feat(sessions): focus check — nudge a session that goes down a rabbit hole (v0.447.0)

[vikasprogrammer]

Why

Sessions go down rabbit holes and nobody can tell. The existing verdicts only see a run's shape (circling = the same action repeated, stuck = silence); a rabbit hole is busy, varied work on the wrong thing, so it reads as forward.

What

• Trigger (free): check after 20 allowed actions, then every 20 more, at most every 8 min. Out of band, so the gate never waits.
• Judge (Haiku):
  • It sees the original request plus later human messages, each paired with the agent proposal it answered (so "go" / "merge both" count as approval), and the transcript tail.
  • It states current_ask, then the verdict.
  • A drifted first vote is confirmed by two more; a 2-of-3 majority at confidence ≥ 0.8 is required.
• Effect:
  • The session strip shows a new drifting verdict.
  • An advisory note goes to the agent on its next tool call (allow + additionalContext), suggesting it task_create the side-finding and return.
  • An unattended run that is still drifting at the next check gets one sessionOwner Inbox card per streak.
  • Nothing is ever blocked or stopped.
• Backend: the workspace Anthropic key (pinned to Haiku) → OpenAI-compatible router LLM → pooled claude -p --model claude-haiku-4-5 (runClaudePrompt, extracted from the summarizer).
• Controls:
  • Settings → Governance mode: nudge (default) · observe · off
  • per-agent driftCheck: false
  • AOS_DRIFT=0

Calibration

The replay used production prompts against 30 long instapods transcripts (read-only), 90 checkpoints. Full table in docs/drift-focus-check.md.

• v1 flagged 8/42, mostly false positives. The main cause: <task-notification> / skill bodies were read as human messages, hiding the real redirections. Now filtered.
• Single votes flicker on identical input, hence the confirmation votes.
• Final: 2 genuine rabbit holes confirmed (one from the held-out half: told "the PR is done, needs a founder decision", the agent re-implemented it from scratch; 3/3 votes), 0 false positives at the 0.8 floor.

Test

• scripts/drift-nudge-test.cjs (96 assertions, added to test:governance): parser, input building, cadence, streak outcomes, copy framing, progress precedence, and end-to-end through TerminalManager.gate against a stub Anthropic API (split votes, observe/off, opt-out, garbled judge, no card for interactive runs).
• npm run build, web build and the full npm run test:governance pass.

🤖 Generated with Claude Code