fwd-issue-fix

name: fwd:issue-fix description: Privately work through your assigned GitHub issues overnight — pick the oldest open issue, fix it in an isolated worktree, run tests, commit (no push). Driven by /loop. Strictly read-only on GitHub (no labels, no comments, no PRs) so collaborators can't tell the work was automated; all state local in .claude/issue-loop/state.json. Use when user says "work through my GitHub issues overnight", runs /loop /fwd:issue-fix, or invokes /fwd:issue-fix to fix one issue.

fwd:issue-fix

One open GitHub issue → one worktree → one commit. Driven by /loop. State lives in .claude/issue-loop/state.json.

Stealth principle. This is a private automation. Collaborators must not be able to tell that the work was done autonomously. That means:

• GitHub stays read-only — no label changes, no comments, no closing, no gh pr create.
• Nothing gets pushed — branches stay local until you push them yourself after morning review.
• Commits look human-authored — use a normal commit message (via /fwd:git-commit); do not add Co-Authored-By: Claude or "Generated with Claude Code" footers.

Autonomous-mode principle. This skill runs unattended, often overnight. There is nobody at the keyboard to answer questions. Every branch must resolve deterministically.

• Never call AskUserQuestion. Not for confirmations, not for clarifications, not "just to be safe". The user has explicitly forbidden it for this skill.
• Never call ExitPlanMode. Plan mode prompts the user for approval. Plan internally instead, then act.
• Never use interactive shell flags (-i, --interactive, git rebase -i, etc.).
• When you would normally ask, decide and log. Pick the conservative option — "stop this tick cleanly" beats "modify the user's tree". Decisions taken on the user's behalf are recorded in state.json under decisions[] (preflight does this automatically; you do it via ${CLAUDE_SKILL_DIR}/scripts/log-decision.sh in step 4 if needed).
• Ambiguity is not a question for the user. If the issue is unclear or the right fix requires their input, call finalize.sh blocked <N> "<reason>" — the user reviews blocked issues in the morning.

Quick start

# Standalone — fix one issue:
/fwd:issue-fix

# Overnight — fix issues until queue empty or circuit breaker trips:
/loop /fwd:issue-fix

/loop re-fires this skill repeatedly. Each tick = one issue + fresh context. When no pickable issue remains, the tick exits cleanly with no-work.

Per-tick flow

Run the bundled scripts in this exact order. Stop the tick on the first non-zero exit.

1. Preflight

bash "${CLAUDE_SKILL_DIR}/scripts/preflight.sh"

Verifies gh auth, repo state, circuit breaker (3 consecutive failures = stop). Auto-recovers stale in_progress locks (>60 min old) by marking them blocked. On any non-ok exit, preflight also appends a {situation, action: "skip-tick"} entry to state.json.decisions[] so morning-review can see what happened.

If the first line is ok (or ok — recovered stale lock on issue #N), continue to step 2. For every other output, stop the tick cleanly — report the line and exit. Do NOT prompt the user about how to recover (see "Autonomous-mode principle").

2. Pick next issue

bash "${CLAUDE_SKILL_DIR}/scripts/pick-issue.sh"

Outputs JSON {"number":N,"title":"...","body":"..."} for the oldest open issue assigned to @me that is not already done/blocked/in_progress in state. Empty output = no work; report no-work and stop the tick (do NOT call setup-worktree).

3. Setup worktree

First derive two values from the pick-issue output (title + body):

• <type> — one of fix, feat, chore, docs, refactor, perf, test, build, ci, style. Judge it the way fwd:git-commit does: bug → fix; new capability → feat; tooling/repo housekeeping → chore; docs-only edit → docs; etc. Strip any leading Type: from the issue title before judging.
• <name> — kebab-case slug derived from what's left of the title. Drop articles and the conventional-prefix verb (add/update/fix/etc); keep ≤50 chars. Example: Fix: dropdown stays open after Esc on Safari → dropdown-stays-open-after-esc-safari.

Then:

bash "${CLAUDE_SKILL_DIR}/scripts/setup-worktree.sh" <N> <type> <name>

Branch <type>/<name> off main, worktree .trees/<type>/<name>. Symlinks .claude/ into the worktree (CC#28041 workaround). Locks state to in_progress. Prints absolute worktree path on stdout.

4. Fix the issue (your job, Claude)

• Read full issue context: rtk gh issue view <N> --json title,body,labels,comments.
• cd into the worktree path from step 3.
• Plan internally (no ExitPlanMode) → implement → run tests. Hard cap: 3 implementation attempts per tick. If still failing after 3 attempts, go to step 5b.
• Commit using /fwd:git-commit (or plain rtk git commit). Include Refs #<N> in the message body so reviewers can trace it back. Do not add Co-Authored-By: Claude or other AI-attribution footers — see "Stealth principle" above.
• Never rtk git push. Never mutate GitHub (no gh issue edit/comment/close, no gh pr create).
• Never prompt the user. Re-read the Autonomous-mode principle. If the issue is ambiguous, the repro impossible, or you'd need to ask for a clarification — go to 5b with reason "needs-human-input: <one-line summary>". If a non-blocking design decision arises mid-fix (e.g. choosing between two equally valid approaches), pick the more conservative one, log it with bash "${CLAUDE_SKILL_DIR}/scripts/log-decision.sh" <N> "<situation>" "<action-taken>", and continue.

5a. Success → finalize

bash "${CLAUDE_SKILL_DIR}/scripts/finalize.sh" ok <N>

Verifies tests pass, ≥1 commit on the branch, working tree clean. Marks done, resets circuit breaker, leaves the worktree intact for review.

5b. Failure → block and continue

bash "${CLAUDE_SKILL_DIR}/scripts/finalize.sh" blocked <N> "<short reason>"

Removes the worktree, drops the branch, marks blocked, increments the circuit breaker. Tick exits cleanly so /loop can fire the next one.

Hard limits — do not override

• 3 implementation attempts per issue per tick.
• Tests must pass before finalize.sh ok.
• Circuit breaker: 3 consecutive blocked issues → preflight refuses to start.
• Stale lock recovery: in_progress >60 min → auto-marked blocked.

Boundaries

• GitHub is read-only. Never run gh issue edit/comment/close, gh pr create, gh repo edit. Only gh issue list and gh issue view.
• Never push. Branches stay local until you review them and push yourself.
• No AI-attribution footers in commit messages — collaborators reviewing the diff should see normal authorship.
• Never reattempt an issue already in state with done or blocked. Manual state edit required.
• No PRs. This skill commits — that's it.

Reviewing the night's work

jq -r '.issues|to_entries[]|"\(.key) \(.value.status) \(.value.branch // "")"' .claude/issue-loop/state.json
cd .trees/<type>/<name> && rtk git log --oneline main..HEAD

See REFERENCE.md for state-file shape, configuration env vars, reset procedure, and design rationale.