You are executing a complete PR lifecycle: from fresh task-owned worktree setup, through ulw-loop-driven implementation with evidence-bound manual QA, PR creation, and an unbounded verification loop until the PR is merged. The loop has two gates — CI and Cubic — and a failing gate sends you back into that PR's worktree to fix and re-QA. You keep cycling until every active gate passes at once.
The unit of delivery is the smallest PR that compiles, passes, and stands on its own — not \"one task, one PR.\" A single task routinely splits into several atomic PRs; the lifecycle below describes ONE of them, so apply it to each, and build the independent ones concurrently (Phase 0).
\n\nCreate a fresh isolated worktree for each PR before implementation starts. The user's main working directory is read-only context — it may have uncommitted work, and a branch checkout would destroy it. Isolation also makes parallelism cheap: one worktree per PR, so several build at once without colliding.
\n\nDrive all implementation through the ulw-loop skill (your harness's native ultrawork loop) from inside the worktree. Do not free-hand the work: ulw-loop decomposes the brief into goals with binary success criteria, delegates code edits and QA to right-sized subagents, and — the reason it is mandatory here — forces every success criterion to be proven with evidence-bound manual QA on a real surface, not just a green test suite.
Manual QA is the gate, not the tests. This repo's rule is absolute: a change that reaches OpenCode or Codex is not done until you have driven the real harness (tmux / HTTP / browser / GUI — use the manual-QA channel table in the ulw-loop skill) AND written the evidence to disk. No evidence file means the QA did not happen, and you may NOT commit or push. \"It typechecks\" and \"bun test is green\" are NOT QA.
<pr_creation>
\ngit push -u origin \"$BRANCH_NAME\"\nWrite the PR body in English for a human reviewer who has not followed the implementation thread. It must explain the work in plain terms, group changes by reviewer-relevant area instead of dumping files, and make QA evidence auditable without forcing the reviewer to guess what each log proves. Cite sanitized artifacts; do not paste raw secret-bearing logs, env dumps, tokens, auth headers, or private credentials into the PR.
\nIf the PR body needs screenshots or terminal PNGs, follow docs/reference/github-attachment-upload.md: upload via GitHub user attachments from an authenticated web session, include only the final https://github.com/user-attachments/assets/<uuid> URLs, and never commit temporary images, use release assets, use external hosts, or log cookies/tokens.
gh pr create \\\n --base \"$BASE_BRANCH\" \\\n --head \"$BRANCH_NAME\" \\\n --title \"$PR_TITLE\" \\\n --body \"$(cat <<'EOF'\n## Summary\n[2-4 sentences in plain language: what changed, why it changed, and how observable behavior is different after this PR.]\n\n## Changes\n[Group bullets by reviewer-relevant area, not by file. Each bullet should say what changed and how a reviewer can map it to the diff.]\n\n## QA & Evidence\nFor each automated command or manual QA action:\n- **What was tested:** [command or surface driven, with the behavior it was meant to prove]\n- **Observed result:** [actual result, including before/after when relevant]\n- **Artifact:** [`path/to/sanitized-log-or-report`]\n- **Why sufficient:** [which risk or success criterion this evidence covers]\n\n## Risks & Residuals\n[Map each meaningful risk to the evidence above and state the conclusion: mitigated, accepted, or blocked. Include unavailable gates here with the concrete reason.]\n\n## Related Issues\n[Link to issue if applicable]\nEOF\n)\"\nCapture the PR number:
\nPR_NUMBER=$(gh pr view --json number -q .number)\n</pr_creation>
\nThis is the core of the skill. Every active gate must pass for the PR to be ready. The loop has no iteration cap — keep going until done. Gate ordering is intentional: CI is cheapest/fastest; Cubic is external and asynchronous. Gate B (Cubic) is the one gate that can be SKIPPED rather than satisfied — only when its quota is exhausted; it is never skipped just because it found issues. A failing gate is not a patch-and-push: route back to Phase 1, where fixes get the same scope discipline and, if behavior changed, fresh manual-QA evidence before you re-enter the loop.
\n<verify_loop>
\nwhile true:\n 1. Wait for CI → Gate A\n 2. If CI fails → back to Phase 1: read logs, fix + re-QA, commit, push, continue\n 3. Check Cubic → Gate B\n 4. If Cubic has issues → back to Phase 1: fix + re-QA, commit, push, continue\n 5. If Cubic quota out → record Gate B SKIPPED, stop waiting on it\n 6. All active gates pass → break\nCI is the fastest feedback loop. Subscribe to its completion via monitor — never block a model round-trip on gh pr checks --watch.
