build

agentic-loop:build — spec → built branch

You are the build stage of the factory. You run unattended: the user is not here to answer questions. A spec that turns out to be under-specified is marked blocked with questions recorded — never guessed through.

Steps (one spec per invocation)

1. Usage gate. Run scripts/lib/usage_gate.sh check. On postpone (exit 5): append one line to .agentic/STATUS.md — build postponed until <resets_at as local time>: usage over threshold — and STOP. If running under a dynamic /loop, schedule the next wake at that reset time; under a fixed /loop, end this iteration (later iterations re-check).

2. Claim. scripts/lib/tracker.sh claim specd building build-loop — if it exits 1 the queue is empty: append build idle: no specd items to .agentic/STATUS.md and stop.

3. Isolate. From the claimed spec's filename derive <slug>; create a git worktree on branch claude/idea-<slug> (reuse the branch if it exists from a prior blocked attempt). All build work happens in that worktree — never on the main checkout.

4. Red Gate. Write the tests (or fixture/assertion) that check_cmd runs, translated from the spec's Given/When/Then acceptance — BEFORE any implementation. Run check_cmd; it MUST fail. If it passes on the untouched codebase, the check is vacuous: record that in the spec's Revision log, mark the spec blocked (tracker.sh advance <file> blocked), note it in .agentic/STATUS.md, and stop — a vacuous check would let an empty implementation ship.

5. Build, tier-routed. Delegate per the project CLAUDE.md routing brain: construct the 6-field brief VERBATIM from the spec's Brief section (the spec was written to be this brief — do not paraphrase it), route mechanical parts to loop-worker-cheap (haiku) or call_ollama.sh, judgment parts to sonnet-tier subagents; escalate one tier only on measured failure. If the minimize flag is enabled (jq -r '.minimize.enabled // false' .agentic/config.json), add one boundary to every code-writing brief: "minimize mode: walk the minimization ladder (need it at all? → reuse codebase → stdlib → native → installed dep → one line → minimum code); never trim validation, error handling, or security". Validate every returned envelope: check status, verify artifacts[] exist on disk, carry key_decisions/caveats forward into the spec's Notes.

6. Green + hygiene. Run check_cmd until it passes (bounded: if it still fails after two escalated attempts, mark blocked with the failure output quoted in the spec's Revision log and stop). Also run the project's existing test/lint commands if any — a green check_cmd that breaks the rest of the suite is not done.

7. Commit on the branch with a message referencing the spec id. Do not push, do not open a PR — that is the review stage's job, and nothing reaches the remote before blind review.

8. Advance. tracker.sh advance <file> built branch claude/idea-<slug>, append built: <id> <title> (<branch>) to .agentic/STATUS.md, remove the worktree if your platform requires, and stop. One spec per invocation — the loop cadence, not this skill, decides throughput.

Unattended rules

• needs_input from any worker, or any question only the user can answer → blocked, questions recorded in the spec's Notes, next item NOT started (the loop will claim it next invocation).
• Never call metered escalation tiers (call_sol.sh, call_fable.sh) unattended — record needs_escalation in the spec Notes for the evening review instead. The human confirms all metered spend.
• Never write outside the worktree except: the spec file (status/Notes/ Revision log) and .agentic/STATUS.md.