How to use the Review Fetch Phase agent skill

What does the review-fetch-phase skill make the agent do?

This skill covers only the state acquisition phase of a review workflow: fetch, validate, render and summarize. It does not triage comments or change code. Its output is the input for a later step, which the skill suggests as /review-triage-phase.

The agent runs a fixed sequence of bundled scripts. The result is a set of generated files in an output directory, which the agent reports back to you by path.

• Fetches the PR's review state into review-state.json (schema version 2)
• Validates that JSON before anything else is derived from it
• Writes review-state.md, a Markdown rendering
• Writes summary.txt, a text summary
• Writes review-targets.json, covering unresolved review threads, reviews with body text, and issue comments with body text

Install the review-fetch-phase skill

Download the skill ZIP from this page and unzip it. You can install it in one of two places, depending on whether you want it for all your projects or just one.

For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder. Alternatively, in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

• Personal: ~/.claude/skills/
• Project only: .claude/skills/ inside your repository
• Claude apps: Settings → Capabilities → Skills, then upload

Set up the requirements before you run it

The skill needs Node.js, the GitHub CLI (gh) authenticated with access to the target repository, and Git. The guard script uses git check-ignore to confirm the output is untracked. In a Jujutsu workspace with no git directory, the guard instead accepts a directory under the workspace root's ignored wip/ tree.

The fetch step calls gh api, so it will fail if gh is missing or not logged in with access to the repository.

• Node.js
• GitHub CLI (gh), authenticated, with access to the target repository
• Git, or a Jujutsu workspace

Trigger the skill with a natural request

Once installed, you can ask for this in plain language. Give the agent the PR URL, and optionally an output directory. The skill's argument hint is [pr-url] [output-dir].

If you omit the output directory, the skill uses wip/reviews/<owner>_<repo>_pr-<number>/ in the working tree.

• "Fetch the review state for https://github.com/OWNER/REPO/pull/123"
• "Run only the fetch phase on this PR and summarize the unresolved threads"
• "Pull the review state for https://github.com/OWNER/REPO/pull/123 into wip/reviews/my-run/"

What the output files contain

The agent returns four artifact paths. review-state.json is the canonical file; the other three are regenerated from it. Because the artifacts contain full comment bodies and author logins, treat them as sensitive.

• review-state.json: canonical state, schema version 2
• review-state.md: Markdown rendering
• summary.txt: text summary
• review-targets.json: triage targets for downstream steps

Safety notes: scripts, network and artifacts on disk

This skill runs scripts and uses the network. The fetch script downloads PR data from GitHub through gh. The source states that artifacts must remain untracked in git, and the guard script checks this before generation.

If gh api fails with TLS or certificate errors inside a sandbox (x509 or OSStatus -26276), the skill stops and tells you to rerun outside the sandbox. It never disables TLS verification.

• Review text and author logins are written to disk in plain files
• Keep the output directory ignored by git, or the guard will stop the run
• Do not use this for private repositories whose review text should not be stored locally