  function currentCallId() {
+    return new URLSearchParams(window.location.search).get('call_id') || '';
+  }
+
+  function voiceUserId() {
+    var key = 'voice-user-id';
+    var id = localStorage.getItem(key);
+    if (!id) {
+      id = 'u-' + Math.random().toString(36).slice(2, 12);
+      localStorage.setItem(key, id);
+    }
+    return id;
+  }
+
+  async function postJson(url, body) {
+    var resp = await fetch(url, {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify(body || {}),
+    });
+    if (!resp.ok) throw new Error('HTTP ' + resp.status);
+    return resp.json();
+  }
+
+  async function attachVoiceUser() {
+    return postJson('/api/voice/memory/attach-user', { user_id: voiceUserId() });
+  }
+
+  function renderUiComponent(comp) {
+    if (window.aiui && typeof window.aiui.renderComponent === 'function') {
+      return window.aiui.renderComponent(comp);
+    }
+    var wrap = el('div');
+    if (comp.type === 'code_block') {
+      var pre = el(
+        'pre',
+        'margin:0;padding:16px;border-radius:10px;background:var(--db-card-bg,#1f2937);' +
+          'border:1px solid var(--db-border,#334155);white-space:pre-wrap;font-family:ui-monospace,monospace;font-size:13px'
+      );
+      pre.textContent = comp.code || '';
+      wrap.appendChild(pre);
+      wrap.className = 'db-code-block';
+      return wrap;
+    }
+    if (comp.type === 'text') {
+      wrap.textContent = comp.content || '';
+      return wrap;
+    }
+    wrap.textContent = JSON.stringify(comp);
+    return wrap;
+  }
+
+  function renderUiTree(data, container) {
+    if (!data) return;
+    if (data._components) {
+      data._components.forEach(function (c) {
+        container.appendChild(renderUiComponent(c));
+      });
+      return;
+    }
+    if (data.type === 'layout' && data.children) {
+      data.children.forEach(function (c) {
+        container.appendChild(renderUiComponent(c));
+      });
+    }
+  }
+
+  function findTranscriptCodeEl(root) {
+    var blocks = root.querySelectorAll('.db-code-block code');
+    if (!blocks.length) {
+      blocks = root.querySelectorAll('.db-code-block pre');
+    }
+    return blocks.length ? blocks[blocks.length - 1] : null;
+  }
+
+  async function fetchJson(url, retries) {
+    var attempts = retries == null ? 1 : retries;
+    var lastErr = null;
+    for (var i = 0; i < attempts; i += 1) {
+      try {
+        var resp = await fetch(url);
+        if (!resp.ok) throw new Error('HTTP ' + resp.status);
+        return resp.json();
+      } catch (err) {
+        lastErr = err;
+        if (i < attempts - 1) {
+          await new Promise(function (r) {
+            setTimeout(r, 1500);
+          });
+        }
+      }
+    }
+    throw lastErr || new Error('fetch failed');
+  }
+
+  async function resolveCallId() {
+    var fromQuery = currentCallId();
+    if (fromQuery) return fromQuery;
+    var data = await fetchJson('/api/voice/calls');
+    var calls = data.calls || [];
+    var live = calls.find(function (c) {
+      return String(c.status || '').toLowerCase().indexOf('progress') >= 0;
+    });
+    return (live && live.call_id) || (calls[0] && calls[0].call_id) || '';
+  }
+
+  async function loadWebSdk(cfg) {
+    var candidates = [];
+    var configured = cfg.sdk_url || '';
+    if (configured && configured.indexOf('dist/vapi.js') < 0) {
+      candidates.push(configured);
+    }
+    candidates.push('https://esm.sh/@vapi-ai/[email protected]');
+
+    var lastErr = null;
+    for (var i = 0; i < candidates.length; i++) {
+      try {
+        var mod = await import(/* webpackIgnore: true */ candidates[i]);
+        var Sdk = mod.default || mod.Vapi;
+        if (typeof Sdk === 'function') return Sdk;
+        lastErr = new Error('Module loaded but no Vapi class');
+      } catch (err) {
+        lastErr = err;
+        console.warn('[voice plugin] SDK load failed for', candidates[i], err);
+      }
+    }
+    throw new Error(
+      'Could not load voice web SDK' + (lastErr && lastErr.message ? ': ' + lastErr.message : '')
+    );
+  }
+
+  function registerVoiceViews() {
+    window.aiui.registerView(
+      'call-detail',
+      async function (container) {
+        container.innerHTML = '';
+        var wrap = el(
+          'div',
+          'padding:20px;display:flex;flex-direction:column;gap:16px;color:var(--db-text,#eee)'
+        );
+        var uiHost = el('div', 'display:flex;flex-direction:column;gap:12px');
+        var statusLine = el(
+          'div',
+          'font-size:12px;color:var(--db-text-dim,#94a3b8)',
+          'Loading PraisonAIUI components…'
+        );
+        wrap.appendChild(uiHost);
+        wrap.appendChild(statusLine);
+        container.appendChild(wrap);
+
+        var es = null;
+        var lines = [];
+        var transcriptEl = null;
+
+        function renderTranscript(fullText) {
+          if (!transcriptEl) return;
+          if (fullText) {
+            transcriptEl.textContent = fullText;
+            return;
+          }
+          transcriptEl.textContent = lines.length ? lines.join('\n') : '(empty)';
+        }
+
+        function appendLine(role, text) {
+          lines.push(role + ': ' + text);
+          renderTranscript('');
+        }
+
+        try {
+          var callId = await resolveCallId();
+          if (!callId) {
+            renderUiTree(
+              {
+                type: 'layout',
+                children: [
+                  {
+                    type: 'alert',
+                    variant: 'info',
+                    title: 'Call detail',
+                    message: 'No calls yet — start from ElevenLabs talk or GPT Realtime.',
+                  },
+                ],
+              },
+              uiHost
+            );
+            statusLine.textContent = 'Idle';
+            return;
+          }
+          var uiData = await fetchJson('/api/voice/calls/' + encodeURIComponent(callId) + '/ui');
+          renderUiTree(uiData.layout, uiHost);
+          transcriptEl = findTranscriptCodeEl(uiHost);
+          var detail = await fetchJson('/api/voice/calls/' + encodeURIComponent(callId));
+          if (detail.transcript) {
+            lines = detail.transcript.split('\n').filter(Boolean);
+            renderTranscript(detail.transcript);
+          }
+          statusLine.textContent = 'Live feed connecting…';
+          es = new EventSource(
+            '/api/voice/calls/' + encodeURIComponent(callId) + '/live-transcript'
+          );
+          es.onmessage = function (ev) {
+            try {
+              var payload = JSON.parse(ev.data);
+              if (payload.type === 'snapshot' && payload.transcript) {
+                lines = payload.transcript.split('\n').filter(Boolean);
+                renderTranscript(payload.transcript);
+                statusLine.textContent = 'Live · snapshot loaded';
+                return;
+              }
+              if (payload.type === 'transcript') {
+                if (payload.transcriptType === 'final') {
+                  appendLine(payload.role || 'unknown', payload.text || '');
+                } else if (transcriptEl) {
+                  transcriptEl.textContent =
+                    (lines.length ? lines.join('\n') + '\n' : '') +
+                    (payload.role || 'unknown') +
+                    ': ' +
+                    (payload.text || '') +
+                    ' …';
+                }
+                statusLine.textContent = 'Live · ' + (payload.transcriptType || 'update');
+              }
+            } catch (err) {
+              console.warn('[voice plugin] SSE parse error', err);
+            }
+          };
+          es.onerror = function () {
+            statusLine.textContent = 'Live feed disconnected (call may have ended)';
+          };
+        } catch (err) {
+          statusLine.textContent = 'Failed to load call: ' + err.message;
+        }
+
+        return function cleanup() {
+          if (es) es.close();
+        };
+      },
+      function () {}
+    );
+
+    window.aiui.registerView(
+      'web-talk',
+      async function (container) {
+        container.innerHTML = '';
+        var wrap = el('div', 'padding:24px;max-width:640px;color:var(--db-text,#eee)');
+        wrap.appendChild(el('h2', 'margin:0 0 12px 0', 'Web voice'));
+        wrap.appendChild(
+          el(
+            'p',
+            'margin:0 0 20px 0;line-height:1.6;color:var(--db-text-dim,#94a3b8)',
+            'Click Start talking, then allow the microphone when Chrome asks. The prompt does not appear until you click the button.'
+          )
+        );
+        var status = el('div', 'margin-bottom:16px;font-size:13px', 'Loading config…');
+        var btn = el(
+          'button',
+          'padding:12px 20px;border-radius:10px;border:none;background:var(--db-accent,#6366f1);' +
+            'color:#fff;font-size:15px;font-weight:600;cursor:pointer',
+          'Start talking'
+        );
+        btn.disabled = true;
+        wrap.appendChild(status);
+        wrap.appendChild(btn);
+        container.appendChild(wrap);
+
+        var cfg = null;
+        var client = null;
+
+        try {
+          cfg = await fetchJson('/api/voice/web-config');
+          if (!cfg.enabled) {
+            status.textContent =
+              cfg.error ||
+              'Web talk is not configured. Set VOICE_PUBLIC_API_KEY, VOICE_ASSISTANT_ID, VOICE_WEB_SDK_URL, VOICE_WEB_SDK_GLOBAL.';
+            btn.style.display = 'none';
+            return;
+          }
+          status.textContent = 'Ready — click Start talking, then allow microphone.';
+          btn.disabled = false;
+        } catch (err) {
+          status.textContent = 'Config error: ' + err.message;
+          return;
+        }
+
+        btn.addEventListener('click', async function () {
+          if (client && typeof client.stop === 'function' && btn.textContent === 'End call') {
+            client.stop();
+            client = null;
+            status.textContent = 'Call ended';
+            btn.textContent = 'Start talking';
+            return;
+          }
+          btn.disabled = true;
+          status.textContent = 'Loading web SDK…';
+          try {
+            var Sdk = await loadWebSdk(cfg);
+            client = new Sdk(cfg.public_key);
+            if (typeof client.start === 'function') {
+              await client.start(cfg.assistant_id);
+            } else if (typeof client.startAssistant === 'function') {
+              await client.startAssistant(cfg.assistant_id);
+            } else {
+              throw new Error('SDK has no start() method');
+            }
+            status.textContent = 'In call — allow microphone if Chrome asks. Speak now.';
+            btn.textContent = 'End call';
+            btn.disabled = false;
+          } catch (err) {
+            status.textContent = 'Start failed: ' + err.message;
+            btn.disabled = false;
+          }
+        });
+      }
+    );
+
+    window.aiui.registerView(
+      'eleven-talk',
+      async function (container) {
+        container.innerHTML = '';
+        var wrap = el('div', 'padding:24px;max-width:640px;color:var(--db-text,#eee)');
+        wrap.appendChild(el('h2', 'margin:0 0 12px 0', 'ElevenLabs voice'));
+        wrap.appendChild(
+          el(
+            'p',
+            'margin:0 0 20px 0;line-height:1.6;color:var(--db-text-dim,#94a3b8)',
+            'Speech Engine runs STT/TTS in the browser. Your PraisonAI agent handles replies on the server.'
+          )
+        );
+        var status = el('div', 'margin-bottom:16px;font-size:13px', 'Loading config…');
+        var btn = el(
+          'button',
+          'padding:12px 20px;border-radius:10px;border:none;background:var(--db-accent,#6366f1);' +
+            'color:#fff;font-size:15px;font-weight:600;cursor:pointer',
+          'Start conversation'
+        );
+        btn.disabled = true;
+        wrap.appendChild(status);
+        wrap.appendChild(btn);
+        container.appendChild(wrap);
+
+        var conversation = null;
+        var cfg = null;
+
+        try {
+          cfg = await fetchJson('/api/voice/speech-engine/config', 8);
+          if (!cfg.enabled) {
+            status.textContent =
+              cfg.error ||
+              'Speech Engine not ready. Run: pip install -U "elevenlabs>=2.47.0" then .\\start_dev.ps1 -Restart. ' +
+                'API key needs convai_write permission in ElevenLabs dashboard.';
+            btn.style.display = 'none';
+            return;
+          }
+          var toolList = (cfg.tools && cfg.tools.length ? cfg.tools.join(', ') : 'get_current_time, echo_message');
+          status.textContent =
+            'Ready — LLM ' +
+            (cfg.model || 'gpt-4o-mini') +
+            ' · tools: ' +
+            toolList +
+            '. Click Start conversation, then allow microphone.';
+          btn.disabled = false;
+        } catch (err) {
+          status.textContent = 'Config error: ' + err.message;
+          return;
+        }
+
+        btn.addEventListener('click', async function () {
+          if (conversation && btn.textContent === 'End conversation') {
+            try {
+              await conversation.endSession();
+            } catch (e) {
+              /* ignore */
+            }
+            conversation = null;
+            status.textContent = 'Conversation ended';
+            btn.textContent = 'Start conversation';
+            return;
+          }
+          btn.disabled = true;
+          status.textContent = 'Loading ElevenLabs client…';
+          try {
+            var sdkUrl = cfg.client_sdk_url || 'https://esm.sh/@elevenlabs/client';
+            var mod = await import(/* webpackIgnore: true */ sdkUrl);
+            var Conversation = mod.Conversation;
+            if (!Conversation || typeof Conversation.startSession !== 'function') {
+              throw new Error('ElevenLabs client SDK missing Conversation.startSession');
+            }
+            var tokenResp = await fetchJson('/api/voice/speech-engine/token');
+            if (!tokenResp.token) throw new Error('No conversation token returned');
+            await navigator.mediaDevices.getUserMedia({ audio: true });
+            await attachVoiceUser();
+            var startOpts = {
+              conversationToken: tokenResp.token,
+              onConnect: function () {
+                status.textContent =
+                  'Connected — speak, then pause ~1s so the agent knows you finished.';
+                btn.textContent = 'End conversation';
+                btn.disabled = false;
+              },
+              onDisconnect: function (details) {
+                var reason = details && details.reason ? details.reason : '';
+                if (reason === 'agent' || reason === 'unknown') {
+                  reason =
+                    'Speech Engine server did not respond in time. Restart .\\start_dev.ps1 -Restart and try again.';
+                } else if (!reason) {
+                  reason =
+                    'Connection closed (ElevenLabs could not reach your Speech Engine WebSocket).';
+                }
+                status.textContent = 'Disconnected — ' + reason;
+                btn.textContent = 'Start conversation';
+                btn.disabled = false;
+                conversation = null;
+              },
+              onError: function (error) {
+                status.textContent =
+                  'Error: ' + (error && error.message ? error.message : String(error));
+                btn.disabled = false;
+              },
+              onStatusChange: function (state) {
+                if (state && state.status) {
+                  status.textContent = 'Status: ' + state.status;
+                }
+              },
+              onModeChange: function (mode) {
+                if (!mode || !mode.mode) return;
+                if (mode.mode === 'listening') {
+                  status.textContent = 'Listening… speak, then pause when done.';
+                } else if (mode.mode === 'speaking') {
+                  status.textContent = 'Agent speaking…';
+                } else {
+                  status.textContent = 'Mode: ' + mode.mode;
+                }
+              },
+              onMessage: function (msg) {
+                if (msg && msg.message) {
+                  console.log('[eleven-talk]', msg.source || 'unknown', msg.message);
+                }
+              },
+            };
+            if (cfg.first_message) {
+              startOpts.overrides = {
+                agent: { firstMessage: cfg.first_message },
+              };
+            }
+            conversation = await Conversation.startSession(startOpts);
+          } catch (err) {
+            status.textContent = 'Start failed: ' + err.message;
+            btn.disabled = false;
+          }
+        });
+      }
+    );
+
+    window.aiui.registerView(
+      'realtime-talk',
+      async function (container) {
+        container.innerHTML = '';
+        var wrap = el('div', 'padding:24px;max-width:640px;color:var(--db-text,#eee)');
+        wrap.appendChild(el('h2', 'margin:0 0 12px 0', 'GPT Realtime voice'));
+        wrap.appendChild(
+          el(
+            'p',
+            'margin:0 0 20px 0;line-height:1.6;color:var(--db-text-dim,#94a3b8)',
+            'Speech-to-speech via OpenAI gpt-realtime-2.1-mini (WebRTC).'
+          )
+        );
+        var status = el('div', 'margin-bottom:16px;font-size:13px', 'Loading config…');
+        var btn = el(
+          'button',
+          'padding:12px 20px;border-radius:10px;border:none;background:var(--db-accent,#6366f1);' +
+            'color:#fff;font-size:15px;font-weight:600;cursor:pointer',
+          'Start conversation'
+        );
+        btn.disabled = true;
+        wrap.appendChild(status);
+        wrap.appendChild(btn);
+        container.appendChild(wrap);
+
+        var pc = null;
+        var dc = null;
+        var audioEl = null;
+        var localStream = null;
+        var callId = null;
+        var cfg = null;
+
+        async function saveTranscript(role, text) {
+          if (!callId || !text) return;
+          try {
+            await postJson('/api/voice/realtime/transcript', {
+              call_id: callId,
+              role: role,
+              text: text,
+            });
+          } catch (e) {
+            console.warn('[realtime-talk] transcript save failed', e);
+          }
+        }
+
+        function handleRealtimeEvent(event) {
+          if (!event || !event.type) return;
+          if (
+            event.type === 'conversation.item.input_audio_transcription.completed' &&
+            event.transcript
+          ) {
+            saveTranscript('user', event.transcript);
+            status.textContent = 'You: ' + event.transcript;
+          }
+          if (event.type === 'response.output_audio_transcript.done' && event.transcript) {
+            saveTranscript('agent', event.transcript);
+            status.textContent = 'Agent: ' + event.transcript;
+          }
+          if (event.type === 'response.output_audio_transcript.delta' && event.delta) {
+            status.textContent = 'Agent speaking…';
+          }
+          if (event.type === 'input_audio_buffer.speech_started') {
+            status.textContent = 'Listening…';
+          }
+        }
+
+        async function endSession() {
+          if (dc) {
+            try {
+              dc.close();
+            } catch (e) {
+              /* ignore */
+            }
+            dc = null;
+          }
+          if (pc) {
+            try {
+              pc.close();
+            } catch (e) {
+              /* ignore */
+            }
+            pc = null;
+          }
+          if (localStream) {
+            localStream.getTracks().forEach(function (t) {
+              t.stop();
+            });
+            localStream = null;
+          }
+          if (audioEl && audioEl.parentNode) {
+            audioEl.parentNode.removeChild(audioEl);
+            audioEl = null;
+          }
+          if (callId) {
+            try {
+              await postJson('/api/voice/realtime/end', { call_id: callId });
+            } catch (e) {
+              /* ignore */
+            }
+          }
+          callId = null;
+          status.textContent = 'Conversation ended.';
+          btn.textContent = 'Start conversation';
+          btn.disabled = false;
+        }
+
+        try {
+          cfg = await fetchJson('/api/voice/realtime/config', 8);
+          if (!cfg.enabled) {
+            status.textContent =
+              cfg.error || 'Set OPENAI_API_KEY in .env, then restart .\\start_dev.ps1';
+            btn.style.display = 'none';
+            return;
+          }
+          status.textContent = 'Ready (' + (cfg.model || 'gpt-realtime-2.1-mini') + ').';
+          btn.disabled = false;
+        } catch (err) {
+          status.textContent = 'Config error: ' + err.message;
+          return;
+        }
+
+        btn.addEventListener('click', async function () {
+          if (pc && btn.textContent === 'End conversation') {
+            await endSession();
+            return;
+          }
+          btn.disabled = true;
+          status.textContent = 'Starting WebRTC session…';
+          try {
+            var callResp = await postJson('/api/voice/realtime/new-call', {
+              user_id: voiceUserId(),
+            });
+            callId = callResp.call_id;
+            localStream = await navigator.mediaDevices.getUserMedia({ audio: true });
+            pc = new RTCPeerConnection();
+            audioEl = document.createElement('audio');
+            audioEl.autoplay = true;
+            audioEl.style.display = 'none';
+            wrap.appendChild(audioEl);
+            pc.ontrack = function (e) {
+              if (audioEl) audioEl.srcObject = e.streams[0];
+            };
+            localStream.getTracks().forEach(function (track) {
+              pc.addTrack(track, localStream);
+            });
+            dc = pc.createDataChannel('oai-events');
+            dc.addEventListener('open', function () {
+              if (cfg.first_message) {
+                dc.send(
+                  JSON.stringify({
+                    type: 'response.create',
+                    response: {
+                      instructions:
+                        'Greet the user by saying exactly: ' + cfg.first_message,
+                    },
+                  })
+                );
+              }
+            });
+            dc.addEventListener('message', function (e) {
+              try {
+                handleRealtimeEvent(JSON.parse(e.data));
+              } catch (err) {
+                console.warn('[realtime-talk] event parse failed', err);
+              }
+            });
+            var offer = await pc.createOffer();
+            await pc.setLocalDescription(offer);
+            var sdpResp = await fetch(
+              '/api/voice/realtime/session?call_id=' + encodeURIComponent(callId),
+              {
+                method: 'POST',
+                headers: { 'Content-Type': 'application/sdp' },
+                body: offer.sdp,
+              }
+            );
+            if (!sdpResp.ok) {
+              var errText = await sdpResp.text();
+              throw new Error(errText || 'SDP exchange failed');
+            }
+            var answerSdp = await sdpResp.text();
+            await pc.setRemoteDescription({ type: 'answer', sdp: answerSdp });
+            status.textContent = 'Connected — speak when ready.';
+            btn.textContent = 'End conversation';
+            btn.disabled = false;
+          } catch (err) {
+            status.textContent = 'Start failed: ' + err.message;
+            await endSession();
+            btn.disabled = false;
+          }
+        });
+      }
+    );
+
+    window.aiui.registerView(
+      'config',
+      async function (container) {
+        container.innerHTML = '';
+        var wrap = el('div', 'padding:24px;max-width:900px;color:var(--db-text,#eee)');
+        wrap.appendChild(el('h2', 'margin:0 0 8px 0', 'Voice stack doctor'));
+        wrap.appendChild(
+          el(
+            'p',
+            'margin:0 0 16px 0;font-size:13px;color:var(--db-text-dim,#94a3b8)',
+            'Checks .env keys, local ports, APIs, and Speech Engine tunnel reachability.'
+          )
+        );
+        var status = el('div', 'margin-bottom:16px;font-size:13px', 'Running checks…');
+        var tableHost = el('div', '');
+        wrap.appendChild(status);
+        wrap.appendChild(tableHost);
+        container.appendChild(wrap);
+
+        function statusColor(st) {
+          if (st === 'pass') return '#22c55e';
+          if (st === 'warn') return '#eab308';
+          return '#ef4444';
+        }
+
+        try {
+          var data = await fetchJson('/api/voice/doctor', 12);
+          var summary = data.summary || {};
+          status.textContent =
+            'Passed ' +
+            (summary.passed || 0) +
+            ' · Warnings ' +
+            (summary.warnings || 0) +
+            ' · Failed ' +
+            (summary.failed || 0);
+          var rows = (data.checks || []).map(function (c) {
+            return [
+              c.name || '—',
+              c.status || '—',
+              c.detail || '—',
+            ];
+          });
+          renderUiTree(
+            {
+              type: 'layout',
+              children: [
+                {
+                  type: 'table',
+                  headers: ['Check', 'Status', 'Detail'],
+                  rows: rows.length ? rows : [['—', '—', '—']],
+                },
+              ],
+            },
+            tableHost
+          );
+          var badges = tableHost.querySelectorAll('td');
+          badges.forEach(function (cell) {
+            if (cell.textContent === 'pass' || cell.textContent === 'warn' || cell.textContent === 'fail') {
+              cell.style.color = statusColor(cell.textContent);
+              cell.style.fontWeight = '600';
+            }
+          });
+        } catch (err) {
+          status.textContent = 'Doctor failed: ' + err.message;
+        }
+      },
+      function () {}
+    );
+
+    console.log(
+      '[voice plugin] Registered call-detail + web-talk + eleven-talk + realtime-talk + config views'
+    );
+
+    var path = (window.location.pathname || '').replace(/^\//, '');
+    if (
+      (path === 'web-talk' ||
+        path === 'call-detail' ||
+        path === 'eleven-talk' ||
+        path === 'realtime-talk') &&
+      typeof window.aiui.selectPage === 'function'
+    ) {
+      window.aiui.selectPage(path);
+    } else if (
+      path === 'web-talk' ||
+      path === 'call-detail' ||
+      path === 'eleven-talk' ||
+      path === 'realtime-talk'
+    ) {
+      var n = 0;
+      var t = setInterval(function () {
+        n += 1;
+        if (typeof window.aiui.selectPage === 'function') {
+          clearInterval(t);
+          window.aiui.selectPage(path);
+        } else if (n > 50) {
+          clearInterval(t);
+        }
+      }, 100);
+    }
+  }
+
+  whenReady(registerVoiceViews);
+})();
@@ -0,0 +1,115 @@
+"""HTTP routes for OpenAI Realtime WebRTC voice."""
+
+from __future__ import annotations
+
+import logging
+
+from starlette.requests import Request
+from starlette.responses import JSONResponse, PlainTextResponse
+
+from integrations.voice.realtime import (
+    RealtimeConfigError,
+    load_realtime_settings,
+    mark_session_started,
+    new_call_id,
+    persist_transcript_line,
+    realtime_enabled,
+    create_webrtc_session,
+)
+
+logger = logging.getLogger(__name__)
+
+
+async def api_realtime_config(_request: Request) -> JSONResponse:
+    try:
+        settings = load_realtime_settings()
+        enabled = realtime_enabled()
+        error = None
+    except RealtimeConfigError as exc:
+        settings = None
+        enabled = False
+        error = str(exc)
+    return JSONResponse(
+        {
+            "enabled": enabled,
+            "error": error,
+            "model": settings.model if settings else None,
+            "voice": settings.voice if settings else None,
+            "reasoning_effort": settings.reasoning_effort if settings else None,
+            "first_message": settings.first_message if settings else None,
+        }
+    )
+
+
+async def api_realtime_new_call(request: Request) -> JSONResponse:
+    from integrations.voice.session_memory import bind_call_user, voice_user_session_id
+
+    try:
+        payload = await request.json()
+    except Exception:  # noqa: BLE001
+        payload = {}
+    call_id = new_call_id()
+    mark_session_started(call_id)
+    user_id = str(payload.get("user_id") or "").strip()
+    praison_session_id = None
+    if user_id:
+        praison_session_id = bind_call_user(call_id, user_id)
+    return JSONResponse(
+        {
+            "call_id": call_id,
+            "user_id": user_id or None,
+            "praison_session_id": praison_session_id or (voice_user_session_id(user_id) if user_id else None),
+        }
+    )
+
+
+async def api_realtime_session(request: Request) -> PlainTextResponse | JSONResponse:
+    """Unified WebRTC interface — browser SDP in, OpenAI SDP answer out."""
+    call_id = request.query_params.get("call_id", "").strip()
+    body = await request.body()
+    sdp_offer = body.decode("utf-8").strip()
+    if not sdp_offer:
+        return JSONResponse({"error": "Missing SDP offer body"}, status_code=400)
+    try:
+        sdp_answer = await create_webrtc_session(sdp_offer, call_id=call_id or None)
+    except RealtimeConfigError as exc:
+        return JSONResponse({"error": str(exc)}, status_code=503)
+    except Exception as exc:  # noqa: BLE001
+        logger.exception("Realtime WebRTC session failed")
+        return JSONResponse({"error": str(exc)}, status_code=502)
+    if call_id:
+        mark_session_started(call_id)
+    return PlainTextResponse(sdp_answer, media_type="application/sdp")
+
+
+async def api_realtime_transcript(request: Request) -> JSONResponse:
+    try:
+        payload = await request.json()
+    except Exception:  # noqa: BLE001
+        return JSONResponse({"error": "Invalid JSON"}, status_code=400)
+    call_id = str(payload.get("call_id") or "").strip()
+    role = str(payload.get("role") or "user").strip()
+    text = str(payload.get("text") or "").strip()
+    if not call_id or not text:
+        return JSONResponse({"error": "call_id and text required"}, status_code=400)
+    from integrations.voice.session_memory import record_voice_turn
+
+    persist_transcript_line(call_id, role, text)
+    if role == "user":
+        await record_voice_turn(call_id, user_text=text)
+    else:
+        await record_voice_turn(call_id, agent_text=text)
+    return JSONResponse({"ok": True})
+
+
+async def api_realtime_end(request: Request) -> JSONResponse:
+    try:
+        payload = await request.json()
+    except Exception:  # noqa: BLE001
+        return JSONResponse({"error": "Invalid JSON"}, status_code=400)
+    call_id = str(payload.get("call_id") or "").strip()
+    if call_id:
+        from integrations.voice.call_finalize import finalize_call
+
+        await finalize_call(call_id, status="ended (customer-ended-call)")
+    return JSONResponse({"ok": True})
@@ -0,0 +1,3 @@
+httpx
+elevenlabs>=2.47.0
+openai
@@ -0,0 +1,172 @@
+"""Create or update ElevenLabs Speech Engine with the public WebSocket URL.
+
+Usage:
+    cd examples/python/voice-agent
+    .\\start_dev.ps1          # sets PUBLIC_API_BASE_URL first
+    py -3.13 setup_speech_engine.py
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+from pathlib import Path
+
+_ROOT = Path(__file__).resolve().parent
+ENV_PATH = _ROOT / ".env"
+
+
+def _load_env() -> None:
+    if not ENV_PATH.is_file():
+        return
+    for raw in ENV_PATH.read_text(encoding="utf-8").splitlines():
+        line = raw.strip()
+        if not line or line.startswith("#") or "=" not in line:
+            continue
+        key, _, value = line.partition("=")
+        key, value = key.strip(), value.strip()
+        if key:
+            os.environ[key] = value
+
+
+def _write_env_value(key: str, value: str) -> None:
+    lines = ENV_PATH.read_text(encoding="utf-8").splitlines() if ENV_PATH.is_file() else []
+    updated = False
+    new_lines: list[str] = []
+    for line in lines:
+        if line.strip().startswith(f"{key}="):
+            new_lines.append(f"{key}={value}")
+            updated = True
+        else:
+            new_lines.append(line)
+    if not updated:
+        new_lines.append(f"{key}={value}")
+    ENV_PATH.write_text("\n".join(new_lines) + "\n", encoding="utf-8")
+    os.environ[key] = value
+
+
+def _http_to_ws(base: str) -> str:
+    base = base.strip().rstrip("/")
+    if base.startswith("https://"):
+        return "wss://" + base[len("https://") :] + "/ws"
+    if base.startswith("http://"):
+        return "ws://" + base[len("http://") :] + "/ws"
+    raise RuntimeError(f"Unsupported URL: {base}")
+
+
+def _public_ws_url() -> str:
+    # Prefer dedicated Speech Engine tunnel (sidecar on 8002).
+    speech_base = os.getenv("SPEECH_ENGINE_PUBLIC_URL", "").strip().rstrip("/")
+    if speech_base:
+        return _http_to_ws(speech_base)
+    base = os.getenv("PUBLIC_API_BASE_URL", "").strip().rstrip("/")
+    if not base:
+        raise RuntimeError(
+            "Set SPEECH_ENGINE_PUBLIC_URL or PUBLIC_API_BASE_URL in .env "
+            "(run .\\start_dev.ps1 first)"
+        )
+    return _http_to_ws(base)
+
+
+async def _patch_overrides(client: object, engine_id: str) -> None:
+    """Enable client first-message override (not exposed on speech_engine.update yet)."""
+    wrapper = getattr(client, "_client_wrapper", None)
+    if wrapper is None:
+        return
+    try:
+        await wrapper.httpx_client.request(
+            f"v1/speech-engine/{engine_id}",
+            method="PATCH",
+            json={"overrides": {"first_message": True}},
+            headers={"content-type": "application/json"},
+        )
+        print("Enabled first_message client override")
+    except Exception as exc:  # noqa: BLE001
+        print(f"Note: could not enable first_message override ({exc})")
+
+
+async def _ensure_engine() -> str:
+    api_key = os.getenv("ELEVENLABS_API_KEY", "").strip()
+    if not api_key:
+        raise RuntimeError("Set ELEVENLABS_API_KEY in .env")
+
+    try:
+        from elevenlabs import AsyncElevenLabs
+    except ImportError as exc:
+        raise RuntimeError(
+            "Install Speech Engine SDK: pip install 'elevenlabs>=2.47.0'"
+        ) from exc
+
+    if not hasattr(AsyncElevenLabs(api_key="check"), "speech_engine"):
+        raise RuntimeError(
+            "elevenlabs package is too old for Speech Engine. Run: "
+            "pip install -U 'elevenlabs>=2.47.0'"
+        )
+
+    ws_url = _public_ws_url()
+    engine_id = os.getenv("SPEECH_ENGINE_ID", "").strip()
+    client = AsyncElevenLabs(api_key=api_key)
+
+    turn_config = {"turn_eagerness": "eager", "turn_timeout": 5.0}
+    voice_id = os.getenv("ELEVENLABS_VOICE_ID", "").strip() or "cjVigY5qzO86Huf0OWal"
+    tts_config = {"model_id": "eleven_flash_v2", "voice_id": voice_id}
+
+    if engine_id:
+        updated = await client.speech_engine.update(
+            speech_engine_id=engine_id,
+            speech_engine={"ws_url": ws_url},
+            turn=turn_config,
+            tts=tts_config,
+        )
+        resolved = getattr(updated, "speech_engine_id", None) or getattr(updated, "engine_id", None) or engine_id
+        print(f"Updated Speech Engine: {resolved}")
+        print(f"WebSocket URL: {ws_url}")
+        print(f"TTS voice_id: {voice_id}")
+        await _patch_overrides(client, str(resolved))
+        return str(resolved)
+
+    created = await client.speech_engine.create(
+        name="PraisonAI Voice",
+        speech_engine={"ws_url": ws_url},
+        turn=turn_config,
+        tts=tts_config,
+        overrides={"first_message": True},
+    )
+    resolved = getattr(created, "engine_id", None) or getattr(created, "id", None)
+    if not resolved:
+        raise RuntimeError("Speech Engine create did not return an engine_id")
+    print(f"Created Speech Engine: {resolved}")
+    print(f"WebSocket URL: {ws_url}")
+    print(f"TTS voice_id: {voice_id}")
+    return str(resolved)
+
+
+async def main_async() -> int:
+    _load_env()
+    try:
+        engine_id = await _ensure_engine()
+    except RuntimeError as exc:
+        print(f"Setup failed: {exc}")
+        return 1
+    except Exception as exc:  # noqa: BLE001
+        message = str(exc)
+        if "convai_write" in message or "missing_permissions" in message:
+            print("Setup failed: ElevenLabs API key lacks convai_write permission.")
+            print(
+                "In ElevenLabs dashboard: API Keys -> edit key -> ElevenAgents: enable Write."
+            )
+        else:
+            print(f"Setup failed: {exc}")
+        return 1
+
+    _write_env_value("SPEECH_ENGINE_ID", engine_id)
+    print("Speech Engine auto-setup complete.")
+    return 0
+
+
+def main() -> int:
+    return asyncio.run(main_async())
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
@@ -28,7 +28,7 @@ def _load_env() -> None:
            continue
        key, _, value = line.partition("=")
        key, value = key.strip(), value.strip()
-        if key and key not in os.environ:
+        if key:
            os.environ[key] = value


