…e /compact

New bundled hook that fires on 'command:compact' and injects a system message
reminding the agent to:
- Update daily logs with session progress
- Update PROJECT-STATE.md files touched this session
- Update PROJECTS.md last-touched dates
- Create handoff if mid-complex-work

This ensures the agent has a chance to persist context before compaction
discards conversation history.

4 tests: inject on compact, skip non-compact, skip non-command, content validation.

Fixes critical issue found by 3-way code review (Opus + Sonnet consensus):

1. Wire triggerInternalHook('command:compact') in commands-compact.ts
   - The pre-compaction-flush handler was dead code — command:compact was
     never dispatched. Now fires before compaction, matching the pattern
     used by /new and /reset in commands-core.ts.
   - Hook messages are enqueued as system events for the agent to process
     in the compacted session's first turn.

2. Add 10-second timeout on plugin hook execution (compact.ts)
   - Prevents a hanging plugin from blocking compaction indefinitely.
   - Uses Promise.race with timeout rejection.

3. Add 2000-message threshold for token estimation (compact.ts)
   - Skips expensive token estimation for very large sessions.
   - tokenCount will be undefined for sessions >2000 messages.

4. Add architecture comment explaining dual hook systems
   - Internal hooks (command:compact) fire in commands-compact.ts
   - Plugin hooks (before/after_compaction) fire in compact.ts
   - Both are non-fatal with independent error handling.

5. Improve hook error logging with session context

Code review: docs/CODE-REVIEW-OPUS.md, docs/CODE-REVIEW-SONNET.md,
docs/CODE-REVIEW-SYNTHESIS.md

Changes by gpt-5.2-codex (Codex CLI), reviewed by Will:

1. Fix flush message wording — now post-compaction reminder, not
   misleading 'imminent' warning (handler.ts)
2. Remove unused 'vi' import, fix createHookRunner() to use proper
   PluginRegistry type instead of [] (compact-hooks.test.ts)
3. Update test assertions for new message wording (handler.test.ts)
4. Move tokensAfter computation before after_compaction hook call,
   pass it as tokenCount instead of undefined (compact.ts)

9/9 hook tests passing.
Co-authored-by: Codex CLI <codex@openai.com>

Resolved 2 conflicts in skills watcher (refresh.ts, refresh.test.ts):
- Merged upstream Python venv/cache ignore patterns with our binary extension patterns
- Combined both test suites with vi.hoisted() for mock compatibility
- Used dynamic imports in tests to work with chokidar mock

8 files auto-merged cleanly (verified by Codex investigation agent):
- bash-process-registry.ts, bash-tools.exec.ts, compact.ts
- subagent-registry.ts, zod-schema.ts, auth.ts
- config-reload.ts, memory/manager.ts

Our 7 custom commits preserved:
- Skills watcher binary file ignore + configurable patterns
- Compaction hooks (before/after) wiring
- Pre-compaction flush bundled hook
- Resource leak fixes (child unref, sweeper, concurrency guards)
- Auth/config-reload safety improvements
- Health command FD diagnostics
- Config schema for watchIgnoredPatterns

Investigation: Agent A (Codex, 80K tokens) reviewed all 10 merge files.
Risk rating: 4/10 (moderate-low).
Test suite: 7,114 passed, 0 failed (1,027 test files).