# Subscribe to CI completion — the monitor event wakes the session when checks finish.\n# Do NOT use `gh pr checks --watch` as a blocking tool call; it burns a full model\n# round-trip (~29s) on every poll. Instead, register a monitor and end the turn:\nmonitor({\n description: \"CI completion for PR $PR_NUMBER\",\n command: \"gh pr checks $PR_NUMBER --watch --fail-fast\",\n filter: \"completed|fail|cancel\"\n})\n# → end turn; the monitor's matching line arrives as an injected event.\n# For a single midpoint status peek (at most once), use:\n# gh pr checks \"$PR_NUMBER\" # one-shot, no --watch\nOn failure: Get the failed run logs to understand what broke:
\n# Find the failed run\nRUN_ID=$(gh run list --branch \"$BRANCH_NAME\" --status failure --json databaseId --jq '.[0].databaseId')\n\n# Get failed job logs\ngh run view \"$RUN_ID\" --log-failed\nRead the logs, then fix per the iteration discipline below.
\nCubic (cubic-dev-ai[bot]) is an automated review bot that comments on PRs. It does NOT use GitHub's APPROVED review state — instead it posts comments with issue counts and confidence scores.
Approval signal: The latest Cubic comment contains **No issues found** and confidence **5/5**.
Issue signal: The comment lists issues with file-level detail.
\nQuota-exhausted signal: Cubic posts a usage/quota/limit message instead of a review, or no Cubic review appears within the bounded wait below. This is the ONLY case where you skip Gate B and proceed — record it as SKIPPED in the final report, never silently. Issues are never a reason to skip.
\n# Get the latest Cubic review\nCUBIC_REVIEW=$(gh api \"repos/${REPO}/pulls/${PR_NUMBER}/reviews\" \\\n --jq '[.[] | select(.user.login == \"cubic-dev-ai[bot]\")] | last | .body')\n\nif echo \"$CUBIC_REVIEW\" | grep -q \"No issues found\"; then\n echo \"Cubic: APPROVED\"\nelif echo \"$CUBIC_REVIEW\" | grep -qiE \"quota|usage limit|rate limit|out of (credits|reviews)|upgrade your plan\"; then\n echo \"Cubic: SKIPPED (quota exhausted)\" # Gate B satisfied-by-skip; do not loop on it\nelse\n echo \"Cubic: ISSUES FOUND\"\n echo \"$CUBIC_REVIEW\"\nfi\nOn issues: Cubic's review body contains structured issue descriptions. Parse them, determine which are valid (some may be false positives), and fix the valid ones per the iteration discipline below.
\nCubic reviews are triggered automatically on PR updates. After pushing a fix, subscribe to the new review arriving — never spin a for _ in $(seq 1 30) polling loop that burns model round-trips.
# Subscribe to a NEW Cubic review after push. The monitor exits when a review\n# newer than PUSH_TIME appears, or times out (quota exhausted → Gate B SKIPPED).\nPUSH_TIME=$(date -u +%Y-%m-%dT%H:%M:%SZ)\nmonitor({\n description: \"Cubic review for PR $PR_NUMBER\",\n command: \"LATEST=$(gh api repos/${REPO}/pulls/${PR_NUMBER}/reviews --jq '[.[] | select(.user.login == \"cubic-dev-ai[bot]\")] | last | .submitted_at // empty'); [ -n \\\"$LATEST\\\" ] && [ \\\"$LATEST\\\" > \\\"$PUSH_TIME\\\" ] && echo NEW_REVIEW || echo WAITING\",\n filter: \"NEW_REVIEW\",\n timeout_ms: 600000 # 10 min bound — if no review arrives, Gate B is SKIPPED (quota exhausted)\n})\n# → end turn; if the monitor times out without NEW_REVIEW, treat Gate B as SKIPPED.\n# For a single midpoint peek (at most once), use:\n# gh api \"repos/${REPO}/pulls/${PR_NUMBER}/reviews\" --jq '[.[] | select(.user.login == \"cubic-dev-ai[bot]\")] | last | .submitted_at'\nEach iteration through the loop:
\nAvoid the temptation to \"improve\" unrelated code during fix iterations. Scope creep in the fix loop makes debugging harder and can introduce new failures.
\n</verify_loop>
\nOnce all active gates pass (Cubic may be SKIPPED on quota):
\n<merge_cleanup>
\nEnabling auto-merge is the default - do it unless the user explicitly told you not to merge. Auto-merge hands the merge to GitHub, which lands the PR the moment every required gate is green, so you never sit and babysit checks. It does NOT bypass the gates: if a gate fails, GitHub will not merge, which routes you back to Phase 1 to fix and re-QA like any other failing gate.
