Gh code review
Validated, agent-friendly GitHub pull request reviews with inline, file, and suggestion threads.
npx -y skills add wildcard/gh-code-review --skill gh-code-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 13 days oldThe repository was created 13 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Review GitHub pull request changes and publish coherent inline, multiline, file-level, and apply-able suggestion comments as one formal review. Use when asked to review a PR, leave code review comments, submit line or file comments, approve, request changes, inspect review threads, reply, resolve discussions, or manage reviewed-file state. Prefer the gh code-review extension; use the bundled gh api fallback when installation is declined or unavailable. Never replace a formal review with a flat gh pr comment.
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
5.3 KB, ~1.0k tokens by cl100k_base, as published. Nobody here has run it
GitHub Code Review
Use a formal pull-request review so every finding is anchored, individually followable, and submitted with one clear decision. The extension is deterministic transport and validation; you remain responsible for review judgment.
Choose the transport
- Run
gh code-review capabilities. - If the command exists, use it.
- If it is absent, offer
gh extension install wildcard/gh-code-review. Never install without the user's consent. - If installation is declined or unavailable, run
scripts/gh_code_review_fallback.py. The fallback supports JSON manifests and the coreinspect,validate, andsubmittransaction. - Read references/cli-reference.md only for advanced lifecycle operations. Read references/direct-api.md when the extension cannot be used.
- Read references/live-examples.md when the user asks for an exact prompt, a walkthrough, or public evidence.
Review workflow
1. Inspect before judging
Resolve the repository and PR, then read normalized review context:
gh code-review inspect 42 -R owner/repo --unresolved
Use --include bodies when full comment text is needed and --include patches only when local diff access is unavailable. Always inspect existing threads before proposing or posting findings.
2. Review the changed code
Confirm each finding against the current diff and surrounding implementation. Avoid duplicating an existing thread or restating a resolved finding without new evidence.
Choose the smallest useful anchor:
line: one changed or context line.- range: one coherent block within a single diff hunk.
file: a concern about the entire changed file that has no honest line anchor.
Use a replacement only when the exact corrected code is known. Suggestions may replace a different number of lines, but they must target RIGHT-side line or range subjects. Do not attach suggestions to deleted lines or whole files.
Write findings so a maintainer can act:
- State the concrete failure or risk.
- Explain when it occurs and why it matters.
- Keep unrelated concerns in separate threads.
- Prefer a suggestion block for a small, exact correction.
- Do not expose hidden reasoning, credentials, or local-only evidence.
3. Build a manifest
Copy references/review-manifest.md and fill the current head SHA returned by inspect. Use unique, stable client_id values.
The default event for unspecified review feedback is COMMENT. Use APPROVE or REQUEST_CHANGES only when the user explicitly authorizes that decision. Do not infer approval from “looks good,” and do not infer a changes request from the presence of findings.
Agent metadata such as severity, category, confidence, rule ID, and evidence remains local; the transport never posts it.
4. Validate before any write
gh code-review validate --input review.json
Validation must pass against the live PR head, diff locations, permissions, existing comments, and pending-review state. Treat HEAD_CHANGED as a required re-inspection. Never move a finding to a nearby valid line merely to satisfy the API.
If event derivation was explicitly requested, preview it:
gh code-review validate --input review.json --derive-event
Derived decisions are visible before submission and still require --confirm-decision for APPROVE or REQUEST_CHANGES.
5. Submit one coherent review
Only submit when the user asked to post or submit the review:
gh code-review submit --input review.json
For an explicitly authorized approval or changes request:
gh code-review submit --input review.json --confirm-decision
Do not use gh pr comment as a fallback. If submission is partially applied, report the recovery receipt and pending review ID exactly; do not blindly retry.
6. Verify
Inspect again and confirm:
- the review state matches the requested event;
- each expected thread exists and is not outdated;
- apply-able suggestions render correctly;
- no duplicate or flat PR comment was created.
Report the review URL, decision, number of threads, and any remaining limitation.
Safety rules
- Reading a PR does not authorize a write. “Post,” “submit,” “leave comments,” “approve,” or “request changes” does.
- Default an unspecified posting decision to
COMMENT. - Require explicit user authorization for
APPROVEandREQUEST_CHANGES. - Never install the extension silently.
- Never abandon a pending review, delete a comment, resolve a thread, dismiss a review, or change viewed state without explicit confirmation.
- Review dismissal creates a visible timeline explanation; warn and require confirmation.
- Never write secrets or hidden provenance markers into comment bodies.
Gives 0 of the 12 instructions most code review skills give in ~1.0k tokens
Counted across 610 of the 674 authors here whose files we hold, read 2026-08-06
- push back with technical reasoning if wrongin 60 of 610, across 24 files
- ask for clarification on unclear itemsin 51 of 610, across 16 files
- fix critical issues immediatelyin 45 of 610, across 29 files
- implement one item at a timein 45 of 610, across 11 files
- group findings by severityin 44 of 610, across 43 files
- verify feedback against the codebasein 42 of 610, across 8 files
- dispatch a code reviewer subagentin 39 of 610, across 23 files
- fix important issues before proceedingin 37 of 610, across 22 files
- test each fix individuallyin 35 of 610, across 7 files
- reply in github comment threadsin 33 of 610, across 5 files
- check for security vulnerabilitiesin 31 of 610, across 27 files
- factualize corrections without over-explainingin 30 of 610, across 2 files
Said here and by no other author read
- inspect unresolved threads before posting findings
- anchor each finding to the smallest useful subject
- target right-side line ranges for suggestions
- keep unrelated concerns in separate threads
- fill the current head SHA in the manifest
- use unique stable client ID values
Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.