feat(voice-chat): replace read_shared_context with request_slow_brain tool

[lancy]

Summary

• Replace the passive read_shared_context tool with request_slow_brain for explicit task delegation — writes talker/worker-request events to shared context via fire-and-forget logContextEvent
• Rewrite TALKER_INSTRUCTIONS with capability awareness: the model now delegates tool-requiring tasks instead of saying "I can't do that"
• Update VOICE_CHAT_WORKER_PROMPT to prioritize explicit talker/worker-request events over autonomous observations

Changes

File 1: turbo/apps/platform/src/signals/voice-chat/voice-chat-session.ts

1. SESSION_TOOLS — Replaced read_shared_context with request_slow_brain (required task parameter)
2. TALKER_INSTRUCTIONS — Rewritten with delegation rules, capability framing, identity awareness, and result handling guidelines
3. handleFnCall$ — Replaced async read handler with sync fire-and-forget: logContextEvent(fetchFn, sid, "talker", "worker-request", task) + static confirmation

File 2: turbo/apps/web/src/lib/zero/integration-prompt.ts
4. VOICE_CHAT_WORKER_PROMPT — Added "Explicit Requests" section after "When to Act" for talker/worker-request prioritization

Test plan

• grep "read_shared_context" returns zero results in production code
• grep "request_slow_brain" finds it in SESSION_TOOLS and handleFnCall$
• grep "worker-request" finds it in handleFnCall$ and worker prompt
• Lint, type-check, format pass
• All existing platform tests pass (1611 tests)
• All integration-prompt tests pass (17 tests)
• Manual: Start voice chat, ask for tool-requiring task → Talker calls request_slow_brain → slow brain works → result injected back

Closes #8745

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 0% with 4 lines in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
...tform/src/signals/voice-chat/voice-chat-session.ts	0.00%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!

[lancy]

Code Review: PR #8748 (Round 1) — LGTM 🎉

All findings reviewed — no critical or high-priority issues.

Summary

Clean replacement of the passive read_shared_context tool with the active request_slow_brain delegation tool. Two files changed:

• voice-chat-session.ts: Replaced SESSION_TOOLS, TALKER_INSTRUCTIONS, and handleFnCall$ logic. Handler simplified from async fetch+parse to sync fire-and-forget logContextEvent.
• integration-prompt.ts: Added "Explicit Requests" section to worker prompt for talker/worker-request event prioritization.

Reviewed Findings (No Action Required)

Finding	Severity	Disposition
await on sync handleFnCall$ at caller site	P1→P2	Harmless — JS await on non-Promise resolves immediately; removing it would change caller code outside the PR's intent
JSON.parse(args) as { task: string } trust	P2	Acceptable — model sends structured args per schema definition; JSON.parse throws on malformed input which is fine

Verdict: LGTM ✅

No critical or high-priority issues remaining. This PR is ready for merge.

---

Completed after 1 round(s) of automated review-fix loop

[github-actions]

⚡ Lighthouse — App

Category	Score
🔴 Performance	33
🟢 Accessibility	93
🟠 Best Practices	79

Tested URL: https://pr-8748-app.vm6.ai/ · Full report

[github-actions]

⚡ Lighthouse — Web

Category	Score
🟠 Performance	55
🟢 Accessibility	96
🟠 Best Practices	79
🟢 SEO	92

Tested URL: https://pr-8748-www.vm6.ai/ · Full report

[lancy]

Acceptance Report: #8745 — feat: replace read_shared_context with request_slow_brain tool

Epic: #8742 — voice-chat dual-brain collaboration
PR: #8748
Reviewer: Tech Lead (via /epic-review)

---

1. Requirements (Definition of Done)

#	Requirement	Status	Notes
1	read_shared_context fully removed from SESSION_TOOLS and handleFnCall$	✅ PASS	All 3 occurrences removed (tool definition, instructions ref, handler)
2	request_slow_brain tool implemented — writes talker/worker-request	✅ PASS	Uses logContextEvent fire-and-forget, source "talker", type "worker-request"
3	TALKER_INSTRUCTIONS rewritten with capability awareness	✅ PASS	Identity, delegation rules, behavioral rules, injection handling, style
4	VOICE_CHAT_WORKER_PROMPT updated for explicit request priority	✅ PASS	New "Explicit Requests" section with priority treatment
5	Talker delegates instead of saying "I can't"	⏭️ SKIP	Runtime behavior — verified by prompt design
6	Slow brain handles both explicit + autonomous	⏭️ SKIP	Runtime behavior — verified by prompt design
7	All existing tests pass	✅ PASS	Pre-commit checks passed (lint, types, format, tests)
8	Lint, type-check, format pass	✅ PASS	Verified by agent

2. Code Review

Dimension	Status	Findings
Correctness	✅	handleFnCall$ correctly changed async→sync (logContextEvent is fire-and-forget). await set(handleFnCall$, ...) in caller resolves immediately — safe. Tool schema has required: ["task"] so parse is safe.
Project Standards	✅	No any, no eslint-disable, no defensive try-catch, YAGNI-compliant
Consistency	✅	Uses existing logContextEvent pattern, as const on SESSION_TOOLS, markdown section structure in worker prompt
Security	✅	No injection risks — uses existing API route with validation. No secrets, no auth changes.
Performance	✅	Removed unnecessary async/await. Fire-and-forget log instead of synchronous fetch+parse. Net improvement.

3. Architectural Alignment

Check	Status	Notes
Source of truth correct	✅	Tool writes talker/worker-request via context API, matching VALID_SOURCES and VALID_TYPES
No residual old patterns	✅	read_shared_context only exists in voice-chat-session.ts (3 occurrences), all replaced in diff
Injection message format aligned	✅	TALKER_INSTRUCTIONS references [Background...] — matches all formatInjectionMessage prefixes from #8744
Worker prompt correctly extended	✅	"Explicit Requests" section placed between "When to Act" and "Writing to Shared Context"

4. Independence

Check	Status
Compiles standalone	✅
Dependency on #8744 met	✅ (#8746 merged)
No forward dependencies	✅

5. Issues Found

Must Fix (blocking acceptance)

None.

Should Fix (non-blocking)

None.

Nits

None.

Verdict

ACCEPTED

Clean, minimal implementation (+44/-28) across 2 files. Exactly matches the approved plan and Epic #8742 design decisions. The async→sync cleanup on handleFnCall$ is a nice improvement. Injection message format in TALKER_INSTRUCTIONS correctly aligns with formatInjectionMessage prefixes from #8744.