Patch release 1.0.32: multi-reviewer simplicity checks land in /implement and /execute-plan, /review and /pr-babysit gain stack and ship workflows, and /design / /learn pick up publish and unattended-run support.
Changelog
[new]/design --publish — After the write–review loop finishes, optionally publish the final design doc (Notion page first; secret gist if Notion is unavailable). Local file stays the source of truth; publish is opt-in only.
[new]Simplicity specialist on every /implement and /execute-plan review — A dedicated simplicity reviewer is always pinned (YAGNI, one-caller helpers, unused abstractions). Effort 1 is now general + simplicity; higher effort still auto-selects security, tests, and plan-alignment specialists and can run up to 7 reviewers per PR.
[changed]/execute-plan --effort range is 1–5 — Per-PR multi-reviewer fan-out matches /implement (individual review files, merge, specialization tags, non-fatal specialist failures).
[new]/review modes and modifiers — Main mode (whole branch vs base on a clean tree), stack mode (one reviewer per PR in a Graphite/stacked chain), --submit (publish as COMMENT), and --approve (publish + approve). Bare /review uses local mode when the tree is dirty and main mode when it is clean. Default PR reviews stay PENDING.
[new]/pr-babysit optional ship cycle — add --ship, ship on|off, and ship can land merge-ready PRs bottom-up when every job has passed and no human or bot comments remain open. Shipping is opt-in; the fix path still never merges by default.
[changed]/pr-babysit review handling — Treats bot/automated findings (Bugbot, Copilot, CodeRabbit, Semgrep, Dependabot, etc.) like human threads: fix or reply substantively, then resolve. No per-cycle fix cap; failed attempts recover the worktree safely and retry next cycle.
[changed]/learn for unattended and safer curation — Headless//loop runs default to report mode with --since-last, block on state.py wait so the workflow finishes, present the report before any picker, skip retire actions on narrow windows (disuse_evidence), and group same-disposition unused surfaces into one decision.
[fix]PR watcher transport errors — watch_pr.py retries timeouts, resets, DNS, and gateway errors with backoff and jitter instead of treating them as PR events.
[changed]Subagent message rows — Transcript shows one-line Message rows for steers/queues/interjects (expand for full text); plan-mode prompt accepts complete pager commands without treating them as revision notes.
[changed]Marketplace plugins — Updated modern-web-guidance, pstack (setup-pstack description now covers reasoning budget), and vercel (vercel-functions description).
Under the hood
Optional request compression path (GROK_REQUEST_COMPRESSION) and related sampler support.