\n# This repository requires merge commits. Never use --squash or --rebase.\n# --auto arms auto-merge: GitHub merges as soon as all required checks pass.\ngh pr merge \"$PR_NUMBER\" --merge --auto --delete-branch\n# If the repo has not enabled the auto-merge feature, --auto errors; once the gates\n# are green, fall back to a direct merge: gh pr merge \"$PR_NUMBER\" --merge --delete-branch\nThen subscribe to the merge completing — never block a model round-trip on an until [ ... MERGED ] polling loop:
# Subscribe to merge completion. The monitor exits when gh pr view returns MERGED.\nmonitor({\n description: \"Merge completion for PR $PR_NUMBER\",\n command: \"[ \\\"$(gh pr view $PR_NUMBER --json state -q .state)\\\" = \\\"MERGED\\\" ] && echo MERGED || echo WAITING\",\n filter: \"MERGED\",\n timeout_ms: 1800000 # 30 min bound for auto-merge to land after all gates pass\n})\n# → end turn; the MERGED event wakes the session for the cleanup step.\n# If the monitor times out, check merge state once: gh pr view \"$PR_NUMBER\" --json state -q .state\nIf the user opted out of merging, skip the merge but STILL run the cleanup below: the worktree is removed either way.
\nBefore removing the worktree, copy .omo/ state back. When .omo/ is gitignored, files written there during worktree execution are not committed or merged — they would be lost on worktree removal.
# Sync .omo state from worktree to main repo (preserves task state, plans, notepads)\nif [ -d \"$WORKTREE_PATH/.omo\" ]; then\n mkdir -p \"$ORIGINAL_DIR/.omo\"\n cp -r \"$WORKTREE_PATH/.omo/\"* \"$ORIGINAL_DIR/.omo/\" 2>/dev/null || true\nfi\nThe worktree served its purpose — remove it to avoid disk bloat:
\ncd \"$ORIGINAL_DIR\" # Return to original working directory\ngit worktree remove \"$WORKTREE_PATH\"\n# Prune any stale worktree references\ngit worktree prune\nSummarize what happened:
\n## PR Complete\n\n- **PR**: #{PR_NUMBER} — {PR_TITLE}\n- **Branch**: {BRANCH_NAME} → {BASE_BRANCH}\n- **Iterations**: {N} verification loops\n- **Gates**: CI pass | Cubic {pass | SKIPPED (quota exhausted)}\n- **Merged**: {yes | no — left for you to merge, as requested}\n- **Worktree**: cleaned up\n</merge_cleanup>
\n<failure_recovery>
\nIf you hit an unrecoverable error (e.g., merge conflict with base branch, infrastructure failure):
\nFor merge conflicts:
\ncd \"$WORKTREE_PATH\"\ngit fetch origin \"$BASE_BRANCH\"\ngit rebase \"origin/$BASE_BRANCH\"\n# Resolve conflicts, then continue the loop\n</failure_recovery>
\n| Violation | \nWhy it fails | \nSeverity | \n
|---|---|---|
| Working in main worktree instead of isolated worktree | \nPollutes user's working directory, may destroy uncommitted work | \nCRITICAL | \n
| Committing or pushing without manual-QA evidence on disk | \n\"Tests pass\" never proves the feature works; the repo forbids it for OpenCode/Codex-touching changes | \nCRITICAL | \n
| Pushing directly to dev/master | \nBypasses review entirely | \nCRITICAL | \n
| Skipping CI gate after code changes | \nCubic may pass on stale code | \nCRITICAL | \n
| Skipping Cubic because it found issues | \nOnly an exhausted quota justifies a skip; real issues must be fixed and re-pushed | \nHIGH | \n
| Fixing unrelated code during verification loop | \nScope creep causes new failures | \nHIGH | \n
| Deleting worktree on failure | \nUser loses ability to inspect/resume | \nHIGH | \n
| Ignoring Cubic false positives without justification | \nCubic issues should be evaluated, not blindly dismissed | \nMEDIUM | \n
| Bundling independent slices into one big PR | \nAtomic review dies — a 2000-line PR gets rubber-stamped, regressions hide, and one bad slice blocks all the others | \nHIGH | \n
| Giant single commits | \nHarder to isolate failures, violates git-master principles | \nMEDIUM | \n
| Not running local checks before push | \nWastes CI time on obvious failures | \nMEDIUM | \n