rigortype/rigor/rigor-queue-release

Rigor queue — release pre-clear

Orchestrator-led, multi-turn, in-session queue. Primary entry: user runs pi in the Rigor repo (packages in .pi/settings.json) then /queue-release or /skill:rigor-queue-release. Same project session continues with pi -c.

Not a one-shot architect→lane.

See also: `../../roles/orchestrator.md`, `../../contracts/README.md` (QueueTurn), project agents under .pi/agents/ (rigor-lane, rigor-reviewer).

Requires project package npm:pi-subagents (listed in .pi/settings.json).

Hard rules (always)

• A release goal / milestone / vX.Y.Z wording is NOT authorization to

cut a release. Never seal changelog, bump VERSION / lib/rigor/version.rb, open release/x.y.z, or run /rigor-release-prep unless the user explicitly invoked release-prep.

• When the queue is clear enough, say the cut is one /rigor-release-prep

away — do not start it.

• Lanes: push head SHA and stop; this parent owns CI. No parallel host

make verify. Issues = backlog (ADR-98).

• GitHub API: sparse CI polls (gh pr view --json statusCheckRollup), ≤1/min/PR.
• Never mutate release metadata in this queue.
• Managed worktree fanout needs a clean source checkout (excluding

.pi/subagents/ runtime state). Do not auto-merge worktree patches into master without a human.

Turn protocol (in this same session)

Every turn:

1. Restate queue goal + constraints (target version as context;

non-authorization rule).

1. Show ranked backlog (open Issues) with evidence (gh issue list /

gh issue view). Rank: bugs/blockers > tidy before that cut.

1. Propose exactly one next unit of work (unless the user asked for a

parallel batch — see spawn below).

1. Wait for user: next / do #N / skip / stop

(JA: 次 / やる #N / スキップ / 止めて / ストップ). Also accept spawn / spawn N / 全部やれ / parallel for fanout.

1. On go — prefer pi-subagents spawn (do not tell the user to open

another terminal for run-role.sh when the package is available):

Caveat (docs): managed worktree: true is documented for workflowScript children (runs.run / runs.all), not as a reliable direct { agent, task } knob. Prefer the workflowScript patterns below.

Single lane (next / do #N):

``text subagent({ async: true, worktree: true, workflowScript: return runs.run("lane", { agent: "rigor-lane", task: args.task, worktree: true }) , args: { task: "<LaneInput text>" } }) ``

Optional: pass model: "opencode-go/deepseek-v4.1-flash" (or another resolved Flash-class id) on the outer call / child if the agent frontmatter model does not resolve for the user's providers. Pin via MODEL= / subagents.agentOverrides when unsure.

Parallel batch (spawn N / 全部やれ / parallel): one top-level async workflow that fans out disjoint units:

``text subagent({ async: true, worktree: true, workflowScript: return runs.all([ { key: "i1", agent: "rigor-lane", task: args.t1, worktree: true }, { key: "i2", agent: "rigor-lane", task: args.t2, worktree: true } ]) , args: { t1: "<LaneInput #1>", t2: "<LaneInput #2>" } }) ``

Fallback if pi-subagents / subagent is unavailable: produce LaneInput and tell the user to run ./agents/pi-harness/scripts/run-role.sh lane (separate terminal).

1. After children return: collect head SHAs and handoff manifests

(artifactPaths / handoff JSON). Parent owns CI via sparse polls — do not let children CI-watch or sleep-loop. Refresh queue, print Queue: N remaining | Next candidate: … | say next/stop.

1. Stay in this session — do not require a wrapper restart. Resume later with

pi -c in the same project.

Shared lane traps (feed into LaneInput / spawn)

From docs/notes/20260921-queue-release-lane-experience.md:

• Pin Flash registry id on spawn (devin/swe-2-flash for non-complex work,

opencode-go/deepseek-v4.1-flash or current).

• Child ~30m is enough for code, not for dual-project corpus check twice —

tell lanes to skip full corpus; collect counts in PR body / residual, or run measurement out-of-band.

• Remind --body-file for any gh pr/issue create (parent and children).
• Worktree bundle: main vendor/bundle via untracked .bundle/config.
• Prefer change-named PR branches; changelog after PR number.

Approved gate

After CI is green on a draft PR:

1. Default: spawn co-primaries rigor-reviewer-grok (thinking: xhigh) and

rigor-reviewer-swe2 (devin/swe-2:high). Both must Approved.

1. Complex implementation: also spawn rigor-reviewer-opus (thinking: high);

both must Approved.

1. Complex design: spawn rigor-reviewer (Fable:medium) — reserve for

architecture / API / inference-shape; do not use on routine tidies.

1. On Approved, require pr_body_draft + pr_comment_drafts from the

final reviewer. Apply / post them (or hand to human) before treating the PR as merge-ready — Approved without matching PR text is incomplete.

Never merge from this queue skill; human or a later ship step owns merge.

When to call architect / lane vs stay here

Finish phrases

• After proposing: wait (no terminal finish until stop or empty).
• Queue clear / stop with cut appropriate:

Queue clear for context cut — one /rigor-release-prep away

• Blocked: Blocked — need human
• User stop: Queue paused — resume with pi -c then /queue-release

Non-goals

• Not a release cut; not /rigor-release-prep
• Not sealing changelog or bumping version files
• Not opening release/x.y.z
• Not parallel host make verify or long CI sleep loops
• Not a second backlog outside Issues (ADR-98)
• Not auto-merging managed worktree patches without a human