@@ -117,6 +117,119 @@ def ensure_server_secret() -> str:
    return secret


+def ensure_assistant_model(api_key: str, api_base: str, assistant_id: str) -> dict[str, Any]:
+    assistant = _request("GET", f"/assistant/{assistant_id}", api_key=api_key, api_base=api_base)
+    if not isinstance(assistant, dict):
+        raise RuntimeError("Could not load assistant")
+    model = dict(assistant.get("model") or {})
+    target = os.getenv("VOICE_AGENT_MODEL", "gpt-4o-mini")
+    if model.get("provider") != "openai" or model.get("model") != target:
+        model["provider"] = "openai"
+        model["model"] = target
+        patched = _request(
+            "PATCH",
+            f"/assistant/{assistant_id}",
+            api_key=api_key,
+            api_base=api_base,
+            body={"model": model},
+        )
+        if patched is None:
+            raise RuntimeError(f"Failed to set assistant model to {target}")
+        print(f"Set assistant model to openai/{target}")
+    return model
+
+
+def ensure_agent_model_env() -> str:
+    """Ensure VOICE_AGENT_MODEL is set in .env and return the target model."""
+    target = os.getenv("VOICE_AGENT_MODEL", "").strip() or "gpt-4o-mini"
+    if not os.getenv("VOICE_AGENT_MODEL", "").strip():
+        _write_env_value("VOICE_AGENT_MODEL", target)
+        print(f"Set VOICE_AGENT_MODEL={target} in .env")
+    return target
+
+
+def ensure_openai_credential(api_key: str, api_base: str, openai_api_key: str) -> tuple[str, str]:
+    """Create or reuse an OpenAI-compatible credential in the provider account."""
+    existing = _request("GET", "/credential", api_key=api_key, api_base=api_base)
+    if isinstance(existing, list):
+        preferred = sorted(
+            [item for item in existing if isinstance(item, dict) and item.get("id")],
+            key=lambda item: 0 if str(item.get("provider") or "") == "openai" else 1,
+        )
+        for item in preferred:
+            provider = str(item.get("provider") or "")
+            if provider in ("openai", "custom-llm"):
+                cred_id = str(item["id"])
+                print(f"Reusing credential ({provider}): {cred_id}")
+                updated = _request(
+                    "PATCH",
+                    f"/credential/{cred_id}",
+                    api_key=api_key,
+                    api_base=api_base,
+                    body={
+                        "provider": provider,
+                        "apiKey": openai_api_key,
+                        "name": "Praison Voice OpenAI",
+                    },
+                )
+                if updated is None:
+                    raise RuntimeError("Failed to update OpenAI credential")
+                return cred_id, provider
+
+    for provider in ("openai", "custom-llm"):
+        created = _request(
+            "POST",
+            "/credential",
+            api_key=api_key,
+            api_base=api_base,
+            body={
+                "provider": provider,
+                "apiKey": openai_api_key,
+                "name": "Praison Voice OpenAI",
+            },
+        )
+        if isinstance(created, dict) and created.get("id"):
+            cred_id = str(created["id"])
+            print(f"Created credential ({provider}): {cred_id}")
+            return cred_id, provider
+
+    raise RuntimeError(
+        "Failed to create OpenAI credential. Add the key manually in the provider dashboard "
+        "under Provider Keys / OpenAI, or enable gpt-4o-mini on your OpenAI project."
+    )
+
+
+def ensure_openai_on_assistant(
+    api_key: str, api_base: str, assistant_id: str, openai_credential_id: str, *, provider: str
+) -> None:
+    assistant = _request("GET", f"/assistant/{assistant_id}", api_key=api_key, api_base=api_base)
+    if not isinstance(assistant, dict):
+        raise RuntimeError("Could not load assistant for OpenAI setup")
+
+    target_model = os.getenv("VOICE_AGENT_MODEL", "gpt-4o-mini")
+    if provider == "custom-llm":
+        model = dict(assistant.get("model") or {})
+        model["model"] = target_model
+        model["provider"] = "custom-llm"
+        model["url"] = "https://api.openai.com/v1"
+        model["metadataSendMode"] = "off"
+        model.pop("credentialId", None)
+    else:
+        # Native OpenAI — drop custom-llm-only fields from prior config.
+        model = {"provider": "openai", "model": target_model}
+
+    patched = _request(
+        "PATCH",
+        f"/assistant/{assistant_id}",
+        api_key=api_key,
+        api_base=api_base,
+        body={"model": model},
+    )
+    if patched is None:
+        raise RuntimeError("Failed to attach OpenAI credential to assistant")
+    print(f"Configured assistant model ({provider}): {target_model}")
+
+
 def ensure_tools(api_key: str, api_base: str, assistant_id: str) -> list[str]:
    existing = _request("GET", "/tool", api_key=api_key, api_base=api_base)
    by_name: dict[str, str] = {}
@@ -230,8 +343,18 @@ def main() -> int:

    secret = ensure_server_secret()
    webhook_url = f"{public_base}/webhooks/voice"
+    target_model = ensure_agent_model_env()
+    openai_api_key = os.getenv("OPENAI_API_KEY", "").strip()

    try:
+        if openai_api_key:
+            cred_id, cred_provider = ensure_openai_credential(api_key, api_base, openai_api_key)
+            ensure_openai_on_assistant(
+                api_key, api_base, assistant_id, cred_id, provider=cred_provider
+            )
+            print(f"Assistant model target: {target_model}")
+        else:
+            print("OPENAI_API_KEY not set — skipping provider OpenAI credential setup")
        ensure_tools(api_key, api_base, assistant_id)
        ensure_webhook(api_key, api_base, assistant_id, webhook_url, secret)
    except RuntimeError as exc:
@@ -0,0 +1,110 @@
+"""HTTP + WebSocket routes for ElevenLabs Speech Engine."""
+
+from __future__ import annotations
+
+import logging
+import os
+
+from starlette.requests import Request
+from starlette.responses import JSONResponse
+from starlette.websockets import WebSocket, WebSocketDisconnect
+
+from integrations.voice.speech_engine import (
+    SpeechEngineConfigError,
+    get_engine_resource,
+    get_webrtc_token,
+    handle_session_close,
+    handle_session_init,
+    handle_user_transcript,
+    load_speech_engine_settings,
+    speech_engine_enabled,
+)
+
+logger = logging.getLogger(__name__)
+
+
+async def api_speech_engine_config(_request: Request) -> JSONResponse:
+    try:
+        settings = load_speech_engine_settings()
+        enabled = speech_engine_enabled()
+        error = None
+    except SpeechEngineConfigError as exc:
+        settings = None
+        enabled = False
+        error = str(exc)
+    tools: list[str] = []
+    if enabled:
+        try:
+            from integrations.voice.agent_runner import list_registered_tool_names
+
+            tools = list_registered_tool_names()
+        except Exception:  # noqa: BLE001
+            tools = ["get_current_time", "echo_message"]
+    return JSONResponse(
+        {
+            "enabled": enabled,
+            "error": error,
+            "engine_id": settings.engine_id if settings else None,
+            "ws_url": settings.public_ws_url if settings else None,
+            "client_sdk_url": "https://esm.sh/@elevenlabs/client",
+            "first_message": settings.first_message if settings else None,
+            "model": os.getenv("VOICE_AGENT_MODEL", "gpt-4o-mini"),
+            "tools": tools,
+        }
+    )
+
+
+async def api_speech_engine_token(_request: Request) -> JSONResponse:
+    try:
+        token = get_webrtc_token()
+    except SpeechEngineConfigError as exc:
+        return JSONResponse({"error": str(exc)}, status_code=503)
+    except Exception as exc:  # noqa: BLE001
+        logger.exception("Speech Engine token mint failed")
+        return JSONResponse({"error": str(exc)}, status_code=502)
+    return JSONResponse({"token": token})
+
+
+async def speech_engine_websocket(websocket: WebSocket) -> None:
+    try:
+        engine = await get_engine_resource()
+    except SpeechEngineConfigError as exc:
+        logger.warning("Speech Engine WS rejected: %s", exc)
+        return
+
+    headers = dict(websocket.headers)
+    if hasattr(engine, "verify_request") and not engine.verify_request(headers):
+        logger.warning(
+            "Speech Engine WS unauthorized (missing or invalid ElevenLabs JWT header)"
+        )
+        return
+
+    await websocket.accept()
+    logger.info("Speech Engine WebSocket accepted from ElevenLabs")
+    session = engine.create_session(websocket, debug=False)
+
+    async def on_transcript(transcript, sess):  # noqa: ANN001
+        await handle_user_transcript(transcript, sess)
+
+    async def on_init(conversation_id, sess):  # noqa: ANN001
+        await handle_session_init(conversation_id, sess)
+
+    async def on_close(sess):  # noqa: ANN001
+        await handle_session_close(sess)
+
+    def on_error(err, sess=None):  # noqa: ANN001
+        logger.error("Speech Engine session error: %s", err)
+
+    session.on("user_transcript", on_transcript)
+    session.on("init", on_init)
+    session.on("close", on_close)
+    session.on("error", on_error)
+
+    try:
+        await session.run()
+    except WebSocketDisconnect:
+        logger.info("Speech Engine WebSocket disconnected")
+        await handle_session_close(session)
+    except Exception:  # noqa: BLE001
+        logger.exception("Speech Engine WebSocket session failed")
+        await handle_session_close(session)
@@ -0,0 +1,89 @@
+"""Standalone ElevenLabs Speech Engine WebSocket server (port 8002).
+
+ElevenLabs connects here for transcripts; the dashboard stays on 8001.
+Official SDK path: engine.serve() — more reliable than in-app Starlette /ws.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+from pathlib import Path
+
+_ROOT = Path(__file__).resolve().parent
+if str(_ROOT) not in __import__("sys").path:
+    __import__("sys").path.insert(0, str(_ROOT))
+
+
+def _load_env() -> None:
+    env_path = _ROOT / ".env"
+    if not env_path.is_file():
+        return
+    for raw in env_path.read_text(encoding="utf-8").splitlines():
+        line = raw.strip()
+        if not line or line.startswith("#") or "=" not in line:
+            continue
+        key, _, value = line.partition("=")
+        key, value = key.strip(), value.strip()
+        if key:
+            os.environ[key] = value
+
+
+async def _main() -> None:
+    _load_env()
+    api_key = os.getenv("ELEVENLABS_API_KEY", "").strip()
+    engine_id = os.getenv("SPEECH_ENGINE_ID", "").strip()
+    port = int(os.getenv("SPEECH_ENGINE_PORT", "8002"))
+    disable_auth = os.getenv("SPEECH_ENGINE_DISABLE_AUTH", "").lower() in (
+        "1",
+        "true",
+        "yes",
+    )
+
+    if not api_key or not engine_id:
+        raise SystemExit("Set ELEVENLABS_API_KEY and SPEECH_ENGINE_ID in .env")
+
+    from elevenlabs import AsyncElevenLabs
+
+    from integrations.voice.speech_engine import (
+        handle_session_close,
+        handle_session_disconnect,
+        handle_session_init,
+        handle_user_transcript,
+    )
+
+    client = AsyncElevenLabs(api_key=api_key)
+    engine = await client.speech_engine.get(engine_id)
+
+    async def on_transcript(transcript, session):  # noqa: ANN001
+        await handle_user_transcript(transcript, session)
+
+    async def on_init(conversation_id, session):  # noqa: ANN001
+        await handle_session_init(conversation_id, session)
+
+    async def on_close(session):  # noqa: ANN001
+        await handle_session_close(session)
+
+    async def on_disconnect(session):  # noqa: ANN001
+        await handle_session_disconnect(session)
+
+    async def on_error(err, session):  # noqa: ANN001
+        print(f"[speech sidecar] error: {err}", flush=True)
+
+    debug = os.getenv("SPEECH_ENGINE_DEBUG", "true").lower() in ("1", "true", "yes")
+    print(f"[speech sidecar] listening on ws://127.0.0.1:{port}/ws (debug={debug})", flush=True)
+    await engine.serve(
+        port=port,
+        path="/ws",
+        debug=debug,
+        disable_auth=disable_auth,
+        on_init=on_init,
+        on_transcript=on_transcript,
+        on_close=on_close,
+        on_disconnect=on_disconnect,
+        on_error=on_error,
+    )
+
+
+if __name__ == "__main__":
+    asyncio.run(_main())
@@ -18,13 +18,14 @@ function Load-DotEnv {
        $line = $_.Trim()
        if (-not $line -or $line.StartsWith("#") -or $line -notmatch "=") { return }
        $key, $value = $line -split "=", 2
-        if ($key -and -not [string]::IsNullOrWhiteSpace($key) -and -not (Get-Item "Env:$key" -ErrorAction SilentlyContinue)) {
+        $key = $key.Trim()
+        if ($key -and -not [string]::IsNullOrWhiteSpace($key)) {
            Set-Item -Path "Env:$key" -Value $value.Trim()
        }
    }
 }

-function Test-PortListening([int]$p) {
+function Test-AppHttpReady([int]$p) {
    foreach ($path in @("/health/live", "/health", "/api/voice/config")) {
        try {
            $r = Invoke-WebRequest -Uri "http://127.0.0.1:$p$path" -UseBasicParsing -TimeoutSec 15
@@ -34,6 +35,33 @@ function Test-PortListening([int]$p) {
    return $false
 }

+function Test-TcpPortListening([int]$p) {
+    try {
+        $conn = Get-NetTCPConnection -LocalPort $p -State Listen -ErrorAction SilentlyContinue | Select-Object -First 1
+        if ($conn) { return $true }
+    } catch { }
+    try {
+        $client = New-Object System.Net.Sockets.TcpClient
+        $async = $client.BeginConnect("127.0.0.1", $p, $null, $null)
+        $ok = $async.AsyncWaitHandle.WaitOne(500)
+        if ($ok -and $client.Connected) {
+            $client.Close()
+            return $true
+        }
+        $client.Close()
+    } catch { }
+    return $false
+}
+
+function Wait-PortReleased([int]$p, [int]$maxSec = 20) {
+    $deadline = (Get-Date).AddSeconds($maxSec)
+    while ((Get-Date) -lt $deadline) {
+        if (-not (Test-TcpPortListening $p)) { return $true }
+        Start-Sleep -Seconds 1
+    }
+    return -not (Test-TcpPortListening $p)
+}
+
 function Stop-PortProcess([int]$p) {
    try {
        $conns = Get-NetTCPConnection -LocalPort $p -State Listen -ErrorAction SilentlyContinue
@@ -69,10 +97,10 @@ function Start-VoiceApp([int]$p) {
        -RedirectStandardOutput $appLog -RedirectStandardError $appErr
    $deadline = (Get-Date).AddSeconds(120)
    while ((Get-Date) -lt $deadline) {
-        if (Test-PortListening $p) { break }
+        if (Test-AppHttpReady $p) { break }
        Start-Sleep -Seconds 3
    }
-    if (-not (Test-PortListening $p)) {
+    if (-not (Test-AppHttpReady $p)) {
        $errText = ""
        if (Test-Path $appErr) { $errText = Get-Content $appErr -Raw -ErrorAction SilentlyContinue }
        if ($errText -match "Uvicorn running") {
@@ -80,7 +108,7 @@ function Start-VoiceApp([int]$p) {
            Start-Sleep -Seconds 10
        }
    }
-    if (-not (Test-PortListening $p)) {
+    if (-not (Test-AppHttpReady $p)) {
        if (Test-Path $appErr) {
            Write-Host "App startup log:" -ForegroundColor Red
            Get-Content $appErr -Tail 20 | ForEach-Object { Write-Host $_ }
@@ -90,6 +118,164 @@ function Start-VoiceApp([int]$p) {
    Write-Host "App ready: http://127.0.0.1:$p"
 }

+function Get-CloudflaredPidFile([string]$suffix) {
+    return Join-Path $Root ".cloudflared$suffix.pid"
+}
+
+function Stop-CloudflaredForSuffix([string]$suffix) {
+    $pidFile = Get-CloudflaredPidFile $suffix
+    try {
+        if (Test-Path $pidFile) {
+            $procId = [int](Get-Content $pidFile -Raw).Trim()
+            if ($procId) {
+                Stop-Process -Id $procId -Force -ErrorAction SilentlyContinue
+            }
+            Remove-Item $pidFile -Force -ErrorAction SilentlyContinue
+        }
+        Start-Sleep -Seconds 1
+    } catch { }
+}
+
+function Stop-CloudflaredTunnel {
+    Stop-CloudflaredForSuffix ""
+}
+
+function Test-TunnelUrl([string]$baseUrl, [int]$retries = 2) {
+    if (-not $baseUrl) { return $false }
+    for ($i = 0; $i -lt $retries; $i++) {
+        try {
+            $null = Invoke-WebRequest -Uri "$baseUrl/api/voice/config" -UseBasicParsing -TimeoutSec 8
+            return $true
+        } catch {
+            if ($i -lt ($retries - 1)) { Start-Sleep -Seconds 2 }
+        }
+    }
+    return $false
+}
+
+function Test-SpeechTunnelUrl([string]$baseUrl, [int]$retries = 12) {
+    if (-not $baseUrl) { return $false }
+    $paths = @("", "/ws")
+    for ($i = 0; $i -lt $retries; $i++) {
+        foreach ($path in $paths) {
+            $uri = if ($path) { "$baseUrl$path" } else { $baseUrl }
+            try {
+                $null = Invoke-WebRequest -Uri $uri -UseBasicParsing -TimeoutSec 10
+                return $true
+            } catch {
+                $code = $null
+                try {
+                    if ($_.Exception.Response) {
+                        $code = [int]$_.Exception.Response.StatusCode
+                    }
+                } catch { }
+                # Sidecar returns 404 on / — that still means the tunnel is up.
+                if ($code -and $code -ge 200 -and $code -lt 500) {
+                    return $true
+                }
+            }
+        }
+        if ($i -lt ($retries - 1)) {
+            Write-Host "  Speech tunnel not ready yet (attempt $($i + 1)/$retries)..."
+            Start-Sleep -Seconds 3
+        }
+    }
+    return $false
+}
+
+function Update-EnvValue([string]$key, [string]$value) {
+    $envFile = Join-Path $Root ".env"
+    if (-not (Test-Path $envFile)) { return }
+    $lines = Get-Content $envFile
+    $updated = $false
+    $newLines = foreach ($line in $lines) {
+        if ($line -match "^\s*$key=") {
+            $updated = $true
+            "$key=$value"
+        } else { $line }
+    }
+    if (-not $updated) { $newLines += "$key=$value" }
+    Set-Content -Path $envFile -Value ($newLines -join "`n")
+    Set-Item -Path "Env:$key" -Value $value
+}
+
+function Start-SpeechSidecar([int]$p = 8002, [switch]$Force) {
+    if ((Test-TcpPortListening $p) -and -not $Force) {
+        Write-Host "Speech Engine sidecar already listening on port $p"
+        return
+    }
+    Write-Host "Starting Speech Engine sidecar on port $p..."
+    Stop-PortProcess $p
+    if (-not (Wait-PortReleased $p)) {
+        Write-Host "Port $p still in use — forcing stop again..." -ForegroundColor Yellow
+        Stop-PortProcess $p
+        Start-Sleep -Seconds 3
+    }
+    $log = Join-Path $Root ".speech-sidecar.log"
+    $logErr = "${log}.err"
+    Clear-LogFile $log
+    Clear-LogFile $logErr
+    $env:PYTHONUNBUFFERED = "1"
+    Start-Process -FilePath "py" -ArgumentList "-3.13", "-u", "speech_engine_sidecar.py" `
+        -WorkingDirectory $Root -WindowStyle Minimized `
+        -RedirectStandardOutput $log -RedirectStandardError $logErr
+    $deadline = (Get-Date).AddSeconds(90)
+    while ((Get-Date) -lt $deadline) {
+        if (Test-TcpPortListening $p) {
+            Write-Host "Speech Engine sidecar ready on port $p"
+            return
+        }
+        if (Test-Path $log) {
+            $tail = Get-Content $log -Tail 1 -ErrorAction SilentlyContinue
+            if ($tail) { Write-Host "  $tail" }
+        }
+        Start-Sleep -Seconds 2
+    }
+    Write-Host "Sidecar logs:" -ForegroundColor Red
+    if (Test-Path $log) {
+        Get-Content $log -Tail 15 -ErrorAction SilentlyContinue | ForEach-Object { Write-Host $_ }
+    }
+    if (Test-Path $logErr) {
+        Get-Content $logErr -Tail 15 -ErrorAction SilentlyContinue | ForEach-Object { Write-Host $_ }
+    }
+    throw "Speech Engine sidecar did not start on port $p (ElevenLabs API fetch can take up to 90s — check .speech-sidecar.log)"
+}
+
+function Start-CloudflaredTunnel([int]$p, [string]$logSuffix = "") {
+    Write-Host "Starting Cloudflare tunnel for port $p..."
+    $log = Join-Path $Root ".tunnel$logSuffix.log"
+    $logErr = "${log}.err"
+    Stop-CloudflaredForSuffix $logSuffix
+    Clear-LogFile $log
+    Clear-LogFile $logErr
+    $proc = Start-Process -FilePath "cloudflared" -ArgumentList "tunnel", "--url", "http://127.0.0.1:$p" `
+        -RedirectStandardOutput $log -RedirectStandardError $logErr -WindowStyle Minimized -PassThru
+    if ($proc -and $proc.Id) {
+        Set-Content -Path (Get-CloudflaredPidFile $logSuffix) -Value $proc.Id
+    }
+    $deadline = (Get-Date).AddSeconds(90)
+    $newUrl = $null
+    while ((Get-Date) -lt $deadline -and -not $newUrl) {
+        Start-Sleep -Seconds 2
+        foreach ($path in @($logErr, $log)) {
+            if (-not (Test-Path $path)) { continue }
+            $match = Select-String -Path $path -Pattern "https://[a-z0-9]+(-[a-z0-9]+)+\.trycloudflare\.com" | Select-Object -First 1
+            if ($match) {
+                $newUrl = $match.Matches[0].Value
+                break
+            }
+        }
+    }
+    if (-not $newUrl) { throw "Could not read tunnel URL from $logErr" }
+    if ($logSuffix) {
+        Write-Host "Tunnel ready ($logSuffix): $newUrl"
+    } else {
+        Update-PublicUrl $newUrl
+        Write-Host "Tunnel ready: $newUrl"
+    }
+    return $newUrl
+}
+
 function Update-PublicUrl([string]$baseUrl) {
    $envFile = Join-Path $Root ".env"
    $content = Get-Content $envFile -Raw
@@ -147,7 +333,7 @@ Ensure-ServerSecret
 $env:VOICE_AGENT_PORT = "$Port"

