feat: add configurable activity-aware ACP request timeout

[brittianwarner]

Summary

• The ACP client timeout is now activity-aware: the timer resets whenever any inbound JSON-RPC message arrives from the agent process, making it an inactivity threshold rather than a wall-clock cap
• Default timeout bumped from 120s to 15 minutes (900,000ms) to accommodate long-running inference
• New public acpTimeoutMs option on both AgentOsOptions (VM-wide default) and CreateSessionOptions (per-session override)
• Input validation rejects acpTimeoutMs <= 0 with a clear error message

Problem

Agents actively streaming session/update notifications during long prompts would hit the fixed 120s timeout and get killed despite being healthy and making progress. Server logs confirmed: Pi process alive (exitCode=null), actively streaming (session/update x20), times out at exactly 120s.

Solution

_resetPendingTimers() clears and recreates all pending request timers whenever any valid JSON-RPC message (response, notification, or inbound request) arrives in _startReading(). This makes the timeout measure silence, not total elapsed time:

• Active agent: timer resets constantly, never hits the cap
• Waiting-for-inference agent: 15min silence threshold is generous
• Dead/hung agent: detected within the timeout window after last activity

Changes

packages/core/src/acp-client.ts

• DEFAULT_TIMEOUT_MS: 120,000 → 900,000 (15 min)
• PendingRequest interface: added method field for timeout error messages
• _pending map key narrowed from number | string | null to number (only auto-increment IDs)
• New _resetPendingTimers(): _closed guard, clears + recreates timers, JSDoc
• Response lookup uses type-safe narrowing for numeric msg.id
• Constructor validates timeoutMs > 0

packages/core/src/agent-os.ts

• AgentOsOptions.acpTimeoutMs?: number — VM-wide default
• CreateSessionOptions.acpTimeoutMs?: number — per-session override
• _acpTimeoutMs private field, set in create(), threaded to AcpClient in createSession()

packages/core/tests/acp-protocol.test.ts

• "activity-aware timeout resets on inbound notifications" — 1200ms of 300ms-interval notifications succeeds under 2s timeout (1700ms CI margin)
• "activity-aware timeout still fires after activity then silence" — notification at T+400ms, timeout at T+900ms, wall-clock assertion elapsed >= 750ms proves the timer was actually reset

packages/core/README.md

• Updated stale AgentOsOptions and CreateSessionOptions field lists

Verified

• Type-check: clean
• All 4 timeout/activity tests pass (existing 3 + 2 new)
• No new lint errors
• Timer reference safety: no double-fire, no leaks, no races (single-threaded JS)
• _rejectAll() on close() clears timers created by _resetPendingTimers

Follow-ups

• Mirror acpTimeoutMs in Rivet actor layer (actor parity requirement)

[brittianwarner]

https://app.graphite.com/github/pr/rivet-dev/rivet/4552/feat(agent-os)-add-acpTimeoutMs-passthrough-test-fixture-and-tests?onboarding_state=not-authorized