What happened
An agent was driving a pull request to green: pushing commits, watching gh pr checks, and fixing whatever went red. When two CI jobs failed, it read their logs with gh run view --job <id> --log, piped the output through grep, and still had to guess which lines mattered. It never invoked the sling skill, even though sling was installed, authenticated and healthy (sling doctor --agent exit 0, every check ok).
Once the skill was invoked, the same diagnosis took two commands:
$ sling why <job-id> --agent
{"classification": "step_failure",
"summary": "The step 12 \"Verify the guards can actually fail (mutation registry)\" step failed (exit code 1).", …}
$ sling logs <job-id> --grep 'SURVIVED|BROKEN|FAIL' --limit 30 --agent
{"lines": [{"line_number": 1540, "log_data": " SURVIVED <mutant name> <- THE GUARD DID NOT FIRE"}], …}Why the skill did not trigger
The description is written around a user asking: it says to invoke the skill "whenever the user asks why a job or run failed, what made a run slow, … or pastes a GitHub Actions URL". In this session no user asked. The agent hit the failure itself while watching CI on its own pull request, so nothing in the description matched, and it fell through to the gh habits it already had.
That is the common case for coding agents: most CI failures an agent meets are red checks during its own work (a review loop, a "watch CI until green" step, a pre-merge check), not questions a user typed.
Proposed fix
- Trigger on agent-initiated CI triage, not only on user questions. Extend the description to cover "a check you are watching on a pull request goes red", "
gh pr checksshows a failure", and "a CI job you pushed failed", with the same instruction: invoke this skill before reading any logs withgh. - Name the anti-pattern in the description.
gh run view --log/--log-failedpiped throughgrepis the exact path agents take instead. Saying "instead ofgh run view --log" in the description gives the trigger something concrete to match. - Lead the routing table with the PR case. "The checks on a specific PR →
gh pr checks, thensling why" sits low in the table today. A red check during PR work is the most frequent entry point and should be the first row.
Related: #102 covers a different path to the same outcome (one sling verb is down, and the agent abandons sling for raw gh logs).
Activity
0No activity yet.