What it does
This skill guides GitHub CLI event evaluations from case selection through execution, result interpretation, and cleanup checks. It directs the user to consult the runbook and credential and execution-mode references, then select a trial that matches the requested scenario. Results are separated into task completion, event routing, polling, and semantic validity, with behavioral failures distinguished from infrastructure failures. The workflow also covers replaying saved evidence and running smoke checks without a model. Reports include the scenario, model, budget, grader outcomes, evidence that the agent waited for notifications, artifact locations, and cleanup status.
When it helps
Use it to run a new evaluation, compare skill-on and skill-off results, reproduce a failure, check grader regressions, or verify a command or API protocol. It can also guide a smoke test and confirmation that cleanup finished.
Example requests
- “Run this event evaluation and compare the skill-on and skill-off results.”
- “Replay the saved evaluation artifacts and check the grader results.”
- “Run a smoke test and verify cleanup.”
Notes
The workflow requires the working-tree ./bin/gh and an explicitly authorized sandbox organization and live-resource run. Credential roles must be checked using the prescribed references; tokens must not be printed, and scopes or credentials must not be broadened or created silently. The procedure does not fetch, push, commit, publish, or change Git remotes. Cleanup should be checked in the runner report and unfinished journals; retries use the original owning identities.
This overview was written by AI (OpenAI Codex) from the public SKILL.md. Check the original below for the exact behavior.