Code Review

Code Review - Iterative Review

An iterative, human-in-loop code review skill that identifies gaps, presents them for user selection, and fixes them in cycles until clean.

Review Modes

Two modes. Pick based on user intent — detail for each mode is covered in Step 0 below.

• PR Review Mode (default): diff-based review against a base branch
• Holistic Review Mode: audit an entire codebase or scoped directories

Subagent Mode (argument)

The skill accepts one optional argument via $ARGUMENTS controlling whether scans/fixes delegate to Task subagents:

Default: auto. If $ARGUMENTS is empty or not recognized, operate in auto mode.

auto decision heuristics:

• Holistic scan over >50 files or multiple ecosystem packages → delegate to Explore subagent
• Fix touches >2 files or needs cross-cutting refactor → delegate to general-purpose subagent
• Small PR diff / single-file scope / known surface → stay in main context
• Tight budget / urgency signal from user → stay in main context

Respect the argument throughout the session. Once subagent or no-subagent is set, apply it consistently across Step 0, Step 4, and Step 5. Do not flip regime mid-run.

Subagent choice by task type (for subagent and auto when delegating):

• Read-only scans / research → Explore subagent (optimized, read-only)
• Multi-file fixes / write work → general-purpose subagent (or a domain-specific subagent if one is available)

Validation (Step 2g) always runs in the main context regardless of this argument — it directly reads files to confirm agent-reported gaps. no-subagent does not disable validation; it disables the upstream delegation that produces the gaps.

Workflow Overview

┌──────────────────────────────────────────────┐
│  1. Gather input                             │
│     PR mode: diff (base vs HEAD)             │
│     Holistic mode: scope from ecosystem      │
└──────────────┬───────────────────────────────┘
               ▼
┌──────────────────────────────────────────────┐
│  2. Review with context → 2a..2g             │
│     (business, diff, tech, integrations,     │
│      dep grounding, identify, validate)      │
└──────────────┬───────────────────────────────┘
               ▼
┌──────────────────────────────────────────────┐
│  3. Present validated gap list               │
│     AskUserQuestion — 2 options:             │
│       · Apply all fixes                      │
│       · Done — looks good                    │
│     (free-form "Other" → selective fix)      │
└──────────────┬───────────────────────────────┘
               ▼
      ┌────────┴─────────────┬──────────────┐
      │                      │              │
   "Done"          "Apply all fixes"    "Other" subset
      │                      │              │
      ▼                      ▼              ▼
   [EXIT]         ┌──────────────────────────────┐
                  │  4. Apply Fixes              │
                  │     Pre-gate checklist       │
                  │     ├─ Tracked Fix Loop      │
                  │     │  (all, or subset ≥3)   │
                  │     └─ Selective Fix         │
                  │        (subset 1-2)          │
                  │     Failure → Retry/Skip/    │
                  │              Abort           │
                  └──────────┬───────────────────┘
                             ▼
                  ┌──────────────────────────────┐
                  │  4.5 Next-Pass Gate          │
                  │      AskUserQuestion:        │
                  │        · Run next pass       │
                  │        · Stop — I'm done     │
                  └────┬─────────────────┬───────┘
                       │                 │
                  "Run next pass"   "Stop"
                       │                 │
                       ▼                 ▼
            ┌──────────────────────┐  ┌────────┐
            │ 5. Re-review → Loop  │  │  EXIT  │
            │    to Step 3         │  │        │
            └──────────────────────┘  └────────┘

Instructions

Step 0: Determine Review Mode

First, determine whether this is a PR Review or Holistic Review:

PR Review (user mentions PR, branch, changes, diff):

• Proceed to Step 1 to get the diff
• Focus on changed code + related context

Holistic Review (user mentions codebase, audit, project review):

• Skip Step 1 (no diff needed)
• Detect the project's source root from ecosystem signals before defaulting to src/:
  • Node/TS: package.json → check main/exports/workspaces; fall back to src/, app/, packages/*/src/
  • Python: pyproject.toml → [tool.*] or [project].packages; fall back to top-level package dirs
  • Go: go.mod → module root; conventional dirs cmd/, internal/, pkg/
  • Java/Kotlin: pom.xml or build.gradle(.kts) → src/main/java, src/main/kotlin
  • Rust: Cargo.toml → src/, workspace members
  • Monorepo: iterate over workspace members
• If ambiguous, ask the user
• Use Glob to enumerate relevant files per ecosystem
• Subagent usage depends on $ARGUMENTS (see "Subagent Mode" section above):
  • subagent → always delegate area scans to Explore subagents in parallel
  • no-subagent → enumerate and read in main context, no Task delegation
  • auto (default) → delegate to Explore when scope is large (>50 files or multi-package), otherwise stay in-context

For holistic reviews, use this workflow:

1. Detect scope from ecosystem signals (manifest files), or ask user
2. Use Glob to list relevant files per ecosystem
3. For large codebases, chunk into areas (api/, components/, utils/, cmd/, pkg/, etc.)
4. Delegate area scans to Explore subagents in parallel when independent
5. Validate reported gaps (Step 2g) — never trust agent output blindly
6. Present all validated gaps
7. Fix selected gaps
8. Re-review affected areas

Step 1: Get the PR Diff (PR Review Mode Only)

First, detect the current branch, resolve repo metadata, and get the diff. Prefer gh when the branch is already pushed as a PR, fall back to raw git otherwise.

Always re-check the current branch — do not rely on conversation memory, the branch may have changed:

git branch --show-current

Resolve repo (when gh is available):

REPO=$(gh repo view --json nameWithOwner --jq '.nameWithOwner' 2>/dev/null)
# e.g., "owner/repo" — use for any subsequent gh api calls

If the branch has an open PR, prefer gh for changed files and base branch:

# PR number for current branch (empty if no PR)
PR=$(gh pr view --json number --jq '.number' 2>/dev/null)

# Base branch from PR metadata
BASE=$(gh pr view --json baseRefName --jq '.baseRefName' 2>/dev/null)

# Changed files (robust in monorepos — first path segment often indicates package/module)
gh pr view "$PR" --json files --jq '.files[].path'

Otherwise, detect base branch from git:

# 1. Upstream tracking branch, if set
git rev-parse --abbrev-ref --symbolic-full-name @{u} 2>/dev/null

# 2. Default branch from remote (origin/HEAD)
git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@'

# 3. Fallback: try common defaults
for b in main master develop trunk; do git show-ref --verify --quiet "refs/heads/$b" && echo "$b" && break; done

Store the resolved base in a variable, then diff. Do not hardcode main:

BASE=<detected-base>
git diff "${BASE}...HEAD"

If detection is ambiguous or fails, ask the user.

Step 2: Review with Context

IMPORTANT: Do not blindly review the diff in isolation. Always gather context - both technical AND business logic.

2a. Understand the Business & Project Context

Before diving into code, understand what this code is supposed to do AND what conventions it must follow:

1. Read documentation: Check README, AGENTS.md, inline comments, PR description
2. Read per-package conventions (monorepos): Each package/service may have its own AGENTS.md, CLAUDE.md, CONTRIBUTING.md — read the one closest to the changed files, not only the repo root
3. Read linter/formatter configs before flagging style or convention gaps: .editorconfig, .eslintrc*, .prettierrc*, ruff.toml / [tool.ruff] in pyproject.toml, .rubocop.yml, checkstyle.xml, golangci.yml, biome.json, .swiftlint.yml, etc. Respect what the project already enforces — do not flag rules the config explicitly disables
4. Understand the feature: What user problem does this solve? What's the expected behavior?
5. Identify business rules: Are there validation rules, workflows, or domain constraints?
6. Check edge cases: What happens with invalid input? Empty states? Concurrent access?

Ask yourself:

• Does this code correctly implement the intended business logic?
• Are there missing business rules or edge cases?
• Would a domain expert find issues with this implementation?

2b. Analyze the Diff

Understand what changed technically:

• New functions/classes added
• Modified logic
• Changed function signatures
• New dependencies

2c. Gather Technical Context

For each changed file, investigate:

1. Callers: Who calls this function? Are they passing correct arguments?
2. Dependencies: What does this code depend on? Are those contracts met?
3. Types: Are type definitions correct? Any missing optional parameters?
4. Related files: Are there related files that should be updated together?
5. Data flow: How does data flow through this code? Any transformations that could corrupt data?

Monorepo / multi-package awareness:

When the diff touches multiple packages, services, or modules (detected from distinct first-segment paths in the changed-files list, or from workspace members in the manifest), investigate per-package. Each package may pin different versions, follow different conventions, and have its own AGENTS.md / CLAUDE.md. Do not apply global assumptions across packages.

Use Grep and Read tools to explore:

# Find callers of a changed function
grep -r "functionName(" --include="*.ts"

# Check type definitions
grep -r "interface.*TypeName" --include="*.ts"

# Find related business logic
grep -r "orderStatus\|paymentState" --include="*.ts"

2d. Trace External Integrations (End-to-End)

When code connects to external systems (frontend ↔ backend ↔ database ↔ third-party APIs), trace the full integration path. Verify contracts, error handling, auth, and pagination match across boundaries — contract drift between tiers is a common source of silent bugs.

