Super Build — Builder lane and legacy headless GitHub-Project executor

Inside super-board run, Super Build is the Builder lane: a lane agent in the in-session wave workflow (workflows/super-board-wave.js) that follows "super-board integration" below. Run on its own, /super-build is the legacy headless path (opt-in): you are the orchestrator, the source of truth is the Ready column of the configured GitHub Project board, and your job is to dispatch every issue in Ready (in board order) as a headless claude -p worker in its own git worktree, open a PR for each finished branch, and move its card to QA so the Tester and Reviewer take it from there. Super Build never merges — merging happens only through the Reviewer's merge gate (scripts/super-board-merge-gate.sh).

Curation contract: the human moves cards into Ready when they should be worked. The loop never reads Backlog, In Progress, or Done cards. If Ready is empty, the loop reports "queue empty" and exits cleanly — no error.

This skill is single-purpose: EXECUTE curated GitHub Project Ready issues. It is INDEPENDENT of /super-qa / Super QA (the autonomous bug-bash loop that hardens shipped code). They share no state unless a super-board run sequences them through the board. Run Super Build to make forward progress on the issue queue; run Super QA to bug-bash the existing codebase. They CAN run concurrently in different orchestrator sessions, but expect merge conflicts if both touch the same files.

Role boundary

Super Build is the canonical builder. Do not use a separate build-feature workflow. If the user wants a one-off feature, first create or curate a GitHub issue/card with clear acceptance criteria, move it to Ready, then let Super Build execute it.

Super Build may implement, test, commit, push, open PRs, and move project cards to QA. It never merges, closes issues, or moves cards to Done — the Reviewer does that through the merge gate. It should not invent new product scope beyond the issue body. If the issue needs product/design/security judgment, apply the human-gate or WIP-partial path instead of guessing.

GitHub budget — every "move card" is .claude/bin/super-board-card.sh --config <cfg> move <N> <Status> (1 GraphQL point; gh project item-list / field-list cost ~100–200). Comments, labels and PR reads go over REST; rate-limit-etiquette.md → "Price list" has the calls.

Configuration

The orchestrator needs to know which project board to read.

Standalone mode (legacy)

Reads Ready from the repo's feature project. Resolution order:

1. BUILD_LOOP_PROJECT env var — project number (e.g. 2)
2. BUILD_LOOP_OWNER env var — project owner (e.g. EricTechPro)
3. Auto-discovery fallback: if env vars unset, run gh project list --owner $(gh repo view --json owner -q .owner.login) --format json and pick the project whose title matches the repo name, or the only open project if there's exactly one.

Surface: 🎯 Reading project: <owner>/<title> (#<n>), column "Ready".

Algorithm

0. Pre-flight

Before reading the board:

• Confirm the main worktree is clean: git status --porcelain must be empty.
• Confirm gh auth status works and the repo remote points at the expected GitHub repo.
• Confirm jq, git, and claude are available.
• Resolve and print BUILD_LOOP_OWNER and BUILD_LOOP_PROJECT.
• Run git fetch origin and identify the base branch from the current checkout; do not assume main.

1. Read the source column

Source column is Ready (override with BUILD_LOOP_SOURCE_COLUMN):

SOURCE_COLUMN="${BUILD_LOOP_SOURCE_COLUMN:-Ready}"

.claude/bin/super-board-card.sh --owner "$BUILD_LOOP_OWNER" --number "$BUILD_LOOP_PROJECT" items \
  | jq --arg col "$SOURCE_COLUMN" '.items[] | select(.status == $col and .content.type == "Issue")'

super-board-card.sh items is the gh project item-list shape for ~1 GraphQL point per 100 cards (item-list spends 101) — without bodies.