 $needsRestart = $Restart.IsPresent
-if (Test-PortListening $Port) {
+if (Test-TcpPortListening $Port) {
    if ($needsRestart) {
        Stop-PortProcess $Port
    } else {
@@ -166,42 +352,20 @@ if (Test-PortListening $Port) {
    }
 }

-if ($needsRestart -or -not (Test-PortListening $Port)) {
-    Start-VoiceApp $Port
-}
-
 $public = $env:PUBLIC_API_BASE_URL
-try {
-    $null = Invoke-WebRequest -Uri "$public/api/voice/config" -UseBasicParsing -TimeoutSec 5
+$mainTunnelRecreated = $false
+if (Test-TunnelUrl $public) {
    Write-Host "Tunnel OK: $public"
-} catch {
-    Write-Host "Starting Cloudflare tunnel..."
-    $log = Join-Path $Root ".tunnel.log"
-    $logErr = "${log}.err"
-    if (Test-Path $log) { Remove-Item $log -Force }
-    if (Test-Path $logErr) { Remove-Item $logErr -Force }
-    Start-Process -FilePath "cloudflared" -ArgumentList "tunnel", "--url", "http://127.0.0.1:$Port" `
-        -RedirectStandardOutput $log -RedirectStandardError $logErr -WindowStyle Minimized
-    $deadline = (Get-Date).AddSeconds(90)
-    $newUrl = $null
-    while ((Get-Date) -lt $deadline -and -not $newUrl) {
-        Start-Sleep -Seconds 2
-        foreach ($path in @($logErr, $log)) {
-            if (-not (Test-Path $path)) { continue }
-            $match = Select-String -Path $path -Pattern "https://[a-z0-9-]+\.trycloudflare\.com" | Select-Object -First 1
-            if ($match) {
-                $newUrl = $match.Matches[0].Value
-                break
-            }
-        }
-    }
-    if (-not $newUrl) { throw "Could not read tunnel URL from $logErr" }
-    Update-PublicUrl $newUrl
-    Write-Host "Tunnel ready: $newUrl"
-    if ($needsRestart -or -not (Test-PortListening $Port)) {
-        Stop-PortProcess $Port
-        Start-VoiceApp $Port
-    }
+} else {
+    $null = Start-CloudflaredTunnel $Port
+    $mainTunnelRecreated = $true
+}
+if ($needsRestart -or -not (Test-TcpPortListening $Port)) {
+    Start-VoiceApp $Port
+} elseif ($mainTunnelRecreated) {
+    Write-Host "Main tunnel URL changed — restarting app to reload .env"
+    Stop-PortProcess $Port
+    Start-VoiceApp $Port
 }

 $config = Invoke-RestMethod -Uri "http://127.0.0.1:$Port/api/voice/config"
@@ -218,4 +382,47 @@ if ($LASTEXITCODE -eq 0) {
    Show-ProviderSetupSteps -webhookUrl $webhook -secret $env:VOICE_SERVER_URL_SECRET
 }

+if ($env:ELEVENLABS_API_KEY) {
+    Write-Host ""
+    $speechSidecarOk = Test-TcpPortListening 8002
+    $speechTunnelOk = Test-SpeechTunnelUrl $env:SPEECH_ENGINE_PUBLIC_URL
+    if (-not $speechTunnelOk) {
+        Write-Host "Speech tunnel down — starting sidecar + new Cloudflare tunnel..." -ForegroundColor Cyan
+        Start-SpeechSidecar 8002 -Force
+        $speechTunnel = Start-CloudflaredTunnel 8002 "-speech"
+        Update-EnvValue "SPEECH_ENGINE_PUBLIC_URL" $speechTunnel
+        Write-Host "Speech Engine tunnel: $speechTunnel"
+        Write-Host "Waiting for speech tunnel to accept connections..."
+        Start-Sleep -Seconds 5
+        Write-Host "Configuring ElevenLabs Speech Engine..." -ForegroundColor Cyan
+        py -3.13 setup_speech_engine.py
+        if ($LASTEXITCODE -ne 0) {
+            throw "Speech Engine setup failed — run: py -3.13 setup_speech_engine.py"
+        }
+        Write-Host "Speech Engine configured (WebSocket /ws + browser token)." -ForegroundColor Green
+        Load-DotEnv
+        Stop-PortProcess $Port
+        Start-VoiceApp $Port
+    } elseif ($Restart.IsPresent -or -not $speechSidecarOk) {
+        Write-Host "Restarting speech sidecar to reload code (keeping existing tunnel)..." -ForegroundColor Yellow
+        Start-SpeechSidecar 8002 -Force
+    } else {
+        Write-Host "Speech stack OK (sidecar :8002 + tunnel reachable)" -ForegroundColor Green
+    }
+    if (-not (Test-TcpPortListening 8002)) {
+        throw "Speech Engine sidecar not listening on 8002"
+    }
+    Load-DotEnv
+    if (-not (Test-SpeechTunnelUrl $env:SPEECH_ENGINE_PUBLIC_URL)) {
+        Write-Host ""
+        Write-Host "WARNING: Speech Engine tunnel not reachable yet: $($env:SPEECH_ENGINE_PUBLIC_URL)" -ForegroundColor Yellow
+        Write-Host "Local dashboard is up at http://127.0.0.1:$Port — wait ~30s, then retry ElevenLabs talk." -ForegroundColor Yellow
+        Write-Host "If it still fails, run: .\start_dev.ps1 -Restart" -ForegroundColor Yellow
+    } else {
+        Write-Host "Speech sidecar: port 8002 | ws_url wss://$($env:SPEECH_ENGINE_PUBLIC_URL -replace 'https://','')/ws"
+    }
+}
+
 Write-Host "Done. Dashboard: http://127.0.0.1:$Port"
+Write-Host "ElevenLabs talk: http://127.0.0.1:$Port/eleven-talk"
+Write-Host "Keep this window open — sidecar + tunnels run in the background."
@@ -7,12 +7,13 @@ import json
 import logging

 from integrations.voice.client import VoiceClient
-from integrations.voice.config import VoiceConfigError, load_voice_settings
+from integrations.voice.config import VoiceConfigError, get_web_sdk_settings, load_voice_settings
+from integrations.voice.live import transcript_hub
 from integrations.voice.processor import process_voice_webhook
 from integrations.voice.store import VoiceCallStore
 from integrations.voice.verify import VerificationError, verify_webhook_request
 from starlette.requests import Request
-from starlette.responses import JSONResponse, PlainTextResponse, Response
+from starlette.responses import JSONResponse, PlainTextResponse, Response, StreamingResponse

 logger = logging.getLogger(__name__)

@@ -21,6 +22,14 @@ logger = logging.getLogger(__name__)
 _background_tasks: set[asyncio.Task] = set()


+def _schedule_webhook_processing(event_type: str, payload: dict, *, settings) -> None:
+    task = asyncio.create_task(
+        asyncio.to_thread(process_voice_webhook, event_type, payload, settings=settings)
+    )
+    _background_tasks.add(task)
+    task.add_done_callback(_background_tasks.discard)
+
+
 async def api_voice_config_status(_request: Request) -> JSONResponse:
    try:
        settings = load_voice_settings()
@@ -36,6 +45,24 @@ async def api_voice_config_status(_request: Request) -> JSONResponse:
            "error": error,
            "webhook_url": settings.webhook_url if settings else None,
            "default_assistant_id": settings.default_assistant_id if settings else None,
+            "web_talk_enabled": bool(settings.public_api_key and settings.default_assistant_id),
+        }
+    )
+
+
+async def api_voice_web_config(_request: Request) -> JSONResponse:
+    try:
+        settings = load_voice_settings(require_key=False)
+    except VoiceConfigError as exc:
+        return JSONResponse({"enabled": False, "error": str(exc)})
+    sdk_url, sdk_global = get_web_sdk_settings()
+    return JSONResponse(
+        {
+            "enabled": bool(settings.public_api_key and settings.default_assistant_id and sdk_url),
+            "public_key": settings.public_api_key,
+            "assistant_id": settings.default_assistant_id,
+            "sdk_url": sdk_url or None,
+            "sdk_global": sdk_global or "Vapi",
        }
    )

@@ -69,15 +96,62 @@ async def api_voice_create_call(request: Request) -> JSONResponse:


 async def api_voice_list_calls(_request: Request) -> JSONResponse:
-    return JSONResponse({"calls": VoiceCallStore.list_calls()})
+    from integrations.voice.call_finalize import analytics_for_record
+
+    calls = []
+    for record in VoiceCallStore.list_calls():
+        row = dict(record)
+        row["analytics"] = analytics_for_record(record)
+        calls.append(row)
+    return JSONResponse({"calls": calls})
+
+
+async def api_voice_doctor(_request: Request) -> JSONResponse:
+    import os
+
+    from integrations.voice.doctor import run_voice_doctor
+
+    app_port = int(os.getenv("VOICE_AGENT_PORT", "8001"))
+    sidecar_port = int(os.getenv("SPEECH_ENGINE_SIDECAR_PORT", "8002"))
+    return JSONResponse(run_voice_doctor(app_port=app_port, sidecar_port=sidecar_port))


 async def api_voice_call_detail(request: Request) -> JSONResponse:
+    from integrations.voice.call_finalize import analytics_for_record
+
    call_id = request.path_params["call_id"]
    record = VoiceCallStore.get_call(call_id)
    if not record:
        return JSONResponse({"error": "not found"}, status_code=404)
-    return JSONResponse(record)
+    out = dict(record)
+    out["analytics"] = analytics_for_record(record)
+    return JSONResponse(out)
+
+
+async def api_voice_call_detail_ui(request: Request) -> JSONResponse:
+    from integrations.voice.call_detail_ui import build_call_detail_layout
+
+    call_id = request.path_params["call_id"]
+    record = VoiceCallStore.get_call(call_id)
+    if not record:
+        return JSONResponse({"error": "not found"}, status_code=404)
+    layout = build_call_detail_layout(record)
+    return JSONResponse({"call_id": call_id, "layout": layout})
+
+
+async def api_voice_live_transcript(request: Request) -> StreamingResponse:
+    call_id = request.path_params["call_id"]
+
+    async def event_stream():
+        record = VoiceCallStore.get_call(call_id)
+        if record and record.get("transcript"):
+            yield transcript_hub.snapshot_sse(
+                {"type": "snapshot", "transcript": record["transcript"]}
+            )
+        async for event in transcript_hub.subscribe(call_id):
+            yield transcript_hub.snapshot_sse(event)
+
+    return StreamingResponse(event_stream(), media_type="text/event-stream")


 async def webhook_voice(request: Request) -> Response:
@@ -105,9 +179,5 @@ async def webhook_voice(request: Request) -> Response:
        result = await asyncio.to_thread(process_voice_webhook, event_type, payload, settings=settings)
        return JSONResponse(result or {})

-    task = asyncio.create_task(
-        asyncio.to_thread(process_voice_webhook, event_type, payload, settings=settings)
-    )
-    _background_tasks.add(task)
-    task.add_done_callback(_background_tasks.discard)
+    _schedule_webhook_processing(event_type, payload, settings=settings)
    return PlainTextResponse("ok", status_code=200)
@@ -95,6 +95,9 @@ nav:
    - Templates & Layouts: concepts/templates.md
    - Slots & Components: concepts/slots.md
    - Providers: concepts/providers.md
+  - Frontend:
+    - BeautifulUI Adapter: frontend/beautifului-adapter-architecture.md
+    - BeautifulUI Upstream Pin: frontend/BEAUTIFULUI_UPSTREAM.md
  - Features:
    - Dashboard: features/dashboard.md
    - Gateway Chat: features/gateway-chat.md
@@ -129,12 +132,15 @@ nav:
    - Feature Inventory: features/feature-inventory.md
    - Gap Analysis: features/gap-analysis.md
    - Integration Analysis: features/integration-analysis.md
+    - BeautifulUI Integration: features/beautifului-integration.md
    - Standardization: features/standardization-verification.md
    - Comparison (OpenClaw): features/comparison-openclaw.md
    - What Makes It Different: features/differentiation.md
    - How to Add a Feature: features/how-to-add-a-feature.md
    - Message Elements: features/message-elements.md
    - Steps: features/steps.md
+  - Frontend:
+    - BeautifulUI Playbook: frontend/beautifului.md
  - API Reference:
    - Python CLI: api/cli.md
    - Python API: api/python.md
@@ -1,6 +1,6 @@
 [project]
 name = "aiui"
-version = "0.3.143"
+version = "0.3.147"
 description = "YAML-driven website generator - CLI and compiler"
 readme = "README.md"
 license = {text = "MIT"}
@@ -1,6 +1,27 @@
-# React + TypeScript + Vite
+# PraisonAIUI Frontend (React + TypeScript + Vite)

-This template provides a minimal setup to get React working in Vite with HMR and some ESLint rules.
+The BeautifulUI dashboard source. Built with React 19 + Vite 7 + Tailwind v4.
+
+## Build & CI
+
+- **Node.js:** 22 (matches the `frontend` job in `.github/workflows/ci.yml`)
+- **Lockfile:** `src/frontend/package-lock.json` — install with `npm ci`
+
+```bash
+cd src/frontend
+npm ci
+npm run build       # SPA build → dist/ (dashboard ship path)
+npm run build:lib   # library build for the @praisonaiui/react package
+npm run lint        # eslint
+```
+
+`npm run build` output in `dist/` is synced to `src/praisonaiui/templates/frontend/`
+(the committed bundle served by `aiui build`). The CI `frontend` job runs
+`npm ci && npm run build` on PRs to `main` and fails on TypeScript/Vite errors.
+
+## Vite template notes
+
+This project uses a minimal setup to get React working in Vite with HMR and some ESLint rules.

 Currently, two official plugins are available:

