Deep Review — Parallel Competency-Based Code Review

Inspired by Memento workflow engine's parallel review pattern. Philosophy: one focused expert per domain > one generalist checking everything.

Step 0: Determine scope

BASE=$(gh pr view --json baseRefName -q .baseRefName 2>/dev/null || git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@refs/remotes/origin/@@' || echo "main")
echo "BASE: $BASE"
BASE_REF=$(git rev-parse --verify --quiet "refs/remotes/origin/$BASE" 2>/dev/null || git rev-parse --verify --quiet "$BASE" 2>/dev/null || true)
if [ -z "$BASE_REF" ]; then
  echo "No locally available base ref for $BASE; cannot perform a read-only diff." >&2
  exit 2
fi
echo "BASE_REF: $BASE_REF"
echo "=== DIFF STATS ==="
git diff "$BASE_REF" --stat
echo "=== CHANGED FILES ==="
git diff "$BASE_REF" --name-only
echo "=== DIFF SIZE ==="
git diff "$BASE_REF" --shortstat

Store the BASE branch name, resolved local base ref, and list of changed files. Report-only review does not run git fetch, pull, or another command that updates Git state. If the accepted task explicitly requires a fresh remote base and permits that preparation mutation, run git fetch origin "$BASE" --quiet without suppressing failure before resolving BASE_REF; otherwise state that the review used the locally available ref.

If there is no diff, stop: "Nothing to review — no changes against $BASE."

Step 1: Scoping — select relevant competencies

Based on the changed files, select ONLY the competencies that are relevant. Do NOT run all 8 for a 3-file CSS change.

Default to a review-only task. Enter review-and-fix only when the user clearly authorizes both review and implementation (for example, "review and fix"). In either mode, the review phase remains read-only; review-and-fix continues with the approved remediation in this same task after findings are triaged.

Competency selection rules

Select every materially relevant competency and no others. Diff size is a signal to inspect scope, not a reason to add unrelated reviewers. A single competency is valid when the changed risk has one material dimension.

Output the selected competencies with a one-line justification each:

Selected competencies (4 of 8):
  ✓ security — auth middleware modified
  ✓ database — new migration + 3 query changes
  ✓ concurrency — BullMQ worker modified
  ✓ error-handling — 4 new try/catch blocks in API routes
  ✗ architecture — no new modules, existing patterns
  ✗ performance — no hot paths touched
  ✗ frontend — no UI changes
  ✗ testing — test files updated alongside logic

Step 2: Get the diff content

git diff "$BASE_REF"

Read the full diff. You need it to construct focused prompts for each competency agent.

Also identify which files are relevant to each selected competency — each agent should receive ONLY the files relevant to its domain, not the entire diff.

Step 3: Launch parallel competency reviews

For each selected competency, launch an Agent tool call in parallel. Each agent:

• Gets ONLY the files relevant to its competency
• Has a focused checklist (from the competency definitions below)
• Returns findings in structured format
• Works in isolated context (no bias from other competency reviews)

CRITICAL: Launch ALL agents in a SINGLE message (parallel tool calls). Do NOT launch them sequentially.

If a reviewer must run tests, give it an isolated writable worktree and a unique writable scratch directory. Capture git status --porcelain=v1, git diff --exit-code, and git diff --cached --exit-code before and after the review. Do not place a test-running reviewer in an OS-level read-only sandbox: pytest and similar tools legitimately create temporary files. If only a strict read-only sandbox is available, provide immutable test receipts for inspection and say that the reviewer did not execute the tests.

Agent prompt template

For each competency, use this prompt structure (fill in {COMPETENCY}, {CHECKLIST}, {FILES}):

You are a {COMPETENCY} specialist reviewing code changes. Your ONLY job is {COMPETENCY} — ignore everything else.

## Review authority and writable scratch
Repository root: {REPO_ROOT}
Scratch directory: {REVIEW_SCRATCH}

You are a reviewer, not an implementer. Do not create, modify, delete, format,
or install anything in {REPO_ROOT}; do not alter Git or worktree state. If a
review command genuinely needs a file, write only under {REVIEW_SCRATCH}.
Return findings in the required response format; do not create a project report
or patch. If tests cannot run without writing elsewhere, inspect the supplied
immutable test receipts and label execution NOT_RUN.

