Auto Fix Github

Auto Fix GitHub

Fix a GitHub issue end to end without disturbing the user’s active worktree. Start by proving the issue still needs work. If it does, implement the smallest correct fix, add regression coverage, run the required checks, review the change against project rules, and open a PR that links back to the original issue.

Arguments

• {issueId} (required) — the GitHub issue number, for example 1234
• {repo} (optional) — owner/name; if omitted, infer from the current git remote
• --force (optional) — bypass the in-progress concurrency check; use when intentionally taking over an issue that another auto-skill or human contributor already claimed

Workflow

0. In-progress concurrency check (claim the issue)

Auto-skills MUST NOT clobber each other. Before doing anything else, decide whether you may claim this issue.

CURRENT_USER=$(gh api user --jq '.login')
gh issue view {issueId} --repo {owner}/{repo} --json assignees,labels,number,title,comments

An issue is considered already in progress when ANY of the following is true:

• It carries the in-progress label
• It has at least one assignee whose login is not $CURRENT_USER (a human contributor or another bot has already taken it)
• A claim comment newer than 30 minutes exists from another actor (look for the 🤖 start marker)
• An open PR already references it via Fixes #{issueId} / Closes #{issueId} (handled in step 2 below — but the lock check still applies)

Decision tree:

Stale lock recovery:

• If the in-progress label is older than 60 minutes and the assignee did not push or comment in that window, treat it as expired. Still ask the user before overriding unless --force was set.

Claim the issue (only after the check above passes)

gh issue edit {issueId} --repo {owner}/{repo} --add-assignee "$CURRENT_USER"
gh issue edit {issueId} --repo {owner}/{repo} --add-label "in-progress"
gh issue comment {issueId} --repo {owner}/{repo} --body "🤖 \`auto-fix-github\` started by @${CURRENT_USER} at $(date -u +%Y-%m-%dT%H:%M:%SZ). Other auto-skills will skip this issue until the lock is released."

The release step happens at the end of step 11 — the lock MUST be released even on failure. Use a trap or finally-block so a crash still clears the label and posts a completion comment.

1. Resolve repository and fetch the issue

If {repo} is not provided, infer it from the current checkout:

gh repo view --json nameWithOwner,defaultBranchRef
gh issue view {issueId} --repo {owner}/{repo} --json number,title,body,state,author,url,labels,assignees,comments

Capture at least:

• repository name
• issue title, URL, state, author
• issue body and recent comments

Base branch: Always use develop as the base branch for issue-work branches and PRs, regardless of the repository's default branch. The develop branch is the active integration branch; main receives merges from develop at release time.

2. Check whether the issue is already solved or already has a solution in progress

Do this before creating a worktree or writing code.

Recommended checks:

gh issue view {issueId} --repo {owner}/{repo} --json state
gh search prs --repo {owner}/{repo} "#{issueId}" --state open --json number,title,url,state
gh search prs --repo {owner}/{repo} "#{issueId}" --state merged --json number,title,url,state
git fetch origin develop
git log origin/develop --grep="#{issueId}" --oneline

Also inspect issue comments for phrases like:

• fixed by
• duplicate of
• superseded by
• links to PRs or commits

Stop early when any of these are true:

• the issue is already closed with a credible fix
• an open PR already appears to solve the issue
• the default branch already contains a fix for the issue

If you stop, report what you found and include the relevant PR or commit link instead of duplicating work.

3. Triage the issue before coding

Read enough project context to avoid blind fixes:

• relevant AGENTS.md files from the root router
• related specs in .ai/specs/ or .ai/specs/enterprise/
• .ai/lessons.md

Then reduce the issue to:

• expected behavior
• actual behavior
• likely root cause
• affected module or package
• the smallest safe fix scope

If the issue is ambiguous, try to infer the intended behavior from code, tests, specs, or issue comments before asking the user.

4. Create an isolated issue-fix worktree

Never implement the fix in the repository’s primary worktree.

REPO_ROOT=$(git rev-parse --show-toplevel)
GIT_DIR=$(git rev-parse --git-dir)
GIT_COMMON_DIR=$(git rev-parse --git-common-dir)
WORKTREE_PARENT="$REPO_ROOT/.ai/tmp/auto-fix-github"
CREATED_WORKTREE=0

if [ "$GIT_DIR" != "$GIT_COMMON_DIR" ]; then
  WORKTREE_DIR="$PWD"
else
  WORKTREE_DIR="$WORKTREE_PARENT/issue-{issueId}-$(date +%Y%m%d-%H%M%S)"
  mkdir -p "$WORKTREE_PARENT"
  git fetch origin develop
  git worktree add --detach "$WORKTREE_DIR" "origin/develop"
  CREATED_WORKTREE=1
fi

cd "$WORKTREE_DIR"
BRANCH_PREFIX="fix"
# Switch to feat only when the issue is clearly an enhancement or new capability,
# not a corrective change to existing behavior.
git checkout -B "${BRANCH_PREFIX}/issue-{issueId}-{slug}" "origin/develop"
yarn install --mode=skip-build

If --mode=skip-build is unavailable, run plain yarn install.

