issue-handler

name	issue-handler
description	Own a GitHub issue end-to-end in strike-cli — worktree, implement, test, PR, spawn review agents that comment on the PR, address feedback in a loop, CI, merge. Use when asked to handle/fix/close an issue, ship a PR from an issue, or babysit issue work through merge to main.

Issue handler (strike-cli)

Own the issue end-to-end with due diligence on research, testing, and validation. Shipping is ownership: commit → push → PR → review-agent loop → CI green → merge → cleanup. All code edits happen inside a git worktree, never the primary checkout.

For agents developing strike. Authoring skills/docs may use this pipeline; apply when the user frames work as an issue (or asks you to own issue→main).

Preconditions

• gh auth status ok; git repo; network; issue number/URL known.
• Stop-and-ask: no write access; gh missing; issue already has an open PR you did not create (attach or ask — no duplicate). Dirty primary is fine for git worktree add from origin/main — do not clean/stash primary WIP.

Workflow

0. Orient — gh issue view N (+ comments). Read AGENTS.md + docs/ARCHITECTURE.md for boundaries. Note issue headers (wave/depends/conflicts/priority) if present.
1. Research — map issue→packages; read code + nearest tests. Optional local .plan/features.md / FEATURES_STATUS.md if present (never block if missing). State acceptance criteria bullets. Stop-and-ask if ambiguous.
2. Worktree — create/enter; confirm pwd + branch before edits.
3. Plan — smallest correct change; no drive-by refactors. Assign risk tier A/B/C (see Verification tiers).
4. Implement — worktree only; load sibling skills when domain matches.
5. Test — load write-go-tests when behavior changes.
6. Validate — load test-and-validate for the tier gate; smoke when applicable.
7. Sync main — if the PR will be open >~1h or origin/main moved, git fetch origin main && git merge origin/main then re-run tier gate before/at push.
8. Ship — commit, push origin/<branch>, PR with Fixes #N.
9. Review loop — tier-scaled (below); post findings; address; re-review until merge-ready.
10. Babysit CI — fix branch-related failures; re-push; re-enter review loop after material code changes.
11. Merge — only when merge gates pass (below).
12. Cleanup — remove worktree after merge; leave cwd outside deleted tree.

Worktree setup

MAIN_ROOT=$(dirname "$(git rev-parse --path-format=absolute --git-common-dir)")
WT_PARENT="$(dirname "$MAIN_ROOT")/strike-cli-worktrees"
BRANCH="worktree-<slug>"   # e.g. worktree-24-model-prefix; 2–4 kebab words; include issue # if helpful
mkdir -p "$WT_PARENT"
git fetch origin main
# Resume if yours; else create. Never delete others' worktrees.
if [ -d "$WT_PARENT/$BRANCH" ] && git worktree list | grep -q "$WT_PARENT/$BRANCH"; then
  # On $BRANCH and yours (clean or only your WIP): cd and continue.
  # Stop-and-ask: foreign worktree, unexpected dirty state you don't own, or ambiguous collision.
  cd "$WT_PARENT/$BRANCH"
else
  git worktree add -b "$BRANCH" "$WT_PARENT/$BRANCH" origin/main
  cd "$WT_PARENT/$BRANCH"
fi

Before every edit session:

pwd   # must be $WT_PARENT/$BRANCH
git rev-parse --abbrev-ref HEAD   # must be $BRANCH

Refuse Write/Edit under primary checkout for issue implementation.

Verification tiers (local ship gate)

Source of truth: root AGENTS.md Verification tiers. Summary:

Tier	When	Local gate
A	Docs/skills/markdown only	gofmt if any .go; no full test suite
B	Normal code (default)	gofmt → TUI generate if _src → web-check if web/ → make test && make vet && make build
C	tool / permission / auth / session / engine concurrency / protocol wire / sandbox	Tier B + go test -race ./... -count=1 + focused package tests

CI runs go test -race ./... on every PR — do not always re-run full local race on A/B.

Smoke: load skill smoke when the diff touches cmd/, harness/engine, TUI input/keymap/app, session, or auth.

Contract freeze: default keybind or config-schema changes must update docs/keybinds.md / docs/config.md and call out migration in the PR body.

Flakes: see test-and-validate flake policy — do not block unrelated merges on known env-only flakes; file wave: 0 bugs.

Load test-and-validate for report format. Never weaken/delete tests for green.

Commit and push

git status && git diff && git log --oneline -10
git add <intended paths only>
git commit -m "$(cat <<'EOF'
type(scope): concise summary

Fixes #<N>
EOF
)"
git push -u origin HEAD

Conventional commits: feat/fix/docs/refactor/test/chore. No secrets. No force-push main. No hook skips. Prefer new commits over amend/force on shared PR branch.

