Crit

Crit

Fast adversarial review: cleanup pre-pass (reuse / quality / efficiency), then 2 parallel reviewers — one for correctness/security, one adversarial. Covers the highest-value dimensions without the overhead of 5-7 specialists.

NEVER review inline. Always dispatch subagents via the Agent tool.

Step 1: Scope

Resolve BASE: gh stack view --json 2>/dev/null | jq -r '.trunk // empty' || gt parent 2>/dev/null || gt trunk 2>/dev/null || git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's|refs/remotes/||'. Args override.

Bugfix detection: If commit messages contain "fix"/"bugfix"/"hotfix", classify files as production vs test. ALL test-only → verdict FAIL with Critical: "Bugfix contains no production code changes."

Step 2: Cleanup Pre-pass

ALWAYS run unless #<PR> input. Spawn 3 agents in ONE message, each with the full diff. Wait for all three, aggregate findings, and apply each fix directly. Skip false positives without arguing, then proceed to Step 3.

Agent A — Code Reuse

You are reviewing a diff for missed opportunities to reuse existing code.

## Focus
1. **Existing utilities and helpers**: search utility directories, shared modules, and files adjacent to the changed ones. Flag any new function that duplicates existing functionality — name the existing function to use instead.
2. **Inline logic that could use an existing utility**: hand-rolled string manipulation, manual path handling, custom environment checks, ad-hoc type guards.

Output: table with File:Line | Issue | Replacement. Brief summary.

Agent B — Code Quality

You are reviewing a diff for hacky patterns.

## Focus
1. **Redundant state**: duplicates existing state, cached values that could be derived, observers/effects that could be direct calls.
2. **Parameter sprawl**: new parameters added instead of generalizing or restructuring existing ones.
3. **Copy-paste with slight variation**: near-duplicate blocks that should unify behind a shared abstraction.
4. **Leaky abstractions**: internal details exposed, existing abstraction boundaries broken.
5. **Stringly-typed code**: raw strings where constants, enums (string unions), or branded types already exist.
6. **Unnecessary JSX nesting**: wrapper Boxes/elements that add no layout value — check if inner component props (flexShrink, alignItems, etc.) already provide the needed behavior.
7. **Unnecessary comments**: comments explaining WHAT, narrating the change, or referencing the task/caller — keep only non-obvious WHY.

Output: table with File:Line | Issue | Fix. Brief summary.

Agent C — Efficiency

You are reviewing a diff for efficiency problems.

## Focus
1. **Unnecessary work**: redundant computations, repeated file reads, duplicate network/API calls, N+1 patterns.
2. **Missed concurrency**: independent operations run sequentially when they could run in parallel.
3. **Hot-path bloat**: new blocking work added to startup or per-request/per-render hot paths.
4. **Recurring no-op updates**: state/store updates inside polling loops, intervals, or event handlers that fire unconditionally — add a change-detection guard. If a wrapper takes an updater/reducer callback, verify it honors same-reference returns so callers' early-return no-ops aren't silently defeated.
5. **Unnecessary existence checks**: pre-checking file/resource existence before operating (TOCTOU) — operate directly and handle the error.
6. **Memory**: unbounded data structures, missing cleanup, event listener leaks.
7. **Overly broad operations**: reading entire files when only a portion is needed, loading all items when filtering for one.

Output: table with File:Line | Issue | Fix. Brief summary.

Step 3: Context

ct tool gitcontext --base $BASE --stat → diff-stat, changed-files, log. Fetch PR context via gh pr view if available.

Step 4: Dispatch 2 Reviewers

Spawn both in ONE message. Pass raw diffs, not summaries.

Large diffs (>3000 lines): Truncate files with >200 lines of diff to first 50 + last 50 lines.

Reviewer 1 — Correctness & Security

You are an adversarial correctness and security reviewer.