Rules:

• If you are already in a linked worktree, reuse it instead of creating a nested worktree.
• The repository’s main worktree must remain untouched.
• All debugging, code changes, testing, and PR prep happen inside the isolated worktree.
• Always clean up the temporary worktree at the end, but only if you created it in this run.

Cleanup sequence:

cd "$REPO_ROOT"
if [ "$CREATED_WORKTREE" = "1" ]; then
  git worktree remove --force "$WORKTREE_DIR"
fi

5. Reproduce or anchor the bug

Before fixing, anchor the issue in code or tests.

Preferred order:

1. Reproduce via an existing failing unit or integration test.
2. Reproduce via a targeted command or local code path.
3. If reproduction is expensive or indirect, encode the missing behavior as a failing unit test first.

Do not skip the reproduction step unless the issue is a trivial static defect and the intended fix is self-evident.

6. Implement the minimal fix

Fix the issue with the smallest defensible code change.

Rules:

• Do not refactor unrelated code.
• Do not broaden scope “while you’re here”.
• Preserve existing contracts unless the issue explicitly requires a compatibility-managed change.
• Prefer modifying the narrowest module or function that owns the bug.

7. Add regression tests (mandatory, autonomous)

Every issue fix MUST include test coverage. This is non-negotiable and must be done autonomously — never skip tests or ask the user whether to add them.

Minimum requirement:

• add or update unit tests that fail before the fix and pass after it

Escalate beyond unit tests when needed:

• add integration tests for risky user flows, permissions, tenant isolation, workflows, or multi-module behavior

Test requirements:

• tests must prove the issue is fixed
• tests must be self-contained
• tests should target the smallest meaningful scope
• tests must pass before the fix is pushed — iterate until they do

8. Run the fix-validation loop

Do not stop after one edit. Keep iterating until the issue is fixed and the change is reviewable.

Per iteration:

1. Run unit tests for every changed package or module.
2. Run typecheck for every changed package or module.
3. If i18n files or user-facing strings changed, run:
   • yarn i18n:check-sync
   • yarn i18n:check-usage
4. If module structure, generated files, entities, or routes changed, run the required generators and follow-up checks:
   • yarn generate
   • yarn build:packages
   • yarn db:generate when entity schema changed
   • yarn template:sync when template-covered files changed
5. Re-read the diff and remove any accidental scope creep.
6. Grep changed non-test files for raw em.findOne(/em.find( — replace with findOneWithDecryption/findWithDecryption. This is a hard rule from AGENTS.md.

Before publishing, run the full CI/CD verification gate from the code-review skill:

• yarn build:packages
• yarn generate
• yarn build:packages
• yarn i18n:check-sync
• yarn i18n:check-usage
• yarn typecheck
• yarn test
• yarn build:app

If the full gate is too expensive to run immediately while debugging, do targeted checks first, but the full gate must pass before you open or update the PR unless a real blocker prevents it.

9. Run code review and backward-compatibility review on your own fix

Before publishing, run the change through the same review discipline as an incoming PR.

Use .ai/skills/code-review/SKILL.md and BACKWARD_COMPATIBILITY.md.

You must explicitly verify:

• no frozen or stable contract surface was broken without the deprecation protocol
• no API response fields were removed
• no event IDs, widget spot IDs, ACL IDs, import paths, or DI names were broken
• no tenant isolation or encryption rules were violated — grep changed files for raw em.findOne(/em.find( in production code; every hit must use findOneWithDecryption/findWithDecryption instead
• the fix remains minimal and does not introduce unrelated churn

If your self-review finds new issues, fix them and repeat the validation loop.

10. Commit and push the fix branch

Only publish after the latest fix state:

• includes regression tests
• passes the required validation
• passes self-review and BC checks

Suggested branch naming:

• fix/issue-{issueId}-{slug}
• feat/issue-{issueId}-{slug} when the issue is clearly an enhancement or feature request rather than a defect

Suggested commit style:

• fix(issue #{issueId}): {short summary}

PR title convention: Use conventional-commit-style prefixes scoped to the affected module or area:

• fix(<area>): <short summary> (#issueId) — for bug fixes (most issue fixes)
• feat(<area>): <short summary> (#issueId) — for new features
• refactor(<area>): <short summary> (#issueId) — for refactors
• security(<area>): <short summary> (#issueId) — for security fixes

Where <area> is the primary affected module or package (e.g., auth, feature_toggles, catalog, ui, shared). Examples:

• fix(auth): prevent privilege escalation via role name spoofing (#1427)
• feat(catalog): add bulk product import endpoint (#1500)

Push with tracking:

git push -u origin "$(git branch --show-current)"

11. Open the PR and release the issue lock

Open a PR against develop using the current repository.

