diff --git a/CLAUDE.md b/CLAUDE.md
index c7fd4ac4..b9fb0b5e 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -200,10 +200,14 @@ Apply before finalizing any multi-phase plan. Trust-but-verify between scout →
 10. **Context key style convention** — check existing `context.go` pattern before introducing new key types. Mixed = code smell.
 11. **Alias/shim coverage** — enumerate ALL exported symbols via `go doc <pkg>`. Add compile-time signature guards.
 12. **Verify pass MANDATORY after rewrite** — spawn fresh Explore/grep to audit planner output. Don't trust self-validation.
+13. **No fabricated Go identifiers** — every type/function/package referenced in plan must be grep-verified to exist (`grep -rn '^func <name>\|^type <name>' <pkg>`). Plausible-sounding APIs (`Keyring`, `StartSpan`, `Validator`, `security.SSRF`) are RED FLAGS — conventions differ per codebase. Rule of thumb: if you can't cite `file_path:line_number` for a symbol, you're fabricating. Apply especially when planner says "reuse existing X" — re-verify X exists before writing into plan.
+14. **Plausible-but-wrong API families** — watch for: OTel-style `StartSpan/EndSpan` (codebase may be emit-based); wrapper types like `Keyring/Vault/Manager` (codebase may use free functions); `internal/security` or `internal/auth` packages (often scattered instead of centralized). When unsure, `go doc <pkg>` lists actual exported surface.
 
 **Pattern to avoid:** user asks → planner writes → report "done".
 **Safer pattern:** user asks → scout → planner writes → audit-verify → report.
 
+**Concrete red-team practice:** After planner completes, run `code-reviewer`/`brainstormer` agent in audit mode with explicit instruction "spot-check 15+ factual claims against live codebase". Caught fabricated `internal/security/`, `crypto.Keyring`, `tracing.StartSpan` in Agent Hooks plan (see `plans/260414-2229-agent-hooks-system/reports/audit-260414-2310-plan-review.md`).
+
 ## Post-Implementation Checklist
 
 After implementing or modifying Go code, run these checks:
diff --git a/docs/00-architecture-overview.md b/docs/00-architecture-overview.md
index ea0c75a3..6faa71db 100644
--- a/docs/00-architecture-overview.md
+++ b/docs/00-architecture-overview.md
@@ -475,6 +475,29 @@ Six distinct workspace scenarios:
 - **L0 Auto-Injection**: Top-K vault entries injected into system prompt as "relevant context from vault"
 - **Filesystem Sync**: Vault entries exported as `.md` files for manual editing, re-imported with change tracking
 
+### Audio & Voice System (ElevenLabs + Streaming TTS)
+
+**Provider architecture** (`internal/audio/`):
+- **`Manager`**: Central orchestrator dispatching TTS/STT/Music/SFX requests to pluggable providers
+- **`TTSProvider` interface**: Core text-to-speech contract (blocking, buffered response)
+- **`StreamingTTSProvider` interface**: Optional interface for ElevenLabs `/stream` endpoint (chunked audio via `io.ReadCloser`)
+- **Implementations**: ElevenLabs, OpenAI, Edge, MiniMax (phase-gated; STT/Music/SFX partial)
+
+**Voice discovery & caching**:
+- **Voice cache** (`internal/audio/voice_cache.go`): In-memory LRU (cap 1000 tenants, TTL 1h) shared by HTTP `/v1/voices` + WS `voices.list` handlers. Thread-safe with `sync.Mutex` (LRU updates require write lock)
+- **Cache miss recovery**: HTTP handler auto-fetches from ElevenLabs; WS handler requires prior cache warm (provider resolution deferred to Phase 3)
+- **Agent audio context** (`store.WithAgentAudio` / `AgentAudioFromCtx`): Immutable snapshot bundle (AgentID + OtherConfig JSONB) injected by dispatcher before tool dispatch; consumed by `TtsTool.Execute` for voice/model resolution
+
+**Agent-level configuration** (`agents.other_config` JSONB):
+- `tts_voice_id`: ElevenLabs voice ID (e.g., "pMsXgVXv3BLzUgSXRplE")
+- `tts_model_id`: Model choice (eleven_v3, eleven_flash_v2_5, eleven_multilingual_v2, eleven_turbo_v2_5)
+- Resolution precedence: CLI args → agent config → tenant override → provider default
+
+**Web UI voice picker** (`ui/web/src/components/voice-picker.tsx`):
+- Combobox with BM25 search, preview playback button (HTML `<audio>`)
+- Handles preview CDN 403 (expiry) via `onError` → auto-refresh cache
+- Embedded in PromptSettingsSection, bound to `other_config.tts_voice_id`
+
 ---
 
 ## Cross-References
diff --git a/docs/03-tools-system.md b/docs/03-tools-system.md
index 89eedd24..49732880 100644
--- a/docs/03-tools-system.md
+++ b/docs/03-tools-system.md
@@ -812,10 +812,17 @@ func (t *MyTool) Execute(ctx context.Context, params MyParams) (*Result, error)
 #### tts (tenant override shape)
 ```json
 {
-  "primary": "elevenlabs"
+  "primary": "elevenlabs",
+  "default_voice_id": "pMsXgVXv3BLzUgSXRplE",
+  "default_model": "eleven_flash_v2_5"
 }
 ```
 
