ci-sweeper

UTILITY SKILL — CI failure triage and minimal fix, not merge or release.

Input

• Interactive (required): natural-language request; optional failure context (URLs, logs, job names) — constraints in ## Constraints or category-scope.md
• Automation (optional): detect JSON in prompt — from a caller or optional skill detect script; not required for interactive runs. Read agent_edit_allowed and agent_edit_target from ## Constraints per category-automation-envelope.md

Path allowlist, when present, arrives in ## Constraints.

Output Specification

Triage report per common-output-format.md. Survey shape when files are not edited; apply shape when edited — within category-scope.md.

Execution Scope

USE FOR:

• Classify CI failures; apply minimal lint/workflow/shell/doc fixes

DO NOT USE FOR:

• Infra outages, secrets, or runner capacity issues (classify as Watch)
• Refactors >5 files or auth/payment/credential paths
• Merge PRs or push to default branch

Reference Files Guide

• common-checklist.md (always read)
• common-output-format.md (always read)
• category-scope.md (always read)
• category-detect-scope.md (read on interactive path)
• category-input-schema.md (read when detect JSON is present or the optional detect script is run)
• category-run-ledger.md (read when ignored[] is non-empty)
• category-automation-envelope.md (read on automation path)
• common-troubleshooting.md (read on failure)
• category-pr-body-links.md (read when synthesizing the PR body)
• assets/pr-body-template.md and assets/pr-body-template-survey.md (read when synthesizing the PR body)

Workflow

Resolve agent_edit_allowed before classifying failures:

When intent is unclear, survey and offer to apply — edits are harder to undo than a report.

When agent_edit_allowed is true, resolve agent_edit_target: on the interactive path use fix (this skill); on the automation path read agent_edit_target from ## Constraints. Do not branch on other caller metadata outside ## Constraints.

1. Resolve scope (category-scope.md). Parse detect JSON when present; otherwise resolve the detect universe from the prompt (category-detect-scope.md), run this skill's optional detect script with the resolved --scope (default all), then Agent-complement beyond mechanical output; if tools for detect are unavailable, gather failures from the user request and available CI/log context instead of silent no-op. Load category-input-schema.md when parsing detect output. On detect script non-zero exit, read stdout and stop.
2. On the automation path, read category-automation-envelope.md for Constraints, PR templates, and Session Metrics.
3. IF detect reports skip OR no actionable failures after gathering → emit survey no-op per common-output-format.md (Overview explains no-op); on automation path append ## Session Metrics per category-automation-envelope.md; stop.
4. Classify each failure per common-checklist.md: fixable regression within scope, or Watch-class (flake, infra, env, or a fix requiring more than five files). Note ignored[] in Overview when non-empty.
5. IF agent_edit_allowed is false → emit survey shape (### Candidates for regressions, ### Watch for Watch-class failures); on automation path load assets/pr-body-template-survey.md at synthesis and append ## Session Metrics per category-automation-envelope.md; stop — do not edit files.
6. ELSE IF agent_edit_allowed is true AND agent_edit_target is not fix → emit survey shape; note expected agent_edit_target: fix in Overview; stop — do not edit files.
7. ELSE IF no failure is a fixable regression (every failure is Watch-class) → emit survey shape with ### Watch per common-output-format.md; on automation path load assets/pr-body-template-survey.md at synthesis and append ## Session Metrics per category-automation-envelope.md; stop — do not edit files.
8. ELSE (agent_edit_allowed is true AND agent_edit_target is fix) → fix fixable regressions within category-scope.md. When more than three failures are present, fix only the first regression and defer the rest. Watch-class failures get no edits.
9. After edits, run the repository's existing check for each fixed job (the command the failing job runs, when available locally) and record commands and outcomes in ## Verification; never claim a check passed when it failed or was not run.
10. Emit apply shape per common-output-format.md: fixed regressions under ### Changes; deferred regressions and Watch-class failures under ### Deferred. Reconcile Changes / Deferred with git diff --name-only; on automation path load assets/pr-body-template.md at synthesis and append ## Session Metrics (including validation commands) per category-automation-envelope.md.

Error Handling

Examples

• Prompt: Triage failing CI on the integration branch
• Result: Survey report per references/common-output-format.md; apply one-line regression fixes only when agent_edit_allowed is true and agent_edit_target is fix.