Closes #1180

The self-review loop still charged every PR a fixed minimum of three sequential reviewer stages after #1159 and #1173: a deep-review round 1 (7 to 24 agents, whatever the PR touched), a forced second round even when deep-review confirmed nothing, and a standalone prose pass. This shortens the loop while keeping the adversarial depth where a bug actually ships into end-user apps.

What changed

In .claude/skills/webjs-start-work/SKILL.md, ### How the loop works:

• Round 1 is tiered by the diff's paths, a path check, never a judgment call. A diff touching shipped source (packages/*/src/, packages/cli/lib/, packages/cli/templates/) takes the deep-review workflow as round 1, unchanged. Every other diff (docs, website, blog, skills, hooks, tests-only, examples) takes ONE broad fresh reviewer using the delta-verifier template with the full diff as its question. A mixed diff takes the deep tier (one shipped-source path is enough), the user asking for a deep review always forces it, and a scope change that pulls shipped source into a light-tier PR re-runs round 1 at the deep tier. The tier is path-based for the same reason the substantive gate is surface-based: membership is checkable from the diff, so the agent motivated to go light cannot misjudge it. This supersedes docs: make the self-review loop fast, parallel, and stall-proof #1159's uniform-round-1 rule at the owner's direction; the misjudgment concern that rule addressed stays closed because the tier reads paths, not importance.
• The minimum is ONE round. A round 1 of either tier that finds nothing substantive converges the loop by itself: on the deep tier its findings were already adversarially jury-refuted, and on the light tier nothing in the diff is shipped source, so the old two-round minimum added a round to exactly the cleanest case. A fix still always buys a fresh delta round.
• The converging round's prose-tier findings replace the standalone prose pass. Every reviewer already reports and tags prose findings alongside substantive ones, so the round that finds nothing substantive IS the prose sweep: its prose fixes apply without re-review and the loop ends. The gate's exception survives: a prose fix that touches a substantive surface gets one delta-scoped check of that fix alone.
• The refuter gate, liveness and watchdog rules, worktree safety, posting mechanics, the substantive gate and its asymmetric tag verification, the 5-round budget, the BLOCKED and OVER BUDGET exits, the deferral ledger, and the deferred suites plus CI read after convergence are all untouched except where they cross-referenced the removed pass or minimum.

In .claude/workflows/deep-review.js: default maxAgents drops 24 to 16 (still clamped 8 to 60, still overridable per run), dynamic lenses cap at three instead of six (scout prompt, schema maxItems, both slices), and whenToUse scopes round 1 to shipped-source diffs. The six fixed lenses, the model split, the jury sizes, and every fail-open path are unchanged.

In .claude/hooks/route-skills.sh: the code-review directive now states the tiered round 1 and the new exit, staying in lockstep with the skill.

In test/hooks/review-loop-exit.test.mjs: the guard assertions pinning the prose pass and the two-round minimum are replaced by assertions pinning the tier rule, the one-round minimum, the converging-round prose application, and the trimmed workflow defaults, so a revert of any piece reds the test.

What a loop costs now

A docs or skill PR: one broad reviewer when clean, plus a delta round per fix. A shipped-source PR: deep-review at 16 agents, plus a delta round per fix. Nothing pays a forced extra round or a separate prose reviewer anymore.

Test plan

• node --test test/hooks/review-loop-exit.test.mjs test/hooks/route-skills.test.mjs: 20/20 pass (the review-loop guard, and the router test including "every skill the hook can route to is committed in-repo").
• Counterfactual: the guard test's new assertions fail against the pre-change files (they pin The minimum is ONE round, the tier bullet, the 16-agent default, and the absence of the standalone prose pass in either spelling). Verified whole-file against the parent commit and hunk-level for the skill's failure-handling hunk and the hook's comment hunk, both red when reverted alone.
• deep-review.js validity in the async workflow context is covered by the existing guard test ("stays valid in the async workflow context").
• Both copies of the skill (in-repo and ~/.claude) verified byte-identical with diff -q after the change.

Doc surfaces

Tests: the guard tests above. Markdown / docs site / scaffold / MCP / editor plugins / marketing / dogfood apps / version bumps: N/A because nothing in packages/*/src or any user-facing surface changed; this edits the Claude Code skill, the committed review workflow, and the routing hook, which ship to agents working this repo, not to anything the framework publishes.