+**Fields:**
+- `primary`: Provider selection (elevenlabs, openai, edge, minimax)
+- `default_voice_id`: Tenant-level voice ID fallback (ElevenLabs); overridable per-agent via `agent.other_config.tts_voice_id`
+- `default_model`: Tenant-level model choice (eleven_v3, eleven_flash_v2_5, eleven_multilingual_v2, eleven_turbo_v2_5); overridable per-agent via `agent.other_config.tts_model_id`
+
 **Secret vs non-secret split:**
 - Non-secret (provider priorities, max_results, allowed_domains): `builtin_tool_tenant_configs.settings` (editable via UI)
 - Secret (API keys): `config_secrets` table (encrypted, tenant-scoped)
diff --git a/docs/17-changelog.md b/docs/17-changelog.md
index c4c69e44..faf47762 100644
--- a/docs/17-changelog.md
+++ b/docs/17-changelog.md
@@ -48,6 +48,22 @@ Unified audio provider management via new `internal/audio/` package with pluggab
 
 **Impact**: Existing TTS flows fully compatible. New code can import `internal/audio` directly. STT/Music/SFX wiring deferred to Phase 3-4.
 
+#### ElevenLabs Audio Manager Enhancements — Phase 2 (2026-04-14)
+
+Voice discovery and agent-level audio config via new backend endpoints, in-memory cache, and web UI picker. Bundles producer/consumer context pattern (`store.WithAgentAudio`) for seamless voice/model resolution throughout the tool execution pipeline.
+
+**What changed:**
+- **Voice cache** (`internal/audio/voice_cache.go`): In-memory LRU (cap 1000 tenants) with TTL 1h, shared between HTTP + WS handlers, thread-safe under concurrent access
+- **Streaming TTS interface** (`internal/audio/types.go`): New `StreamingTTSProvider` optional interface for ElevenLabs `/v1/text-to-speech/{voice_id}/stream` chunked playback
+- **ElevenLabs enhancements**: Model allowlist (11_v3, eleven_flash_v2_5, eleven_multilingual_v2, eleven_turbo_v2_5), `SynthesizeStream()` method, `ListVoices()` via `/v1/voices`
+- **Agent audio context** (`internal/store/context.go`): New `WithAgentAudio` / `AgentAudioFromCtx` bundle; producer wires snapshot at dispatcher level (internal/agent/), consumer (`TtsTool.Execute`) reads voice/model overrides from agent config
+- **Agent config extension** (`agents.other_config` JSONB): New `tts_voice_id` and `tts_model_id` fields with resolution precedence: args → agent → tenant → provider default
+- **HTTP + WS endpoints**: GET /v1/voices (cached), POST /v1/voices/refresh (admin-only), WS method `voices.list` + `voices.refresh`
+- **Web voice picker** (`ui/web/src/components/voice-picker.tsx`): Combobox with search, preview button (HTML audio + onError → refresh), embedded in PromptSettingsSection
+- **i18n**: 10 new frontend keys (voice_label, voice_placeholder, voice_refresh, voice_preview, model_label, etc.) + 2 new backend keys (MsgTtsUnknownModel, MsgVoicesListFailed) across en/vi/zh
+
+**Impact**: Existing TTS callers fully compatible (backward-compat via Phase 1 alias layer). Web UI gains voice discovery + per-agent voice/model overrides. Zero breaking changes. Integration test validates producer+consumer context flow.
+
 #### Trace Stop/Abort Redesign — Cascading 4-Layer Fix (2026-04-14)
 
 The Stop button on the traces page now reliably aborts running traces. Previous implementation had independent race conditions across HTTP streaming, agent router, trace persistence, and UI polling; this redesign fixes all four layers atomically.
diff --git a/docs/18-http-api.md b/docs/18-http-api.md
index 24630b7b..ea0e7ac1 100644
--- a/docs/18-http-api.md
+++ b/docs/18-http-api.md
@@ -798,7 +798,64 @@ Workspace file management.
 
 ---
 
-## 22. Media
+## 22. Voices & Audio
+
+Voice discovery for TTS providers (ElevenLabs). All endpoints are tenant-scoped and require tenant admin or operator role.
+
+| Method | Path | Description |
+|--------|------|-------------|
+| `GET` | `/v1/voices` | List available voices (in-memory cached, TTL 1h) |
+| `POST` | `/v1/voices/refresh` | Force refresh voice cache (admin-only) |
+
+### `GET /v1/voices`
+
+Fetch available TTS voices for the current tenant's provider.
+
+**Query Parameters:**
+- None
+
+**Response** (200 OK):
+```json
+[
+  {
+    "voice_id": "pMsXgVXv3BLzUgSXRplE",
+    "name": "Alice",
+    "preview_url": "https://...",
+    "category": "premade",
+    "labels": {
+      "use_case": "conversational",
+      "accent": "american"
+    }
+  }
+]
+```
+
+**Errors:**
+- 401: Missing or invalid token
+- 403: Insufficient permissions (requires tenant admin/operator)
+- 500: Provider error (e.g., ElevenLabs API unreachable)
+
+**Caching:**
+- Responses are in-memory cached per tenant with TTL 1h
+- Cache is shared across all HTTP + WebSocket handlers
+- Cache miss triggers immediate fetch from provider
+
+### `POST /v1/voices/refresh`
+
+Invalidate the voice cache for the current tenant, forcing a fresh fetch on the next request.
+
+**Admin-only endpoint.** Useful after voice updates or CDN expiry issues.
+
+**Request body:** (empty)
+
+**Response** (202 Accepted):
+```json
+{ "message": "voice cache invalidated" }
+```
+
+---
+
+## 23. Media
 
 | Method | Path | Description |
 |--------|------|-------------|
@@ -807,7 +864,7 @@ Workspace file management.
 
 ---
 
-## 23. Files
+## 24. Files
 
 | Method | Path | Description |
 |--------|------|-------------|
@@ -818,7 +875,7 @@ Auth via Bearer token or `?token=` query param (for `<img>` tags). MIME type aut
 
 ---
 
-## 24. API Keys
+## 25. API Keys
 
 Admin-only endpoints for managing gateway API keys. See [20 — API Keys & Auth](20-api-keys-auth.md) for the full authentication and authorization model.
 
@@ -856,7 +913,7 @@ Admin-only endpoints for managing gateway API keys. See [20 — API Keys & Auth]
 
 ---
 
-## 25. OAuth
+## 26. OAuth
 
 | Method | Path | Description |
 |--------|------|-------------|
@@ -943,7 +1000,7 @@ Notes:
 
 ---
 
-## 26. Edition
+## 27. Edition
 
 | Method | Path | Description |
 |--------|------|-------------|
@@ -951,7 +1008,7 @@ Notes:
 
 ---
 
-## 27. Tenants
+## 28. Tenants
 
 Multi-tenant management (admin only).
 
@@ -967,7 +1024,7 @@ Multi-tenant management (admin only).
 
 ---
 
-## 28. System Configs
+## 29. System Configs
 
 Key-value system configuration store.
 
@@ -980,7 +1037,7 @@ Key-value system configuration store.
 
 ---
 
-## 29. Team Workspace & Attachments
+## 30. Team Workspace & Attachments
 
 | Method | Path | Description |
 |--------|------|-------------|
@@ -990,7 +1047,7 @@ Key-value system configuration store.
 
 ---
 
-## 30. Shell Deny Groups
+## 31. Shell Deny Groups
 
 | Method | Path | Description |
 |--------|------|-------------|
@@ -998,7 +1055,7 @@ Key-value system configuration store.
 
 ---
 
-## 31. System
+## 32. System
 
 | Method | Path | Description |
 |--------|------|-------------|
@@ -1017,7 +1074,7 @@ Key-value system configuration store.
 
 ---
 
-## 32. MCP Bridge
+## 33. MCP Bridge
 
 Exposes GoClaw tools to Claude CLI via streamable HTTP at `/mcp/bridge`. Only listens on localhost. Protected by gateway token with HMAC-signed context headers.
 
diff --git a/docs/19-websocket-rpc.md b/docs/19-websocket-rpc.md
index c2257c47..15379845 100644
--- a/docs/19-websocket-rpc.md
+++ b/docs/19-websocket-rpc.md
@@ -501,6 +501,68 @@ Multi-tenant management (admin only).
 
 ---
 
+## 17.1. Voices (Voice Discovery)
+
+Discover available TTS voices for the tenant's configured provider.
+
+| Method | Description |
+|--------|-------------|
+| `voices.list` | Fetch available voices (in-memory cached, TTL 1h) |
+| `voices.refresh` | Force cache invalidation (admin-only) |
+
+### `voices.list` Request
+
+```json
+{
+  "method": "voices.list",
+  "id": 1
+}
+```
+
+**Response** (200 OK):
+```json
+{
+  "id": 1,
+  "result": [
+    {
+      "voice_id": "pMsXgVXv3BLzUgSXRplE",
+      "name": "Alice",
+      "preview_url": "https://...",
+      "category": "premade",
+      "labels": {
+        "use_case": "conversational",
+        "accent": "american"
+      }
+    }
+  ]
+}
+```
+
+**Errors:**
+- `code: -1`: Provider error (e.g., ElevenLabs API unreachable)
+- `code: -2`: Cache miss + no provider context available (desktop edition in Phase 2; HTTP handler resolves provider dynamically)
+
+### `voices.refresh` Request
+
+Admin-only. Invalidate tenant cache, forcing fresh fetch on next list.
+
+```json
+{
+  "method": "voices.refresh",
+  "id": 2
+}
+```
+
+**Response** (200 OK):
+```json
+{
+  "id": 2,
+  "result": { "message": "voice cache invalidated" }
+}
+```
+
+---
+
 ## 18. Browser Automation
 
 | Method | Description |