@@ -29,7 +29,8 @@
    "build": "tsc -b && vite build",
    "build:lib": "tsc -b && vite build --mode lib",
    "lint": "eslint .",
-    "preview": "vite preview"
+    "preview": "vite preview",
+    "test": "node --experimental-strip-types --test \"src/agent-ui/**/*.test.ts\""
  },
  "keywords": [
    "react",
@@ -0,0 +1,13 @@
+<!doctype html>
+<!-- Spike #293 dev harness. Open with `npm run dev` → http://localhost:5173/spike.html -->
+<html lang="en">
+  <head>
+    <meta charset="UTF-8" />
+    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+    <title>BeautifulUI spike · PraisonAIUI #293</title>
+  </head>
+  <body>
+    <div id="root"></div>
+    <script type="module" src="/src/spike/main.tsx"></script>
+  </body>
+</html>
@@ -0,0 +1,104 @@
+/**
+ * Unit tests for the Agent-UI adapter (mock events -> stable contracts).
+ *
+ * Dependency-free: uses the Node built-in test runner so it needs no extra
+ * toolchain. Run with `npm test` (from src/frontend), which invokes the Node
+ * test runner with `--experimental-strip-types` so the TypeScript types are
+ * stripped on Node 22+. The frontend CI gate (see issue #295) wires this in.
+ */
+
+import { test } from 'node:test'
+import assert from 'node:assert/strict'
+
+import { mapApprovalToCardProps, mapChatEventToUi } from './adapter.ts'
+
+test('mapChatEventToUi normalizes a streaming token frame', () => {
+    const event = mapChatEventToUi({
+        type: 'run_content',
+        token: 'Hello',
+        agent_name: 'Researcher',
+        session_id: 's1',
+        run_id: 'r1',
+    })
+    assert.equal(event.kind, 'run-token')
+    assert.equal(event.text, 'Hello')
+    assert.equal(event.agentName, 'Researcher')
+    assert.equal(event.sessionId, 's1')
+    assert.equal(event.runId, 'r1')
+})
+
+test('mapChatEventToUi maps a started tool call to a running chip', () => {
+    const event = mapChatEventToUi({
+        type: 'tool_call_started',
+        name: 'search',
+        description: '🔎 Searching the web',
+        icon: '🔎',
+        step_number: 2,
+        tool_call_id: 'tc-1',
+    })
+    assert.equal(event.kind, 'tool-call')
+    assert.deepEqual(event.toolCall, {
+        id: 'tc-1',
+        label: '🔎 Searching the web',
+        icon: '🔎',
+        stepNumber: 2,
+        status: 'running',
+        error: undefined,
+    })
+})
+
+test('mapChatEventToUi marks completed tool calls as done', () => {
+    const event = mapChatEventToUi({ type: 'tool_call_completed', name: 'search' })
+    assert.equal(event.toolCall?.status, 'done')
+    // Falls back to the tool name for id and label when unspecified.
+    assert.equal(event.toolCall?.id, 'search')
+    assert.equal(event.toolCall?.label, 'search')
+})
+
+test('mapChatEventToUi treats team_* frames the same as their solo variant', () => {
+    assert.equal(mapChatEventToUi({ type: 'team_run_content', token: 'x' }).kind, 'run-token')
+    assert.equal(mapChatEventToUi({ type: 'team_run_error', error: 'boom' }).kind, 'run-error')
+})
+
+test('mapChatEventToUi preserves unknown frame types without throwing', () => {
+    const event = mapChatEventToUi({ type: 'future_event', foo: 'bar' })
+    assert.equal(event.kind, 'unknown')
+    assert.equal(event.rawType, 'future_event')
+})
+
+test('mapChatEventToUi defaults run errors to a readable message', () => {
+    const event = mapChatEventToUi({ type: 'run_error' })
+    assert.equal(event.kind, 'run-error')
+    assert.equal(event.error, 'Unknown error')
+})
+
+test('mapApprovalToCardProps maps a pending approval item', () => {
+    const card = mapApprovalToCardProps({
+        id: 'a-1',
+        tool_name: 'delete_file',
+        agent_name: 'Ops',
+        risk_level: 'high',
+        risk_icon: '🔥',
+        description: 'Deletes a file',
+        arguments: { path: '/tmp/x' },
+        // Approvals API emits epoch seconds (time.time()), not an ISO string.
+        created_at: 1758585600,
+    })
+    assert.equal(card.id, 'a-1')
+    assert.equal(card.toolName, 'delete_file')
+    assert.equal(card.agentName, 'Ops')
+    assert.equal(card.riskLevel, 'high')
+    assert.equal(card.riskIcon, '🔥')
+    assert.equal(card.argumentsJson, '{\n  "path": "/tmp/x"\n}')
+    // Epoch seconds are normalized to an ISO-8601 string for display.
+    assert.equal(card.createdAt, '2025-09-23T00:00:00.000Z')
+})
+
+test('mapApprovalToCardProps falls back for missing/invalid fields', () => {
+    const card = mapApprovalToCardProps({ id: 'a-2', risk_level: 'nonsense' })
+    assert.equal(card.toolName, 'unknown')
+    assert.equal(card.agentName, 'Unknown')
+    assert.equal(card.riskLevel, 'medium')
+    assert.equal(card.riskIcon, '⚠️')
+    assert.equal(card.argumentsJson, '{}')
+})
@@ -0,0 +1,177 @@
+/**
+ * PraisonAIUI Agent-UI adapter
+ *
+ * Pure mapping functions from wire payloads to stable PraisonAIUI contracts.
+ * This is the single boundary between transport shapes (WebSocket frames from
+ * chat.js, REST payloads from /api/approvals/*) and the UI layer. Vendored
+ * BeautifulUI components are rendered from these contracts elsewhere in
+ * `src/agent-ui/`; consumers never import upstream prop types directly.
+ *
+ * Functions are intentionally side-effect free so they can be unit tested
+ * with mock events (see adapter.test.ts).
+ */
+
+import type {
+    AgentStreamEvent,
+    ApprovalPromptModel,
+    ApprovalRiskLevel,
+    ToolCallChipModel,
+} from './contracts'
+
+/** Raw WebSocket frame as parsed in chat.js (`JSON.parse(event.data)`). */
+export type RawChatFrame = Record<string, unknown>
+
+/** Raw approval item as returned by `/api/approvals/pending`. */
+export type RawApprovalItem = Record<string, unknown>
+
+const VALID_RISK_LEVELS: readonly ApprovalRiskLevel[] = [
+    'low',
+    'medium',
+    'high',
+    'critical',
+]
+
+function asString(value: unknown): string | undefined {
+    return typeof value === 'string' ? value : undefined
+}
+
+function asNumber(value: unknown): number | undefined {
+    return typeof value === 'number' ? value : undefined
+}
+
+/**
+ * Normalize a `created_at` value to an ISO-8601 string. The approvals API
+ * emits epoch seconds (`time.time()`), while some payloads already carry an
+ * ISO string — accept both, drop anything else.
+ */
+function asTimestamp(value: unknown): string | undefined {
+    if (typeof value === 'string') return value
+    if (typeof value === 'number' && Number.isFinite(value)) {
+        return new Date(value * 1000).toISOString()
+    }
+    return undefined
+}
+
+/** Map a raw tool-call frame to a stable chip view model. */
+function mapToolCall(
+    frame: RawChatFrame,
+    status: ToolCallChipModel['status'],
+): ToolCallChipModel {
+    const name = asString(frame.name) ?? 'tool'
+    return {
+        id: asString(frame.tool_call_id) ?? name,
+        label: asString(frame.description) ?? name,
+        icon: asString(frame.icon),
+        stepNumber: asNumber(frame.step_number),
+        status,
+        error: asString(frame.error),
+    }
+}
+
+/**
+ * Map a raw WebSocket chat frame to a normalized {@link AgentStreamEvent}.
+ *
+ * Unknown frame types are preserved as `kind: 'unknown'` with `rawType` set,
+ * so new upstream event types degrade gracefully instead of throwing.
+ */
+export function mapChatEventToUi(frame: RawChatFrame): AgentStreamEvent {
+    const rawType = asString(frame.type)
+    const sessionId = asString(frame.session_id)
+    const runId = asString(frame.run_id)
+    const agentName = asString(frame.agent_name)
+
+    switch (rawType) {
+        case 'run_started':
+            return { kind: 'run-started', sessionId, runId, agentName }
+        case 'run_content':
+        case 'team_run_content':
+            return {
+                kind: 'run-token',
+                sessionId,
+                runId,
+                agentName,
+                text: asString(frame.token),
+            }
+        case 'run_completed':
+        case 'team_run_completed':
+            return {
+                kind: 'run-completed',
+                sessionId,
+                runId,
+                agentName,
+                text: asString(frame.content),
+            }
+        case 'run_error':
+        case 'team_run_error':
+            return {
+                kind: 'run-error',
+                sessionId,
+                runId,
+                agentName,
+                error: asString(frame.error) ?? 'Unknown error',
+            }
+        case 'tool_call_started':
+        case 'team_tool_call_started':
+            return {
+                kind: 'tool-call',
+                sessionId,
+                runId,
+                agentName,
+                toolCall: mapToolCall(frame, 'running'),
+            }
+        case 'tool_call_completed':
+        case 'team_tool_call_completed':
+            return {
+                kind: 'tool-call',
+                sessionId,
+                runId,
+                agentName,
+                toolCall: mapToolCall(frame, 'done'),
+            }
+        case 'reasoning_step':
+        case 'team_reasoning_step':
+            return {
+                kind: 'reasoning-step',
+                sessionId,
+                runId,
+                agentName,
+                text: asString(frame.step),
+            }
+        default:
+            return { kind: 'unknown', sessionId, runId, agentName, rawType }
+    }
+}
+
+function normalizeRisk(value: unknown): ApprovalRiskLevel {
+    return VALID_RISK_LEVELS.includes(value as ApprovalRiskLevel)
+        ? (value as ApprovalRiskLevel)
+        : 'medium'
+}
+
+/**
+ * Map a raw approval item from `/api/approvals/pending` to a stable
+ * {@link ApprovalPromptModel}. Arguments are pretty-printed for display and
+ * an unknown/missing risk level falls back to `medium`.
+ */
+export function mapApprovalToCardProps(
+    item: RawApprovalItem,
+): ApprovalPromptModel {
+    let argumentsJson = '{}'
+    try {
+        argumentsJson = JSON.stringify(item.arguments ?? {}, null, 2)
+    } catch {
+        // Circular or non-serializable args — degrade to empty object.
+        argumentsJson = '{}'
+    }
+
+    return {
+        id: asString(item.id) ?? '',
+        toolName: asString(item.tool_name) ?? 'unknown',
+        agentName: asString(item.agent_name) ?? 'Unknown',
+        riskLevel: normalizeRisk(item.risk_level),
+        riskIcon: asString(item.risk_icon) ?? '⚠️',
+        description: asString(item.description),
+        argumentsJson,
+        createdAt: asTimestamp(item.created_at),
+    }
+}
@@ -0,0 +1,69 @@
+/**
+ * PraisonAIUI Agent-UI contracts
+ *
+ * Stable, framework-agnostic view models owned by PraisonAIUI. These are the
+ * ONLY types that non-adapter code (chat views, dashboards, tests) should
+ * depend on. BeautifulUI / vendored component prop types must never leak
+ * outside `src/agent-ui/` — the adapter maps to these contracts instead so
+ * the upstream implementation can be swapped without touching consumers.
+ *
+ * See docs/frontend/beautifului-adapter-architecture.md.
+ */
+
+/** Discriminant for a normalized agent stream event. */
+export type AgentStreamEventKind =
+    | 'run-started'
+    | 'run-token'
+    | 'run-completed'
+    | 'run-error'
+    | 'tool-call'
+    | 'reasoning-step'
+    | 'unknown'
+
+/**
+ * Normalized agent stream event — the stable shape UI code consumes,
+ * decoupled from the raw WebSocket frame `type` strings handled in chat.js.
+ */
+export interface AgentStreamEvent {
+    kind: AgentStreamEventKind
+    /** Owning session, when the frame carried one. */
+    sessionId?: string
+    /** Run identifier, when the frame carried one. */
+    runId?: string
+    /** Human-facing agent name, when present. */
+    agentName?: string
+    /** Streaming token or final content, depending on `kind`. */
+    text?: string
+    /** Populated for `kind === 'tool-call'`. */
+    toolCall?: ToolCallChipModel
+    /** Populated for `kind === 'run-error'`. */
+    error?: string
+    /** Raw frame `type` for events we do not yet normalize (`kind === 'unknown'`). */
+    rawType?: string
+}
+
+/** View model for a tool-call chip, independent of any component library. */
+export interface ToolCallChipModel {
+    id: string
+    label: string
+    icon?: string
+    stepNumber?: number
+    status: 'running' | 'done' | 'error'
+    error?: string
+}
+
+/** Risk levels surfaced by the approvals REST API (`/api/approvals/*`). */
+export type ApprovalRiskLevel = 'low' | 'medium' | 'high' | 'critical'
+
+/** View model for an approval prompt/card, independent of any component library. */
+export interface ApprovalPromptModel {
+    id: string
+    toolName: string
+    agentName: string
+    riskLevel: ApprovalRiskLevel
+    riskIcon: string
+    description?: string
+    /** Pretty-printed tool arguments, ready for display. */
+    argumentsJson: string
+    createdAt?: string
+}
@@ -0,0 +1,19 @@
+/**
+ * Agent-UI boundary barrel.
+ *
+ * Consumers import PraisonAIUI's stable agent-UI contracts and adapters from
+ * here. Vendored BeautifulUI components live under `src/agent-ui/vendors/` and
+ * are re-exported from this folder only — no upstream prop types should be
+ * imported outside `src/agent-ui/`.
+ */
+
+export type {
+    AgentStreamEvent,
+    AgentStreamEventKind,
+    ApprovalPromptModel,
+    ApprovalRiskLevel,
+    ToolCallChipModel,
+} from './contracts'
+
+export type { RawApprovalItem, RawChatFrame } from './adapter'
+export { mapApprovalToCardProps, mapChatEventToUi } from './adapter'
@@ -0,0 +1,112 @@
+# Spike #293 — BeautifulUI primitives in PraisonAIUI
+
+Time-boxed evaluation proving BeautifulUI's AI-native primitives can build inside
+PraisonAIUI's React 19 + Vite + Tailwind v4 toolchain and render against **mock,
+PraisonAIUI-shaped** agent event streams.
+
+**This folder is a spike. It is intentionally isolated from the shipped bundle:**
+nothing here is imported by `src/main.tsx`, and none of it is synced into
+`src/praisonaiui/templates/frontend`. Removing `src/frontend/src/spike/` and
+`src/frontend/spike.html` fully reverts the spike.
+
+## How to run
+
+```bash
+cd src/frontend
+npm install
+npm run dev          # then open http://localhost:5173/spike.html
+```
+
+`SpikePage.tsx` drives the three primitives from `mockEvents.ts`, which uses the
+real `SSEEvent` contract from `src/types.ts` (`run_content` deltas,
+`reasoning_step` / `tool_call_started` / `tool_call_completed`, and an
+`approval_request` payload shaped like a row from `GET /api/approvals`).
+
+## What was vendored (upstream pin)
+
+- Registry index: `https://www.beautifului.dev/r/registry.json`
+- Component JSON: `https://www.beautifului.dev/r/<name>.json`
+- Upstream repo: `slev12397/beautiful-ui` @ **`44a274e598395ab61e7c96c26fda2758780253b7`** (2026-09-22)
+- License: **MIT** (upstream repo)
+
+| Primitive | Registry item | Source | Registry deps | npm deps |
+|---|---|---|---|---|
+| StreamingText | `streaming-text` | `beautifui/primitives/StreamingText.tsx` (10.3 KB) | foundation | **none** (`react` only) |
+| ThinkingState | `thinking-state` | `beautifui/primitives/ThinkingState.tsx` (11.3 KB) | foundation | **none** (`react` only) |
+| ApprovalCard | `approval-card` | `beautifui/primitives/ApprovalCard.tsx` (16.5 KB) | foundation, button, glide-menu | **none** (`react` only) |
+| Button (dep) | `button` | `beautifui/atoms/Button.tsx` (2.0 KB) | foundation | `class-variance-authority` *(already a PraisonAIUI dep)* |
+| GlideMenu (dep) | `glide-menu` | `beautifui/primitives/GlideMenu.tsx` (1.9 KB) | foundation | **none** (`react` only) |
+| foundation | `foundation` | `beautifui/foundation.css` (17.5 KB) | — | `shadow-plugin` *(see deviation)* |
+
+### npm packages pulled by the registry & licence check
+
+For the three chosen primitives (+ their two registry deps) the registry pulls
+**zero new runtime npm packages**. The only npm dependency any of them declares is
+`class-variance-authority` (via `button`), which PraisonAIUI already ships.
+
+- `liveline`, `glimm` — **not pulled.** These are referenced only by chart-style
+  components (InsightCards) that we did not adopt. No licence review needed for this spike.
+- `@central-icons-react` — **avoided.** All three primitives render icons as
+  **inline SVG** (no icon package at all), so no `lucide-react` substitution was
+  even required. This is the cleanest possible outcome for the icon concern.
+
+## Deviations from a raw `shadcn add`
+
+1. **`shadow-plugin` removed.** `foundation.css` imports `shadow-plugin/unprefixed`
+   (an extra Tailwind plugin dependency). To keep the spike lightweight we replaced
+   it with a small local `@theme` block defining the `--shadow-*` scale. Visual
+   shadows are approximate but faithful.
+2. **Unbalanced-brace fix (upstream bug).** The flattened `foundation.css` emits a
+   **stray extra `}`** around `.records-footer-hint` (a `@media` block that lost its
+   opening line in the registry export). This makes Tailwind v4 fail with
+   `Missing opening {`. We removed the stray brace. **This is a real upstream blocker
+   worth reporting to `beautiful-ui`.**
+3. **Imports rewritten** from `@/components/*` to `@/spike/beautifui/*` so the spike
+   stays isolated and cannot collide with the existing shadcn `components/ui/button.tsx`.
+4. **`"use client"` directives dropped** (Vite SPA, not RSC).
+
+## Build result
+
+`vite build` of the spike entry (`spike.html`) succeeds:
+
+```
+✓ 39 modules transformed
+spike.css   97.82 kB │ gzip: 17.48 kB   (Tailwind + foundation)
+spike.js   247.94 kB │ gzip: 78.52 kB   (React runtime + 3 primitives + 2 deps)
+```
+
+- **Bundle size:** the primitives are cheap — self-contained TSX with no extra npm
+  deps. The JS weight is dominated by the React runtime; the incremental cost of the
+  three primitives is on the order of ~40 KB raw / low-teens KB gzip. The CSS is
+  full Tailwind + the ~17 KB foundation token/keyframe layer.
+- **Motion / `prefers-reduced-motion`:** foundation ships two
+  `@media (prefers-reduced-motion: reduce)` guards (global animation kill-switch +
+  a StreamText-specific one). Motion is handled responsibly out of the box.
+- **Dark mode parity:** foundation defines its **own** token namespace (`--page`,
+  `--canvas`, `--ink`, `--accent`, …) with its own `.dark` block, *separate* from
+  PraisonAIUI's shadcn tokens (`--background`/`--foreground` via `hsl(var(--…))`).
+  Good news: **no token collision** with the dashboard. Trade-off: the primitives do
+  **not** automatically inherit `chat.css`/YAML-driven theme colours — they'd need a
+  token bridge (map `--ink`→`--foreground`, `--surface`→`--card`, `--accent`→`--primary`)
+  before they visually match the shipped chat surface in both themes.
+
+## Assessment
+
+**Fit score: 7/10** for a spike, with caveats for production migration.
+
+**Blockers / gaps:**
+- Upstream `foundation.css` brace bug (fixed here; report upstream).
+- No automatic theme inheritance — needs a token bridge to match `chat.css` dark/light.
+- Primitives embed **demo content** (ice-cream sales copy, fake sources). They accept
+  props (`content`, `rows`, `questions`, `onSubmitted`, …) so they *can* be driven by
+  real events — proven in `SpikePage` — but the built-in defaults must be replaced.
+- PraisonAIUI already has native equivalents: `chat/ThinkingSteps.tsx`,
+  `chat/ToolCallDisplay.tsx`, streaming in `chat/ChatMessages.tsx`. Adopting
+  BeautifulUI should be weighed against enhancing these rather than duplicating them.
+
+**Recommended primitive order for migration (lowest risk → highest):**
+1. **StreamingText** — pure token-stream UX, `react`-only, maps 1:1 to `run_content`.
+2. **ThinkingState** — reasoning/tool trace, `react`-only, maps to
+   `reasoning_step` + `tool_call_*`; overlaps existing `ThinkingSteps.tsx`.
+3. **ApprovalCard** — highest value (HITL) but most surface: pulls `button` +
+   `glide-menu` and needs wiring to `POST /api/approvals/{id}/resolve`.
@@ -0,0 +1,90 @@
+/* Spike #293 — BeautifulUI primitives driven by PraisonAIUI-shaped mock events.
+ *
+ * Dev-only harness. Run with `npm run dev` and open /spike.html. This page is
+ * NOT imported by the shipped library entry (main.tsx) and is NOT synced into
+ * src/praisonaiui/templates/frontend — per the issue's non-goals it must not
+ * touch the production bundle.
+ */
+import { useEffect, useState } from "react";
+import StreamingText from "@/spike/beautifui/primitives/StreamingText";
+import ThinkingState from "@/spike/beautifui/primitives/ThinkingState";
+import ApprovalCard from "@/spike/beautifui/primitives/ApprovalCard";
+import {
+  streamingTimeline,
+  tokensFromDeltas,
+  rowsFromReasoning,
+  REASONING_STREAM,
+  APPROVAL_REQUEST,
+} from "@/spike/mockEvents";
+
+function Panel({ title, children }: { title: string; children: React.ReactNode }) {
+  return (
+    <section style={{ marginBottom: 40 }}>
+      <h2 style={{ font: "600 13px/1.4 ui-monospace, monospace", opacity: 0.6, marginBottom: 12 }}>{title}</h2>
+      {children}
+    </section>
+  );
+}
+
+export default function SpikePage() {
+  const [dark, setDark] = useState(false);
+  const [log, setLog] = useState<string[]>([]);
+
+  useEffect(() => {
+    document.documentElement.classList.toggle("dark", dark);
+  }, [dark]);
+
+  // Emulate the SSE timeline so we exercise real delta pacing (logs only —
+  // the primitives below are fed the fully-resolved token/row arrays).
+  useEffect(() => {
+    const frames = streamingTimeline();
+    const timers = frames.map((f) =>
+      setTimeout(() => setLog((l) => [...l.slice(-4), `${f.event.type}: ${f.event.token ?? ""}`]), f.at),
+    );
+    return () => timers.forEach(clearTimeout);
+  }, []);
+
+  const tokens = tokensFromDeltas(streamingTimeline());
+  const rows = rowsFromReasoning(REASONING_STREAM);
+
+  return (
+    <div style={{ maxWidth: 720, margin: "0 auto", padding: 32, fontFamily: "system-ui, sans-serif" }}>
+      <header style={{ display: "flex", justifyContent: "space-between", alignItems: "center", marginBottom: 32 }}>
+        <div>
+          <h1 style={{ font: "700 20px/1.2 system-ui" }}>BeautifulUI spike</h1>
+          <p style={{ opacity: 0.6, fontSize: 13 }}>Issue #293 · mock PraisonAIUI event streams</p>
+        </div>
+        <button onClick={() => setDark((d) => !d)} style={{ padding: "6px 12px", borderRadius: 8 }}>
+          {dark ? "☀ light" : "🌙 dark"}
+        </button>
+      </header>
+
+      <Panel title="StreamingText ← run_content deltas">
+        <StreamingText content={tokens} loop={false} sources={[]} followUps={[]} />
+      </Panel>
+
+      <Panel title="ThinkingState ← reasoning_step / tool_call_*">
+        <ThinkingState rows={rows} active="Thinking" done={`Ran ${rows.length} steps`} />
+      </Panel>
+
+      <Panel title="ApprovalCard ← approval_request (/api/approvals)">
+        <ApprovalCard
+          questions={APPROVAL_REQUEST.questions}
+          onSubmitted={(answers) =>
+            // In production this maps to POST /api/approvals/{id}/resolve.
+            setLog((l) => [
+              ...l.slice(-4),
+              `resolve ${APPROVAL_REQUEST.id}: ${JSON.stringify(answers)}`,
+            ])
+          }
+        />
+      </Panel>
+
+      <Panel title="event log (last 5)">
+        <pre style={{ font: "12px/1.5 ui-monospace, monospace", opacity: 0.7, whiteSpace: "pre-wrap" }}>
+          {log.join("\n") || "…"}
+        </pre>
+      </Panel>
+    </div>
+  );
+}
@@ -0,0 +1,48 @@
+/* Vendored from BeautifulUI registry: https://www.beautifului.dev/r/button.json
+ * Spike #293 — DO NOT ship in templates/frontend. Imports rewritten to @/spike/beautifui/*.
+ * License: MIT (upstream https://github.com/slev12397/beautiful-ui). */
+import type { ButtonHTMLAttributes } from "react";
+import { cva, type VariantProps } from "class-variance-authority";
+import { cn } from "@/lib/utils";
+
+const filledShadow = "shadow-[inset_0_1px_0_rgba(255,255,255,0.14)]";
+
+/* Pill-shaped by default — the app's core button style. Explicit symmetric
+ * padding (not a fixed height) so the top/bottom spacing is always equal. */
+export const buttonVariants = cva(
+  `inline-flex items-center justify-center font-medium select-none
+   transition-[transform,background-color,opacity] duration-150 ease-out
+   active:scale-[0.96] disabled:opacity-50 disabled:pointer-events-none`,
+  {
+    variants: {
+      variant: {
+        primary: `bg-ink text-canvas hover:opacity-90 dark:bg-ink dark:text-canvas ${filledShadow}`,
+        secondary: "bg-surface text-ink shadow-btn hover:bg-inset aria-expanded:bg-hover",
+        ghost: "bg-hover-2 text-ink hover:bg-line-strong",
+        accent: `bg-accent text-white hover:bg-accent-ink ${filledShadow}`,
+        success: `bg-green text-white hover:brightness-95 ${filledShadow}`,
+        /* transparent until hovered — for dense toolbars/action rows */
+        quiet: "text-ink hover:bg-hover",
+      },
+      size: {
+        /* compact toolbar pill — fixed height, lighter weight */
+        xs: "h-7 rounded-full px-2.5 text-[12px] font-normal leading-none gap-1",
+        /* canonical action pill — 27px tall, roomy sides */
+        sm: "h-[27px] px-3 text-[13px] leading-none rounded-full gap-1.5",
+        md: "px-4 py-[9px] text-sm leading-none rounded-full gap-2",
+      },
+    },
+    defaultVariants: { variant: "secondary", size: "md" },
+  },
+);
+
+export type ButtonVariant = NonNullable<VariantProps<typeof buttonVariants>["variant"]>;
+
+export function Button({
+  variant,
+  size,
+  className,
+  ...props
+}: ButtonHTMLAttributes<HTMLButtonElement> & VariantProps<typeof buttonVariants>) {
+  return <button className={cn(buttonVariants({ variant, size }), className)} {...props} />;
+}
@@ -0,0 +1,659 @@
+/* Vendored from BeautifulUI foundation: https://www.beautifului.dev/r/foundation.json
+ * Spike #293. shadow-plugin dependency neutralized (see note below). License: MIT. */
+@import "tailwindcss";
+/* shadow-plugin (flornkm) — smooth, layered shadow stacks. The unprefixed
+ * entry maps them onto Tailwind's --shadow-xs…2xl scale; our semantic tokens
+ * below build on those values so the whole app picks up the smooth shadows. */
+/* @import "shadow-plugin/unprefixed"; -- removed for spike */
+/* SPIKE deviation (#293): upstream imports `shadow-plugin/unprefixed` (extra npm dep).
+ * To keep the spike lightweight we define the --shadow-* scale locally instead. */
+@theme inline {
+  --shadow-2xs: 0 1px 0 0 oklch(0 0 0 / 0.03);
+  --shadow-xs: 0 1px 2px 0 oklch(0 0 0 / 0.05);
+  --shadow-sm: 0 1px 2px 0 oklch(0 0 0 / 0.06), 0 1px 1px -1px oklch(0 0 0 / 0.04);
+  --shadow-md: 0 2px 4px -1px oklch(0 0 0 / 0.07), 0 1px 2px -1px oklch(0 0 0 / 0.05);
+  --shadow-lg: 0 8px 16px -4px oklch(0 0 0 / 0.10), 0 2px 4px -2px oklch(0 0 0 / 0.06);
+  --shadow-xl: 0 16px 32px -8px oklch(0 0 0 / 0.12);
+  --shadow-2xl: 0 24px 48px -12px oklch(0 0 0 / 0.18);
+}
+
+
+@custom-variant dark (&:where(.dark, .dark *));
+
+/* GENERIC FOUNDATION — tokens, @theme, base rules, the primitive-*
+ * spacing utilities, and shared keyframes below are needed by EVERY
+ * component. Blocks tagged COMPONENT-SPECIFIC further down belong to
+ * one primitive each. */
+/* ──────────────────────────────────────────────────────────
+ * Beautiful UI — design tokens
+ * Cool near-white canvas, white cards, hairline borders, layered
+ * single-digit-opacity shadows, a neutral ink ramp, and semantic
+ * color used sparingly as a condiment.
+ * ────────────────────────────────────────────────────────── */
+
+:root {
+  /* surfaces — cool, blue-tinted neutrals */
+  --page: oklch(0.985 0.001 286.376);
+  --canvas: oklch(0.961 0.002 247.84);
+  --surface: oklch(1 0 0);
+  --inset: oklch(0.979 0.002 247.839);
+  --hover: oklch(0.97 0.002 247.839);
+  --hover-2: oklch(0.933 0.003 247.86);
+
+  /* ink ramp */
+  --ink: oklch(0.247 0.006 258.361);
+  --ink-2: oklch(0.506 0.01 264.477);
+  --ink-3: oklch(0.695 0.009 264.505);
+
+  /* borders — solid and crisp, not alpha */
+  --line: oklch(0.946 0.003 264.542);
+  --line-strong: oklch(0.912 0.005 258.326);
+  --line-soft: oklch(0.966 0.002 264.542);
+  --grid-line: color-mix(in srgb, var(--line) 78%, transparent);
+  --field: oklch(0.961 0.001 286.375);
+  --stripe: oklch(0.405 0 0 / 0.075);
+  --stripe-bg: oklch(0.97 0 0);
+
+  /* accent */
+  --accent: oklch(0.626 0.205 254.947);
+  --accent-ink: oklch(0.556 0.187 255.617);
+  --accent-tint: oklch(0.96 0.019 252.878);
+
+  /* semantic */
+  --green: oklch(0.603 0.155 150.883);
+  --green-tint: oklch(0.958 0.017 159.118);
+  --orange: oklch(0.689 0.179 49.902);
+  --orange-tint: oklch(0.964 0.021 67.581);
+  --red: oklch(0.621 0.192 23.042);
+  --red-tint: oklch(0.956 0.017 17.462);
+
+  /* chart overlays — neutral system colors with enough contrast on either theme */
+  --tooltip-bg: oklch(0.272 0.008 264.435);
+  --tooltip-fg: oklch(0.976 0.002 247.839);
+  --tooltip-muted: oklch(0.731 0.008 260.731);
+  --tooltip-border: oklch(0.356 0.007 264.474);
+
+  /* shadows — our solid hairline ring + shadow-plugin's smooth layered stacks
+   * (--shadow-xs…lg come from shadow-plugin/unprefixed) so every surface reads
+   * consistently across the app. */
+  --shadow-hairline: 0 0 0 1px var(--line);
+  --shadow-btn: 0 0 0 1px var(--line-strong), var(--shadow-xs);
+  --shadow-card: 0 0 0 1px var(--line), var(--shadow-sm);
+  --shadow-raised: 0 0 0 1px var(--line), var(--shadow-md);
+  --shadow-overlay: 0 0 0 1px var(--line), var(--shadow-lg);
+  --shadow-inset-field: inset 0 1px 2px oklch(0 0 0 / 0.12);
+}
+
+.dark {
+  --page: oklch(0.209 0.004 264.477);
+  --canvas: oklch(0.231 0.004 264.487);
+  --surface: oklch(0.26 0.006 271.191);
+  --inset: oklch(0.243 0.004 264.492);
+  --hover: oklch(0.289 0.006 271.22);
+  --hover-2: oklch(0.318 0.007 274.747);
+
+  --ink: oklch(0.964 0.002 247.839);
+  --ink-2: oklch(0.731 0.008 260.731);
+  --ink-3: oklch(0.541 0.01 264.484);
+
+  --line: oklch(0.308 0.006 258.354);
+  --line-strong: oklch(0.356 0.007 264.474);
+  --line-soft: oklch(0.278 0.006 258.354);
+  --field: oklch(0.293 0.006 271.223);
+  --stripe: oklch(1 0 0 / 0.055);
+  --stripe-bg: oklch(0.226 0.004 264.485);
+
+  --accent: oklch(0.68 0.173 253.301);
+  --accent-ink: oklch(0.788 0.113 248.33);
+  --accent-tint: oklch(0.68 0.173 253.301 / 0.16);
+
+  --green: oklch(0.705 0.154 153.814);
+  --green-tint: oklch(0.705 0.154 153.814 / 0.14);
+  --orange: oklch(0.746 0.156 55.642);
+  --orange-tint: oklch(0.746 0.156 55.642 / 0.14);
+  --red: oklch(0.666 0.18 21.433);
+  --red-tint: oklch(0.666 0.18 21.433 / 0.14);
+
+  --tooltip-bg: oklch(0.182 0.004 264.459);
+  --tooltip-fg: oklch(0.964 0.002 247.839);
+  --tooltip-muted: oklch(0.731 0.008 260.731);
+  --tooltip-border: oklch(0.308 0.006 258.354);
+
+  --shadow-hairline: 0 0 0 1px var(--line);
+  /* Elevation rings go to a subtle *light* hairline in dark mode: the ring
+   * paints over the surface behind the element, so a low-alpha white edge
+   * dissolves into the shadow as one continuous stroke (per smooth-shadow-ring),
+   * graduated by how far the surface floats. */
+  --shadow-btn: 0 0 0 1px oklch(1 0 0 / 0.1), 0 1px 2px oklch(0 0 0 / 0.3);
+  --shadow-card:
+    0 0 0 1px oklch(1 0 0 / 0.11), 0 1px 2px oklch(0 0 0 / 0.2),
+    0 2px 6px oklch(0 0 0 / 0.2);
+  --shadow-raised:
+    0 0 0 1px oklch(1 0 0 / 0.13), 0 2px 10px oklch(0 0 0 / 0.22);
+  --shadow-overlay:
+    0 0 0 1px oklch(1 0 0 / 0.15), 0 8px 28px oklch(0 0 0 / 0.34);
+  --shadow-inset-field: inset 0 1px 2px oklch(0 0 0 / 0.4);
+}
+
+@theme inline {
+  /* colors */
+  --color-page: var(--page);
+  --color-canvas: var(--canvas);
+  --color-surface: var(--surface);
+  --color-inset: var(--inset);
+  --color-ink: var(--ink);
+  --color-ink-2: var(--ink-2);
+  --color-ink-3: var(--ink-3);
+  --color-hover: var(--hover);
+  --color-hover-2: var(--hover-2);
+  --color-field: var(--field);
+  --color-line: var(--line);
+  --color-line-strong: var(--line-strong);
+  --color-accent: var(--accent);
+  --color-accent-ink: var(--accent-ink);
+  --color-accent-tint: var(--accent-tint);
+  --color-green: var(--green);
+  --color-green-tint: var(--green-tint);
+  --color-orange: var(--orange);
+  --color-orange-tint: var(--orange-tint);
+  --color-red: var(--red);
+  --color-red-tint: var(--red-tint);
+
+  /* shadows */
+  --shadow-hairline: var(--shadow-hairline);
+  --shadow-btn: var(--shadow-btn);
+  --shadow-inset-field: var(--shadow-inset-field);
+  --shadow-card: var(--shadow-card);
+  --shadow-raised: var(--shadow-raised);
+  --shadow-overlay: var(--shadow-overlay);
+
+  /* type */
+  --font-sans: var(--font-inter), ui-sans-serif, system-ui, sans-serif;
+  --font-mono: var(--font-mono-face), ui-monospace, "SF Mono", monospace;
+
+  /* radii — chips 6, controls 8, cards 10, windows 14, pills 999 */
+  --radius-chip: 6px;
+  --radius-control: 8px;
+  --radius-card: 10px;
+  --radius-window: 14px;
+
+  /* easing — strong curves; built-ins are too weak */
+  --ease-out-strong: cubic-bezier(0.23, 1, 0.32, 1);
+  --ease-in-out-strong: cubic-bezier(0.77, 0, 0.175, 1);
+  --ease-link: cubic-bezier(0.16, 1, 0.3, 1);
+}
+
+/* ── base ─────────────────────────────────────────────── */
+
+html {
+  -webkit-font-smoothing: antialiased;
+  -moz-osx-font-smoothing: grayscale;
+  text-rendering: optimizeLegibility;
+}
+
+body {
+  background-color: var(--stripe-bg);
+  background-image: repeating-linear-gradient(
+    -45deg,
+    transparent 0,
+    transparent 7px,
+    var(--stripe) 7px,
+    var(--stripe) 8px
+  );
+  /* Pin the stripe pattern to the viewport so it stays put on scroll. */
+  background-attachment: fixed;
+  color: var(--ink);
+  font-feature-settings: "cv11", "ss01";
+  font-size: 14px;
+  letter-spacing: -0.01em;
+  line-height: 1.5;
+}
+
+::selection {
+  background: var(--accent-tint);
+  color: var(--ink);
+}
+
+/* ── interface hygiene (interfaces.dev cheat sheet) ────── */
+
+/* headings balance; descriptions avoid orphans (skip long-form — we have none) */
+h1,
+h2,
+h3,
+h4 {
+  text-wrap: balance;
+}
+
+p {
+  text-wrap: pretty;
+}
+
+/* underlines clear descenders */
+a {
+  text-underline-position: from-font;
+  text-decoration-skip-ink: auto;
+}
+
+/* theme switches swap every token at once — freeze transitions so the flip
+ * is one clean repaint instead of hundreds of mismatched color fades */
+.theme-switching *,
+.theme-switching *::before,
+.theme-switching *::after {
+  transition: none !important;
+}
+
+/* touch: :hover sticks after a tap, so neutralize row/header hovers there */
+@media (hover: none) {
+  .records-row:hover > .records-cell {
+    background: transparent;
+  }
+
+  .records-row:hover > td.is-colsel {
+    background: color-mix(in srgb, var(--accent) 4%, var(--surface));
+  }
+
+  .records-header-button:hover {
+    background: transparent;
+    color: var(--ink-2);
+  }
+}
+
+/* Keep the component nav scrollable but hide its scrollbar. */
+.component-nav-scroll {
+  scrollbar-width: none; /* Firefox */
+  -ms-overflow-style: none; /* legacy Edge */
+}
+.component-nav-scroll::-webkit-scrollbar {
+  display: none; /* WebKit */
+}
+
+@media (min-width: 64rem) {
+  .component-nav-scroll {
+    -webkit-mask-image: linear-gradient(
+      to bottom,
+      oklch(0 0 0) 0,
+      oklch(0 0 0) calc(100% - 6rem),
+      transparent 100%
+    );
+    mask-image: linear-gradient(
+      to bottom,
+      oklch(0 0 0) 0,
+      oklch(0 0 0) calc(100% - 6rem),
+      transparent 100%
+    );
+  }
+}
+
+/* Mobile demo gestures should operate controls and visualizations rather than
+ * selecting surrounding labels. Text entry and code remain intentionally
+ * selectable. */
+@media (hover: none) and (pointer: coarse) {
+  .primitive-showcase {
+    -webkit-user-select: none;
+    user-select: none;
+    -webkit-touch-callout: none;
+  }
+
+  .primitive-showcase :is(input, textarea, [contenteditable="true"], pre, code) {
+    -webkit-user-select: text;
+    user-select: text;
+    -webkit-touch-callout: default;
+  }
+
+  .primitive-showcase :is(a, button, [role="button"]) {
+    -webkit-tap-highlight-color: transparent;
+  }
+
+  .primitive-showcase :is(a, button, [role="button"]):focus-visible {
+    outline: none;
+  }
+
+  .primitive-showcase :is(img, svg, canvas) {
+    -webkit-user-drag: none;
+  }
+
+  .insight-chart-stage {
+    touch-action: pan-y;
+  }
+}
+
+/* Tailwind v4 resets buttons to cursor:default — anything
+ * clickable must invite the click */
+button:not(:disabled),
+[role="button"]:not([aria-disabled="true"]),
+label:has(input),
+a {
+  cursor: pointer;
+}
+
+
+:focus-visible {
+  outline: 2px solid var(--accent);
+  outline-offset: 2px;
+  border-radius: 4px;
+}
+
+input:focus-visible,
+textarea:focus-visible,
+select:focus-visible,
+[contenteditable]:focus-visible {
+  outline: none;
+}
+
+.animated-underline {
+  position: relative;
+  display: inline-block;
+}
+
+.animated-underline::after {
+  position: absolute;
+  right: 0;
+  bottom: -1px;
+  left: 0;
+  height: 1px;
+  content: "";
+  background: currentColor;
+  transform: scaleX(0);
+  transform-origin: left center;
+  transition: transform 280ms var(--ease-link);
+}
+
+a:hover .animated-underline::after,
+a:focus-visible .animated-underline::after {
+  transform: scaleX(1);
+}
+
+/* ══════════════════════════════════════════════════════════
+ * COMPONENT-SPECIFIC — StreamingText / ContextCards
+ * Favicon source avatars.
+ * Copy this block only alongside StreamingText / ContextCards; it is not part of the
+ * generic foundation above.
+ * ══════════════════════════════════════════════════════════ */
+.source-avatar {
+  box-shadow: 0 0 0 1px oklch(0.21 0.034 263.436 / 0.1);
+}
+
+.dark .source-avatar {
+  box-shadow: 0 0 0 1px oklch(1 0 0 / 0.08);
+}
+
+/* ── primitive spacing system ─────────────────────────────
+ * Shared optical offsets for small demo cards. Use these
+ * instead of hand-tuning 10/12/14px per component.
+ */
+
+.primitive-card-pad {
+  padding: 12px;
+}
+
+.primitive-card-bar {
+  padding: 10px 12px;
+}
+
+.primitive-card-footer {
+  padding: 10px;
+}
+
+.primitive-table-cell {
+  padding: 10px 12px;
+}
+
+.primitive-icon-button {
+  display: inline-flex;
+  width: 28px;
+  height: 28px;
+  align-items: center;
+  justify-content: center;
+  border-radius: var(--radius-control);
+}
+
+/* ── shared keyframes ─────────────────────────────────── */
+
+@keyframes shimmer-text {
+  from {
+    background-position: 150% center;
+  }
+  to {
+    background-position: -50% center;
+  }
+}
+
+@keyframes fade-up {
+  from {
+    opacity: 0;
+    transform: translateY(8px);
+  }
+  to {
+    opacity: 1;
+    transform: translateY(0);
+  }
+}
+
+.home-reveal {
+  will-change: opacity, transform, filter;
+}
+
+/* The sidebar is one persistent 224px tree clipped by a 52px shell.
+ * Every rail glyph remains centered at x=26 in both states. */
+  .records-footer-hint {
+    display: none;
+  }
+/* SPIKE fix (#293): upstream foundation.css emits a stray extra `}` here — a
+ * flattened @media block that lost its opening line. Removed so Tailwind v4
+ * can parse the stylesheet. See spike report. */
+
+/* ── filter table electric status colors ─────────────────── */
+
+/* ══════════════════════════════════════════════════════════
+ * COMPONENT-SPECIFIC — FilterTable
+ * Electric status-pill colors.
+ * Copy this block only alongside FilterTable; it is not part of the
+ * generic foundation above.
+ * ══════════════════════════════════════════════════════════ */
+.filter-status-todo,
+.filter-status-progress,
+.filter-status-done {
+  border: 1px solid transparent;
+}
+
+/* ══════════════════════════════════════════════════════════
+ * COMPONENT-SPECIFIC — InsightCards
+ * Liveline chart tooltip + cursor.
+ * Copy this block only alongside InsightCards; it is not part of the
+ * generic foundation above.
+ * ══════════════════════════════════════════════════════════ */
+.insight-chart-stage {
+  overflow: hidden;
+}
+
+.insight-chart-cursor {
+  position: absolute;
+  z-index: 4;
+  top: 0;
+  bottom: 0;
+  width: 1px;
+  background: var(--ink);
+  opacity: 0.26;
+  pointer-events: none;
+}
+
+.insight-chart-tooltip-anchor {
+  position: absolute;
+  z-index: 5;
+  top: 8px;
+  transform: translateX(-50%);
+  pointer-events: none;
+}
+
+.insight-chart-tooltip {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+  padding: 5px 9px;
+  border: 1px solid var(--line-strong);
+  border-radius: 8px;
+  color: var(--tooltip-fg);
+  background: var(--tooltip-bg);
+  box-shadow: var(--shadow-overlay);
+  font-size: 11.5px;
+  font-variant-numeric: tabular-nums;
+  white-space: nowrap;
+}
+
+.insight-chart-tooltip-item {
+  display: inline-flex;
+  align-items: center;
+  gap: 6px;
+}
+
+.insight-chart-tooltip-dot {
+  width: 8px;
+  height: 8px;
+  flex: 0 0 8px;
+  border-radius: 50%;
+}
+
+/* Status pills share the same base-hue → color-mix derivation as .records-tag,
+ * so both components stay consistent across light and dark. */
+.filter-status-todo {
+  --tag-base: oklch(0.757 0.153 66.401);
+}
+.filter-status-progress {
+  --tag-base: oklch(0.671 0.118 219.351);
+}
+.filter-status-done {
+  --tag-base: oklch(0.652 0.131 162.865);
+}
+
+.filter-status-todo,
+.filter-status-progress,
+.filter-status-done {
+  color: color-mix(in srgb, var(--tag-base) 92%, var(--ink));
+  background: color-mix(in srgb, var(--tag-base) 20%, var(--surface));
+  border-color: color-mix(in srgb, var(--tag-base) 34%, var(--surface));
+}
+
+/* Dark mode — a richer, saturated same-hue fill
+ * with bright same-hue text and no visible border (light mode keeps the soft
+ * tinted pastels above). Applies to both the records grid and status pills. */
+.dark .records-tag,
+.dark .filter-status-todo,
+.dark .filter-status-progress,
+.dark .filter-status-done {
+  color: color-mix(in srgb, var(--tag-base) 86%, white);
+  background: color-mix(in srgb, var(--tag-base) 34%, var(--surface));
+  border-color: color-mix(in srgb, var(--tag-base) 34%, var(--surface));
+}
+
+@keyframes fade-in {
+  from {
+    opacity: 0;
+  }
+  to {
+    opacity: 1;
+  }
+}
+
+@keyframes eq-bounce {
+  0%,
+  100% {
+    transform: scaleY(0.35);
+  }
+  50% {
+    transform: scaleY(1);
+  }
+}
+
+/* ── StreamText primitive ─────────────────────────────────
+ * Fast character reveal: the newest few characters carry a
+ * soft, masked blur so the leading edge resolves smoothly as
+ * text flows in. The caret holds solid while streaming, then
+ * blinks once the string settles. */
+/* ══════════════════════════════════════════════════════════
+ * COMPONENT-SPECIFIC — StreamText
+ * The streaming caret + tail.
+ * Copy this block only alongside StreamText; it is not part of the
+ * generic foundation above.
+ * ══════════════════════════════════════════════════════════ */
+.stream-tail {
+  filter: blur(1.6px);
+  -webkit-mask-image: linear-gradient(to right, oklch(0 0 0) 20%, oklch(0 0 0 / 0.2));
+  mask-image: linear-gradient(to right, oklch(0 0 0) 20%, oklch(0 0 0 / 0.2));
+}
+
+.stream-caret {
+  display: inline-block;
+  width: 2px;
+  height: 1.05em;
+  margin-left: 1.5px;
+  translate: 0 -0.5px;
+  border-radius: 1px;
+  background: var(--ink);
+  vertical-align: text-bottom;
+  animation: caret-blink 1s step-end infinite;
+}
+
+.stream-caret.is-streaming {
+  animation: none;
+}
+
+@keyframes caret-blink {
+  0%,
+  100% {
+    opacity: 1;
+  }
+  50% {
+    opacity: 0;
+  }
+}
+
+@media (prefers-reduced-motion: reduce) {
+  .stream-caret {
+    animation: none;
+  }
+  .stream-tail {
+    filter: none;
+    -webkit-mask-image: none;
+    mask-image: none;
+  }
+}
+
+@keyframes pop-in {
+  from {
+    opacity: 0;
+    transform: scale(0.95);
+  }
+  to {
+    opacity: 1;
+    transform: scale(1);
+  }
+}
+
+@keyframes spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+
+/* pixel-grid loader — brighten plateau, staggered per cell */
+@keyframes pixel-on {
+  0%, 100% {
+    opacity: 0.15;
+  }
+  18%, 42% {
+    opacity: 1;
+  }
+  62% {
+    opacity: 0.15;
+  }
+}
+
+@media (prefers-reduced-motion: reduce) {
+  *,
+  *::before,
+  *::after {
+    animation-duration: 0.01ms !important;
+    animation-iteration-count: 1 !important;
+    transition-duration: 0.01ms !important;
+  }
+}
@@ -0,0 +1,420 @@
+/* Vendored from BeautifulUI registry: https://www.beautifului.dev/r/approval-card.json
+ * Spike #293 — DO NOT ship in templates/frontend. Imports rewritten to @/spike/beautifui/*.
+ * License: MIT (upstream https://github.com/slev12397/beautiful-ui). */
+import { useEffect, useLayoutEffect, useRef, useState, type CSSProperties } from "react";
+import { Button } from "@/spike/beautifui/atoms/Button";
+import GlideMenu from "@/spike/beautifui/primitives/GlideMenu";
+
+/* ─────────────────────────────────────────────────────────
+ * APPROVAL CARD (human-in-the-loop)
+ * One question at a time. The stack slides vertically as you
+ * move between questions (the card's height animates to fit),
+ * the step counter rolls like an odometer, and the footer uses
+ * pill actions — a quiet Skip and a dark Continue with a ⏎.
+ * Single-choice answers auto-advance; multi-select waits.
+ * ───────────────────────────────────────────────────────── */
+
+export type ApprovalQuestion = {
+  q: string;
