execute-loop

Execute loop

This skill packages CLAUDE.md's Execute loop section for delivery via a plugin. If this project's coordinator uses the file-copy install, the project-root CLAUDE.md already carries this exact content under its own "Execute loop" heading — this skill is a second, parallel delivery path for the same rules, not a replacement. See CLAUDE.md's Role section for the standing boundary this loop runs inside: the coordinator dispatches every build, verify, and investigative step via the Agent tool and never performs one itself, including the rebase and test run that back the push gate below. See CLAUDE.md's Model routing section for which tier ("build" vs "verifier" vs "cheapest") each dispatch below uses — every Agent call sets model explicitly, never inheriting the coordinator's own tier.

Per task pulled from docs/plan.md

1. Spawn a build agent (the build tier per CLAUDE.md's Model routing) with a self-contained prompt carrying acceptance criteria plus the required verification step — see coordinator-kit:verification-standard for what that step actually needs to establish.
2. Spawn an independent verifier agent (the verifier tier) for any non-trivial task — adversarial, not a rubber stamp. It re-derives and re-checks the acceptance criteria; it does not just re-read the build agent's own claims.
3. If verification fails: re-prompt or respawn the build agent with the specific gap. Repeat until acceptance criteria are actually met — but cap it at 2 failed re-prompt/respawn cycles on the same gap. On the 3rd failure on that same gap, stop retrying and escalate instead of continuing to loop (see coordinator-kit:escalation).
4. Update docs/coordination/STATE.md (build → verify → fix → re-verify, commit hashes, disclosed caveats).
5. Commit and push once two conditions both hold — this is the push gate: — The build agent's own report establishes zero new failures versus the base commit on rebased work. Its brief requires it to rebase its work onto latest main, re-run the complete suite on the rebased result (including any DB-gated integration tests against a real local database, no self-skip/mocked mode), and report the actual results: zero new failures versus the base commit, diffing the failure sets and naming the pre-existing failure set in the report (on a fresh greenfield repo that pre-existing set is simply empty). — An independent verifier has passed the acceptance criteria, or the task was exempt under coordinator-kit:verification-standard's non-trivial heuristic (a pure config/copy/comment tweak with no logic or behavior change; when unsure, treat it as non-trivial).