Open PR

gh pr create --base main --head "$BRANCH" --title "type(scope): summary" --body "$(cat <<'EOF'
## Summary
- …

## Issue
Fixes #<N>

## Tier
A | B | C

## Verification
- <exact commands from tier gate>
- smoke: ran | skipped (reason)
EOF
)"

Focused PR; no AI wall of text. If PR exists: gh pr view.

Review-agent loop (required for B/C; optional for A)

After the PR exists, do not merge until the tier’s review bar is met. You own dispatching reviewers, posting findings on the PR, fixing code, and repeating.

Pass budget (risk-scaled)

Tier	Review passes	Merge bar
A	0–1 optional	Owner judgment; no blocking findings if a pass ran
B	1 clean pass on current HEAD	0 blocking; should-fix fixed or explicitly deferred with reason on the PR
C	1–2 passes on current HEAD	0 blocking and 0 should-fix

Stall ceiling: 5 passes max — if still dirty, stop-and-ask (do not merge). Five is a cap, not a target.

When to run

• Tier B/C: once after first push + PR open (may run in parallel with CI).
• Again after any push that changes production, test, skill, agent, workflow, or other merge-bound files — including review/CI fixes.
• Skip re-review only for pure changelog typos that cannot affect behavior or process — when unsure, re-review.
• A clean review on an older SHA does not satisfy merge gates after a new push.

Resolve PR identity

PR=$(gh pr view --json number -q .number)
HEAD_SHA=$(gh pr view --json headRefOid -q .headRefOid)
REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)

Spawn reviewers

Preferred: dispatch a read-only reviewer subagent:

• Task subagent_type: reviewer when the host provides it, or
• the in-repo agent at .claude/agents/strike-reviewer.md

Fallback if Task/reviewer is unavailable: you run the same rubric in-process — still load the diff, still post findings via gh. Never skip the loop on Tier B/C.

Optionally spawn a second pass focused on tests/security if Tier C or the first pass is large.

Reviewer prompt must include:

• PR number/URL, base branch (main), and current headRefOid
• Issue number and acceptance criteria
• Hard requirement: run gh pr diff $PR (or receive the full patch) and read surrounding callers/tests before any verdict — never approve from title/URL alone
• Instruction: correctness bugs first (concrete failure scenario), then needless complexity / missed reuse; no impact-free style nits
• Instruction: return ranked findings with path:line, severity blocking | should-fix | nit, failure scenario, and the head SHA actually reviewed
• Instruction: do not edit files; findings only; output must be paste-ready verbatim

Also pull human/GitHub review state:

gh pr view "$PR" --json reviews,comments,reviewDecision,headRefOid
gh api "repos/$REPO/pulls/$PR/comments"
gh api "repos/$REPO/issues/$PR/comments"

Post findings as PR review comments

Post reviewer output on the PR (not only chat). Post the reviewer’s ranked findings verbatim — do not omit, soften, or downgrade severity. If you disagree with a blocking/should-fix item, still post it and stop-and-ask (or reply on the thread with the dispute); silent drop is forbidden.

Event policy (handler is usually the PR author):

• Automated/handler-authored reviews always use "event": "COMMENT" (never REQUEST_CHANGES or APPROVE — GitHub rejects self-approve and self-request-changes leaves merge gates stuck).
• Encode severity in the review body and inline comment text (**blocking:**, **should-fix:**, **nit:**).
• Reserve GitHub APPROVE / REQUEST_CHANGES for distinct human reviewers only.

Preferred — single COMMENT review with inline notes (only for lines in the PR diff hunk on $HEAD_SHA):

HEAD_SHA=$(gh pr view "$PR" --json headRefOid -q .headRefOid)
gh api "repos/$REPO/pulls/$PR/reviews" --method POST --input - <<EOF
{
  "commit_id": "$HEAD_SHA",
  "event": "COMMENT",
  "body": "Review pass N on $HEAD_SHA\n\n(verbatim summary of blocking / should-fix / nits)",
  "comments": [
    {
      "path": "path/to/file.go",
      "line": 42,
      "side": "RIGHT",
      "body": "**blocking:** …\n\nFailure scenario: …"
    }
  ]
}
EOF

If an inline line is not part of the diff hunk, the API 422s — put that finding in the top-level review body instead (do not invent lines).

Fallback — top-level PR comment when inline mapping is impractical:

gh pr comment "$PR" --body "$(cat <<EOF
## Review pass N (head \`$HEAD_SHA\`)
### Blocking
- \`file:line\` — … (failure scenario)

### Should-fix
- …

### Nits (optional)
- …
EOF
)"

Clean pass (meets tier bar) — still required on the PR for B/C: A chat-only “LGTM” does not count. Always record the clean pass against $HEAD_SHA:

HEAD_SHA=$(gh pr view "$PR" --json headRefOid -q .headRefOid)
gh pr comment "$PR" --body "$(cat <<EOF
## Review pass N (head \`$HEAD_SHA\`) — clean
0 blocking, 0 should-fix (or should-fix deferred per tier B). Merge checklist may proceed for this SHA only.
EOF
)"

When building gh api JSON for inline reviews, encode bodies with jq -n --arg body "$text" so quotes/newlines do not break the payload.