See references/integration-tracing.md for per-boundary checklists (FE→BE, BE→DB, BE→external API), an 8-item watch list (contract/type/auth/pagination mismatches), and a worked trace example.

2e. Dependency & Compatibility Grounding

Internal knowledge about package versions, APIs, and deprecations is often outdated. When the diff touches a dependency manifest, introduces a third-party import, or uses a version-sensitive API, ground against the exact pinned version in the manifest/lockfile — not the latest release, not general knowledge.

Trigger when: manifest changed, new third-party import, or version-sensitive API usage (deprecated methods, renamed exports, signature changes).

Workflow (brief): Read manifest → WebSearch for breaking changes / deprecations / CVEs on that version → WebFetch official changelog or migration guide → cross-check usage against the version's actual API surface.

See references/dependency-grounding.md for the stack → manifest lookup table (Node, Python, Go, Java, Kotlin, Rust, Ruby, PHP, Elixir + runtime pinning files), the full grounding workflow, what-to-flag checklist, and an example.

Label findings under the Dependency Compatibility category (see Step 2f and gap-categories.md §3).

2f. Identify Gaps

Look for gaps across ALL categories:

1. Business Logic - Missing validations, incorrect workflows, wrong domain rules, edge cases not handled
2. Integration Issues - Contract mismatches, missing fields, type mismatches between frontend/backend/database/external APIs
3. Dependency Compatibility - Usage of APIs not available in pinned version, missed breaking changes on bump, peer-dep mismatches, runtime incompatibility, pinned version with known CVE (see Step 2e)
4. Logic Errors - Bugs, incorrect conditions, off-by-one errors, race conditions
5. Security Issues - Injection vulnerabilities, auth gaps, data exposure
6. Performance Problems - N+1 queries, unnecessary loops, memory leaks
7. Error Handling - Missing try/catch, unhandled edge cases, silent failures
8. Code Style - Inconsistent naming, formatting, patterns
9. Missing Tests - Untested code paths, edge cases
10. Documentation Gaps - Missing comments for complex logic, outdated docs

Include related pre-existing issues if they:

• Affect the correctness of the new changes
• Are in callers that need to be updated (e.g., missing new required parameter)
• Create security/performance issues when combined with new code

For each gap found, note:

• File and line number
• Category
• Brief description
• Severity (high/medium/low)
• Whether it's in the diff or related code

See references/gap-categories.md for detailed definitions.

2g. Validate Gaps Before Presenting

CRITICAL: Before presenting any gap to the user, you MUST validate it by reading the actual code.

When using Task agents (Explore or general-purpose) to scan different parts of a codebase in parallel, agents may report false positives or misunderstand context. Never trust agent claims blindly.

Validation Process:

For each gap reported (by you or by a sub-agent):

1. Read the actual file at the reported line number using the Read tool
2. Verify the issue exists - confirm the code actually has the problem described
3. Check the context - ensure the surrounding code doesn't already handle the case
4. Confirm line numbers - agent-reported line numbers may be approximate or outdated

Discard gaps that:

• Reference files or lines that don't exist
• Describe code that doesn't match reality when you read it
• Are already handled by surrounding context the agent missed
• Are duplicates of other gaps (same issue, different wording)

Example Validation:

Agent reports: "[HIGH] src/auth.ts:45 - SQL injection in query"

Validation steps:
1. Read src/auth.ts lines 40-50
2. Check: Is there actually a query at line 45?
3. Check: Does it use string concatenation with user input?
4. Check: Is it already using parameterized queries?

If the code uses parameterized queries → DISCARD the gap
If the code has injection vulnerability → KEEP the gap

Only present validated gaps to the user.

Step 3: Present Gaps to User

Use the AskUserQuestion tool to present gaps as a numbered list with selection options. Invoke the tool verbatim — do not paraphrase the question as prose in chat. The structured panel must appear; a plain-text question is a bug, not an alternative.

First, display the gaps found:

## PR Review - Iteration 1

Found 4 gaps in your PR:

1. [HIGH] src/auth.ts:45 - Security: SQL injection vulnerability in query
2. [MED] src/api.ts:123 - Error Handling: Missing try/catch around API call
3. [MED] src/utils.ts:67 - Logic: Off-by-one error in loop condition
4. [LOW] src/components/Button.tsx:12 - Style: Inconsistent naming convention

Then call the AskUserQuestion tool with only two options. The "Other" free-form box is provided automatically by the tool and is the fallback for any subset or filter the user wants. Do not substitute a chat prompt for this tool call.