+  type: "radio" | "check";
+  options: string[];
+};
+
+const QUESTIONS: ApprovalQuestion[] = [
+  {
+    q: "How many flavors should we launch?",
+    type: "radio",
+    options: ["Three (core line)", "Five (full case)", "Just one hero"],
+  },
+  {
+    q: "Which mix-ins should we stock?",
+    type: "check",
+    options: ["Chocolate chips", "Waffle bits", "Sprinkles"],
+  },
+  {
+    q: "Which market do we enter first?",
+    type: "radio",
+    options: ["Food trucks", "Grocery freezers", "Scoop shops"],
+  },
+];
+
+export type ApprovalLabels = {
+  skip: string;
+  continue: string;
+  send: string;
+  customPlaceholder: string;
+  sentMessage: string;
+};
+
+const DEFAULT_LABELS: ApprovalLabels = {
+  skip: "Skip",
+  continue: "Continue",
+  send: "Send",
+  customPlaceholder: "Something else…",
+  sentMessage: "Answers sent",
+};
+
+const ROLL_MS = 400;
+const SLIDE = "360ms cubic-bezier(0.22, 1, 0.36, 1)";
+
+/* odometer digits — each character that changes rolls up (or down) */
+function RollingDigits({ value }: { value: string }) {
+  const prevRef = useRef(value);
+  const [oldVal, setOldVal] = useState(value);
+  const [newVal, setNewVal] = useState(value);
+  const [rolling, setRolling] = useState(false);
+  const [shifted, setShifted] = useState(false);
+  const [dir, setDir] = useState<"up" | "down">("up");
+
+  useEffect(() => {
+    if (prevRef.current === value) return;
+    const from = prevRef.current;
+    prevRef.current = value;
+    const fromN = parseInt(from, 10);
+    const toN = parseInt(value, 10);
+    setDir(Number.isFinite(fromN) && Number.isFinite(toN) && toN < fromN ? "down" : "up");
+    setOldVal(from);
+    setNewVal(value);
+    setRolling(true);
+    setShifted(false);
+
+    let raf2 = 0;
+    const raf1 = requestAnimationFrame(() => {
+      raf2 = requestAnimationFrame(() => setShifted(true));
+    });
+    const done = setTimeout(() => {
+      setRolling(false);
+      setOldVal(value);
+      setShifted(false);
+    }, ROLL_MS);
+
+    return () => {
+      cancelAnimationFrame(raf1);
+      cancelAnimationFrame(raf2);
+      clearTimeout(done);
+    };
+  }, [value]);
+
+  const chars = rolling ? newVal : oldVal;
+
+  return (
+    <>
+      {Array.from({ length: chars.length }, (_, i) => {
+        const o = oldVal[i] ?? "";
+        const n = chars[i] ?? "";
+        if (!rolling || o === n) {
+          return <span key={`${i}-${n}`}>{n}</span>;
+        }
+        const top = dir === "down" ? n : o;
+        const bottom = dir === "down" ? o : n;
+        const restY = dir === "down" ? "0" : "-1em";
+        const startY = dir === "down" ? "-1em" : "0";
+        return (
+          <span
+            key={`${i}-${o}-${n}-${dir}`}
+            style={{ display: "inline-block", position: "relative", overflow: "hidden", height: "1em", lineHeight: "1em", verticalAlign: "-0.05em" }}
+          >
+            <span
+              style={{
+                display: "flex",
+                flexDirection: "column",
+                transition: "transform 350ms cubic-bezier(0.4, 0, 0.2, 1)",
+                transform: `translateY(${shifted ? restY : startY})`,
+              }}
+            >
+              <span style={{ height: "1em", lineHeight: "1em" }}>{top}</span>
+              <span style={{ height: "1em", lineHeight: "1em" }}>{bottom}</span>
+            </span>
+          </span>
+        );
+      })}
+    </>
+  );
+}
+
+function Ico({ path, size = 14, sw = 2 }: { path: React.ReactNode; size?: number; sw?: number }) {
+  return (
+    <svg width={size} height={size} viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth={sw} strokeLinecap="round" strokeLinejoin="round" aria-hidden>
+      {path}
+    </svg>
+  );
+}
+
+export default function ApprovalCard({
+  questions = QUESTIONS,
+  labels,
+  onSubmitted,
+  onAnswerChange,
+  resettable = true,
+}: {
+  questions?: ApprovalQuestion[];
+  labels?: Partial<ApprovalLabels>;
+  onSubmitted?: (answers: Record<number, number[]>) => void;
+  onAnswerChange?: (questionIndex: number, answer: number[]) => void;
+  resettable?: boolean;
+  variant?: string;
+} = {}) {
+  const t = { ...DEFAULT_LABELS, ...labels };
+  const [qi, setQi] = useState(0);
+  const [answers, setAnswers] = useState<Record<number, number[]>>({});
+  const [custom, setCustom] = useState<Record<number, string>>({});
+  const [sent, setSent] = useState(false);
+  const [open, setOpen] = useState(true);
+
+  const advanceTimer = useRef<ReturnType<typeof setTimeout> | null>(null);
+  const questionRefs = useRef<(HTMLDivElement | null)[]>([]);
+  const measured = useRef(false);
+  const [viewportH, setViewportH] = useState<number | undefined>(undefined);
+  const [trackY, setTrackY] = useState(0);
+  const [animate, setAnimate] = useState(false);
+  // Until the first question is measured, render only the active one so the
+  // initial (and SSR) height is Q1's height — not all questions stacked, which
+  // would flash to full height and then shrink on mount.
+  const [ready, setReady] = useState(false);
+
+  const last = qi === questions.length - 1;
+  const selected = answers[qi] ?? [];
+  const hasAnswer = selected.length > 0 || Boolean(custom[qi]?.trim());
+
+  const sync = (withAnim: boolean) => {
+    const item = questionRefs.current[qi];
+    if (!item) return;
+    const reduce = typeof window !== "undefined" && window.matchMedia("(prefers-reduced-motion: reduce)").matches;
+    setViewportH(item.offsetHeight);
+    setTrackY(item.offsetTop);
+    setAnimate(withAnim && !reduce);
+  };
+
+  useLayoutEffect(() => {
+    const withAnim = measured.current;
+    measured.current = true;
+    sync(withAnim);
+    setReady(true);
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [qi, answers, custom, open, sent]);
+
+  useEffect(() => {
+    const id = requestAnimationFrame(() => sync(measured.current));
+    return () => cancelAnimationFrame(id);
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [qi]);
+
+  useEffect(() => () => { if (advanceTimer.current) clearTimeout(advanceTimer.current); }, []);
+
+  const goTo = (next: number) => {
+    if (advanceTimer.current) clearTimeout(advanceTimer.current);
+    setQi(Math.min(Math.max(next, 0), questions.length - 1));
+  };
+
+  const send = () => {
+    if (advanceTimer.current) clearTimeout(advanceTimer.current);
+    setSent(true);
+    onSubmitted?.(answers);
+  };
+
+  const advance = () => {
+    if (last) send();
+    else goTo(qi + 1);
+  };
+
+  const toggle = (index: number) => {
+    const type = questions[qi].type;
+    setAnswers((current) => {
+      const picked = current[qi] ?? [];
+      const next = type === "radio"
+        ? [index]
+        : picked.includes(index)
+          ? picked.filter((item) => item !== index)
+          : [...picked, index];
+      onAnswerChange?.(qi, next);
+      return { ...current, [qi]: next };
+    });
+    if (type === "radio") {
+      setCustom((current) => ({ ...current, [qi]: "" }));
+      if (advanceTimer.current) clearTimeout(advanceTimer.current);
+      advanceTimer.current = setTimeout(() => {
+        if (last) send();
+        else setQi((current) => Math.min(questions.length - 1, current + 1));
+      }, 480);
+    }
+  };
+
+  const reset = () => {
+    setQi(0);
+    setAnswers({});
+    setCustom({});
+    setSent(false);
+    setOpen(true);
+    measured.current = false;
+  };
+
+  if (!open) {
+    return (
+      <button type="button" onClick={() => setOpen(true)} className="rounded-control bg-surface px-3 py-2 text-[12.5px] font-medium text-ink shadow-btn transition-colors duration-150 hover:bg-hover">
+        Open approval
+      </button>
+    );
+  }
+
+  if (sent) {
+    return (
+      <div className="flex w-full max-w-80 items-center gap-3" style={{ animation: "pop-in 260ms cubic-bezier(0.23,1,0.32,1) both" }}>
+        <span className="inline-flex items-center gap-1.5 rounded-full bg-green-tint py-1 pr-2.5 pl-1 text-[12.5px] font-medium text-green">
+          <span className="flex size-4.5 items-center justify-center rounded-full bg-green text-white">
+            <svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="3" strokeLinecap="round" strokeLinejoin="round"><path d="M20 6L9 17l-5-5" /></svg>
+          </span>
+          {t.sentMessage}
+        </span>
+        {resettable && (
+          <button type="button" onClick={reset} className="text-[12px] font-medium text-ink-3 transition-colors duration-150 hover:text-ink">
+            Start over
+          </button>
+        )}
+      </div>
+    );
+  }
+
+  return (
+    <div className="w-full max-w-80">
+      <div className="relative overflow-hidden rounded-card bg-surface shadow-card" style={{ animation: "fade-up 380ms cubic-bezier(0.23,1,0.32,1) both" }}>
+        <button
+          type="button"
+          aria-label="Dismiss"
+          onClick={() => setOpen(false)}
+          className="primitive-icon-button absolute right-2.5 top-2.5 z-10 text-ink-3 transition-colors duration-100 hover:bg-hover hover:text-ink"
+        >
+          <Ico size={14} sw={2.2} path={<path d="M18 6L6 18M6 6l12 12" />} />
+        </button>
+        <div className="primitive-card-pad">
+          {/* the question itself is the heading */}
+          <div
+            className="overflow-hidden"
+            style={{ height: viewportH, transition: animate ? `height ${SLIDE}` : undefined }}
+            aria-live="polite"
+          >
+            <div
+              style={{
+                display: "flex",
+                flexDirection: "column",
+                gap: 26,
+                transform: `translate3d(0, ${-trackY}px, 0)`,
+                transition: animate ? `transform ${SLIDE}` : undefined,
+                willChange: "transform",
+              }}
+            >
+              {questions.map((question, qIdx) => {
+                const active = qIdx === qi;
+                // Before the first measure, mount only the active question so the
+                // card opens at its real height instead of flashing to full height.
+                if (!ready && !active) return null;
+                const picked = answers[qIdx] ?? [];
+                const questionStyle: CSSProperties = {
+                  opacity: active ? 1 : 0,
+                  transition: animate ? `opacity ${SLIDE}` : undefined,
+                  pointerEvents: active ? undefined : "none",
+                };
+                return (
+                  <div
+                    key={qIdx}
+                    ref={(el) => { questionRefs.current[qIdx] = el; }}
+                    aria-hidden={active ? undefined : true}
+                    style={questionStyle}
+                  >
+                    <div className="pr-7 text-[14px] font-medium text-ink">{question.q}</div>
+                    <GlideMenu className="mt-2.5 flex flex-col gap-1" highlightClassName="inset-x-0 rounded-control bg-hover">
+                      {question.options.map((option, i) => {
+                        const on = picked.includes(i);
+                        return (
+                          <button
+                            key={option}
+                            type="button"
+                            data-menu-row
+                            aria-pressed={on}
+                            tabIndex={active ? 0 : -1}
+                            onClick={() => { if (active) toggle(i); }}
+                            className="relative z-10 flex items-center gap-1.5 rounded-control pl-1 pr-2 py-1 text-left transition-colors duration-100"
+                          >
+                            <span
+                              className={`flex size-4 shrink-0 items-center justify-center transition-colors duration-200
+                                ${question.type === "radio" ? "rounded-full" : "rounded-[5px]"}
+                                ${on ? "bg-ink text-canvas" : "shadow-[inset_0_0_0_1.5px_var(--line-strong)] text-transparent"}`}
+                            >
+                              {question.type === "radio" ? (
+                                <span className="size-1.5 rounded-full bg-canvas transition-transform duration-200" style={{ transform: on ? "scale(1)" : "scale(0)" }} />
+                              ) : (
+                                <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="3" strokeLinecap="round" strokeLinejoin="round"><path d="M20 6L9 17l-5-5" /></svg>
+                              )}
+                            </span>
+                            <span className={`text-[13px] leading-none transition-colors duration-200 ${on ? "text-ink" : "text-ink-2"}`}>
+                              {option}
+                            </span>
+                          </button>
+                        );
+                      })}
+                      <label data-menu-row className="relative z-10 flex items-center gap-1.5 rounded-control pl-1 pr-2 py-1 transition-colors duration-100">
+                        <input
+                          value={custom[qIdx] ?? ""}
+                          tabIndex={active ? 0 : -1}
+                          onChange={(event) => {
+                            if (!active) return;
+                            setCustom((current) => ({ ...current, [qIdx]: event.target.value }));
+                            if (question.type === "radio") setAnswers((current) => ({ ...current, [qIdx]: [] }));
+                          }}
+                          onKeyDown={(event) => {
+                            if (event.key === "Enter" && hasAnswer) {
+                              event.preventDefault();
+                              advance();
+                            }
+                          }}
+                          placeholder={t.customPlaceholder}
+                          aria-label="Custom answer"
+                          className="min-w-0 flex-1 bg-transparent pl-1.5 text-[13px] text-ink outline-none placeholder:text-ink-3"
+                        />
+                      </label>
+                    </GlideMenu>
+                  </div>
+                );
+              })}
+            </div>
+          </div>
+        </div>
+
+        {/* footer — step nav (rolling counter) + pill actions */}
+        <div className="primitive-card-footer flex items-center justify-between gap-3">
+          <div className="flex items-center gap-1 text-ink-3">
+            <button
+              type="button"
+              aria-label="Previous question"
+              disabled={qi <= 0}
+              onClick={() => goTo(qi - 1)}
+              className="flex size-[18px] items-center justify-center rounded-[5px] transition-colors duration-100 enabled:hover:text-ink disabled:opacity-30"
+            >
+              <Ico size={14} path={<path d="M18 15l-6-6-6 6" />} />
+            </button>
+            <span className="inline-flex items-center text-[12px] font-medium tabular-nums text-ink-3" style={{ letterSpacing: "-0.1px", lineHeight: 1 }}>
+              <RollingDigits value={`${qi + 1} / ${questions.length}`} />
+            </span>
+            <button
+              type="button"
+              aria-label="Next question"
+              disabled={last}
+              onClick={() => goTo(qi + 1)}
+              className="flex size-[18px] items-center justify-center rounded-[5px] transition-colors duration-100 enabled:hover:text-ink disabled:opacity-30"
+            >
+              <Ico size={14} path={<path d="M6 9l6 6 6-6" />} />
+            </button>
+          </div>
+
+          <div className="-mr-0.5 flex items-center gap-1.5">
+            <Button variant="ghost" size="sm" onClick={() => (last ? setOpen(false) : goTo(qi + 1))}>
+              {t.skip}
+            </Button>
+            <Button variant="accent" size="sm" disabled={!hasAnswer} onClick={advance}>
+              {last ? t.send : t.continue}
+            </Button>
+          </div>
+        </div>
+      </div>
+    </div>
+  );
+}
@@ -0,0 +1,60 @@
+/* Vendored from BeautifulUI registry: https://www.beautifului.dev/r/glide-menu.json
+ * Spike #293 — DO NOT ship in templates/frontend. Imports rewritten to @/spike/beautifui/*.
+ * License: MIT (upstream https://github.com/slev12397/beautiful-ui). */
+import { useRef, useState, type ReactNode } from "react";
+
+type GlideMenuProps = {
+  children: ReactNode;
+  className?: string;
+  highlightClassName?: string;
+  rowSelector?: string;
+};
+
+/** A single hover layer that glides between interactive menu rows. */
+export default function GlideMenu({
+  children,
+  className = "",
+  highlightClassName = "inset-x-0 rounded-[8px] bg-hover",
+  rowSelector = "[data-menu-row]",
+}: GlideMenuProps) {
+  const ref = useRef<HTMLDivElement>(null);
+  const [box, setBox] = useState<{ top: number; height: number } | null>(null);
+  const [visible, setVisible] = useState(false);
+
+  const moveTo = (target: EventTarget | null) => {
+    const container = ref.current;
+    if (!(target instanceof Element) || !container) return;
+    const row = target.closest(rowSelector);
+    if (!(row instanceof HTMLElement) || !container.contains(row)) return;
+    const containerRect = container.getBoundingClientRect();
+    const rowRect = row.getBoundingClientRect();
+    setBox({ top: rowRect.top - containerRect.top, height: rowRect.height });
+    setVisible(true);
+  };
+
+  return (
+    <div
+      ref={ref}
+      onMouseOver={(event) => moveTo(event.target)}
+      onMouseLeave={() => setVisible(false)}
+      onFocusCapture={(event) => moveTo(event.target)}
+      onBlurCapture={(event) => {
+        if (!ref.current?.contains(event.relatedTarget as Node | null)) setVisible(false);
+      }}
+      className={`group/glide-menu relative ${className}`}
+    >
+      <span
+        aria-hidden
+        className={`pointer-events-none absolute ${highlightClassName}`}
+        style={{
+          top: box?.top ?? 0,
+          height: box?.height ?? 0,
+          opacity: box && visible ? 1 : 0,
+          transition:
+            "top 220ms cubic-bezier(0.23,1,0.32,1), height 220ms cubic-bezier(0.23,1,0.32,1), opacity 150ms ease",
+        }}
+      />
+      {children}
+    </div>
+  );
+}
@@ -0,0 +1,253 @@
+/* Vendored from BeautifulUI registry: https://www.beautifului.dev/r/streaming-text.json
+ * Spike #293 — DO NOT ship in templates/frontend. Imports rewritten to @/spike/beautifui/*.
+ * License: MIT (upstream https://github.com/slev12397/beautiful-ui). */
+import { useEffect, useState } from "react";
+
+/* ─────────────────────────────────────────────────────────
+ * STREAMING TEXT
+ * Words resolve out of blur, inline citations appear in
+ * context, then actions and follow-up prompts become usable.
+ * ───────────────────────────────────────────────────────── */
+
+const WORD_MS = 55;
+const HOLD_MS = 3400;
+
+/* one streamed word, or a `cite` placeholder that renders an inline source chip */
+export type StreamingToken = { text: string; cite?: boolean };
+
+const TOKENS: StreamingToken[] = [
+  ..."Pistachio is your fastest-growing flavor — sales are up 23% this month and margins beat vanilla by 8 points."
+    .split(" ")
+    .map((text) => ({ text })),
+  { text: "", cite: true },
+  ..."Stone-fruit flavors are trending in the same range."
+    .split(" ")
+    .map((text) => ({ text })),
+];
+
+const FOLLOW_UPS = [
+  "Which flavors sell best in winter",
+  "Compare gelato and soft serve margins",
+];
+
+const SOURCE_IMAGES = {
+  scoop:
+    "data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 64 64'%3E%3Crect width='64' height='64' rx='16' fill='%231f7a5f'/%3E%3Cpath d='M20 36c0 7 5.4 12 12 12s12-5 12-12H20Z' fill='%23fff'/%3E%3Ccircle cx='32' cy='25' r='11' fill='%23bff3dd'/%3E%3Cpath d='M24 24c4-7 13-7 17 0' fill='none' stroke='%231f7a5f' stroke-width='4' stroke-linecap='round'/%3E%3C/svg%3E",
+  trends:
+    "data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 64 64'%3E%3Crect width='64' height='64' rx='16' fill='%232f6fec'/%3E%3Cpath d='M15 43 27 31l8 7 14-18' fill='none' stroke='%23fff' stroke-width='7' stroke-linecap='round' stroke-linejoin='round'/%3E%3Ccircle cx='49' cy='20' r='5' fill='%23bfe0ff'/%3E%3C/svg%3E",
+  market:
+    "data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 64 64'%3E%3Crect width='64' height='64' rx='16' fill='%23e56d24'/%3E%3Cpath d='M17 45V25h8v20h-8Zm11 0V16h8v29h-8Zm11 0V30h8v15h-8Z' fill='%23fff'/%3E%3Cpath d='M16 49h32' stroke='%23ffd6b8' stroke-width='4' stroke-linecap='round'/%3E%3C/svg%3E",
+};
+
+/* one cited source rendered as an inline chip and in the sources list */
+export type StreamingSource = { name: string; domain: string; href: string; image: string };
+
+const SOURCES: StreamingSource[] = [
+  { name: "Scoop Data", domain: "scoopdata.io", href: "https://scoopdata.io/", image: SOURCE_IMAGES.scoop },
+  { name: "Trends Index", domain: "trends.google.com", href: "https://trends.google.com/trends/", image: SOURCE_IMAGES.trends },
+  { name: "Market Basket", domain: "marketbasket.io", href: "https://marketbasket.io/", image: SOURCE_IMAGES.market },
+];
+
+function sourceImage(source: StreamingSource) {
+  return source.image;
+}
+
+function SourceChip({ source }: { source?: StreamingSource }) {
+  if (!source) return null;
+  return (
+    <a
+      href={source.href}
+      target="_blank"
+      rel="noreferrer"
+      className="ml-0 mr-1 inline-flex h-4.5 translate-y-[-1px] items-center gap-1 rounded-[5px]
+        bg-inset pr-[3px] pl-[3px] align-middle font-mono text-[10.5px] text-ink-2 shadow-hairline
+        transition-colors duration-150 hover:bg-hover hover:text-ink"
+      style={{ animation: "pop-in 250ms cubic-bezier(0.23,1,0.32,1) both" }}
+    >
+      <img src={sourceImage(source)} alt="" className="source-avatar size-3 rounded-[3px]" />
+      <span>{source.domain}</span>
+    </a>
+  );
+}
+
+const ACTION_ICONS: React.ReactNode[] = [
+  <g key="copy"><rect x="9" y="9" width="12" height="12" rx="2.5" /><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1" /></g>,
+  <path key="retry" d="M21 12a9 9 0 1 1-2.64-6.36M21 3v6h-6" />,
+  <path key="up" d="M7 10v12M15 5.88L14 10h5.83a2 2 0 0 1 1.92 2.56l-2.33 8A2 2 0 0 1 17.5 22H4a2 2 0 0 1-2-2v-8a2 2 0 0 1 2-2h2.76a2 2 0 0 0 1.79-1.11L12 2a3.13 3.13 0 0 1 3 3.88z" />,
+  <path key="down" d="M17 14V2M9 18.12L10 14H4.17a2 2 0 0 1-1.92-2.56l2.33-8A2 2 0 0 1 6.5 2H20a2 2 0 0 1 2 2v8a2 2 0 0 1-2 2h-2.76a2 2 0 0 0-1.79 1.11L12 22a3.13 3.13 0 0 1-3-3.88z" />,
+];
+
+export type StreamingLabels = {
+  /** label on the collapsed sources toggle */
+  sources: string;
+  /** heading above the follow-up prompts */
+  followUps: string;
+};
+
+const DEFAULT_LABELS: StreamingLabels = {
+  sources: "10 sources",
+  followUps: "Follow-ups",
+};
+
+export default function StreamingText({
+  content = TOKENS,
+  sources = SOURCES,
+  followUps = FOLLOW_UPS,
+  labels,
+  loop = true,
+  fill = false,
+  onDone,
+  onFollowUp,
+}: {
+  variant?: string;
+  /** the streamed tokens; `cite` tokens render an inline source chip */
+  content?: StreamingToken[];
+  /** cited sources shown in the chip, avatar stack, and expanded list */
+  sources?: StreamingSource[];
+  /** follow-up prompt suggestions shown once the stream completes */
+  followUps?: string[];
+  /** prominent copy strings */
+  labels?: Partial<StreamingLabels>;
+  /** restart the stream after a hold; turn off when embedding in a real thread */
+  loop?: boolean;
+  /** fill the parent width instead of the gallery's fixed measure */
+  fill?: boolean;
+  onDone?: () => void;
+  /** fired when a follow-up prompt is chosen */
+  onFollowUp?: (text: string, index: number) => void;
+} = {}) {
+  const l = { ...DEFAULT_LABELS, ...labels };
+  const [count, setCount] = useState(0);
+  const [sourcesOpen, setSourcesOpen] = useState(false);
+  const done = count >= content.length;
+
+  useEffect(() => {
+    if (done && !loop) {
+      onDone?.();
+      return;
+    }
+    const t = setTimeout(
+      () => setCount((c) => (c >= content.length ? 0 : c + 1)),
+      done ? HOLD_MS : WORD_MS,
+    );
+    return () => clearTimeout(t);
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [count, done, loop]);
+
+  return (
+    <div className={fill ? "w-full" : "min-h-[15.5rem] w-full max-w-95"}>
+      <p className="text-[13px] leading-relaxed text-ink">
+        {content.slice(0, count).map((token, i) =>
+          token.cite ? (
+            <SourceChip key={i} source={sources[0]} />
+          ) : (
+            <span key={i} className="inline">
+              {token.text}{" "}
+            </span>
+          ),
+        )}
+        {!done && (
+          <span
+            className="ml-0.5 inline-block h-3 w-0.5 translate-y-0.5 rounded-full bg-ink"
+            style={{ animation: "fade-in 150ms ease-out both" }}
+          />
+        )}
+      </p>
+
+      {/* action icons row */}
+      <div
+        className="mt-2 flex items-center gap-0.5 transition-opacity duration-400"
+        style={{ opacity: done ? 1 : 0, pointerEvents: done ? "auto" : "none" }}
+      >
+        {ACTION_ICONS.map((icon, i) => (
+          <button
+            key={i}
+            type="button"
+            aria-label="Action"
+            className="flex size-6 items-center justify-center rounded-[6px] text-ink-3
+              transition-colors duration-100 hover:bg-hover-2 hover:text-ink-2"
+          >
+            <svg width="15" height="15" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.8" strokeLinecap="round" strokeLinejoin="round">
+              {icon}
+            </svg>
+          </button>
+        ))}
+        <button
+          type="button"
+          aria-expanded={sourcesOpen}
+          onClick={() => setSourcesOpen((current) => !current)}
+          className="ml-1.5 flex items-center gap-1.5 rounded-[6px] px-1 py-0.5 text-left transition-colors duration-150 hover:bg-hover"
+        >
+          <span className="flex -space-x-1">
+            {sources.map((source) => (
+              <img
+                key={source.domain}
+                src={sourceImage(source)}
+                alt=""
+                className="source-avatar size-3.5 rounded-full bg-surface shadow-[0_0_0_1.5px_var(--canvas)]"
+              />
+            ))}
+          </span>
+          <span className="text-[12px] text-ink-2">{l.sources}</span>
+        </button>
+      </div>
+
+      <div
+        className="grid transition-[grid-template-rows,opacity] duration-300"
+        style={{
+          gridTemplateRows: done && sourcesOpen ? "1fr" : "0fr",
+          opacity: done && sourcesOpen ? 1 : 0,
+          transitionTimingFunction: "cubic-bezier(0.23, 1, 0.32, 1)",
+        }}
+      >
+        <div className="overflow-hidden">
+          <div className="mt-1.5 flex flex-col rounded-[10px] bg-inset p-1 shadow-hairline">
+            {sources.map((source) => (
+              <a
+                key={source.domain}
+                href={source.href}
+                target="_blank"
+                rel="noreferrer"
+                className="flex items-center gap-2 rounded-[6px] px-1.5 py-1 text-[12px] text-ink-2 transition-colors duration-150 hover:bg-hover hover:text-ink"
+              >
+                <img src={sourceImage(source)} alt="" className="source-avatar size-4 rounded-[4px]" />
+                <span className="animated-underline">{source.name}</span>
+                <span className="ml-auto font-mono text-[10.5px] text-ink-3">{source.domain}</span>
+              </a>
+            ))}
+          </div>
+        </div>
+      </div>
+
+      {/* follow-ups */}
+      <div
+        className="mt-2.5 transition-opacity duration-400"
+        style={{ opacity: done ? 1 : 0, pointerEvents: done ? "auto" : "none" }}
+      >
+        <p className="text-[12px] font-medium text-ink-2">{l.followUps}</p>
+        <div className="mt-0.5 flex flex-col">
+          {followUps.map((text, i) => (
+            <button
+              key={text}
+              onClick={() => onFollowUp?.(text, i)}
+              className="-mx-1.5 flex items-center gap-2 rounded-[7px] border-b border-line
+                px-1.5 py-1.5 text-left text-[12.5px] text-ink transition-colors
+                duration-100 hover:bg-hover-2"
+              style={
+                done
+                  ? { animation: `fade-up 350ms cubic-bezier(0.23,1,0.32,1) ${i * 90}ms both` }
+                  : { opacity: 0 }
+              }
+            >
+              <svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="var(--ink-3)" strokeWidth="2" strokeLinecap="round" strokeLinejoin="round" className="shrink-0">
+                <path d="M9 10l-5 5 5 5" />
+                <path d="M20 4v7a4 4 0 0 1-4 4H4" />
+              </svg>
+              {text}
+            </button>
+          ))}
+        </div>
+      </div>
+    </div>
+  );
+}
@@ -0,0 +1,302 @@
+/* Vendored from BeautifulUI registry: https://www.beautifului.dev/r/thinking-state.json
+ * Spike #293 — DO NOT ship in templates/frontend. Imports rewritten to @/spike/beautifui/*.
+ * License: MIT (upstream https://github.com/slev12397/beautiful-ui). */
+import { useEffect, useLayoutEffect, useRef, useState, type ReactNode } from "react";
+
+/* ─────────────────────────────────────────────────────────
+ * THINKING — expandable agent trace, four variants
+ *
+ *   Steps      step list with spinner → muted checks
+ *   Reasoning  prose reasoning that expands, then settles
+ *   Search     web-search trace: query + sources read
+ *   Coding     tool trace: files read, edits, commands
+ *
+ * The trace runs once, settles, and remains expandable.
+ * ───────────────────────────────────────────────────────── */
+
+const STAGES = [800, 600, 1800, 2600, 1600];
+
+function useSequence(steps: number[]) {
+  const [stage, setStage] = useState(0);
+  useEffect(() => {
+    if (stage >= steps.length - 1) return;
+    const t = setTimeout(() => setStage((s) => s + 1), steps[stage]);
+    return () => clearTimeout(t);
+  }, [stage, steps]);
+  return stage;
+}
+
+type Row = {
+  primary: string;
+  secondary?: string;
+  mono?: boolean;
+  add?: number;
+  del?: number;
+  href?: string;
+};
+
+const VARIANTS: Record<
+  string,
+  { active: string; done: string; rows: Row[]; query?: string }
+> = {
+  Steps: {
+    active: "Thinking",
+    done: "Thought for 4 seconds",
+    rows: [
+      { primary: "Reading flavor briefs" },
+      { primary: "Scanning supplier lists" },
+      { primary: "Comparing tasting notes", secondary: "6 flavors" },
+      { primary: "Writing the scoop report" },
+    ],
+  },
+  Reasoning: {
+    active: "Thinking",
+    done: "Thought for 4 seconds",
+    rows: [
+      { primary: "Summer demand spikes for stone-fruit flavors — peach and apricot lead." },
+      { primary: "I should check cone inventory before promoting a waffle-bowl special." },
+    ],
+  },
+  Search: {
+    active: "Searching the web",
+    done: "Searched the web",
+    query: "best waffle cone supplier",
+    rows: [
+      { primary: "Joy Cone", secondary: "joycone.com", href: "https://joycone.com/fs_products/waffle-cones/" },
+      { primary: "WebstaurantStore", secondary: "webstaurantstore.com", href: "https://www.webstaurantstore.com/ice-cream-shop-supplies.html" },
+      { primary: "The Konery", secondary: "thekonery.com", href: "https://www.thekonery.com/" },
+    ],
+  },
+  Coding: {
+    active: "Running tools",
+    done: "Ran 3 tools",
+    rows: [
+      { primary: "Read", secondary: "flavors.ts", mono: true },
+      { primary: "Edit", secondary: "ChurnSchedule.tsx", mono: true, add: 74, del: 41 },
+      { primary: "Run", secondary: "npm run freeze", mono: true },
+    ],
+  },
+};
+
+function Dot({ tone }: { tone: string }) {
+  return (
+    <span className={`flex size-3.5 shrink-0 items-center justify-center rounded-full text-white ${tone}`}>
+      <svg width="9" height="9" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="2.5">
+        <circle cx="12" cy="12" r="9" />
+        <path d="M3.5 12h17M12 3a14 14 0 0 1 0 18M12 3a14 14 0 0 0 0 18" />
+      </svg>
+    </span>
+  );
+}
+
+const TONES = ["bg-accent", "bg-orange", "bg-green"];
+
+export default function ThinkingState({
+  variant = "Steps",
+  onSettled,
+  rows,
+  active,
+  done,
+  icon,
+}: {
+  variant?: string;
+  onSettled?: () => void;
+  /** override the built-in trace content (keeps the primitive reusable) */
+  rows?: Row[];
+  active?: string;
+  done?: string;
+  /** override the header glyph (defaults to the sparkle) */
+  icon?: ReactNode;
+}) {
+  const stage = useSequence(STAGES);
+  const [manualExpanded, setManualExpanded] = useState<boolean | null>(null);
+  const [selectedTool, setSelectedTool] = useState<string | null>(null);
+  const base = VARIANTS[variant] ?? VARIANTS.Steps;
+  const v = {
+    ...base,
+    rows: rows ?? base.rows,
+    active: active ?? base.active,
+    done: done ?? base.done,
+  };
+  const autoExpanded = stage >= 1 && stage < 4;
+  const expanded = manualExpanded ?? autoExpanded;
+  const working = stage < 3;
+  const visible = stage < 2 ? 0 : stage === 2 ? Math.min(2, v.rows.length) : v.rows.length;
+  const traceRef = useRef<HTMLDivElement>(null);
+  const [lineHeight, setLineHeight] = useState(0);
+  useLayoutEffect(() => {
+    if (traceRef.current) setLineHeight(traceRef.current.offsetHeight);
+  }, [visible, expanded, variant, stage]);
+
+  /* let embedders sequence content after the trace settles */
+  const settledRef = useRef(false);
+  useEffect(() => {
+    if (working || settledRef.current) return;
+    settledRef.current = true;
+    onSettled?.();
+  }, [working, onSettled]);
+
+  return (
+    <div
+      key={variant}
+      className="flex w-full max-w-95 flex-col"
+      style={{
+        minHeight: working || expanded ? 176 : undefined,
+        transition: "min-height 400ms cubic-bezier(0.23,1,0.32,1)",
+      }}
+    >
+      {/* header — shared across variants */}
+      <button
+        type="button"
+        aria-expanded={expanded}
+        onClick={() => setManualExpanded((current) => !(current ?? autoExpanded))}
+        className="-mx-1.5 flex w-fit items-center gap-2 rounded-control px-1.5 py-1
+          transition-colors duration-100 hover:bg-hover-2"
+      >
+        {icon ? (
+          <span className="flex shrink-0 transition-colors duration-200" style={{ color: working ? "var(--ink-2)" : "var(--ink-3)" }}>
+            {icon}
+          </span>
+        ) : (
+          <svg width="16" height="16" viewBox="0 0 24 24" fill={working ? "var(--ink-2)" : "var(--ink-3)"}>
+            <path d="M12 2l2.4 7.2L22 12l-7.6 2.8L12 22l-2.4-7.2L2 12l7.6-2.8z" />
+          </svg>
+        )}
+        <span role="status" className="contents">
+          {working ? (
+            <span
+              className="bg-clip-text text-[13px] font-medium whitespace-nowrap text-transparent"
+              style={{
+                backgroundImage:
+                  "linear-gradient(90deg, var(--ink-3) 35%, var(--ink) 50%, var(--ink-3) 65%)",
+                backgroundSize: "200% 100%",
+                animation: "shimmer-text 1.4s linear infinite",
+              }}
+            >
+              {v.active}
+            </span>
+          ) : (
+            <span
+              className="text-[13px] font-medium whitespace-nowrap text-ink-2"
+              style={{ animation: "fade-in 350ms ease-out both" }}
+            >
+              {v.done}
+            </span>
+          )}
+        </span>
+        <svg
+          width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="var(--ink-3)" strokeWidth="2.2" strokeLinecap="round" strokeLinejoin="round"
+          className="transition-transform duration-300"
+          style={{ transform: expanded ? "rotate(180deg)" : "rotate(0)" }}
+        >
+          <path d="M6 9l6 6 6-6" />
+        </svg>
+      </button>
+
+      {/* expandable trace */}
+      <div
+        className="grid transition-[grid-template-rows,opacity] duration-400"
+        style={{
+          gridTemplateRows: expanded ? "1fr" : "0fr",
+          opacity: expanded ? 1 : 0,
+          transitionTimingFunction: "cubic-bezier(0.23, 1, 0.32, 1)",
+        }}
+      >
+        <div className="overflow-hidden">
+          <div className="relative mt-1 ml-[5px] pl-4">
+            <span
+              aria-hidden
+              className="absolute left-[3px] w-px bg-line"
+              style={{ top: -8, height: lineHeight ? lineHeight - 2 : 0, transition: "height 500ms cubic-bezier(0.23,1,0.32,1)" }}
+            />
+            <div ref={traceRef} className="flex flex-col gap-1 py-1">
+            {v.query && (
+              <div className="flex h-6 items-center gap-2 px-1.5" style={{ animation: expanded ? "fade-up 300ms cubic-bezier(0.23,1,0.32,1) both" : undefined }}>
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="var(--ink-3)" strokeWidth="2" strokeLinecap="round" className="shrink-0">
+                  <circle cx="11" cy="11" r="7" />
+                  <path d="M21 21l-4.3-4.3" />
+                </svg>
+                <span className="text-[12.5px] text-ink-2">{v.query}</span>
+              </div>
+            )}
+            {v.rows.slice(0, visible).map((row, i) => {
+              const content = (
+                <>
+                {variant === "Search" && <Dot tone={TONES[i % 3]} />}
+                {variant === "Steps" && (
+                  i < visible - 1 || !working ? (
+                    <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="var(--ink-3)" strokeWidth="2.5" strokeLinecap="round" strokeLinejoin="round" className="shrink-0">
+                      <path d="M20 6L9 17l-5-5" />
+                    </svg>
+                  ) : (
+                    <span className="size-3 shrink-0 rounded-full border-[1.5px] border-line-strong border-t-ink-2" style={{ animation: "spin 700ms linear infinite" }} />
+                  )
+                )}
+                <span className={`min-w-0 truncate text-[12.5px] ${variant === "Reasoning" ? "whitespace-normal leading-relaxed text-ink-2" : "font-medium text-ink"} ${variant === "Search" ? "animated-underline" : ""}`}>
+                  {row.primary}
+                </span>
+                {row.secondary && (
+                  <span className={`shrink-0 text-[11.5px] text-ink-3 ${row.mono ? "font-mono" : ""}`}>
+                    {row.secondary}
+                  </span>
+                )}
+                {row.add !== undefined && (
+                  <span className="shrink-0 font-mono text-[11px] tabular-nums">
+                    <span className="text-green">+{row.add}</span>{" "}
+                    <span className="text-red">−{row.del}</span>
+                  </span>
+                )}
+                </>
+              );
+              const rowClass = "flex min-h-7 w-full items-center gap-2 rounded-[6px] px-1.5 py-0.5 text-left";
+              const animation = { animation: `fade-up 320ms cubic-bezier(0.23,1,0.32,1) ${i * 120}ms both` };
+
+              if (variant === "Search") {
+                return (
+                  <a
+                    key={row.primary}
+                    href={row.href}
+                    target="_blank"
+                    rel="noreferrer"
+                    className={`${rowClass} transition-colors duration-150 hover:bg-hover`}
+                    style={animation}
+                  >
+                    {content}
+                  </a>
+                );
+              }
+
+              if (variant === "Coding") {
+                const selected = selectedTool === row.primary;
+                return (
+                  <button
+                    key={row.primary}
+                    type="button"
+                    aria-pressed={selected}
+                    onClick={() => setSelectedTool(selected ? null : row.primary)}
+                    className={`${rowClass} transition-colors duration-150 ${selected ? "bg-inset" : "hover:bg-hover"}`}
+                    style={animation}
+                  >
+                    {content}
+                  </button>
+                );
+              }
+
+              return (
+                <div key={row.primary} className={rowClass} style={animation}>
+                  {content}
+                </div>
+              );
+            })}
+            {variant === "Search" && stage >= 3 && (
+              <span className="text-[12px] text-ink-3" style={{ animation: "fade-in 300ms ease-out both" }}>
+                +7 more
+              </span>
+            )}
+            </div>
+          </div>
+        </div>
+      </div>
+    </div>
+  );
+}
@@ -0,0 +1,11 @@
+/* Spike #293 — dev-only entry. Not part of the production bundle. */
+import { StrictMode } from "react";
+import { createRoot } from "react-dom/client";
+import "@/spike/beautifui/foundation.css";
+import SpikePage from "@/spike/SpikePage";
+
+createRoot(document.getElementById("root")!).render(
+  <StrictMode>
+    <SpikePage />
+  </StrictMode>,
+);
@@ -0,0 +1,91 @@
+/* Spike #293 — PraisonAIUI-shaped mock event stream.
+ *
+ * These events use the real `SSEEvent` contract from `@/types` so the spike
+ * proves the BeautifulUI primitives can render against the *actual* event
+ * vocabulary the chat surface emits (see src/praisonaiui/templates/frontend/
+ * plugins/views/chat.js and RunEventType in src/frontend/src/types.ts).
+ *
+ * Nothing here talks to a live socket — that is an explicit non-goal of the
+ * spike. The timeline below is replayed by SpikePage with setTimeout to
+ * emulate SSE deltas / tool lifecycle / a HITL approval request.
+ */
+import type { SSEEvent } from "@/types";
+
+/** A scripted SSE frame plus the delay (ms) before it fires. */
+export type MockFrame = { at: number; event: SSEEvent };
+
+/** approval_request payload shaped like a row from GET /api/approvals. */
+export type ApprovalRequest = {
+  id: string;
+  tool: string;
+  args: Record<string, unknown>;
+  questions: { q: string; type: "radio" | "check"; options: string[] }[];
+};
+
+/* SSE-style text deltas — streamed word-by-word as `run_content`. */
+const ANSWER =
+  "Pistachio is your fastest-growing flavor — sales are up 23% this month " +
+  "and margins beat vanilla by 8 points. Stone-fruit flavors trend in the same range.";
+
+function contentDeltas(startAt: number, stepMs = 55): MockFrame[] {
+  const words = ANSWER.split(" ");
+  return words.map((w, i) => ({
+    at: startAt + i * stepMs,
+    event: { type: "run_content", token: w + " ", content: w + " " } as SSEEvent,
+  }));
+}
+
+const LAST_DELTA_AT = 300 + ANSWER.split(" ").length * 55;
+
+/* Reasoning + tool lifecycle, mapped to ThinkingState rows. */
+export const REASONING_STREAM: SSEEvent[] = [
+  { type: "reasoning_started" },
+  { type: "reasoning_step", step: "Reading flavor briefs", step_number: 1 },
+  { type: "tool_call_started", name: "sql.query", tool_call_id: "t1", args: { table: "sales" } },
+  { type: "tool_call_completed", name: "sql.query", tool_call_id: "t1", formatted_result: "6 flavors" },
+  { type: "reasoning_step", step: "Comparing tasting notes", step_number: 2 },
+  { type: "tool_call_started", name: "trends.lookup", tool_call_id: "t2", args: { window: "30d" } },
+  { type: "tool_call_completed", name: "trends.lookup", tool_call_id: "t2", formatted_result: "+23%" },
+  { type: "reasoning_completed" },
+];
+
+/* approval_request payload — the /api/approvals shape a HITL gate would push. */
+export const APPROVAL_REQUEST: ApprovalRequest = {
+  id: "appr_9f2",
+  tool: "menu.publish",
+  args: { flavors: 3, season: "winter" },
+  questions: [
+    { q: "How many flavors should we launch?", type: "radio", options: ["Three (core line)", "Five (full case)", "Just one hero"] },
+    { q: "Which mix-ins should we stock?", type: "check", options: ["Chocolate chips", "Waffle bits", "Sprinkles"] },
+  ],
+};
+
+/** Full scripted timeline for the streaming panel. */
+export function streamingTimeline(): MockFrame[] {
+  return [
+    { at: 0, event: { type: "run_started" } as SSEEvent },
+    ...contentDeltas(300),
+    { at: LAST_DELTA_AT + 100, event: { type: "run_completed" } as SSEEvent },
+  ];
+}
+
+/* ── Adapters: PraisonAIUI events → BeautifulUI primitive props ── */
+
+/** run_content deltas → StreamingText tokens (`cite` marks a source chip slot). */
+export function tokensFromDeltas(frames: MockFrame[]): { text: string; cite?: boolean }[] {
+  return frames
+    .filter((f) => f.event.type === "run_content")
+    .map((f) => ({ text: (f.event.token ?? f.event.content ?? "").trim() }));
+}
+
+/** reasoning_step / tool_call_* → ThinkingState rows. */
+export function rowsFromReasoning(events: SSEEvent[]): { primary: string; secondary?: string; mono?: boolean }[] {
+  const rows: { primary: string; secondary?: string; mono?: boolean }[] = [];
+  for (const e of events) {
+    if (e.type === "reasoning_step" && e.step) rows.push({ primary: e.step });
+    else if (e.type === "tool_call_started" && e.name) rows.push({ primary: e.name, mono: true });
+    else if (e.type === "tool_call_completed" && e.name)
+      rows.push({ primary: e.name, secondary: e.formatted_result, mono: true });
+  }
+  return rows;
+}
@@ -1,3 +1,3 @@
 """Version information."""