The PR should:

• link the original issue
• describe the root cause
• describe exactly what changed
• mention the added regression tests
• summarize the checks you ran
• call out BC status when relevant

Recommended body structure:

Fixes #{issueId}

## Problem
- {brief issue summary}

## Root Cause
- {root cause}

## What Changed
- {change 1}
- {change 2}

## Tests
- {unit tests added or updated}
- {other checks}

## Backward Compatibility
- No contract surface changes

If the issue is in another repository or should not auto-close, replace Fixes #{issueId} with a plain issue link.

After creating the PR, normalize its labels immediately:

• apply the review pipeline label
• add skip-qa only for clearly low-risk changes such as docs-only, dependency-only, CI-only, test-only, or trivial typo/single-file maintenance fixes
• do not add needs-qa automatically unless the fix clearly introduces customer-facing behavior that must be manually exercised
• never add both needs-qa and skip-qa
• after each added label, post a short PR comment explaining why it was applied

If another auto-skill will immediately continue on the new PR, that follow-up skill must run the normal PR claim protocol (assignee + in-progress + claim comment) before mutating it.

Suggested label comments:

• review: Label set to \review` because the fix PR is ready for code review.`
• skip-qa: Label set to \skip-qa` because this change is low-risk and does not need manual QA.`

Author handoff on the fixed issue

Once the fix PR exists, hand the issue back to the original issue author for verification, unless the author is the current fixer, a bot account, or otherwise unavailable.

Suggested flow:

ISSUE_AUTHOR=$(gh issue view {issueId} --repo {owner}/{repo} --json author --jq '.author.login')

if [ -n "$ISSUE_AUTHOR" ] && [ "$ISSUE_AUTHOR" != "$CURRENT_USER" ] && [ -n "${PR_URL:-}" ]; then
  gh issue edit {issueId} --repo {owner}/{repo} --remove-assignee "$CURRENT_USER"
  gh issue edit {issueId} --repo {owner}/{repo} --add-assignee "$ISSUE_AUTHOR"
  gh issue comment {issueId} --repo {owner}/{repo} --body "Thanks @${ISSUE_AUTHOR} — a fix PR is ready for this issue: ${PR_URL}. Reassigning the issue to you for recheck/verification of the proposed fix."
fi

Rules:

• Do this only after a concrete fix PR exists, or when you are explicitly handing the issue back after confirming the fix landed elsewhere.
• If the author cannot be assigned (bot/deleted account/permission issue), keep the current assignee and still leave the verification handoff comment.
• Keep this handoff comment separate from the lock-release comment so the timeline clearly shows both the human handoff and the automation completion.

Release the in-progress lock on the issue

Always run this as a finally-block — even if the PR open failed or the run was aborted earlier:

gh issue edit {issueId} --repo {owner}/{repo} --remove-label "in-progress"
gh issue comment {issueId} --repo {owner}/{repo} --body "🤖 \`auto-fix-github\` completed: opened ${PR_URL:-(no PR — fix aborted)}. Lock released."

Rules:

• Once a fix PR is opened, the issue assignee should already be handed back to the original issue author before the lock is released
• If no fix PR exists or the author cannot be reassigned, keep the current assignee and explain the fallback in the handoff/completion comments
• Remove the in-progress label so other auto-skills can act on the issue (e.g., a follow-up reviewer)
• Post a completion comment that links the PR (or notes the abort) so the timeline stays auditable

12. Report back

Summarize:

Issue #{issueId}: {title}
Status: {fixed | already solved | already in progress | blocked}
Branch: {branch}
PR: {url}
Tests: {summary}

If you stopped because a fix already exists, report the existing PR or commit instead of creating a new one.

Rules

• Always run the step 0 in-progress check before any other action; never silently override another actor's claim
• Always release the in-progress lock on the issue at the end of step 11, even if the run fails or is aborted (use a trap/finally)
• Always check whether the issue is already solved before writing code
• Always use an isolated worktree
• Reuse the current linked worktree when already inside one; do not create nested worktrees
• Keep the fix scope minimal
• Every fix MUST include regression tests — this is non-negotiable; never push without tests, never ask whether to add them
• Run targeted tests and typecheck while iterating — all tests must pass before pushing
• Run i18n checks when user-facing strings or locale files changed; auto-fix with --fix if issues are found
• Run the full code-review skill and BC check before publishing; auto-fix any actionable findings from the self-review
• Do not open a PR with known failing required checks unless a real blocker prevents completion and you explain that blocker explicitly
• Link the issue in the PR and explain what changed and why
• New PRs opened by this skill must start in the review pipeline state
• Add skip-qa only for clearly low-risk non-customer-facing fixes; otherwise leave QA routing to the author/reviewer
• When this skill adds PR labels, it must also add a short PR comment explaining why
• After opening the fix PR, always hand the issue back to the original author with an explicit reassignment/comment handoff when possible
• Always clean up any temporary worktree created by the current run
• Branches opened by this skill must use fix/ for corrective work or feat/ for enhancement work; never codex/

Source: fullstackhouse/open-mercato — distributed by TomeVault.