feat(dashboard): Human Intervention metric (#34)

[jeff-r2026]

Summary

Implements Human Intervention metric for the team dashboard (closes #34) — tracking how often a human has to step in during a Claude Code / Cursor / CodeBuddy session. Three signals are detected from the session transcript and hook stream, then rolled up per-session, per-user and into the team digest:

• interrupt — user pressed ESC mid-turn ([Request interrupted by user] in the transcript).
• toolReject — user denied a tool/permission prompt (The tool use was rejected / The user doesn't want to proceed with this tool use).
• correction — a re-prompt within 60s of a Stop that matches a course-correction keyword (不对/重新/redo/undo/…) rather than starting a fresh task.

Wiring (no new dependencies):

• src/dashboard-collector.ts — countInterventions() streams the transcript (memory-bounded, cheap pre-filter), idempotent Stop snapshot, rebuildSessions() exposes per-session badge data, aggregateSessionInterventions().
• src/types.ts — intervention types + detection constants (TRANSCRIPT_INTERRUPT_PREFIX, TRANSCRIPT_REJECT_MARKERS, CORRECTION_KEYWORDS, CORRECTION_WINDOW_MS, scan cap).
• src/team-push.ts — computeInterventionDelta() / mergeInterventionStats() (idempotent deltas into stats/<user>.yaml).
• src/digest.ts — summarizeInterventions() (totals, avg/session, ranked).
• src/dashboard-html.ts — per-session intervention badge.
• docs/usage-guide.md — metric documentation.

Test Plan

All commands run on this branch rebased onto main (a4a344e).

Type check — npx tsc --noEmit → ✅ pass (exit 0)

Unit suite — npx vitest run

Test Files  120 passed (120)
     Tests  1617 passed (1617)

Intervention-specific coverage (4 new unit files + amended collector test, 72 cases):
dashboard-collector.test.ts (countInterventions + rebuildSessions), digest-interventions.test.ts, intervention-report.test.ts, team-push-interventions.test.ts (delta + merge idempotency).

E2E suite — npm run test:e2e (runs under vitest.e2e.config.ts)

Test Files  6 passed (6)
     Tests  44 passed | 23 skipped (67)

Includes intervention-e2e.test.ts (2 tests): full hook STDIN → events.jsonl → rebuildSessions → delta → digest pipeline.

Real end-to-end with the smallest model (claude-haiku-4-5) — beyond the in-process tests, the full chain was exercised against a real Claude Code session and the real compiled CLI binary:

1. Ran a real headless haiku session (claude --model claude-haiku-4-5 -p …) that used the Write tool → captured the real transcript JSONL. Confirmed our parser handles the production transcript schema (user/assistant/tool_use/tool_result + queue-operation/attachment/last-prompt) and reports 0 interventions on a clean session (no false positives).
2. Verified our detection markers are byte-for-byte present in the real Claude Code binary (2.1.196): Request interrupted by user, The tool use was rejected, doesn't want to proceed with this tool use — all FOUND.
3. Confirmed a headless sandbox block ("…was blocked. For security…") is correctly not counted as a human intervention (only genuine user actions count).
4. Drove the real built binary as the hook (node dist/index.js dashboard-report --stdin --tool claude) over SessionStart → PostToolUse → Stop → UserPromptSubmit, against the real transcript with the genuine interrupt + reject markers appended:

Stop event snapshot      : { interrupt: 1, toolReject: 1 }
session card             : { interrupt: 1, toolReject: 1, correction: 1 }  count=3  status=running
team delta               : { sessions: 1, interrupt: 1, toolReject: 1, correction: 1 }
digest summary           : totalSessions=1  totalInterventions=3  avgPerSession=3  ranked[0]=jeff

Notes

• No package.json / version bump (release is tag-triggered).
• Branch rebased onto current main; full suite is 100% green (the older agent-status test-isolation flakiness is not present here).

🤖 Generated with Claude Code