-__version__ = "0.3.143"
+__version__ = "0.3.147"
@@ -44,6 +44,43 @@ def _icon(symbol: str, fallback: str) -> str:
    return symbol if _supports_unicode() else fallback


+def _load_env_files() -> int:
+    """Load environment variables from the standard PraisonAI ``.env`` files.
+
+    Mirrors the PraisonAI CLI convention so ``env:VAR`` references (e.g. in
+    Channels config) resolve when the server was started without the shell
+    exporting the tokens. Reads ``~/.praisonai/.env`` first, then a ``.env`` in
+    the current working directory. Existing process env vars are never
+    overridden, and parsing is intentionally dependency-free (no python-dotenv).
+
+    Returns the number of variables newly set into ``os.environ``.
+    """
+    import os
+
+    loaded = 0
+    for env_path in (Path.home() / ".praisonai" / ".env", Path.cwd() / ".env"):
+        try:
+            if not env_path.is_file():
+                continue
+            for line in env_path.read_text(encoding="utf-8").splitlines():
+                line = line.strip()
+                if not line or line.startswith("#") or "=" not in line:
+                    continue
+                if line.startswith("export "):
+                    line = line[len("export "):]
+                key, value = line.split("=", 1)
+                key = key.strip()
+                value = value.strip().strip('"').strip("'")
+                if key and key not in os.environ:
+                    os.environ[key] = value
+                    loaded += 1
+        except Exception:
+            # Never let env loading break startup — resolution falls back to
+            # the existing process environment.
+            continue
+    return loaded
+
+
 app = typer.Typer(
    name="aiui",
    help="PraisonAIUI - YAML-driven website generator",
@@ -1380,6 +1417,17 @@ def run(
        console.print(f"[red]Error:[/red] App file not found: {app_file}")
        raise typer.Exit(code=1)

