analyze-ci
Analyze failed GitHub Action jobs for a pull request.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 0
- Forks
- 1
01 前言 Front Matter
- name
- analyze-ci
- description
- Analyze failed GitHub Action jobs for a pull request.
- allowed-tools
- [ "Bash(uv run skills analyze-ci:*)" ]
03 SKILL.md 正文
Analyze CI Failures
This skill analyzes logs from failed GitHub Action jobs using Claude.
Prerequisites
- GitHub Token: Auto-detected via
gh auth token, or setGITHUB_TOKENenv var
Usage
# Analyze all failed jobs in a PR
uv run skills analyze-ci <pr_url>
# Analyze specific job URLs directly
uv run skills analyze-ci <job_url> [job_url ...]
# Show debug info (tokens and costs)
uv run skills analyze-ci <pr_url> --debug
Output: A concise failure summary with root cause, error messages, test names, and relevant log snippets.
Local or Reproduced CI Failures
When the task provides a failed repository and a local CI/reproducer script instead of a PR URL, treat the captured CI artifacts and the provided reproducer as the authoritative build contract.
- Before constructing any new local reproduction, inspect all available captured CI artifacts and original build logs, including files such as
*-orig.log, workflow output, and reproducer logs. Record the earliest concrete failing command, failing test node IDs, pass/fail counts, and runtime version(s). If the captured CI reaches pytest/tox and reports a specific code-level failure, that failure is the baseline even if a newly constructed local environment later exposes unrelated packaging or installation problems. - Preserve the original repository state and inspect the build workflow, package metadata, test configuration, runtime constraints, and the exact failing command before editing.
- Run the narrowest faithful reproduction of the captured failure first. Do not replace an existing authoritative failure with an easier-to-trigger command such as an editable install unless that exact command is part of the recorded CI path.
- Separate actionable product failures from infrastructure limitations. Missing interpreters, unavailable services, package-index failures, permissions, and sandbox-specific installer behavior are not evidence that packaging, tox, workflow, or product code is the original root cause unless the captured CI fails at that same stage.
- Do not change packaging metadata, dependency declarations, tox configuration, workflow matrices, or supported-runtime declarations merely to make an ad-hoc local environment run. Such changes require direct evidence from the authoritative CI failure stage.
- Create a reversible snapshot of the last-known-good working tree before every candidate change. Rerun the original failing test and then the relevant full suite, and compare exact before/after failing node IDs and exception classes. If the pass count decreases, a new failing node appears, or a new exception class appears, restore that snapshot immediately. Never debug a regression introduced by the candidate patch on top of that same patch.
- Keep unrelated packaging, dependency, and behavior changes out of the repair. A successful import, installation, or partial test run is not evidence that the CI build is fixed.
- Generate deliverable patches from version-control differences rather than hand-counting unified-diff hunks. Verify each patch with
git apply --checkagainst a clean copy, apply it once, and confirm that the applied working tree is exactly the version that passed. - Treat every interpreter actually exercised by the authoritative CI/reproducer as a blocking gate. Record unavailable interpreters separately as infrastructure limitations rather than hiding them by changing the matrix or tox configuration.
- Do not finish after announcing a next command. Finish only after the requested diagnosis note and version-control-generated patch files exist, each patch passes
git apply --checkagainst a clean copy, and the authoritative build command has completed successfully; otherwise document a separately verified infrastructure blocker.
Evidence hierarchy
When evidence conflicts, use this order:
- Captured original CI/reproducer logs from the failed job.
- The benchmark-provided authoritative build/reproducer script run against the failed repository.
- The repository's own configured test command under the recorded runtime/dependency conditions.
- Narrow diagnostic commands that preserve the same environment.
- Newly constructed local install/test environments.
Lower-ranked evidence may explain a blocker, but must not silently replace a higher-ranked failure as the repair target.
Examples
# Analyze CI failures for a PR
uv run skills analyze-ci https://github.com/mlflow/mlflow/pull/19601
# Analyze specific job URLs directly
uv run skills analyze-ci https://github.com/mlflow/mlflow/actions/runs/12345/job/67890
04 同仓库更多 Skills
video-frame-extraction
Extract frames from video files and save them as images using OpenCV
openai-vision
Analyze images and multi-frame sequences using OpenAI GPT vision models
image-ocr
Extract text content from images using Tesseract OCR via Python, including structured field extraction from receipts and forms
video-frame-extraction
Extract frames from video files and save them as images using OpenCV
openai-vision
Analyze images and multi-frame sequences using OpenAI GPT vision models
image-ocr
Extract text content from images using Tesseract OCR via Python, including structured field extraction from receipts and forms