Rules for posted comments:

• Include severity tag and failure scenario for blocking/should-fix.
• Anchor path:line to the diff on this PR’s current head SHA.
• Do not spam duplicates across passes — reply on the existing thread or mark resolved in the next summary with the new SHA.

Address comments

1. List open review threads / new comments (human + automated).
2. For each blocking (and should-fix on Tier C, or B when not deferring): fix in the worktree, add/adjust tests when behavior changes, run tier gate.
3. Nits: fix if cheap and clearly better; otherwise reply on the thread why deferred (one line).
4. Commit and push (new commit; no amend/force on shared PR branch).
5. Reply on each addressed thread (or in a single PR comment) with what changed (commit shortsha + brief note).
6. Re-run tier gate; watch CI; spawn another review pass on the new headRefOid when the tier requires it.

Merge-ready checklist (all required for the tier)

• Review bar met for tier A/B/C on current headRefOid (posted review/comment cites that SHA)
• That SHA matches gh pr view --json headRefOid at merge time
• Latest pass findings posted verbatim on the PR (when a pass ran)
• All actionable human review comments addressed or explicitly deferred with reason
• reviewDecision is not CHANGES_REQUESTED (human reviewers); do not use self-REQUEST_CHANGES
• CI checks green on the same head SHA
• mergeable=MERGEABLE, not draft, state OPEN
• Local tier gate passed on the same commit

CI watch / fix

gh-only (no Python watchers):

gh pr checks --watch
# or poll:
gh pr view --json state,mergeable,mergeStateStatus,statusCheckRollup,reviewDecision,isDraft
gh pr checks
gh run list --branch "$BRANCH" --limit 5
gh run view <run-id> --log-failed

Class	Action
Branch-related	fix in worktree → commit → push → re-watch → re-enter review loop if tier requires
Flaky/infra	gh run rerun <id> --failed ≤2; then stop-and-ask
Ambiguous	one diagnosis; then stop-and-ask
Actionable review	fix → commit → push → re-enter review loop

Merge

Merge only when the merge-ready checklist above is fully satisfied:

gh pr merge --merge

--merge matches repo history. Do not pass --delete-branch while still checked out on the feature branch in the worktree — remote/local branch deletion happens in Cleanup after git worktree remove. Blocked on review/permissions → stop-and-ask; never force. Hard forbids: force-push; careless reset --hard; merge with failing checks; merge without meeting the tier review bar; close/reopen PR unprompted.

Cleanup (after merge)

From MAIN_ROOT:

cd "$MAIN_ROOT" && git fetch origin main
git worktree remove "$WT_PARENT/$BRANCH"
git branch -d "$BRANCH" 2>/dev/null || true
git worktree prune

Do not leave cwd inside deleted worktree. Only delete branch you created.

Sibling skills

When	Load
Writing/extending *_test.go	write-go-tests
Before claiming done / after impl or CI-fix	test-and-validate
User-visible cmd/engine/tui/session/auth	smoke
Cutting a version tag	release
Any internal/frontend/tui view/panel/modal/theme	tui-components first

Hard rules

1. Own issue end-to-end through merge — including review-agent loop on B/C.
2. Never edit primary checkout for issue implementation.
3. Never claim done without the tier local gate (not always full race).
4. Never push secrets. Never weaken/delete tests for green.
5. Never force-push main; no destructive git on shared history.
6. Smallest correct change; honor AGENTS.md scope. TUI imports: protocol, host, tui/… only.
7. .plan/ optional research only — never required; never treat unscoped roadmap as the issue.
8. Stop-and-ask on ambiguity rather than guess.
9. Never merge with open blocking findings (or tier-C should-fix) from the latest pass on the current head SHA.
10. Always post review findings on the PR via gh (COMMENT review or comment) when a pass runs — chat-only review does not count.
11. Post reviewer findings verbatim; never omit or downgrade severity. Disputes → stop-and-ask.
12. Handler-authored automated reviews use event: COMMENT only — never self-APPROVE or self-REQUEST_CHANGES.

Stop-and-ask

• gh auth/permission failures; unclear/contradictory acceptance criteria
• Foreign/unexpected dirty worktree; ambiguous path/branch collision; CI red outside branch after 2 reruns
• Merge conflicts with main you cannot resolve confidently (prefer git merge origin/main over rebase)
• Review requires product decision; would commit secrets or change CI/security unexpectedly
• Review loop still fails the tier bar after 5 passes
• Desire to drop or downgrade a reviewer’s blocking/should-fix finding

What this skill is not

• Not unscoped .plan/features.md implementer
• Not multi-agent orchestrator requirement (beyond required PR review subagent / strike-reviewer)
• Not Python CI babysitter
• Not a substitute for sibling domain skills