+    # Auto-load standard PraisonAI env files so `env:VAR` references (e.g. in
+    # Channels config) resolve even when the shell did not export the tokens.
+    # This matches the PraisonAI CLI convention and fixes the "Connected but
+    # bot never replies" class of issues caused by empty env-resolved tokens.
+    _loaded = _load_env_files()
+    if _loaded:
+        console.print(
+            f"[green]{_icon('✓', 'OK')}[/green] Loaded {_loaded} environment "
+            f"variable(s) from PraisonAI config"
+        )
+
    is_yaml = app_file.suffix in (".yaml", ".yml")

    if is_yaml:
@@ -155,12 +155,14 @@ class SDKFileDataStore(BaseDataStore):
        role = message.get("role", "user")
        content = message.get("content", "")

-        # UI-specific fields go into metadata
-        metadata: dict[str, Any] = {}
-        if "toolCalls" in message:
-            metadata["toolCalls"] = message["toolCalls"]
-        if "elements" in message:
-            metadata["elements"] = message["elements"]
+        # All non-core fields (toolCalls, elements, platform, sender, icon,
+        # channel_id, …) are preserved in metadata so channel and voice
+        # context survives reload — matching the SQLAlchemy backend.
+        metadata: dict[str, Any] = {
+            k: v
+            for k, v in message.items()
+            if k not in ("role", "content") and v is not None
+        }

        await asyncio.to_thread(
            self._store.add_message,
@@ -290,10 +292,13 @@ class SDKFileDataStore(BaseDataStore):
                "role": msg.role,
                "content": msg.content,
            }
-            # Restore UI-specific fields from metadata
-            if "toolCalls" in meta:
-                d["toolCalls"] = meta["toolCalls"]
-            if "elements" in meta:
-                d["elements"] = meta["elements"]
+            # Restore all persisted metadata (toolCalls, elements, platform,
+            # sender, icon, channel_id, …). Internal sentinel keys never
+            # appear here (the sentinel message is skipped above), but guard
+            # against them defensively without clobbering role/content.
+            for k, v in meta.items():
+                if k in (_META_KEY, _TITLE_KEY, _CUSTOM_META_KEY):
+                    continue
+                d.setdefault(k, v)
            messages.append(d)
        return messages
@@ -315,6 +315,13 @@ class ChannelsFeature(BaseFeatureProtocol):

        Returns None on success, or an error string on failure.
        """
+        # Idempotent start (#277): stop any bot already running for this channel
+        # before creating a new one. Without this, a toggle off/on, a re-add, or
+        # an autostart+manual race leaves a duplicate bot with duplicate event
+        # handlers attached while _live_bots only tracks the latest task.
+        if channel_id in _live_bots:
+            await self._stop_channel_bot(channel_id)
+
        platform = entry["platform"]
        config = entry.get("config", {})
        fallback_errors: List[str] = []
@@ -645,13 +652,24 @@ class ChannelsFeature(BaseFeatureProtocol):
        if hasattr(bot, "_session") and hasattr(bot._session, "chat"):
            _original_chat = bot._session.chat

-            async def _wrapped_chat(agent: Any, user_id: str, text: str) -> str:
+            async def _wrapped_chat(
+                agent: Any,
+                user_id: str,
+                text: str,
+                *args: Any,
+                **kwargs: Any,
+            ) -> str:
                """Wrap agent chat to broadcast responses AND intermediate
                tool/step events to the Chat UI.

                Hooks into the agent's StreamEventEmitter (same pattern as
                PraisonAIProvider._run_direct_mode) to capture tool_call,
                reasoning, and content events during execution.
+
+                Any extra positional/keyword arguments (e.g. ``chat_id``,
+                ``platform``, ``thread_id``) are forwarded untouched to the
+                wrapped session ``chat`` so per-conversation memory/session
+                continuity is preserved instead of raising ``TypeError``.
                """
                from .chat import _enrich_tool_payload, get_chat_manager

@@ -824,7 +842,7 @@ class ChannelsFeature(BaseFeatureProtocol):
                drain_task = asyncio.create_task(_drain_events())

                try:
-                    response = await _original_chat(agent, user_id, text)
+                    response = await _original_chat(agent, user_id, text, *args, **kwargs)
                finally:
                    # Signal drain to stop and clean up callback
                    try:
@@ -905,9 +923,15 @@ class ChannelsFeature(BaseFeatureProtocol):
        try:
            if hasattr(bot, "stop"):
                await bot.stop()
-            # Cancel the task if still running
+            # Cancel the task and await it so the old bot is fully torn down
+            # before this returns — otherwise a restart (#277) can spawn a new
+            # bot while the previous task is still cancelling (a zombie).
            if info and info.get("task") and not info["task"].done():
                info["task"].cancel()
+                try:
+                    await info["task"]
+                except asyncio.CancelledError:
+                    pass
            logger.info(f"Stopped bot for channel '{channel_id}'")
        except Exception as e:
            logger.error(f"Error stopping bot for '{channel_id}': {e}")
@@ -918,6 +942,19 @@ class ChannelsFeature(BaseFeatureProtocol):
        gw = self._get_gateway()
        if gw is not None:
            getattr(gw, "_channel_bots", {}).pop(channel_id, None)
+            # Mirror start-time registration: drop the task from the gateway
+            # too, so a restart (#277) leaves no stale/cancelled task behind.
+            # The dict form is overwritten on restart, but the list form would
+            # otherwise accumulate a cancelled task on every restart.
+            channel_tasks = getattr(gw, "_channel_tasks", None)
+            task = info.get("task") if info else None
+            if isinstance(channel_tasks, dict):
+                channel_tasks.pop(channel_id, None)
+            elif isinstance(channel_tasks, list) and task is not None:
+                try:
+                    channel_tasks.remove(task)
+                except ValueError:
+                    pass

        # Update local state
        ch = _channels.get(channel_id)
@@ -462,6 +462,24 @@ async def _run_and_broadcast(
    collected_tool_calls: Dict[str, Dict] = {}  # Merge STARTED+COMPLETED per tool_call_id
    collected_elements: List[Dict[str, Any]] = []

+    # Register this run so ``abort_run(run_id)`` can cancel it, and announce the
+    # run to WS clients up-front. Without this the ``_active_runs`` registry was
+    # never populated, so the Stop button and ``POST /api/chat/abort`` were
+    # silent no-ops that always returned ``no_active_run`` (issue #274). The
+    # ``run_started`` event also lets the frontend capture ``run_id`` and enable
+    # the Stop button immediately, before the first token arrives.
+    self_task = asyncio.current_task()
+    if self_task is not None:
+        mgr._active_runs[run_id] = self_task
+    started_event: Dict[str, Any] = {
+        "type": "run_started",
+        "session_id": session_id,
+        "run_id": run_id,
+    }
+    if agent_name:
+        started_event["agent_name"] = agent_name
+    await mgr.broadcast(session_id, started_event)
+
    try:
        # Pass image attachments to provider for SDK native multimodal handling
        run_kwargs = {}
@@ -644,6 +662,9 @@ async def _run_and_broadcast(
                "error": str(e),
            },
        )
+    finally:
+        # Always deregister the run, on normal completion, error, or abort.
+        mgr._active_runs.pop(run_id, None)

    # Save assistant response
    if full_response or collected_elements:
@@ -168,7 +168,9 @@ async def run_startup_hooks() -> None:
    """Execute all registered startup hooks.

    Called by the server during application startup.
-    Blocks until all hooks complete.
+    Has a timeout from AIUI_STARTUP_TIMEOUT (default 60s) so a hung hook
+    cannot block the server from reaching the lifespan ``yield`` — startup is
+    marked completed on timeout so ``/health/live`` still responds.
    """
    if _lifecycle_state["startup_completed"]:
        _log.warning("Startup hooks already executed")
@@ -177,15 +179,24 @@ async def run_startup_hooks() -> None:
    _log.info(f"Running {len(_startup_hooks)} startup hooks")
    start_time = time.perf_counter()

-    for hook in _startup_hooks:
-        try:
-            _log.debug(f"Executing startup hook: {getattr(hook, '__name__', str(hook))}")
-            result = hook()
-            if asyncio.iscoroutine(result):
-                await result
-        except Exception as e:
-            _log.error(f"Startup hook failed: {getattr(hook, '__name__', str(hook))}: {e}")
-            # Continue with other hooks - individual failures shouldn't break startup
+    # Get startup timeout from environment (mirrors AIUI_SHUTDOWN_TIMEOUT)
+    timeout = float(os.environ.get("AIUI_STARTUP_TIMEOUT", "60.0"))
+
+    async def _execute_hooks():
+        for hook in _startup_hooks:
+            try:
+                _log.debug(f"Executing startup hook: {getattr(hook, '__name__', str(hook))}")
+                result = hook()
+                if asyncio.iscoroutine(result):
+                    await result
+            except Exception as e:
+                _log.error(f"Startup hook failed: {getattr(hook, '__name__', str(hook))}: {e}")
+                # Continue with other hooks - individual failures shouldn't break startup
+
+    try:
+        await asyncio.wait_for(_execute_hooks(), timeout=timeout)
+    except asyncio.TimeoutError:
+        _log.error(f"Startup hooks timed out after {timeout}s")

    execution_time = time.perf_counter() - start_time
    _lifecycle_state.update(
@@ -9,6 +9,8 @@ Provides first-class MCP client functionality with:

 from __future__ import annotations

+import asyncio
+import contextlib
 import json
 import logging
 import subprocess
@@ -109,50 +111,52 @@ class StdioMCPClient:
        self.args = args
        self.session: Optional[ClientSession] = None
        self._connected = False
-        self._session_manager = None
+        self._exit_stack: contextlib.AsyncExitStack | None = None
        # For test backward compatibility
        self.process: Optional[subprocess.Popen] = None

    async def connect(self) -> bool:
        """Connect via stdio subprocess using official MCP SDK."""
+        # Store the stack on self before entering contexts so that a caller
+        # cancellation (e.g. asyncio.wait_for timeout, which raises the
+        # BaseException CancelledError past the except below) can still find
+        # and close the partially-entered stack via disconnect().
+        exit_stack = contextlib.AsyncExitStack()
+        self._exit_stack = exit_stack
        try:
-            # Use official MCP stdio_client instead of raw subprocess
            server_params = StdioServerParameters(command=self.command, args=self.args)
-
-            # Create a new session using stdio_client context manager
-            session_manager = stdio_client(server_params)
-            self.session = await session_manager.__aenter__()
-
-            # Initialize the MCP session
+            read_stream, write_stream = await exit_stack.enter_async_context(
+                stdio_client(server_params)
+            )
+            self.session = await exit_stack.enter_async_context(
+                ClientSession(read_stream, write_stream)
+            )
            await self.session.initialize()

-            # Store the context manager for cleanup
-            self._session_manager = session_manager
-
            self._connected = True
            logger.info(f"Connected to MCP stdio server: {self.command}")
            return True

        except Exception as e:
            logger.error(f"Failed to connect to MCP stdio server: {e}")
+            await exit_stack.aclose()
+            self._exit_stack = None
            self._connected = False
+            self.session = None
            return False

    async def disconnect(self) -> None:
        """Properly disconnect the MCP session."""
-        if self.session and self._connected:
+        if self._exit_stack:
            try:
-                # Properly exit the context manager
-                if hasattr(self, "_session_manager"):
-                    await self._session_manager.__aexit__(None, None, None)
-                else:
-                    await self.session.close()
+                await self._exit_stack.aclose()
            except Exception as e:
                logger.warning(f"Error during MCP session close: {e}")
            finally:
+                self._exit_stack = None
                self.session = None
                self._connected = False
-                self.process = None  # Reset for backward compatibility
+                self.process = None

    async def list_tools(self) -> List[ToolInfo]:
        """List tools via proper MCP protocol."""
@@ -203,42 +207,45 @@ class SSEMCPClient:
        self.headers = headers or {}
        self.session: Optional[ClientSession] = None
        self._connected = False
-        self._session_manager = None
+        self._exit_stack: contextlib.AsyncExitStack | None = None

    async def connect(self) -> bool:
        """Connect via SSE using official MCP SDK."""
+        # Store the stack on self before entering contexts so a caller
+        # cancellation (e.g. asyncio.wait_for timeout) can still close the
+        # partially-entered stack via disconnect(). See StdioMCPClient.connect.
+        exit_stack = contextlib.AsyncExitStack()
+        self._exit_stack = exit_stack
        try:
-            # Use official MCP sse_client context manager
-            session_manager = sse_client(self.url, headers=self.headers)
-            self.session = await session_manager.__aenter__()
-
-            # Initialize the MCP session
+            read_stream, write_stream = await exit_stack.enter_async_context(
+                sse_client(self.url, headers=self.headers)
+            )
+            self.session = await exit_stack.enter_async_context(
+                ClientSession(read_stream, write_stream)
+            )
            await self.session.initialize()

-            # Store the context manager for cleanup
-            self._session_manager = session_manager
-
            self._connected = True
            logger.info(f"Connected to MCP SSE server at {self.url}")
            return True

        except Exception as e:
            logger.error(f"Failed to connect to MCP SSE server: {e}")
+            await exit_stack.aclose()
+            self._exit_stack = None
            self._connected = False
+            self.session = None
            return False

    async def disconnect(self) -> None:
        """Properly disconnect SSE."""
-        if self.session and self._connected:
+        if self._exit_stack:
            try:
-                # Properly exit the context manager
-                if hasattr(self, "_session_manager") and self._session_manager:
-                    await self._session_manager.__aexit__(None, None, None)
-                else:
-                    await self.session.close()
+                await self._exit_stack.aclose()
            except Exception as e:
                logger.warning(f"Error during MCP SSE session close: {e}")
            finally:
+                self._exit_stack = None
                self.session = None
                self._connected = False

@@ -354,6 +361,13 @@ class MCPClientManager:
            connection_data=server_config,
        )

+        # Disconnect any existing client for this name (reconnect) before
+        # replacing it, otherwise the previous stdio subprocess/streams are
+        # orphaned when the reference is overwritten.
+        previous_client = self._clients.pop(name, None)
+        if previous_client is not None:
+            await self._safe_disconnect(name, previous_client)
+
        # Register server
        _mcp_servers[name] = server
        self._clients[name] = client
@@ -362,8 +376,7 @@ class MCPClientManager:
        await self._notify_status_change(server)

        try:
-            # Attempt connection
-            connected = await client.connect()
+            connected = await asyncio.wait_for(client.connect(), timeout=120.0)
            if connected:
                server.status = MCPStatus.CONNECTED
                server.tools = await client.list_tools()
@@ -374,16 +387,38 @@ class MCPClientManager:
            else:
                server.status = MCPStatus.ERROR
                server.last_error = "Connection failed"
+                await self._cleanup_failed_client(name, client)

+        except asyncio.TimeoutError:
+            logger.error("Timed out connecting MCP server %s", name)
+            server.status = MCPStatus.ERROR
+            server.last_error = "Connection timed out after 120s (check npx/Node and server args)"
+            await client.disconnect()
        except Exception as e:
            logger.exception(f"Failed to connect MCP server {name}")
            server.status = MCPStatus.ERROR
            server.last_error = str(e)
+            await self._cleanup_failed_client(name, client)

        # Notify final status
        await self._notify_status_change(server)
        return server

+    async def _safe_disconnect(self, name: str, client: MCPClientProtocol) -> None:
+        """Disconnect a client, logging (not raising) on failure."""
+        try:
+            await client.disconnect()
+        except Exception:
+            logger.exception(f"Error disconnecting MCP client {name}")
+
+    async def _cleanup_failed_client(self, name: str, client: MCPClientProtocol) -> None:
+        """Release a client whose connection failed and drop it from the registry."""
+        await self._safe_disconnect(name, client)
+        # Only remove if this client is still the registered one (avoid evicting
+        # a newer client from a concurrent reconnect).
+        if self._clients.get(name) is client:
+            del self._clients[name]
+
    async def disconnect_server(
        self, name: str, session_context: Optional[Dict[str, Any]] = None
    ) -> bool:
@@ -252,6 +252,7 @@ window.aiui.registerSurfaceRenderer = function(surfaceId, renderFn) {
 window.aiui.surfaces = SURFACE_REGISTRY;
 window.aiui.version = '1';
 window.aiui.components = COMPONENT_REGISTRY;
+window.aiui.renderComponent = renderComponent;
 window.aiui.registerSlot = function (name, renderFn) { SLOT_REGISTRY[name] = renderFn; };
 window.aiui.slots = SLOT_REGISTRY;
 window.aiui.sdk = aiuiSdk;
@@ -1,65 +1,154 @@
 /**
 * Approvals View — execution approval queue
 * API: /api/approvals
+ *
+ * Pending items render as an Approval Card (human-in-the-loop): risk, question,
+ * optional preset options / custom answer / confidence copy — surfaced only when
+ * the API supplies them, so the card stays a plain approve/deny prompt otherwise.
 */
-import { isSkillApproval } from '/plugins/views/_helpers.js';
+import { isSkillApproval, esc, emptyState } from '/plugins/views/_helpers.js';
+
+const RISK = {
+  low: { icon: '✅', color: '#22c55e' },
+  medium: { icon: '⚠️', color: '#eab308' },
+  high: { icon: '🟠', color: '#f97316' },
+  critical: { icon: '🔴', color: '#ef4444' },
+};
+
+/**
+ * Normalise a raw /api/approvals item into an ApprovalPromptModel-shaped object.
+ * Keeps the field-mapping in one place; only carries what the API actually sends.
+ */
+function toApprovalModel(a) {
+  const id = a.id || a.approval_id || '';
+  const risk = String(a.risk_level || '').toLowerCase();
+  const options = Array.isArray(a.options)
+    ? a.options.map(o => (typeof o === 'string' ? { value: o, label: o } : { value: o.value ?? o.id ?? o.label, label: o.label ?? String(o.value ?? o.id) }))
+    : [];
+  const confidence = typeof a.confidence === 'number' ? Math.round(a.confidence * (a.confidence <= 1 ? 100 : 1)) : null;
+  return {
+    id,
+    title: a.tool_name || a.action || 'Unknown',
+    question: a.reason || a.description || a.question || '',
+    agent: a.agent_name || '',
+    risk,
+    riskMeta: RISK[risk] || { icon: '❓', color: 'var(--db-text-dim)' },
+    createdAt: a.created_at ? new Date(a.created_at * 1000).toLocaleString() : '',
+    options,
+    allowCustom: a.allow_custom_answer === true || (options.length === 0 && a.custom_answer !== false && !!a.question),
+    confidence,
+    isSkill: isSkillApproval(a),
+  };
+}
+
+function approvalCard(m) {
+  const optionsHtml = m.options.length
+    ? `<div class="db-approval-options" role="radiogroup" aria-label="Response options" style="display:flex;flex-wrap:wrap;gap:6px;margin-top:10px">
+        ${m.options.map((o, i) => `
+          <label style="display:inline-flex;align-items:center;gap:6px;font-size:12px;padding:4px 10px;border:1px solid var(--db-border);border-radius:6px;cursor:pointer">
+            <input type="radio" name="opt-${esc(m.id)}" value="${esc(o.value)}" ${i === 0 ? 'checked' : ''}> ${esc(o.label)}
+          </label>`).join('')}
+      </div>`
+    : '';
+  const customHtml = m.allowCustom
+    ? `<input class="approv-custom" data-id="${esc(m.id)}" type="text" placeholder="Custom answer (optional)"
+        aria-label="Custom answer for ${esc(m.title)}"
+        style="margin-top:10px;width:100%;padding:6px 10px;font-size:12px;background:var(--db-bg);color:var(--db-text);border:1px solid var(--db-border);border-radius:6px">`
+    : '';
+  const confidenceHtml = m.confidence != null
+    ? `<div style="font-size:11px;color:var(--db-text-dim);margin-top:6px">Model confidence: ${m.confidence}%</div>`
+    : '';
+  const skillHtml = m.isSkill
+    ? `<a href="/skills?tab=pending&approval_id=${encodeURIComponent(m.id)}" data-skill-studio="1" style="display:inline-block;margin-top:8px;font-size:11px;color:var(--db-accent);text-decoration:none">Skill write? Open Skills Studio →</a>`
+    : '';
+
+  return `
+    <div class="db-card db-approval-card" data-id="${esc(m.id)}" role="group"
+         aria-label="Approval request: ${esc(m.title)}"
+         style="margin-bottom:12px;padding:16px 20px;border-left:3px solid ${m.riskMeta.color}">
+      <div style="display:flex;align-items:flex-start;justify-content:space-between;gap:12px">
+        <div style="flex:1;min-width:0">
+          <div style="font-weight:600;font-size:14px">${m.riskMeta.icon} ${esc(m.title)}</div>
+          ${m.question ? `<div style="font-size:13px;color:var(--db-text);margin-top:6px">${esc(m.question)}</div>` : ''}
+          <div style="font-size:12px;color:var(--db-text-dim);margin-top:4px">Agent: ${esc(m.agent) || '—'}${m.risk ? ` · risk: ${esc(m.risk)}` : ''}</div>
+          ${m.createdAt ? `<div style="font-size:11px;color:var(--db-text-dim)">${esc(m.createdAt)}</div>` : ''}
+          ${optionsHtml}
+          ${customHtml}
+          ${confidenceHtml}
+          ${skillHtml}
+        </div>
+        <div style="display:flex;gap:8px;flex-shrink:0">
+          <button class="approv-yes" data-id="${esc(m.id)}" aria-label="Approve ${esc(m.title)}"
+            style="padding:6px 16px;background:rgba(34,197,94,0.15);color:#22c55e;border:1px solid rgba(34,197,94,0.3);border-radius:6px;cursor:pointer;font-size:12px">✓ Approve</button>
+          <button class="approv-no" data-id="${esc(m.id)}" aria-label="Deny ${esc(m.title)}"
+            style="padding:6px 16px;background:rgba(239,68,68,0.15);color:#ef4444;border:1px solid rgba(239,68,68,0.3);border-radius:6px;cursor:pointer;font-size:12px">✗ Deny</button>
+        </div>
+      </div>
+    </div>`;
+}

 export async function render(container) {
  container.innerHTML = '<div class="db-loading"><div class="db-spinner"></div></div>';

  let pending = [], history = [], policies = {};
  try { const r = await fetch('/api/approvals/pending'); pending = (await r.json()).approvals || []; } catch(e) {}
-  try { const r = await fetch('/api/approvals/history'); history = (await r.json()).approvals || (await r.json()).history || []; } catch(e) {}
+  try { const r = await fetch('/api/approvals/history'); const d = await r.json(); history = d.approvals || d.history || []; } catch(e) {}
  try { const r = await fetch('/api/approvals/policies'); policies = await r.json(); } catch(e) {}

-  const riskIcons = { low: '✅', medium: '⚠️', high: '🟠', critical: '🔴' };
+  const models = pending.map(toApprovalModel);

  container.innerHTML = `
    <div class="db-columns" style="grid-template-columns:repeat(3,1fr);margin-bottom:24px">
-      <div class="db-card"><div class="db-card-title">Pending</div><div class="db-card-value" style="color:#eab308">${pending.length}</div></div>
-      <div class="db-card"><div class="db-card-title">Risk Threshold</div><div class="db-card-value" style="font-size:18px">${policies.risk_threshold || 'high'}</div></div>
+      <div class="db-card"><div class="db-card-title">Pending</div><div class="db-card-value" style="color:#eab308">${models.length}</div></div>
+      <div class="db-card"><div class="db-card-title">Risk Threshold</div><div class="db-card-value" style="font-size:18px">${esc(policies.risk_threshold || 'high')}</div></div>
      <div class="db-card"><div class="db-card-title">Auto-Approve Tools</div><div class="db-card-value" style="font-size:18px">${(policies.auto_approve_tools || []).length}</div></div>
    </div>

    <h3 style="margin:0 0 12px;font-size:15px;font-weight:600">Pending Approvals</h3>
    <div id="approv-pending" style="margin-bottom:28px">
-      ${pending.length === 0 ? '<div class="db-viewer"><pre>No pending approvals</pre></div>' : ''}
+      ${models.length === 0 ? emptyState({ icon: '✅', title: 'No pending approvals', body: 'All clear — nothing needs your decision.' }) : models.map(approvalCard).join('')}
    </div>

    <h3 style="margin:0 0 12px;font-size:15px;font-weight:600">Recent History</h3>
    <div id="approv-history"></div>
  `;

