tsumugi: Skill for Codex

Observe the tsumugi runtime

Gather evidence without changing runtime state. Separate observed facts, interpretation, and unverified hypotheses.

Establish the observation contract

1. Identify the environment, project scope, time window, and decision the evidence should support.
2. Require explicit authorization before contacting production or any private runtime.
3. If access is not authorized or unavailable, produce a read-only query plan instead of guessing.
4. Read docs/ROADMAP.md and the relevant ADR only when the observation supports a product decision.

Select evidence

Read references/metrics.md and select only signals relevant to the question.

For ADR-014 rollout evaluation, also read references/adr-014-evaluation.md. Treat it as an initial use case, not the permanent boundary of this skill.

Prefer evidence in this order:

1. Read-only Admin REST responses and pipeline traces.
2. Prometheus metrics over a defined interval.
3. Scheduler state and persisted maintenance run history.
4. Sanitized logs for exact failure mechanisms.
5. Read-only PostgreSQL queries only when explicitly authorized and higher-level evidence is insufficient.

Do not read .env*, credentials, tokens, private endpoint configuration, or raw sensitive capture content.

Collect and normalize

1. Record the collection timestamp, time window, filters, and source for every result.
2. Compare counters and queue sizes across at least two timestamps when claiming a trend.
3. Record which filters and timestamp columns each source actually applies. Do not assume related endpoints use identical populations.
4. Use aggregate data first. Inspect individual traces only to explain an aggregate anomaly or sample quality.
5. Sanitize user content, prompts, memory narratives, error payloads, and identifiers before reporting.
6. Record missing signals and failed queries as coverage gaps, not healthy results.

Interpret cautiously