fix(monitor): route assistant messages to channels for ACP .claude workdirs (#3286)

[OneStepAt4time]

Summary

Closes #3286. Two coupled root causes prevented message.assistant (and the user-facing equivalent) from reaching any notification channel (Telegram, Slack, webhook, email) — or /v1/sessions/:id/transcript — for sessions whose workdir contains a dotted segment (e.g. .claude/worktrees/...).

Commit 1 — path-utils.ts: computeProjectHash preserved . inside segments. Claude Code replaces every . with -, so …/aegis/.claude/worktrees/x mapped to -…-aegis-.claude-… while CC's on-disk slug is -…-aegis--claude-…. discoverFromFilesystemFallback in session-transcripts.ts (and the equivalent in session-discovery.ts) looked in the wrong directory and never resolved session.jsonlPath.

Commit 2 — monitor.ts: Even with the slug fixed, checkSession guarded forwarding with !this.jsonlWatcher (instance presence) instead of per-session watching state. On the exact poll where the fallback discovered the JSONL and read every existing entry, those entries were dropped — and the watcher then attached at end-of-file and saw no further changes for the content that pre-dated its attachment. Result: 4 historical user prompts forwarded only by coincidence (timing-dependent), 0 assistant entries forwarded. Switching the guard to !this.jsonlWatcher?.isWatching(session.id) lets checkSession flush the discovery-poll backlog before the watcher takes over.

End-to-end verification (with Telegram channel)

Instrumented tgApi and monitor paths, ran a fresh session in .claude/worktrees/verify-tg-trace2, waited 60 s, killed the session. Trace:

[poll]    jsonlPath=false isWatching=false            ← initial state
Transcripts (#1768 fallback): session mapped         ← slug fix worked
[fwd]     role=user × 4                              ← watcher-check fix worked
[poll]    jsonlPath=true  isWatching=true            ← watcher owns it now
[watcher] msgs=1
[fwd]     role=assistant type=text TRACE2_ASSISTANT_OK ← reaches channels.message
[tg]      → sendMessage thread=topic_id text="✨ TRACE2_ASSISTANT_OK"

Aegis's session-ended summary in the topic shows 5 msgs · No errors (4 user + 1 assistant) instead of the previous 0 msgs / 4 msgs.

Same session, no fix applied → no [fwd] role=assistant, summary 0 msgs.

Test plan

• npx tsc --noEmit — clean.
• src/__tests__/path-utils-909.test.ts — 6/6 pass (5 existing + 1 new for the .claude case).
• src/__tests__/monitor-fixes.test.ts, monitor-idle-broadcast.test.ts, monitor-initial-offset.test.ts, jsonl-watcher.test.ts — 41/41 pass.
• Runtime spot-check: computeProjectHash('/Users/…/aegis/.claude/worktrees/foo') now matches the on-disk ~/.claude/projects/…--claude-worktrees-foo.
• End-to-end Telegram delivery: instrumented Aegis + real Telegram API. Assistant text lands in the topic after both fixes; was missing before either fix alone.

Files changed

• src/path-utils.ts — 1-line addition to segment sanitization.
• src/__tests__/path-utils-909.test.ts — 1 assertion for the .claude case.
• src/monitor.ts — 4 lines (guard + 2-line comment explaining the hand-off).

3 files, +11/-3 lines.

Gate notes (pre-existing, not introduced here)

npm run gate continues to fail on dashboard:tokens:gate, dashboard:clickable:gate, and 3 unrelated test cases on pristine origin/develop@c4cab08c. None touch monitor / channel / path-utils.

Aegis version

Developed with: v0.6.7-preview.1

[aegis-gh-agent]

🔍 Review — Requesting changes (CI not green)

Code review:

• ✅ The fix is correct: .replace(/\./g, "-") mirrors Claude Code slug behavior
• ✅ Root cause analysis is thorough — clear explanation of why JSONL lookup fails
• ✅ Test coverage added for .claude and worktree paths
• ✅ Small, focused change (1 line + 2 test assertions)

Blocking issue — Gate 3 (CI green):

• test (ubuntu-latest, 20) is failing — TemplatesPage.test.tsx > deletes a template after confirmation times out at 5000ms
• This appears to be a Node 20-specific flaky test (unrelated to your path-utils change), but CI must be green per merge gate rules

Recommendation: Re-trigger CI (gh run rerun) to see if it passes on a retry. If the flaky test persists, please address it (increase timeout or fix the test) so CI goes green. I'll approve as soon as CI is clean.

Non-blocking observation: PR body mentions "3 pre-existing failures" — but test (ubuntu-latest, 20) is the only one showing in CI. If the others are gate-level, they should be documented.

👁️ Gate 3 must pass.

[aegis-gh-agent]

✅ Approved.

Surgical fix for ACP .claude workdir message routing.

What's good:

• Root cause clear: this.jsonlWatcher truthy check missed sessions where watcher exists but hasn't subscribed to this specific session yet (discovery poll race)
• Fix: this.jsonlWatcher?.isWatching(session.id) — per-session check instead of global instance check
• computeProjectHash dot replacement — .claude segments now produce -claude instead of being passed through, matching Claude Code's slug behavior
• Test added for the dot-replacement case with two path examples
• Comment explains the race condition well

CI fully green. Ready to merge.