robosystems: Command for Claude Code

The logs are a bug detector only if someone reads them. This skill reads them: pull every ERROR and WARNING over a window, collapse them into distinct messages, set aside the known noise, and trace each remaining signal far enough to say whether it is a bug, a transient, or noise nobody has named yet. Pairs with the log-review runbook in local/RoboSystems/runbooks/, which holds this environment's known-noise list, the signals already accepted or routed, and the last run's findings. Read it first: a review that rediscovers last week's accepted noise wastes the pass.

When

• After every production deploy, over the window since it went out (the deploy is the highest-prior cause of a new error class).
• Weekly otherwise, over the last 7 days.
• When the user asks, over the window they name (default 24h).

Scope & guardrails

• Read-only. Logs Insights, filter-log-events, describe-*, gh reads. Fixing what you find is a separate step on a feature branch, proposed and confirmed.
• The findings are sensitive — never commit them. They name user ids, IPs, hostnames and error bodies. Keep raw output in the scratchpad; the durable record is the runbook, which is git-ignored.
• A finding needs a cause, not a pattern. "Errno 5 on the graph tier" is a symptom; the finding is what tore the volume away. Trace each signal to the code or the event that produced it before calling it a bug, and say when you could not.
• Bug bar. Report highs, medium-highs and the mediums worth fixing. Lows stay in the raw output.

1. Pull the window

The application log groups are /robosystems/<env>/{api,worker,dagster,graph-api}. The API and the graph API write JSON with a level field; Dagster's own lines and uvicorn access lines are text with the level word inline (and ANSI colour codes). Run the four queries in parallel with aws logs start-query, polling get-query-results until Complete:

# api, worker: structured
filter level in ["ERROR","CRITICAL","WARNING"]
| stats count(*) as n, min(@timestamp) as first, max(@timestamp) as last
    by level, component, substr(message, 0, 180) as msg
| sort n desc

# dagster, graph-api: text or mixed
filter @message like /ERROR|CRITICAL|WARNING|Traceback|Errno| 5\d\d /
| stats count(*) as n, min(@timestamp) as first, max(@timestamp) as last
    by substr(@message, 0, 200) as msg
| sort n desc