Swarm Review

Adversarial Reviewer — Tiered

Thin dispatch layer. Phase plans live in sibling tier files (tier-low.md, tier-high.md, tier-max.md). This file parse args, resolve baseline, classify diff (classify.md), resolve overlays (overlays.md), optionally gate on meta-plan approval, then hand off to matching tier file. Shared content (worker table, adversarial protocol, output format, constraints) stay here — perspective-by- perspective execution lives in tier files.

Argument syntax

/swarm-review [tier] <target> [--flags]

• tier (optional): low | auto | high | max. Default auto.
• target (optional): branch name; PR ref (<N> / #<N> / PR <N> / pull/<N> / full GitHub PR URL); empty → current branch HEAD.
• flags (convention: flags before positional):
  • --base=<git-ref> — diff baseline; default main. When target resolves to PR, inferred from gh pr view --json baseRefName unless user pass --base explicit.
  • --breadth=minimal|full|adversarial — Stage 2 perspective breadth
  • --reviewer=haiku|sonnet|opus — model for worker-reviewer
  • --doc-reviewer=haiku|sonnet — model for worker-doc-reviewer
  • --rca=on|off — Five Whys systemic-fix analysis
  • --codex / --no-codex — cross-model adversarial pass (code-diff scope)
  • --codex-model=luna|terra|sol — Codex model for cross-model pass (default per tier: high=terra, max=sol; luna if forced at low)
  • --dry-run — meta-plan preview via EnterPlanMode
  • --form — structured AskUserQuestion form; implies --dry-run

--base is pipeline input, not axis. Feeds classifier as diff baseline for whole run. Overlays (breadth, rca, codex) are single-axis pipeline adjustments on top of chosen tier.

Workflow

1. Parse arguments and resolve target

Detect target (ordered, first match wins):

1. Full GitHub PR URL
2. PR <N> / pull/<N> → PR
3. #<N> or bare integer → probe PR via mcp__github__pull_request_read
4. Branch name — verify via git rev-parse --verify <name>
5. Empty → current branch HEAD

PR metadata (labels, base ref, title, body) feeds both baseline resolution and classifier. On fetch fail, fall back to free-text treatment (match /swarm-plan convention).

2. Resolve the baseline

• User pass --base=<ref> → use verbatim (user wins).
• Else target is PR → use gh pr view <N> --json baseRefName -q .baseRefName.
• Else default main.

Fast-fail when git rev-parse --verify <base> fails — print remediation (did you mean origin/main?) and stop. Empty diff → report "nothing to review" and exit clean.

3. Classify (only when tier=auto)

Read classify.md. Compute diff metrics once (git diff <base>...HEAD --name-only for files, --shortstat for lines). Apply tier signals (size + subsystems + structural markers + PR labels). Produce candidate tier + confidence flag + overlay set. When signals split across adjacent tiers, mark low-confidence — meta-plan gate in step 5 resolve it. No mid-flow AskUserQuestion.

4. Resolve overlays

Final config = tier defaults (overlays.md per-tier table) + classifier overlays + user flag overrides. User flags always win.

5. Meta-plan gate (single consolidated approval point)

Fire when ANY of: --dry-run, --form, tier resolved to max, or low-confidence classification. Only user-prompt point.

Write .claude/state/plans/meta-plan_review_[target].md with: Classification (tier + rationale + diff metrics snapshot), Baseline (source: user-flag / PR-base / default), Overlays (with rationale), Workers per perspective, Estimated cost, Not Doing (no auto-fixes, no commits).

Approval UI (always single interaction):

• Default: EnterPlanMode with meta-plan path; resume on approve. Fall back to AskUserQuestion Approve/Edit/Cancel if skill resume after ExitPlanMode unreliable in practice.
• --form: ONE AskUserQuestion call with ≤4 batched axis questions (Tier / Breadth / RCA / Codex), first option "Recommended".

On reject: re-draft with rejection rationale and re-present once.

6. Announce final config (always)

Swarm review
  Tier:       high                             (auto — 8 files, 240 lines, 1 subsystem)
  Baseline:   main                             (default)
  Target:     HEAD (branch: soraka)
  Overlays:   breadth=full, rca=on, codex=off  (tier default)
  Workers:    Stage 1 (2 parallel), Stage 2 (4 parallel)
  Codex diff review: off
  Proceed? (Ctrl+C to abort; re-run with explicit tier to override)

7. Dispatch to tier file

Read matching tier-{low,high,max}.md and execute its phase plan.

Worker assignment (shared across tiers)

See .claude/rules/workflow-swarm.md for worker types, models, tools, focus modes. Tier files select subset of these perspectives.

Max concurrent workers: 8 (per workflow-swarm.md).

Adversarial protocol & output format

Adversarial questioning anchors ("What if this assumption is wrong?", "Under what conditions would this fail?", "What edge cases weren't considered?") apply at every tier. Root-cause Five-Whys analysis is tier-gated via rca axis (see overlays.md).

Output skeleton — tier files add or remove Stage 2 sections per breadth they run. Every tier produces:

## Code Review: [target]

### Summary

- Verdict: Approved | Needs Work | Request Changes
- Tier / Baseline / Diff: N files, +L / -L lines, S subsystems

### Stage 1 — Correctness (spec-compliance, test-coverage)

### Stage 2 — [perspectives run at this tier]

### Cross-Model Adversarial (Codex) # if --codex fired

### Root-Cause Analysis # if rca=on

### Deferred Findings (human judgment required)

Verdict: Approve (Block/High resolved or deferred with reasoning), Needs Work (Warn-tier present), Request Changes (unresolved Block-tier / security / breaking changes / missing tests / arch violations).

Constraints

• NO auto-fixing — review read-only; actionable findings reported, not committed
• NO approving with unresolved Block-tier findings
• NO nitpicking style covered by ESLint/Prettier
• NO mid-flow AskUserQuestion during classification
• NO exceeding 8 parallel workers
• ALWAYS reference specific files and lines
• ALWAYS suggest alternatives, not just problems
• ALWAYS classify every finding as actionable or deferred
• ALWAYS stay within diff scope (<base>...HEAD)

Plan Status block (mutate on round entry + verdict)

Mutate the plan's ## Status block: flip Step to /swarm-review → round N on each round entry; on verdict, set awaiting /finalize (Approve) or awaiting /swarm-execute (review-fix loop) (Needs Work / Request Changes); bump Last update. Skip silently when no current_plan.md / no Status block. Full mutation table → plan-status.md.

Handoff

• To /swarm-execute: when actionable findings exist and user want them fixed. /swarm-execute runs Review-Fix Loop.
• To /builder: with remediation task list for specific finding.
• To /doc-writer: with gap report from worker-doc-reviewer.
• To /architect: for architectural concerns requiring ADR.

$ARGUMENTS