feat(diagnostics): per-call timing, in-flight markers, and a diagnostics endpoint

Instrument the provider router to emit a model_call_start before every model call and a model_call with duration_ms after, plus explicit phase events. A hung/in-flight call is now visible (a start with no matching end) instead of producing no log line until it returns. Add GET /v1/sessions/{id}/diagnostics: a machine-readable summary (for human or AI maintainers) of phase + per-call durations, slowest calls, errors, and any in-flight call with how long it has been waiting — the key signal for "is it hung?". e2e covers it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Blake Bertuccelli-Booth committed on May 22, 2026, 04:50 PM

Showing 7 changed files +236 additions -8 deletions

M README.md +1

M docs/API.md +33 -1

A src/diagnostics.ts +149

M src/pipeline/orchestrator.ts +13 -5

M src/providers/index.ts +19 -2

M src/routes/sessions.ts +15

M test/e2e.sh +6