## Changed files to review
{list each relevant file path}

Read each file listed above using the Read tool. Then apply this checklist:

## Checklist
{CHECKLIST from competency definitions below}

## Output format
For EACH finding, output exactly:

FINDING: {one-line description}
FILE: {path}:{line}
SEVERITY: CRITICAL | HIGH | MEDIUM | LOW
EVIDENCE: {quote the problematic code, max 3 lines}
FIX: {concrete fix suggestion, not vague advice}
CONFIDENCE: HIGH | MEDIUM | LOW

If you find NOTHING — output: "NO_FINDINGS — {COMPETENCY} review clean."

Do NOT pad with compliments. Do NOT report things that are fine. Only problems.

## Admission — screen each candidate BEFORE you publish it

A suspicion is a candidate, not a finding. Put each one through these five before it
reaches the report, and drop it silently if it fails any:

- **Authority** — does a real contract, spec, invariant or accepted requirement say
  this is wrong? "I would have written it differently" is not authority.
- **Reachability** — can the bad path actually be reached by a supported input or
  state? Show how. Unreachable code is at most a tidiness note.
- **Materiality** — if it happens, what breaks for a user or an operator? If the
  answer is "nothing observable", it is not a finding at this severity.
- **Evidence** — can you quote the lines that prove it, without inference? If your
  proof is "this pattern is usually a bug", you have a hypothesis.
- **Remedy cost** — is the fix smaller than the harm? A finding whose remedy is a
  rewrite needs the harm to justify a rewrite.

Report the count you dropped and why, in one line: "screened out 4: 2 unreachable,
1 no authority, 1 immaterial". That line is the evidence you screened at all.

The failure this prevents is specific and tempting. A candidate that generalises your
finding to a second case makes the report sound weightier, and nobody downstream will
check the second case. Verifying it and dropping it costs you one paragraph; leaving
it in costs the reader their trust in the other findings.

Competency checklists

Use the checklist content from the files in competencies/ directory. Read each relevant file before constructing the agent prompt.

If competency files don't exist yet, use the inline checklists below:

security:

• SQL/NoSQL injection (parameterized queries?)
• XSS (user input escaped in output?)
• Auth bypass (middleware on all protected routes?)
• CSRF (tokens validated?)
• Secrets in code (API keys, passwords, tokens hardcoded?)
• Trust boundary violations (LLM/external output used unsanitized?)
• Path traversal (user-controlled file paths?)
• Rate limiting on sensitive endpoints?
• CORS configuration (too permissive?)
• JWT validation (expiry, algorithm, issuer checked?)

performance:

• N+1 queries (loop with DB call inside?)
• Missing indexes on filtered/joined columns?
• Unbounded queries (no LIMIT on user-facing endpoints?)
• Memory leaks (event listeners not cleaned up? growing collections?)
• Bundle impact (new dependencies? tree-shaking?)
• Caching opportunities missed?
• Expensive operations in hot paths?
• Pagination on list endpoints?
• Image/asset optimization?

architecture:

• Separation of concerns (business logic in controllers?)
• Circular dependencies between modules?
• God objects/functions (>200 lines, >5 responsibilities?)
• Abstraction level consistency (mixing HTTP and business logic?)
• Interface contracts (types/schemas for cross-module communication?)
• DRY violations (copy-pasted logic that should be shared?)
• Layer violations (UI calling DB directly?)
• Configuration vs hardcoding?

database:

• Migration safety (reversible? data-preserving? locks?)
• Query performance (JOINs, subqueries, full table scans?)
• Transaction boundaries (operations that should be atomic?)
• Data integrity (constraints, foreign keys, NOT NULL where needed?)
• Index coverage for new queries?
• Connection management (pool exhaustion risk?)
• Schema evolution (backward compatible?)
• Default values for new columns?

concurrency:

• Race conditions (read-modify-write without lock?)
• Deadlock potential (multiple locks in inconsistent order?)
• Idempotency (can the operation be safely retried?)
• Queue handling (ack/nack, retry policy, DLQ?)
• Shared state access (multiple workers writing same resource?)
• Distributed locking (Redis/DB advisory locks where needed?)
• Timeout handling on external calls?
• Graceful shutdown (in-flight requests drained?)

error-handling:

• Swallowed errors (empty catch blocks?)
• Error propagation (errors from dependencies surfaced correctly?)
• User-facing error messages (informative but not leaking internals?)
• Partial failure handling (what if step 3 of 5 fails?)
• Retry logic (exponential backoff? max retries? circuit breaker?)
• Cleanup on failure (resources released? partial writes rolled back?)
• Validation at boundaries (API input, file uploads, webhook payloads?)
• Logging on error paths (enough context to debug? not too noisy?)

frontend:

• Reactivity correctness (Vue: ref vs reactive, computed vs watch?)
• Component responsibility (>300 lines = split candidate?)
• Accessibility (ARIA labels, keyboard nav, contrast?)
• Loading/error states handled?
• Memory leaks (subscriptions, intervals, event listeners cleaned up?)
• i18n (hardcoded strings?)
• Responsive design (mobile breakpoints?)
• XSS in template rendering (v-html with user data?)

testing:

• Logic changed without test update?
• Happy path only (no error/edge case tests?)
• Mocking too much (testing mocks instead of behavior?)
• Test isolation (tests depend on execution order?)
• Assertions quality (testing implementation details vs behavior?)
• Missing integration tests for cross-module changes?
• Flaky patterns (timing, network, random data?)
• Coverage gaps for critical paths (auth, payments, data integrity?)

Step 4: Collect and synthesize

After ALL agents complete, collect their findings and synthesize:

4a: Deduplicate

Multiple competency reviewers may find the same issue. Group findings by file:line and merge duplicates. When merged, note which competencies flagged it (higher confidence when multiple domains agree).

4b: Triage each finding

For each unique finding, assign a triage:

4c: Output the synthesis report

══════════════════════════════════════════════════
  DEEP REVIEW: {branch} → {base}
  {N} findings across {M} competencies
══════════════════════════════════════════════════

COMPETENCIES REVIEWED: {list with ✓}
DIFF SIZE: {N insertions, M deletions, K files}

── FIX ({count}) ─────────────────────────────────
  1. [CRITICAL/security+concurrency] file:line
     Problem: {description}
     Evidence: {code quote}
     Fix: {concrete suggestion}

  2. [HIGH/database] file:line
     ...

── DEFER ({count}) ──────────────────────────────
  3. [MEDIUM/performance] file:line
     Problem: {description}
     Why defer: {rationale}

── ACCEPT ({count}) ─────────────────────────────
  4. [LOW/architecture] file:line
     Finding: {description}
     Why accept: {rationale}

══════════════════════════════════════════════════
  VERDICT: {PASS / PASS_WITH_CAVEATS / NEEDS_FIXES}
══════════════════════════════════════════════════

Step 5: Continue only for explicit review-and-fix

For review-only requests, stop after the synthesis report. Do not edit source, format files, alter Git state, install dependencies, or create a patch; report which FIX findings must be resolved before merge.

For an explicit review-and-fix request, continue in this same task after the read-only synthesis. Preserve the selected scope and verify each remediation with the smallest causal check:

For each FIX finding:

1. If the fix is mechanical (add missing await, add index, add LIMIT, fix typo) — apply it directly. Output: [FIXED] file:line — {what}
2. If the fix requires judgment — use the available evidence to choose and apply the best reversible fix within the already authorized task, then run its causal check. Judgment alone does not require another approval. Ask only when required authority is actually missing or a material user choice cannot be resolved from the request and available evidence; state that exact boundary.

After all FIX items are resolved, output final status:

DEEP REVIEW COMPLETE:
  {N} findings total
  {X} auto-fixed
  {Y} fixed with user input
  {Z} deferred
  {W} accepted

Gotchas

• Bounded delegation: competency agents normally use inline tools to keep review focused. This is a workflow convention, not a universal tool limitation. A bounded child task is allowed when the current host supports it and the task's governing instructions authorize delegation.
• Context size: each agent gets focused file list, not full diff. If a competency touches >20 files, prioritize the most critical ones and note "N additional files not reviewed."
• False positives: parallel agents don't share context, so they may flag things that are addressed in other files. The synthesis step (4a-4c) catches these through cross-referencing.
• Cost: launching 5 parallel agents costs ~5x a single-pass review. This is the trade-off for depth. For quick checks use /review instead.
• Timing: parallel agents complete at different speeds. Wait for ALL before synthesizing.
• Read-only is a role, not a filesystem: reviewers do not edit source. A reviewer asked to execute tests still needs isolated writable scratch and a writable runtime.

Troubleshooting

• pytest reports no usable temporary directory or Access is denied before collection: the reviewer was launched with a contradictory execution contract. Relaunch it in an isolated writable worktree with {REVIEW_SCRATCH}, or have it inspect immutable test receipts and report NOT_RUN; never count the failed command as test evidence.
• Reviewer may have changed the repository: compare the before/after git status --porcelain=v1, git diff --exit-code, and git diff --cached --exit-code receipts. Any new tracked/staged change invalidates the review until reconciled.

When to use /deep-review vs /review