Review gauntlet (serial)

Run the reviewers one after another on the live branch, each the sole editor while it runs — every reviewer reads a settled tree and applies its own fixes. Why serial, and the incidents behind these rules: RATIONALE.md — read when editing this skill, not running it.

1. architecture-first-principles — FIRST, for a diff touching framework packages (@kolu/surface*), adding/reshaping module structure, or whose correctness rests on a concurrency/ordering claim (a race declared closed, an asserted happens-before, a flag one callback sets and another path reads as truth). Skip ONLY for pure-docs or trivially-local diffs — a diff leaning on a happens-before is not trivially-local. Say so either way. Confirmed findings are dispositioned like any stage's: fix now or record where, never "acceptable for scope".
2. /lens-debate — lowy ∥ hickey as parallel subagents, one reconcile-and-apply pass commits the fixes. Pass the rationale.
3. /agent-debate — the selected peer debates to consensus; author rounds commit fix(…).
4. /simplify — the self-applying reuse/simplification pass.
5. code-police — rules + fact-check passes, applying fixes.

PR comments come after the push, never before. Each step commits locally; be-review pushes once after all steps, then posts. The debate skills run with --no-comment and leave their bodies on disk. No PR comment may reference a local-only commit.

Preflight

• Non-empty diff vs <base> (default: the repo default branch). Empty → stop. Reviewers review committed code — commit/stash outstanding work.
• Verify <base> is a remote-tracking ref, mechanically. A caller's --base master names the local master — routinely tens of commits stale, swelling the diff with history nobody wrote. After git fetch origin: git -C "$repoPath" rev-parse --verify --quiet "refs/remotes/$base"; on failure retry as origin/$base; if that fails too, stop and say so.
• Resolve the scope once: MB=$(git merge-base <base> HEAD), START=$(git rev-parse HEAD). Pass MB to every step so each reviews against the identical fork point (later reviewers seeing earlier reviewers' fixes is intended).
• Pin repoPath — the repo under review may not be the cwd (a /be run can carry a companion repo). Thread it into every step and subagent prompt with absolute paths and git -C "$repoPath". A cross-repo step returning clean against a non-empty target diff means repoPath didn't take effect.
• Debate peer (unless --tracks excludes debate): require --agent <claude|codex|grok> and run its auth preflight. On failure, name the login command or exact error and stop — never skip a mandatory track or substitute a fallback peer.

Run the steps

--tracks checks,lens,debate,simplify,police selects steps (default all), in order, each to completion. Preflight already fetched and resolved the base — don't redo it per step. Both review steps run inline in your own turn; act on real signals (a subagent returning, a peer ping), never timers — no ScheduleWakeup polling.

1. lens — /lens-debate with repoPath, base = MB, --no-comment, and the rationale. It commits agreed fixes and leaves its comment at $repoPath/.lens-debate/comment.md — post that file after the push, don't improvise a summary. Statuses: clean/applied (settled); needs-human (adjudicate each finding yourself — decide drop or apply, apply survivors, fold the adjudication into the deferred comment; never report it as settled); merge-base-error (report and move on).
2. debate — capture DEBATE_START=$(git -C "$repoPath" rev-parse HEAD), then invoke /agent-debate: review --agent <selected> --repo "$repoPath" --base MB --no-comment --context <context> --rationale <rationale> (no reasoning-effort flag — the peer runs at its CLI default). When an API-facing shared-surface change trips the drishti companion-repo gate, note in --rationale that the gate defers to §ship (it can only be checked against final post-gauntlet HEAD). On consensus, assemble the deferred comment at $repoPath/.agent-debate/comment.md: peer ⇄ author header, round count (from the verdict-*.json files), a commit table of $DEBATE_START..HEAD (or "peer approved after author disputes" when empty), and a legend of the findings raised (unique by id across the verdicts). Fill both agent names from the actual harness and peer — never hardcode. On reviewer-error/merge-base-error, post no false-consensus body; report the failure and move on.
3. simplify — /simplify scoped to MB; commit what it changed (refactor: simplify <area>, staging only its files).
4. police — /code-police with --no-elegance whenever simplify ran (its elegance pass would re-invoke /simplify over an already-simplified tree); omit the flag only when --tracks excluded simplify. Tell its passes to scope to MB — their default origin/HEAD...HEAD diff is wrong whenever --base isn't the repo default. Apply and commit each fix (fix(police): <title>).

Push, then comment

Settle whether anything is new: git log --oneline $START..HEAD. Then:

• just fmt before any push with new commits — no reviewer guarantees formatting, just check never runs the formatter, and an unformatted tree reds ci::fmt later. Commit any reformat (style: just fmt).
• New commits + a PR exists → git push; only after it succeeds, post the deferred comments (their SHAs are now remote).
• No new commits but a PR exists → post the comments immediately.
• No PR → nothing to push or comment on; findings live in chat and the log.
• A required push fails → do not post; report the failure.

Post one comment per track that produced a body: the lens body, the agent-debate body (gh pr comment -F), and the code-police summary (## [👮 Code-police](https://agency.srid.ca/)). Never merge.

Report

Summarize in chat, covering only the selected tracks and naming any skipped track explicitly: lens status + adjudications; debate peer + consensus/reviewer-error; what simplify changed; police findings and their disposition; whether fixes were pushed; git log --oneline <base>..HEAD + git diff --stat <base>.

ARGUMENTS: