SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsfix-github-issues

fix-github-issues

Coding & development

View repository

Updated 10-23 16:18

About this skill

Fix GitHub Issues

Work the open issues that a review bot filed. Each issue is a container of findings: some are legit, some stale, some wrong. Every finding earns its own verdict from the code on the default branch, never from the issue's headline. An issue closes only when no legit finding is left in it, and a fix counts only once it reaches the default branch.

Project context. Read .agents/github-review.md when the repo has one. Its sections ## Review bots (with the bot's issue label), ## Auto-dismiss, ## Never dismiss, and ## Stack facts (what the bots get wrong) hold this repo's facts; a missing section means the default here. Then follow the repo's AGENTS.md / CLAUDE.md rules for tests, lint, cache clears, and a dirty tree.

1. Build the working set

Find the default branch (written <default> below), run git fetch origin, and verify everything against origin/<default>, not a local branch. List every open issue with the bot label (or the ones the user named), newest first. The calls, and the check that the list is not cut off at its limit, are in references/mechanics.md.

Walk order is newest to oldest, but let evidence reorder it. A newer diff that introduced a defect an older issue already flagged makes the older finding LEGIT, and you fix both in one commit chain. A newer commit that fixed an older finding makes that finding STALE: cite the sha.

Completion criterion: every open issue in scope is listed, newest first, with fewer issues than the limit.

2. Classify every finding

For each issue, read its body and every comment (note any "fixed in " or "false positive" from earlier runs). Read the commits it reviewed. Then write one line per finding:

Finding N: [STALE | AUTO-DISMISS | HALLUCINATION | LEGIT] — <one-line evidence>
  • Finding 1: STALE — fixed in bf707fde3, see app/Foo.php:42
  • Finding 3: HALLUCINATION — rg 'BarPolicy::class' returns 0; symbol does not exist
  • Finding 4: LEGIT — app/Auth.php:88 still uses < instead of <=, exploitable on token boundary

Verification rules:

  • For a "class, method, or permission missing" claim, rg the symbol where the finding says it is missing (that file, call site, or test). If it is there, the finding is HALLUCINATION. That it exists somewhere else does not refute a missing check.
  • For an "X may break Y" claim, trace the real call sites. If nothing calls it, the finding is STALE or HALLUCINATION.
  • A finding whose file:line does not exist at the reviewed commit, or does not hold the code it describes, is HALLUCINATION. A finding on a file the commit did not touch stands when the change reaches it (a changed caller that breaks an unchanged callee); only when nothing in the commit leads there is it HALLUCINATION.
  • A documented design choice in the commit body that contradicts the finding is AUTO-DISMISS.
  • A finding whose body retracts itself ("this technically works", "disregard") is AUTO-DISMISS.

Delegate the legwork on a large set. With more than 5 findings, hand the first pass to read-only workers in batches of about 10. Each brief carries the batch (id, author, reviewed commit, file:line, body), the ref origin/<default>, the line format above, and the verification rules. In Claude Code with this plugin installed, dispatch subagent_type=leek-skills:finding-verifier for each batch in parallel; it keeps the bots' recurring false premises in project memory, under this repo's .claude/. Elsewhere, use any read-only sub-agent on a cheaper model, or classify inline when the harness has none. Then re-verify yourself every line that came back LEGIT or UNSURE. You read that code for the fix anyway.

Second pass. Re-read every LEGIT against the project's auto-dismiss list and downgrade any match. A category on the project's never-dismiss list stays LEGIT. Do this before writing any code.

Completion criterion: every finding in the working set has a line, and the second pass is done.

3. Fix the legit ones

Work on a new branch fix/issues-<run> from origin/<default>, in a worktree when the tree is dirty (Branch or worktree). On a clean tree, record where the user's checkout is before you switch: step 4 puts it back there.

Per LEGIT finding: make the minimal change, with no drive-by refactors. Add a test for logic, auth, or data changes and run it green. Commit it on its own (Conventional Commits, explicit staging). Put fixes #NNN only on the commit that resolves the issue's last legit finding, so the issue closes at the right moment.

Run tests against the repo's configured test database when you can. A run that needs its own database names and registers it by Scratch databases.

A LEGIT finding you cannot fix (a secret rotation, infra, a human decision) keeps the issue OPEN; step 5 names the blocker.

Completion criterion: every fixable LEGIT finding has a commit on fix/issues-<run>, with a green test where it needs one, and every unfixable one is named.

4. Land the fixes

An issue closes only when its fixes #NNN commit reaches <default>, so a local commit fixes nothing yet. Show the user the commits and the issues each one closes, say where they will go, and land them without asking: push to <default> by default; open a pull request instead when branch protection or the repo's rules forbid direct pushes. A user instruction in the input or the conversation ("open a PR", "keep them local") overrides both. The commands are in Land.

Once they are landed, switch the user's checkout back to the branch or commit recorded in step 3 (Restore the checkout). Never leave it on fix/issues-<run>. Do the same if the run stops early for any reason after the switch.

Completion criterion: the commits are on <default>, in an open pull request, or kept local by the user's instruction, and the checkout is back where it started.

5. Post a disposition

Post one comment per issue in the working set, with one bullet per finding (calls in Disposition and close):

- Finding 1 (X): fixed in <sha>
- Finding 2 (Y): stale — superseded by <sha>
- Finding 3 (Z): auto-dismiss — <category>
- Finding 4 (W): hallucination — symbol does not exist at claimed location
  • Fixed on <default>: the push closed it; post the comment after.
  • Nothing legit: post the comment, then close the issue.
  • Fix in an open pull request or kept local: the issue stays open; the comment says where the fix is.
  • A LEGIT finding you cannot fix: the issue stays open; the comment names the blocker.

If you later find a misclassified finding, reopen the issue with a correction note.

Completion criterion: every issue has its comment, and gh issue view N --json state agrees with it.

6. Clean up

Remove nothing before this step. Then remove without asking every worktree and database you created whose fixes are on <default> (or nothing was fixed in it), and list each removal in the report (Clean up). A worktree git refuses to remove without --force holds unsaved work: keep it.

Ask once only about what is left that you created: a refused worktree, and one whose fixes are in an open pull request or kept local (AskUserQuestion where available: a multiSelect: true question with one option per item, nothing pre-ticked, at most 4 per question and further questions for the rest; otherwise a plain question in chat: "Can I delete these too?"). Remove only what the answer selects, never with --force. What stays stays registered, for repository-cleanup. Skip the question when nothing is left.

Completion criterion: everything removable is removed, any question is answered, or you created nothing. Report the tally: fixed, stale, auto-dismiss, hallucination, and blocked. Name each fix not yet on <default>, and each worktree or database kept. Mark the blocked count and each kept item ⚠️; no other marks.