Observability

Production logs are structured JSON; development logs are human-readable. Set TASUKU_LOG_LEVEL=debug temporarily when investigating a run.

Correlation identifiers

Logs include relevant identifiers for:

• HTTP request;
• GitHub delivery;
• organization and repository;
• workflow run and queue job;
• agent task and attempt;
• sandbox and provider.

They do not include secret plaintext or raw webhook bodies.

Durable diagnostics

The dashboard retains evidence beyond process logs:

• ordered workflow events;
• task attempts and timing;
• provider and sandbox metadata;
• input snapshots and sanitized failures;
• findings and validation states;
• GitHub thread and publication identifiers;
• usage, feedback, and learning provenance.

Health signals

Monitor:

1. GET /healthz for process liveness and GET /readyz for Postgres-backed readiness.
2. Postgres availability and storage growth.
3. Worker process uptime and database connection errors.
4. Old queued jobs, expired leases, repeated attempts, and stalled workflow event timestamps.
5. GitHub webhook delivery failures and OAuth errors.
6. Provider latency, quota, authentication, and sandbox creation failures.

/healthz proves only that the server process can answer HTTP. /readyz checks the database with a bounded query and returns 503 while unavailable. Keep liveness probes on /healthz so a temporary database outage does not restart a healthy process; route traffic only when /readyz succeeds. The worker has no HTTP health endpoint and should be monitored through process state, logs, queue age, and lease expiry.

Live events

The workflow event stream can update the dashboard while a run progresses. Persisted events remain the audit source if a client disconnects; the stream is not the only copy.

Use Troubleshooting for common symptoms and failure codes.