Use when finishing a branch or PR after verification and review, including merge, PR publication, keep, discard, cleanup, local sync, and reporting merge or PR URLs with current git state.
Use when fixing or investigating crashes, regressions, flakes, hangs, incorrect outputs, or incidents.
Use when work is complex, long-running, multi-step, cross-boundary, likely to span multiple PRs/sessions, or needs handoff-ready planning under plans/.
Use at a stable checkpoint or closeout after rollback, abandoned/rejected approaches, material rerouting, repeated failure, two+ materially different attempts, a rejected completion claim, or an explicit lessons-learned request; routes learning to task,…
Preflight, AGENTS.md, agent context, skill routing, test routing, subagent handoff, prompt caching readiness. Use before long-running, multi-agent, unfamiliar, high-risk, or cross-service coding tasks. Do not use for small one-file edits.
Use when more than five experiment results have accumulated without synthesis, at a research decision point, or before promoting a research finding into a delivery path. Reads the experiment ledger for knowledge deltas — supported, falsified, still-open — to…
Use for any delivery-mode task that changes code or tests. Routes implementation work by risk and applicable branch skills before editing.
Use for delivery-mode tasks whose purpose is quality improvement on shipped code: closing high-risk coverage gaps, structure-budget findings, flaky tests, missing observability, or smell density on a named path. Sequences a measured baseline, risk-tiered…