For each Ready item:

• Capture content.number (issue #), content.title and labels; read the body over REST: gh api repos/{owner}/{repo}/issues/<N> --jq .body.
• Skip issues with the loop:in-progress label (in flight in another orchestrator) or loop:halted label (manually paused) or human-gated label (requires manual handling).
• Parse body for Depends on: #N1, #N2 lines (case-insensitive). If any dep is still open AND not in the current Ready set, the issue is blocked — leave it for a future run.
• Parse body for an optional Skills: line (e.g. Skills: mattpocock-skills:tdd, verification-before-completion). If absent, the worker uses defaults from the preamble.

Ordering:

1. Board order first — items returns cards in the human's manual board order. Respect that. If the user wants Issue X before Issue Y, they drag X above Y on the board.
2. Tiebreaker (rare): within the same drag-position, sort by priority extracted from title regex ^\[?(P[1-4])\]?: P1 < P2 < P4 < P3 < no priority. Then issue number ascending.

If Ready is empty: print 📭 Project board "Ready" column is empty — nothing to dispatch. and exit 0. This is not an error; it means the human hasn't curated work yet.

1b. Resolve the selected worker skills before locks

Hydrate each selected issue through super-board-github-read.py --kind issue -- issue view <N> --json number,title,body. Save a JSON array with number, Ready status, labels and literal body. Run .claude/bin/super-board-skills-ready.py --root <repo> --host claude --cards <cards.json> --lane build before label creation, stale-lock recovery, comments or dispatch. Exit 10 halts and names the missing helpers. Preserve existing claims and worktrees. No automatic helper installation during dispatch. The dispatcher and worker repeat this check.

2. Ensure required labels exist (idempotent)

Once per run, before first dispatch:

gh label create loop:in-progress --color FFA500 --description "Currently being worked on by /super-build" 2>/dev/null || true
gh label create loop:halted --color B60205 --description "Manually paused — /super-build will skip" 2>/dev/null || true
gh label create human-gated --color B60205 --description "Requires manual handling — /super-build will not auto-execute" 2>/dev/null || true

3. Dispatch wave (worker cap)

Worker cap on this legacy path: max_workers from the super-board config when set, otherwise 3.

Report once per wave: ▶️ Super Build dispatching: Issues [#N1, #N2, #N3] (parallel)

For each issue N in the ready set (up to the cap at a time):

a. Lock the issue:

gh issue edit N --add-label loop:in-progress

b. Announce on the issue (so a human reading the issue page knows it's being worked):

gh issue comment N --body "🤖 Dispatched by /super-build — worker spinning up in worktree \`.claude/worktrees/issue-N\` on branch \`loop/issue-N\`. Skills: <skills-line-from-body-or-defaults>."

c. Dispatch:

bash .claude/skills/super-build/scripts/super-build-dispatch.sh N

…via Bash with run_in_background: true. Capture each shell ID.

The dispatcher (this skill's scripts/super-build-dispatch.sh) handles:

• git worktree add -b loop/issue-N .claude/worktrees/issue-N <base-branch>
• gh issue view N --json title,body,labels to compose the worker prompt
• prepend references/worker-preamble.md + append working-directory footer
• cd into worktree and exec claude -p --dangerously-skip-permissions --output-format stream-json --verbose --max-turns 250
• verify the worker produced a 🔧 [chore] loop: close #N commit on its branch
• exit 0 (success) / 2 (worker non-zero) / 3 (no done-commit) / 4 (HUMAN GATE TRIPPED)

Base branch: the orchestrator's currently-checked-out branch (e.g. frontend-rebuild). Pass via BASE_BRANCH env var to the dispatcher. Don't assume main.

4. Wait + reconcile

Poll BashOutput on each in-flight shell. As each finishes:

After creating either a complete or partial PR below, follow PR author notes before reporting the handoff complete. Add file summaries and critical inline notes on the PR, not in source.

