Engineering Loop

Engineering Loop (single change, one PR)

The standard loop for one ordinary change: spec to merged PR with review and verification. Lightweight counterpart to the epic-dev workflow.

Scale check (before starting)

The loop

1. Read the spec. gh issue view N. If the request is a one-line or ambiguous instruction, restate your interpretation in one sentence with an opt-out ("proceeding on that basis; tell me if you meant X") and continue; do not stall.
2. Explore just enough, then find the pattern to mirror. Read the target area and locate an existing correct implementation of the same shape (a similar route, a similar test, the sibling module). New code mirrors the house pattern; do not invent a new one for an ordinary change. If the change builds on a third-party API you have not used in this repo, probe the installed version first (import and inspect; do not trust memory).
3. Branch. gh issue develop N --checkout (or git checkout -b feature/issue-N-<slug> per repo convention). Never work on the default branch.
4. Pin test first (refactors only). If the change claims to preserve behavior, commit #1 is a characterization test captured against the ORIGINAL code (route inventory, golden output). A pin test written after the move proves nothing.
5. Implement the smallest coherent diff. Match surrounding style. Discard incidental churn before committing (lockfiles, formatting of untouched lines). No backward-compatibility shims when the user confirms zero consumers; remove outright and preserve the substantive invariant instead. For every new guard or error path, decide fail-open vs fail-closed explicitly and say why in the code or PR. If you find a genuine adjacent bug outside the assigned scope: fix it only if it is small and mechanistically connected to the change you are already making, and disclose it under its own heading in the PR with evidence; otherwise file an issue. Never fold it silently into the main diff.
6. Gate every commit. Format, lint, typecheck, test; the bar is zero NEW diagnostics against the MEASURED baseline (measure it; do not trust a documented number, and note discrepancies). Python: uv run ruff format && uv run ruff check --fix && uv run ty check && uv run pytest. JS/TS: bun run biome check --write && bun test (or the project's configured commands). Commit atomically: subject under 50 characters, no emojis, no AI attribution. For changes to decision logic (parsing, approve/deny paths, routing, security gates), green tests alone are not enough: run a repeatable scenario suite and quote the before/after scores in the PR.
7. Long-running steps (benchmarks, big builds, batch jobs over ~10 minutes): detach them per references/background-jobs.md; never leave them as session-tracked shells. Commit expensive-to-reproduce results the moment they land, separately from code.
8. Push and open the PR. Body: what changed, why (link "Closes #N"), and what was tested with commands and counts. No emojis, no AI attribution. Semantic line breaks remain the default for prose source, but GitHub issue and PR bodies are the exception: keep each paragraph on one source line and separate paragraphs with blank lines; do not insert sentence- or clause-level newlines inside a paragraph. Verify any user-supplied factual claim (dates, numbers, attributions) against available evidence before writing it into a PR, issue, or other durable document; if the evidence contradicts the claim, surface the discrepancy instead of transcribing it.
9. Review. Run the pr-review-toolkit skill (or repo review command) on the PR. Address ALL findings: fix, or reject with a one-sentence reason posted to the PR ("false positive" / "intentionally different by design"). No silent drops; a deleted test that held the only coverage of a scenario is a blocking gap, not a nit.
10. CI, then merge. Wait for checks (background watch: gh pr checks N --watch); merge only when green, using the repo's merge convention (regular merge unless the repo says squash; when a local skill/command convention conflicts with the user's standing preference, surface it once BEFORE the first merge, not after). Confirm any irreversible cleanup (deleting remote branches beyond the PR's own, force-removing dirty worktrees) with the user instead of folding it into a broad "proceed".
11. Close out. Close the issue with a substantive comment (evidence, not "done"), update project state/memory files with anything reusable learned, close/remove completed one-off agent threads, and clean up artifacts you created. Retain an agent only for a named recurring role with a concrete next task. If stopping mid-stream instead, write a handoff note with exactly these sections: Shipped (commit/PR refs), Operational state (what is running and how to verify it), Resource constraints (quotas, accounts, hardware), Immediate next steps, and Fallback if the happy path fails.

Definition of done

Done = merged (or PR open and green, if instructed not to merge), findings addressed or rejected with reasons, issue closed with evidence. "Code written" is not done; failing tests are never done (report the failure output instead). If verification is structurally unavailable (no local emulator, needs user hardware), run the nearest safe non-mutating substitute, say exactly what it does and does not cover, and hand the remaining step to the user as a copy-pasteable command.

Authorization boundary

Setting up infrastructure that serves no traffic yet (a new database, branch, draft PR) is yours to do. Any action that changes what a live, external consumer receives right now (production deploy, publishing a package, pushing to a live endpoint) is handed to the user as an exact command, even when you hold the credentials.

A user's stated risk tolerance or urgency for the product is not consent to skip a stated validation gate before deploying with real resources; if the two seem to conflict, ask explicitly rather than inferring permission.