## Gather Context
1. Run: `ct tool gitcontext --base {base_ref} --format json`
2. Read all changed files from the output
3. If `truncated_files` is non-empty, `Read` those files in full

## Assumption Verification (do this BEFORE reviewing)

1. **Boundary semantics**: When code branches on a field from an external system, verify what it actually represents by reading the source definition.
2. **Value correctness across boundaries**: Trace every value crossing a system boundary from producer to consumer. Check tuple/struct destructuring.
3. **Error fallback safety**: Is the default safe? Silent fallback to production URL or permissive auth can be worse than crashing.
4. **Completeness of external interactions**: Paginated/batched APIs — verify all pages handled.
5. **Existing pattern divergence**: Flag reimplementations of existing utilities.
6. **Multi-driver/adapter symmetry**: Verify patterns applied consistently across all changed files.

## Focus
- Edge cases (empty, null, overflow, concurrent access)
- Invalid states, race conditions, resource leaks
- Silent failures, swallowed errors, dangerous fallbacks
- Off-by-one, logic inversions
- Injection, auth/authz gaps, data exposure
- Missing tests for new/changed behavior
- Error type conflation (catch-all handlers losing specificity)
- Input validation gaps

Classify each finding:
- FIX: correctness bugs, security issues, test gaps
- IGNORE: style, subjective, out-of-scope tech debt

Tier: critical | notable | nitpick

Output: table with Tier | Disposition | File:Line | Issue | Suggestion
Then brief summary.

Reviewer 2 — Devil's Advocate

You are an adversarial devil's advocate reviewer. Try to break everything.

## Gather Context
1. Run: `ct tool gitcontext --base {base_ref} --format json`
2. Read all changed files from the output
3. If `truncated_files` is non-empty, `Read` those files in full

## Focus
- **Failure modes**: What happens when dependencies fail?
- **Bad assumptions**: What does this code assume that might not hold?
- **Silent contract changes**: When behavior changes, check all callers.
- **Race conditions**: Trace full execution paths for async/concurrent code.
- **Adversarial input**: Malformed, enormous, deeply nested, special chars.
- **Premise check**: Does the fix actually fix the stated problem?
- **Approach risks**: Solving the right problem?
- **Assumption inversion**: What does each filter/guard INCORRECTLY exclude?
- **Silent data loss**: Operations suppressed during certain states.
- **Over-engineering**: Abstractions with <3 call sites, "might need it later" scaffolding.
- **Architecture**: Incomplete refactors, coupling, unnecessary indirection.
- **Performance**: O(n^2) loops, unbounded growth, N+1 queries, hot-path waste.

Classify each finding:
- FIX: correctness bugs, security issues, test gaps
- IGNORE: style, subjective, out-of-scope tech debt

Tier: critical | notable | nitpick

Output: table with Tier | Disposition | File:Line | Issue | Suggestion
Then brief summary.

Step 5: Consolidate

1. Validate: Spot-check 1-2 claims per reviewer. Read actual code at file:line. Prune false positives aggressively.
2. Deduplicate: Same issue → highest severity.
3. Consensus: Critical from either reviewer survives. Non-critical flagged by both → keep. Flagged by one → keep if confirmed by spot-check, otherwise IGNORE.

Output # Review Summary:

• FIX table: Tier | File:Line | Finding | Recommendation
• IGNORE section (collapsed): findings below threshold
• Verdict: PASS (no FIX items) | CHANGES_REQUESTED (any FIX) | FAIL (any Critical)

Store the review via blueprint_create { kind: "review", topic: "Review: <branch>" }, Edit the scaffold with the findings body, then blueprint_commit.

Step 6: Fix

--auto critical|high|medium|all → auto-fix at or above the given severity. No --auto → ask: Fix all / Fix critical+high / Fix critical only / Skip.

Spawn fix agent → fix, verify, self-check (remove debug artifacts, unused imports), report. Single pass — no re-review loop (use /superreview --loop for iterative fixing).

Step 7: Summary

Output: Fixes Applied, Ignored, Remaining. Suggest /superreview if the user wants deeper analysis.