On dispatcher exit 0 (success):

• git -C .claude/worktrees/issue-N push -u origin loop/issue-N
• gh pr create --draft --base <base-branch> --head loop/issue-N --title "<emoji> [<type>] <scope>: <subject>" --body-file <body.md> — title in commit-subject format, body = the marker-block template from skills/super-board/references/run.md (writing-standard.md § 2), Closes #N in the status card, Docs: bullets in Solution
• gh issue edit N --remove-label loop:in-progress
• gh issue comment N --body "[builder] [report] ✅ built · PR <URL>" (Next line: qa)
• Move the project card to QA. The Tester and Reviewer take it from there; the issue closes when the Reviewer merges through the merge gate.
• Leave the worktree in place — the merge gate (after the merge) or the cleanup-wt SessionStart hook removes it once the work is on the base branch.
• Report: ✅ Super Build issue #N → PR <URL>, card in QA
• Recompute ready set; if new issues are now unblocked, dispatch in the next wave (respecting the worker cap)

On dispatcher exit 5 (intentional WIP-PARTIAL — open a partial PR, do NOT close):

• git -C .claude/worktrees/issue-N push -u origin loop/issue-N
• gh pr create --draft --base <base-branch> --head loop/issue-N --title "🚧 [wip] <scope>: <slice from worker's final message>" — the status card says Refs #N, never Closes, so the merge does not close the issue.
• gh issue edit N --remove-label loop:in-progress --add-label human-gated
• gh issue comment N --body "[builder] [report] ⚠️ partial · PR <URL>\nDid: <slice>\n❌ Not done: <worker's reason for stopping>\nNext: Eric" — the issue stays open with human-gated until the rest lands.
• Report: 🟡 Issue #N partial PR opened — issue stays open (human-gated)
• Continue dispatching the next issue in the wave. A WIP-PARTIAL is NOT a halt. The worker did intentional, scoped work and the orchestrator advances.

On dispatcher exit 2 or 3 (worker failed or no done-commit):

• gh issue edit N --remove-label loop:in-progress
• gh issue comment N --body "[builder] [report] ❌ failed · exit <X>\nDid: worker ran; worktree \.claude/worktrees/issue-N` kept\n❌ Not done: <one line from the log>\nNext: Eric\n\n<details><summary>log tail</summary>\n\n```\n$(tail -50 .planning/super-build-logs/issue-N.log)\n```\n</details>"`
• Report to the user with tail -50 of .planning/super-build-logs/issue-N.log
• Do NOT open a PR; leave the worktree intact for human inspection
• Halt the loop

On dispatcher exit 4 (HUMAN GATE TRIPPED):

• gh issue edit N --remove-label loop:in-progress --add-label human-gated
• gh issue comment N --body "🔴 HUMAN GATE TRIPPED — needs manual handling. See worktree \.claude/worktrees/issue-N`."`
• Report: 🔴 Issue #N tripped HUMAN GATE — needs manual handling
• Halt the loop

5. Final report

When the selected GitHub Project Ready queue is empty, or only blocked cards remain: report a summary listing PRs opened this run (cards in QA), issues not built (human-gated / loop:halted / blocked dependencies), and any halts. Suggest: "Run super-board run to test, review and merge them."

Issue contract

Every Ready issue should be executable without live clarification. Preferred issue body shape:

## Goal
<one-sentence outcome>

## Acceptance Criteria
- [ ] <observable behavior or artifact>
- [ ] <test/verification expectation>

## Notes / Constraints
- Depends on: #123, #456
- Skills: mattpocock-skills:tdd, verification-before-completion
- Human gates: <deploy, destructive DB action, product decision, etc.>

Super Build treats acceptance criteria as the completion contract. Workers must not edit checkboxes to make the issue look complete. If the issue is too vague to execute, label/comment it as human-gated instead of guessing.