-  const pendingEl = container.querySelector('#approv-pending');
-  pending.forEach(a => {
-    const card = document.createElement('div');
-    card.className = 'db-card';
-    card.style.cssText = 'margin-bottom:10px;padding:16px 20px;display:flex;align-items:center;justify-content:space-between';
-    const isSkill = isSkillApproval(a);
-    card.innerHTML = `
-      <div style="flex:1">
-        <div style="font-weight:500;font-size:14px">${riskIcons[a.risk_level] || '❓'} ${a.tool_name || a.action || 'Unknown'}</div>
-        <div style="font-size:12px;color:var(--db-text-dim);margin-top:4px">Agent: ${a.agent_name || '—'} · ${a.reason || ''}</div>
-        <div style="font-size:11px;color:var(--db-text-dim)">${a.created_at ? new Date(a.created_at * 1000).toLocaleString() : ''}</div>
-        ${isSkill ? `<a href="/skills?tab=pending&approval_id=${encodeURIComponent(a.id || a.approval_id || '')}" data-skill-studio="1" style="font-size:11px;color:var(--db-accent);text-decoration:none">Skill write? Open Skills Studio →</a>` : ''}
-      </div>
-      <div style="display:flex;gap:8px">
-        <button class="approv-yes" data-id="${a.id || a.approval_id}" style="padding:6px 16px;background:rgba(34,197,94,0.15);color:#22c55e;border:1px solid rgba(34,197,94,0.3);border-radius:6px;cursor:pointer;font-size:12px">✓ Approve</button>
-        <button class="approv-no" data-id="${a.id || a.approval_id}" style="padding:6px 16px;background:rgba(239,68,68,0.15);color:#ef4444;border:1px solid rgba(239,68,68,0.3);border-radius:6px;cursor:pointer;font-size:12px">✗ Deny</button>
-      </div>
-    `;
-    pendingEl.appendChild(card);
-  });
+  function collectExtra(id) {
+    const card = container.querySelector(`.db-approval-card[data-id="${CSS.escape(id)}"]`);
+    const body = {};
+    if (!card) return body;
+    const opt = card.querySelector(`input[name="opt-${CSS.escape(id)}"]:checked`);
+    if (opt) body.option = opt.value;
+    const custom = card.querySelector('.approv-custom');
+    if (custom && custom.value.trim()) body.custom_answer = custom.value.trim();
+    // Also surface the human's choice as `reason` — the only extra field the
+    // approve/deny endpoints persist today — so the answer isn't dropped while
+    // `option`/`custom_answer` stay for a future richer backend.
+    const reason = body.custom_answer || body.option;
+    if (reason != null && reason !== '') body.reason = String(reason);
+    return body;
+  }

-  container.querySelectorAll('.approv-yes').forEach(b => b.addEventListener('click', async () => { try { await fetch(`/api/approvals/${b.dataset.id}/approve`, {method:'POST'}); render(container); } catch(e){} }));
-  container.querySelectorAll('.approv-no').forEach(b => b.addEventListener('click', async () => { try { await fetch(`/api/approvals/${b.dataset.id}/deny`, {method:'POST'}); render(container); } catch(e){} }));
+  async function decide(id, action) {
+    try {
+      await fetch(`/api/approvals/${encodeURIComponent(id)}/${action}`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify(collectExtra(id)),
+      });
+      render(container);
+    } catch(e) {}
+  }
+
+  container.querySelectorAll('.approv-yes').forEach(b => b.addEventListener('click', () => decide(b.dataset.id, 'approve')));
+  container.querySelectorAll('.approv-no').forEach(b => b.addEventListener('click', () => decide(b.dataset.id, 'deny')));

  const historyEl = container.querySelector('#approv-history');
  (Array.isArray(history) ? history : []).slice(0, 20).forEach(h => {
    const div = document.createElement('div');
    div.style.cssText = 'padding:8px 0;border-bottom:1px solid var(--db-border);font-size:13px;display:flex;justify-content:space-between';
    const approved = h.status === 'approved' || h.approved;
-    div.innerHTML = `<span>${h.tool_name || h.action || '?'} — ${h.agent_name || ''}</span><span style="color:${approved ? '#22c55e' : '#ef4444'}">${approved ? '✓ approved' : '✗ denied'}</span>`;
+    div.innerHTML = `<span>${esc(h.tool_name || h.action || '?')} — ${esc(h.agent_name || '')}</span><span style="color:${approved ? '#22c55e' : '#ef4444'}">${approved ? '✓ approved' : '✗ denied'}</span>`;
    historyEl.appendChild(div);
  });
  if (!history.length) historyEl.innerHTML = '<div style="font-size:13px;color:var(--db-text-dim)">No history</div>';
@@ -116,7 +116,7 @@ function openConnectForm(host, container) {
      <h3 style="margin:0 0 16px;font-size:14px;font-weight:600">Connect MCP Server</h3>
      <div class="db-form-group">
        <label class="db-form-label">Name</label>
-        <input id="mcp-f-name" class="db-form-input" placeholder="my-server">
+        <input id="mcp-f-name" class="db-form-input" value="everything-demo" placeholder="my-server">
      </div>
      <div class="db-form-group">
        <label class="db-form-label">Transport</label>
@@ -128,11 +128,11 @@ function openConnectForm(host, container) {
      <div id="mcp-f-stdio">
        <div class="db-form-group">
          <label class="db-form-label">Command</label>
-          <input id="mcp-f-command" class="db-form-input" placeholder="npx">
+          <input id="mcp-f-command" class="db-form-input" value="npx" placeholder="npx" required>
        </div>
        <div class="db-form-group">
          <label class="db-form-label">Arguments (space-separated)</label>
-          <input id="mcp-f-args" class="db-form-input" placeholder="-y @modelcontextprotocol/server-filesystem /tmp">
+          <input id="mcp-f-args" class="db-form-input" value="-y @modelcontextprotocol/server-everything" placeholder="-y @modelcontextprotocol/server-everything">
        </div>
      </div>
      <div id="mcp-f-url" style="display:none">
@@ -144,6 +144,74 @@ async def test_auto_start_skips_disabled_channels():
    assert mock_start.call_count == 0


+@pytest.mark.asyncio
+async def test_start_channel_bot_is_idempotent_stops_existing_first():
+    """Regression for #277: starting a bot for a channel that already has a
+    live bot must stop the existing one first so no duplicate bot/task lingers.
+    """
+    feature = ChannelsFeature()
+    gw = _GatewayDictTasks()
+    entry = {"id": "ch6", "platform": "discord", "config": {"bot_token": "x" * 30}}
+
+    with patch.object(feature, "_get_gateway", return_value=gw), patch.object(
+        feature, "_create_bot_direct", side_effect=lambda *a, **k: _FakeBot()
+    ), patch.object(feature, "_attach_chat_bridge"):
+        await feature._start_channel_bot("ch6", entry)
+        first = channels_mod._live_bots["ch6"]
+
+        await feature._start_channel_bot("ch6", entry)
+        second = channels_mod._live_bots["ch6"]
+
+    # A new bot/task replaced the first, and only one entry is tracked.
+    assert second is not first
+    assert first["bot"] is not second["bot"]
+    assert list(channels_mod._live_bots.keys()).count("ch6") == 1
+    # The first task was cancelled by the pre-start _stop_channel_bot guard.
+    assert first["task"].cancelled() or first["task"].done()
+    # Gateway has exactly one (fresh) bot reference — no stale entry.
+    assert gw._channel_bots["ch6"] is second["bot"]
+    # ...and exactly one (fresh) task reference in the dict registry.
+    assert gw._channel_tasks["ch6"] is second["task"]
+
+
+@pytest.mark.asyncio
+async def test_restart_leaves_no_stale_task_in_list_gateway():
+    """Regression for #277: with a list-based gateway _channel_tasks, a restart
+    must not accumulate the old cancelled task alongside the fresh one.
+    """
+    feature = ChannelsFeature()
+    gw = _GatewayListTasks()
+    entry = {"id": "ch8", "platform": "discord", "config": {"bot_token": "x" * 30}}
+
+    with patch.object(feature, "_get_gateway", return_value=gw), patch.object(
+        feature, "_create_bot_direct", side_effect=lambda *a, **k: _FakeBot()
+    ), patch.object(feature, "_attach_chat_bridge"):
+        await feature._start_channel_bot("ch8", entry)
+        await feature._start_channel_bot("ch8", entry)
+        current = channels_mod._live_bots["ch8"]
+
+    # Only the fresh task remains in the list — the cancelled one was removed.
+    assert gw._channel_tasks == [current["task"]]
+
+
+@pytest.mark.asyncio
+async def test_start_channel_bot_no_existing_does_not_stop():
+    """Guard must be a no-op on first start (nothing in _live_bots)."""
+    feature = ChannelsFeature()
+    gw = _GatewayDictTasks()
+    entry = {"id": "ch7", "platform": "discord", "config": {"bot_token": "x" * 30}}
+
+    with patch.object(feature, "_get_gateway", return_value=gw), patch.object(
+        feature, "_create_bot_direct", return_value=_FakeBot()
+    ), patch.object(feature, "_attach_chat_bridge"), patch.object(
+        feature, "_stop_channel_bot"
+    ) as mock_stop:
+        error = await feature._start_channel_bot("ch7", entry)
+
+    assert error is None
+    mock_stop.assert_not_called()
+
+
@pytest.mark.asyncio
 async def test_auto_start_failure_sets_not_running_and_error():
    feature = ChannelsFeature()
@@ -0,0 +1,88 @@
+"""Regression coverage for issue #287: the channels chat bridge dropped
+extra keyword arguments (e.g. ``chat_id``, ``platform``) because the
+``_wrapped_chat`` wrapper installed over ``bot._session.chat`` had a fixed
+``(agent, user_id, text)`` signature.
+
+When the bot internals invoked the session with per-conversation kwargs the
+wrapper raised ``TypeError: _wrapped_chat() got an unexpected keyword
+argument 'chat_id'`` and the bot silently failed to reply. The wrapper must
+forward any extra positional/keyword arguments untouched to the original
+session ``chat`` so memory/session continuity is preserved.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+
+from praisonaiui.features.channels import ChannelsFeature
+
+
+class _FakeSession:
+    """Minimal session whose ``chat`` records forwarded arguments."""
+
+    def __init__(self) -> None:
+        self.received_args: tuple = ()
+        self.received_kwargs: dict = {}
+
+    async def chat(self, agent: Any, user_id: str, text: str, *args: Any, **kwargs: Any) -> str:
+        self.received_args = args
+        self.received_kwargs = dict(kwargs)
+        return f"reply to {text}"
+
+
+class _FakeBot:
+    def __init__(self) -> None:
+        self._session = _FakeSession()
+
+
+@pytest.mark.asyncio
+async def test_wrapped_chat_forwards_kwargs() -> None:
+    feature = ChannelsFeature()
+    bot = _FakeBot()
+
+    feature._attach_chat_bridge("chan-1", bot, "telegram")
+
+    # The bridge replaced the session chat with the wrapper.
+    assert bot._session.chat.__name__ == "_wrapped_chat"
+
+    reply = await bot._session.chat(
+        object(),
+        "user-42",
+        "Hi",
+        chat_id="987654321",
+        platform="telegram",
+        user_id_meta="987654321",
+    )
+
+    assert reply == "reply to Hi"
+    assert bot._session.received_kwargs["chat_id"] == "987654321"
+    assert bot._session.received_kwargs["platform"] == "telegram"
+
+
+@pytest.mark.asyncio
+async def test_wrapped_chat_forwards_extra_positional_args() -> None:
+    feature = ChannelsFeature()
+    bot = _FakeBot()
+
+    feature._attach_chat_bridge("chan-2", bot, "slack")
+
+    reply = await bot._session.chat(object(), "user-7", "Hello", "extra")
+
+    assert reply == "reply to Hello"
+    assert bot._session.received_args == ("extra",)
+
+
+@pytest.mark.asyncio
+async def test_wrapped_chat_still_works_without_kwargs() -> None:
+    """No-kwargs path (web UI / existing callers) must keep working."""
+    feature = ChannelsFeature()
+    bot = _FakeBot()
+
+    feature._attach_chat_bridge("chan-3", bot, "discord")
+
+    reply = await bot._session.chat(object(), "user-9", "Ping")
+
+    assert reply == "reply to Ping"
+    assert bot._session.received_kwargs == {}
@@ -0,0 +1,145 @@
+"""Tests for chat run abort / cancellation (issue #274).
+
+Regression coverage for the wiring bug where ``ChatManager._active_runs`` was
+never populated, so ``abort_run`` (Stop button / ``POST /api/chat/abort`` /
+``chat_abort`` WS) always returned ``no_active_run`` even while a run was
+streaming.
+"""
+
+from __future__ import annotations
+
+import asyncio
+
+import pytest
+
+from praisonaiui.provider import RunEvent, RunEventType
+
+# ── Unit: abort_run against the _active_runs registry ────────────────
+
+
+@pytest.mark.asyncio
+async def test_abort_cancels_active_run():
+    from praisonaiui.features.chat import ChatManager
+
+    mgr = ChatManager()
+    run_id = "test-run"
+
+    async def slow_run():
+        await asyncio.sleep(60)
+
+    task = asyncio.create_task(slow_run())
+    mgr._active_runs[run_id] = task
+
+    result = await mgr.abort_run(run_id)
+
+    assert result["status"] == "aborted"
+    assert result["run_id"] == run_id
+    # Give the event loop a tick to process the cancellation.
+    await asyncio.sleep(0)
+    assert task.cancelled() or task.done()
+    # Registry entry removed on abort.
+    assert run_id not in mgr._active_runs
+
+
+@pytest.mark.asyncio
+async def test_abort_unknown_run_returns_no_active_run():
+    from praisonaiui.features.chat import ChatManager
+
+    mgr = ChatManager()
+    result = await mgr.abort_run("does-not-exist")
+    assert result["status"] == "no_active_run"
+    assert result["run_id"] == "does-not-exist"
+
+
+# ── Integration: _run_and_broadcast registers + aborts cleanly ───────
+
+
+class _FakeWS:
+    """Collects broadcast frames pushed by ChatManager.broadcast."""
+
+    def __init__(self) -> None:
+        self.messages: list[dict] = []
+
+    async def send_text(self, data: str) -> None:
+        import json
+
+        self.messages.append(json.loads(data))
+
+
+class _SlowProvider:
+    """Streams a couple of tokens then hangs, keeping the run active."""
+
+    async def run(self, content, *, session_id=None, agent_name=None, **kwargs):
+        yield RunEvent(type=RunEventType.RUN_CONTENT, token="hello ")
+        yield RunEvent(type=RunEventType.RUN_CONTENT, token="world")
+        # Simulate a long generation so the run stays registered.
+        await asyncio.sleep(60)
+        yield RunEvent(type=RunEventType.RUN_COMPLETED, content="hello world")
+
+
+class _FakeDatastore:
+    async def add_message(self, *args, **kwargs):
+        return None
+
+
+@pytest.mark.asyncio
+async def test_run_and_broadcast_registers_and_aborts(monkeypatch):
+    import praisonaiui.server as server
+    from praisonaiui.features.chat import ChatManager, _run_and_broadcast, set_chat_manager
+
+    mgr = ChatManager()
+    set_chat_manager(mgr)
+    ws = _FakeWS()
+    mgr.add_ws_client("c1", ws)
+
+    monkeypatch.setattr(server, "get_provider", lambda: _SlowProvider())
+    monkeypatch.setattr(server, "_datastore", _FakeDatastore())
+
+    task = asyncio.create_task(_run_and_broadcast("hi", "sess-1", "Agent", None))
+
+    # Wait until the run is registered (populated by _run_and_broadcast).
+    for _ in range(100):
+        if mgr._active_runs:
+            break
+        await asyncio.sleep(0.01)
+    assert mgr._active_runs, "run was never registered in _active_runs"
+
+    run_id = next(iter(mgr._active_runs))
+
+    # Frontend contract: a run_started frame is broadcast with the run_id.
+    started = [m for m in ws.messages if m.get("type") == "run_started"]
+    assert started and started[0]["run_id"] == run_id
+
+    # Abort the active run.
+    result = await mgr.abort_run(run_id)
+    assert result["status"] == "aborted"
+
+    # Let the cancellation propagate and _run_and_broadcast finish.
+    await asyncio.sleep(0.05)
+
+    # Clients receive a run_cancelled frame, and the registry is cleaned up.
+    cancelled = [m for m in ws.messages if m.get("type") == "run_cancelled"]
+    assert cancelled and cancelled[0]["run_id"] == run_id
+    assert run_id not in mgr._active_runs
+    assert task.done()
+
+
+@pytest.mark.asyncio
+async def test_run_and_broadcast_cleans_up_on_normal_completion(monkeypatch):
+    import praisonaiui.server as server
+    from praisonaiui.features.chat import ChatManager, _run_and_broadcast, set_chat_manager
+
+    class _QuickProvider:
+        async def run(self, content, *, session_id=None, agent_name=None, **kwargs):
+            yield RunEvent(type=RunEventType.RUN_CONTENT, token="done")
+            yield RunEvent(type=RunEventType.RUN_COMPLETED, content="done")
+
+    mgr = ChatManager()
+    set_chat_manager(mgr)
+    monkeypatch.setattr(server, "get_provider", lambda: _QuickProvider())
+    monkeypatch.setattr(server, "_datastore", _FakeDatastore())
+
+    await _run_and_broadcast("hi", "sess-2", None, None)
+
+    # No leftover registry entries after a normal run.
+    assert mgr._active_runs == {}
@@ -117,6 +117,51 @@ async def test_elements_persisted_in_metadata(adapter):
    assert messages[0]["elements"] == elements


+# ── Channel metadata round-trip ──────────────────────────────────────
+
+
+@pytest.mark.asyncio
+async def test_channel_metadata_round_trip(adapter):
+    """platform/sender/icon/channel_id survive a write→read cycle."""
+    sid = str(uuid.uuid4())
+    await adapter.create_session(sid)
+    await adapter.add_message(sid, {
+        "role": "user",
+        "content": "[Bob] hi",
+        "platform": "discord",
+        "sender": "Bob",
+        "icon": "game",
+        "channel_id": "ch-99",
+    })
+
+    messages = await adapter.get_messages(sid)
+    assert len(messages) == 1
+    assert messages[0]["platform"] == "discord"
+    assert messages[0]["sender"] == "Bob"
+    assert messages[0]["icon"] == "game"
+    assert messages[0]["channel_id"] == "ch-99"
+    # Core fields still intact
+    assert messages[0]["role"] == "user"
+    assert messages[0]["content"] == "[Bob] hi"
+
+
+@pytest.mark.asyncio
+async def test_none_metadata_values_dropped(adapter):
+    """None-valued fields are not persisted (kept out of the payload)."""
+    sid = str(uuid.uuid4())
+    await adapter.create_session(sid)
+    await adapter.add_message(sid, {
+        "role": "assistant",
+        "content": "reply",
+        "platform": "voice",
+        "sender": None,
+    })
+
+    messages = await adapter.get_messages(sid)
+    assert messages[0]["platform"] == "voice"
+    assert "sender" not in messages[0]
+
+
 # ── Title auto-generation ────────────────────────────────────────────


@@ -0,0 +1,68 @@
+"""Unit tests for `_load_env_files` — auto-loading PraisonAI `.env` files.
+
+Regression coverage for issue #286: `env:VAR` channel token references
+resolved to empty strings because `aiui run` never loaded the standard
+`~/.praisonai/.env` file, so bots showed "Connected" but never replied.
+"""
+
+import os
+
+from praisonaiui.cli import _load_env_files
+
+
+def _write_env(path, body):
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(body, encoding="utf-8")
+
+
+def test_loads_home_praisonai_env(tmp_path, monkeypatch):
+    """Variables in ~/.praisonai/.env are loaded into the process env."""
+    monkeypatch.setattr("pathlib.Path.home", lambda: tmp_path)
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.delenv("TELEGRAM_BOT_TOKEN", raising=False)
+
+    _write_env(tmp_path / ".praisonai" / ".env", 'TELEGRAM_BOT_TOKEN="123:AAtoken"\n')
+
+    loaded = _load_env_files()
+
+    assert loaded >= 1
+    assert os.environ["TELEGRAM_BOT_TOKEN"] == "123:AAtoken"
+
+
+def test_does_not_override_existing_env(tmp_path, monkeypatch):
+    """An already-set process variable is never clobbered by the file."""
+    monkeypatch.setattr("pathlib.Path.home", lambda: tmp_path)
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setenv("SLACK_BOT_TOKEN", "xoxb-from-shell")
+
+    _write_env(tmp_path / ".praisonai" / ".env", "SLACK_BOT_TOKEN=xoxb-from-file\n")
+
+    _load_env_files()
+
+    assert os.environ["SLACK_BOT_TOKEN"] == "xoxb-from-shell"
+
+
+def test_ignores_comments_blanks_and_export_prefix(tmp_path, monkeypatch):
+    """Comments/blank lines are skipped and `export ` prefixes are stripped."""
+    monkeypatch.setattr("pathlib.Path.home", lambda: tmp_path)
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.delenv("FOO", raising=False)
+    monkeypatch.delenv("BAR", raising=False)
+
+    _write_env(
+        tmp_path / ".praisonai" / ".env",
+        "# a comment\n\nexport FOO=foo\nBAR='bar'\n",
+    )
+
+    _load_env_files()
+
+    assert os.environ["FOO"] == "foo"
+    assert os.environ["BAR"] == "bar"
+
+
+def test_missing_files_are_harmless(tmp_path, monkeypatch):
+    """No env files present → returns 0 and does not raise."""
+    monkeypatch.setattr("pathlib.Path.home", lambda: tmp_path)
+    monkeypatch.chdir(tmp_path)
+
+    assert _load_env_files() == 0
@@ -246,6 +246,25 @@ class TestLifecycleHooks:
            assert elapsed < 0.5
            assert _lifecycle_state["shutdown_initiated"] is True

+    @pytest.mark.asyncio
+    async def test_startup_timeout(self):
+        """Test startup timeout mechanism — a hung hook must not block forever."""
+
+        @on_app_startup
+        async def slow_hook():
+            await asyncio.sleep(2)  # Simulate hung startup
+
+        # Set short timeout
+        with patch.dict(os.environ, {"AIUI_STARTUP_TIMEOUT": "0.1"}):
+            start_time = time.time()
+            await run_startup_hooks()
+            elapsed = time.time() - start_time
+
+            # Should timeout quickly and still mark startup completed so the
+            # server reaches lifespan yield and /health/live responds
+            assert elapsed < 0.5
+            assert _lifecycle_state["startup_completed"] is True
+
    @pytest.mark.asyncio
    async def test_startup_idempotent(self):
        """Test that startup hooks are only run once."""
@@ -110,18 +110,27 @@ class TestStdioMCPClient:
        """Test successful stdio connection."""
        client = StdioMCPClient("echo", ["test"])

-        with patch("praisonaiui.features.mcp.stdio_client") as mock_stdio_client:
-            # Mock the context manager
-            mock_session_manager = AsyncMock()
+        with (
+            patch("praisonaiui.features.mcp.stdio_client") as mock_stdio_client,
+            patch("praisonaiui.features.mcp.ClientSession") as mock_client_session,
+            patch("praisonaiui.features.mcp.contextlib.AsyncExitStack") as mock_stack_cls,
+        ):
+            mock_read = MagicMock()
+            mock_write = MagicMock()
            mock_session = AsyncMock()
-            mock_session_manager.__aenter__.return_value = mock_session
-            mock_stdio_client.return_value = mock_session_manager
+            mock_stack = AsyncMock()
+            mock_stack.enter_async_context = AsyncMock(
+                side_effect=[(mock_read, mock_write), mock_session]
+            )
+            mock_stack_cls.return_value = mock_stack
+            mock_stdio_client.return_value = MagicMock()
+            mock_client_session.return_value = mock_session

            result = await client.connect()

            assert result is True
            assert client.session == mock_session
-            mock_stdio_client.assert_called_once()
+            assert mock_stack.enter_async_context.await_count == 2
            mock_session.initialize.assert_called_once()

    @pytest.mark.asyncio
@@ -138,16 +147,13 @@ class TestStdioMCPClient:
        """Test stdio disconnect."""
        client = StdioMCPClient("echo", ["test"])

-        # Mock session and session manager
-        mock_session = AsyncMock()
-        mock_session_manager = AsyncMock()
-        client.session = mock_session
-        client._session_manager = mock_session_manager
+        mock_stack = AsyncMock()
+        client._exit_stack = mock_stack
        client._connected = True

        await client.disconnect()

-        mock_session_manager.__aexit__.assert_called_once_with(None, None, None)
+        mock_stack.aclose.assert_called_once()
        assert client.session is None
        assert client.process is None

@@ -156,18 +162,14 @@ class TestStdioMCPClient:
        """Test stdio disconnect with exception handling."""
        client = StdioMCPClient("echo", ["test"])

-        # Mock session that throws exception during disconnect
-        mock_session = AsyncMock()
-        mock_session_manager = AsyncMock()
-        mock_session_manager.__aexit__.side_effect = Exception("Disconnect failed")
-        client.session = mock_session
-        client._session_manager = mock_session_manager
+        mock_stack = AsyncMock()
+        mock_stack.aclose.side_effect = Exception("Disconnect failed")
+        client._exit_stack = mock_stack
        client._connected = True

-        # Should not raise exception
        await client.disconnect()

-        mock_session_manager.__aexit__.assert_called_once()
+        mock_stack.aclose.assert_called_once()
        assert client.session is None
        assert client.process is None

@@ -347,6 +349,61 @@ class TestMCPClientManager:
        with pytest.raises(ValueError, match="Invalid MCP server config"):
            await manager.connect_server(config)

+    @pytest.mark.asyncio
+    async def test_reconnect_same_name_disconnects_old(self):
+        """Reconnecting the same name disconnects the previous client (no leak)."""
+        manager = MCPClientManager()
+
+        old_client = AsyncMock()
+        old_client.disconnect = AsyncMock()
+        manager._clients["srv"] = old_client
+
+        config = {"name": "srv", "command": "echo", "args": []}
+        with patch("praisonaiui.features.mcp.StdioMCPClient") as mock_client_class:
+            new_client = AsyncMock()
+            new_client.connect.return_value = True
+            new_client.list_tools.return_value = []
+            mock_client_class.return_value = new_client
+
+            await manager.connect_server(config)
+
+        old_client.disconnect.assert_awaited_once()
+        assert manager._clients["srv"] is new_client
+
+    @pytest.mark.asyncio
+    async def test_connect_failure_removes_client(self):
+        """A failed connect (returns False) removes the client from the registry."""
+        manager = MCPClientManager()
+        config = {"name": "srv", "command": "echo", "args": []}
+
+        with patch("praisonaiui.features.mcp.StdioMCPClient") as mock_client_class:
+            client = AsyncMock()
+            client.connect.return_value = False
+            mock_client_class.return_value = client
+
+            server = await manager.connect_server(config)
+
+        assert server.status == MCPStatus.ERROR
+        client.disconnect.assert_awaited_once()
+        assert "srv" not in manager._clients
+
+    @pytest.mark.asyncio
+    async def test_connect_exception_removes_client(self):
+        """A connect that raises removes the client from the registry."""
+        manager = MCPClientManager()
+        config = {"name": "srv", "command": "echo", "args": []}
+
+        with patch("praisonaiui.features.mcp.StdioMCPClient") as mock_client_class:
+            client = AsyncMock()
+            client.connect.side_effect = Exception("boom")
+            mock_client_class.return_value = client
+
+            server = await manager.connect_server(config)
+
+        assert server.status == MCPStatus.ERROR
+        client.disconnect.assert_awaited_once()
+        assert "srv" not in manager._clients
+
    @pytest.mark.asyncio
    async def test_disconnect_server_success(self):
        """Test successful server disconnect."""