way-of-working

Post the fresh-session Architect Review a PR's review CI gate requires, verify the gate went green on the head SHA, and file the non-blocking findings. Run in a NEW session that did not author the diff. Never approves, never merges.

Retire a COMPLETED sprint that has passed its HITL Gate and is committed — snapshot its .ai/next-steps.md into .ai/archive/, compact the deep record (move completed narrative and items closed during the sprint, resolved and declined alike, to archive files), advance .ai/state.json to the next sprint, and seed a fresh next-steps.md. Run ONLY on sprint completion; /way-of-working:handoff and /way-of-working:resume never archive.

Run the QA-critic pass on a coding diff — after the local green gate, before /way-of-working:handoff. PROPOSES which read-only critic subagents apply by what the diff touches, and waits for the human to confirm or trim before spawning any — it never auto-fans-out. Aggregates findings for the coder to fix, then re-runs the critics on the fixed tree and iterates to convergence under a bounded stopping rule. Defense-in-depth that runs EARLIER — it is explicitly NOT the repo's review CI gate and never satisfies it.

Serialize the current dev-session state into .ai/ before switching model or session — check the /way-of-working:critic-gate pass ran on any code diff, update .ai/state.json (including hitl_gate, always), regenerate .ai/next-steps.md, and commit/push it as its own docs-only PR (never merged — the human merges). Run this at the END of a session. Does NOT archive a sprint.

Set the live sprint aside: snapshot its cursor as tracked files under .ai/parked/, seed the live cursor for the next sprint or swap in a parked one, and open one docs-only cursor-sync PR. Mechanical — any model. Never merges.

Report the status of a PR's required checks and give an explicit merge-ready verdict WITHOUT merging. Run when asked whether a PR is green / ready to merge, or to poll a PR whose checks are still running. Reports ALL required checks (never a subset), decodes the skipped/BLOCKED/conflict traps, and never merges — the human's merge is the approval.

Rehydrate a fresh dev session from .ai/ externalized state — read the cursor, adopt the assigned persona/model, and state the exact pick-up point. Then start the next_action unattended IF the cursor is clean and unambiguous (hitl_gate NONE OPEN, sprint_status implementing, model matches, no drift); otherwise state the pick-up point and wait. Fails closed — an open, missing, or unreadable gate always waits. Run this at the START of a session working on this repo.

A session-end (or on-demand) retrospective on HOW WE WORKED — mine the live session for friction that RECURRED or cost real rework, then route each finding to its existing home (a backlog item, a memory, or a small skill/doc edit). High signal bar; no speculative suggestions, no new parallel doc, no gratuitous new personas. Run at the end of a working session or when asked to look for workflow improvements. NOT for code diffs (that is /way-of-working:critic-gate) or prose-vs-code drift (that is the docs-consistency subagent).

Close out a finished task — commit the working tree with a conventions-correct message, push to a branch cut from the PR base (never the base itself), and open a PR with a length-checked title. Stops at the open PR — never merges, never --approve, never force-pushes. Run when work is done and you want it on a PR.

Restore a parked sprint from .ai/parked/ to the live cursor behind a hitl_gate that forces re-verification against what merged since, as one docs-only cursor-sync PR. Mechanical — any model may run it. Never merges.