Constraints

• Worker cap: max_workers from the config when set, otherwise 3 on this legacy path (resource throttle). The workflow backend has no cap unless max_workers is set.
• Never auto-touch issues with human-gated label (production cutover, secrets, irreversible ops).
• Never modify code in the main worktree while workers are running. Only run gh commands and push/PR operations.
• Conflicts between concurrent workers' branches are not resolved here — they surface at the merge gate, and the card comes back to the Builder for a rebuild. Don't auto-resolve.
• Report cadence (to the user, in the session): 1 message at start, 1 per dispatch wave, 1 per completion (success/fail), 1 final summary. Don't spam.
• Workers MUST use the right skills. The worker preamble enforces: workers parse the Skills: line from the issue body if present; otherwise they route on the issue's type label (bug → diagnosing-bugs+tdd, feature → implement+tdd+codebase-design, refactor/tech-debt → codebase-design+tdd, docs → skip tdd), always adding ponytail:ponytail (simplest solution first, before any code; inline fallback when the plugin is absent), verification-before-completion and a code-review pass on their own diff. Test mechanics are picked by the localisation ladder, never by label. Decision points walk the decision ladder — acceptance criteria → repo precedent → smallest blast radius → human gate — and stop at the first rung that answers the question. No panel, no vote; mattpocock-skills:grilling is forbidden inside a worker (it waits on a user who is not there) and lives in super-board lint instead. mattpocock-skills:code-review runs once against the worker's own diff before the final commit. See references/decision-policy.md for the skill map, the ladder, the human gates, and the --- decision --- commit trailer.

Worker preamble

The preamble at references/worker-preamble.md is the worker contract: decision policy, HUMAN GATE handling, the per-issue 14-gate contract, and final-commit format. The dispatcher prepends it to every worker prompt. See that file for the verbatim text.

Recovery / re-entry

If the user invokes /super-build after a partial run:

• Re-read gh issue list. Issues already closed are skipped automatically.
• If legacy .worktrees/issue-N exists, stop and inspect it before dispatch. Preserve edits and use git worktree move to the matching .claude/worktrees/issue-N path; never delete the old folder to make a fresh dispatch fit.
• If .claude/worktrees/issue-N exists for an N that's still open: a previous worker was interrupted. Default: notify the user and ask before auto-resuming (a re-dispatch overwrites the previous attempt's branch).
• If the loop:in-progress label is set on issue N but no worktree exists: stale lock from a crashed orchestrator. Remove the label and treat as ready.

Stop conditions

• Selected GitHub Project Ready queue is empty, or all remaining Ready cards are blocked / human-gated / loop:halted.
• Any worker fails (dispatcher exit 2 or 3).
• HUMAN GATE TRIPPED (dispatcher exit 4).
• Push or PR creation fails.
• User interrupts.

Not a stop condition: WIP-PARTIAL (dispatcher exit 5). The orchestrator opens a partial PR, leaves the issue open with human-gated, and continues dispatching the next issue in the wave. WIP-PARTIAL is intentional, bounded work.

Invocation patterns

• Super Build / /super-build → process all GitHub Project Ready issues in board order
• /super-build --only N → execute only issue #N (smoke-test mode, ignores loop:in-progress label on that issue if you set FORCE=1)
• /super-build --dry-run → print the dispatch plan (issues, order, parallelism waves) without invoking workers, without setting labels, without commenting

Companion skill

/super-qa / Super QA is the autonomous bug-bash iteration loop. It is independent of this skill unless a super-board run sequences them through the board. Run Super Build for forward progress on Ready issues; run Super QA to harden the existing codebase by hunting bugs.

super-board integration

When invoked by super-board (env SUPER_BOARD_RUN=1 set by the runner, or invocation contains "super-board run"), follow these rules instead of the standalone defaults:

State protocol

• Read context from: issue body + ALL issue comments + linked PR description + PR comments + PR review threads. NEVER from local state files for inter-lane coordination.
• Respect the worktree path super-board hands you (typically .claude/worktrees/issue-<N>-build/). Don't create your own.
• Respect the single branch super-board hands you (issue-<N>-<slug>). Don't create alternate branches.

Lifecycle (Builder, first pass)

Follow spec .claude/skills/super-board/references/run.md → Builder (first pass). Summary:

1. Worktree off config.base_branch.
2. Branch issue-<N>-<slug> from config.base_branch.
3. Docs check (section below) — if the ticket touches a third-party surface, an upgrade, or auth/billing, read the current official docs first.
4. Implement smallest safe change covering ACs — under the PR size cap (section below).
5. Commit + push (always).
6. Open draft PR with the PR description template from run.md — title in commit-subject format, every marker block filled, Docs: bullets in Solution.
7. Follow PR author notes, then post the [builder] [report] PR comment + the short issue comment with the PR URL (writing-standard.md § 4).
8. Move card Ready/Building → QA.

Keep each PR small (every ticket)

Each PR stays under the merge size cap: merge_policy.auto_max_lines (default 400 changed lines, additions + deletions; lockfiles, generated files, snapshots and migration SQL do not count — see merge_policy.size_exclude). A PR over the cap is never auto-merged: the merge gate parks it in Blocked 🙋 "big PR — please review".

Check git diff --shortstat origin/<base>...HEAD (minus excluded files) before each push. If the change grows past the cap mid-build, stop — do not push on. Commit and push what you have on the branch, leave the PR as a draft, and post a PR comment proposing the split: the vertical slices (each thin end to end, under the cap, with its own ACs), which slice this PR keeps, and /to-tickets as the way to file the rest. Then move the card to Blocked with ❓ ("too big — split proposed in PR #<P>", block-template.md). Never pad past the cap to "finish" the ticket.

Lifecycle (Builder, rebuild)

Triggered when card returns to Ready/Building with loop:rebuild-N label.

1. Read PR review threads on this branch's PR; filter [builder] prefix.
2. For each unresolved [builder] thread: read file:line + suggested fix → apply → resolve thread via:

   gh api graphql -f query='mutation($threadId:ID!){resolveReviewThread(input:{threadId:$threadId}){thread{isResolved}}}' -f threadId="<thread-id>"
   

3. Address any new failure feedback from Tester's latest ❌ comment.
4. Commit + push to same branch. Verify ALL [builder] threads are resolved before exit.
5. Refresh PR author notes, rewrite your PR body blocks (super-board-pr-body.sh), post [builder] [report] PR + issue comments. Move card Ready/Building → QA.

Docs before outside-tool code (every ticket)

Before writing code, decide whether the ticket touches any of:

• a third-party API, SDK, library, CLI or cloud service (calling it, configuring it, adding it);
• a dependency or framework upgrade;
• auth or billing (OAuth/scopes, sessions, tokens, webhooks, payments, plans).

If yes, read the current official docs for the exact surface and version first — context7 when available, otherwise web search to the vendor's own docs, migration guide or changelog. Check the installed version (package.json / lockfile, npm view <pkg> version for a new dep) and read the docs for that major. Repo docs and existing call sites come first for how this repo uses it; vendor docs settle what the tool does. Not Stack Overflow, not old blog posts, not memory.

Cite what you read as Docs: bullets at the end of the PR body's Solution block (URL + the one fact it settled). Docs unreachable → say so there and in the [builder] [report] comment, mark the affected code unverified against current docs, and keep the change minimal. No outside tool touched → Docs: none needed — no third-party surface.

Failure → handoff comment must include root-cause-hash: line

On any failure-handoff comment, include:

root-cause-hash: <sha256 first 12 hex chars>

Hash inputs (joined with |): lane (build) | error class | first 3 unique normalized file:line frames. See .claude/skills/super-board/references/run.md → Root-cause hash.

Never merge

Builder NEVER squash-merges. Reviewer owns merge.

Block exits use the §4 mandatory template

When moving a card to Blocked (or dropping it: closed as not planned, Done, 🤷), populate the full template from .claude/skills/super-board/references/block-template.md. A 1-line "needs creds" comment is a contract violation.