Pluralization at runtime — substitute N with the actual count and adjust wording for singular vs plural:

Template below uses the N ≥ 2 form. Adjust wording for N = 1 before calling the tool.

{
  "questions": [
    {
      "question": "Apply fixes for all N validated gaps?",
      "header": "Fix gaps",
      "multiSelect": false,
      "options": [
        {
          "label": "Apply all fixes",
          "description": "Fix all N gaps sequentially, track progress, stop on failure"
        },
        {
          "label": "Done — looks good",
          "description": "Exit without fixing"
        }
      ]
    }
  ]
}

Fallback: Interpreting free-form input

When the user types into the "Other" box instead of picking an option, parse the intent:

Any subset interpretation routes to the selective fix path in Step 4. Only the clean "Apply all fixes" choice enters the tracked fix loop.

Step 4: Apply Fixes

Step 4 has two execution paths depending on the Step 3 outcome:

• Tracked fix loop (primary) — user picked "Apply all fixes"
• Selective fix (fallback) — user typed a subset/filter in "Other"

Both paths share the pre-gate checklist below.

Pre-gate Checklist

Before any edit, verify all of the following. If any item fails, do not start; surface the problem to the user and wait.

1. Validation gate — every selected gap has passed Step 2g (Validate Gaps). Never fix an un-validated gap.
2. File existence — every referenced file:line exists in the current working tree. Drop or re-validate any that don't.
3. Working tree clean (or explicitly approved) — run git status. If dirty with unrelated changes, invoke the AskUserQuestion tool (verbatim — not prose) with three options: Stash unrelated changes, Commit unrelated changes first, Abort. Do not silently mix unrelated work into review fixes.
4. Working context confirmed:
   • PR mode: git branch --show-current still matches the branch being reviewed
   • Holistic mode: the intended scope path is still the one to edit

Tracked Fix Loop (primary path)

Triggered when the user picks "Apply all fixes". This path guarantees:

• Exhaustive — every validated gap gets fixed, no re-prompt per item
• Sequential — one gap at a time, finish + verify before moving to the next
• Tracked — every gap becomes a visible task the user can watch
• Verified — a task is marked completed only after a post-fix check confirms the change took effect
• Fail-stop — on failure, the loop halts; nothing is silently skipped

Protocol:

1. Create tasks up front. Before applying any edit, TaskCreate one task per selected gap:
   • subject: Fix [SEVERITY] <file:line> — <category>
   • activeForm: Fixing <file:line>
2. Process sequentially. For each gap, in the order presented:
   1. TaskUpdate status=in_progress
   2. Read the file around the reported location
   3. Apply the fix using the Edit tool (or delegate a complex multi-file fix to a Task subagent — see below)
   4. Post-fix verify — re-read the edited region and confirm the change actually addresses the gap. For type/compile-sensitive edits, run a quick check if available (tsc --noEmit, go build, python -m py_compile, linter, etc.).
   5. Only on successful verification: TaskUpdate status=completed
3. On failure — apply the Failure & Resumption Protocol below. Do not continue to the next gap.
4. After the loop — summarize fixes made, then go to the Next-Pass Gate below before running Step 5.

Delegating complex fixes — subject to $ARGUMENTS (see "Subagent Mode"):

• subagent → delegate any multi-file or complex fix to general-purpose
• no-subagent → apply all fixes directly in main context with Read + Edit
• auto (default) → delegate only when the fix touches >2 files or requires cross-cutting refactor

When delegating, the Task call happens inside step 2.3 above; the outer task bookkeeping (create → in_progress → verify → completed) still applies:

Task tool parameters:
- subagent_type: "general-purpose"
- description: "Fix SQL injection in auth.ts"
- prompt: "Fix the SQL injection vulnerability in src/auth.ts at line 45.
          Use parameterized queries instead of string concatenation.
          Only modify what's necessary to fix this specific issue."

Selective Fix (fallback path)

Triggered when the user typed a subset/filter into "Other" (e.g. 1,3, only HIGH, skip #2).

• If the resolved subset has ≥3 gaps — run the same Tracked Fix Loop above, scoped to the subset.
• If the subset has 1–2 gaps — lightweight loop: pre-gate → Read → Edit → verify → summarize. No TaskCreate overhead.

Otherwise the rules are identical: pre-gated, sequentially applied, verified before marking done, and fail-stop on error.

Failure & Resumption Protocol

If a fix can't complete (Edit fails, verification fails, unexpected state, subagent error): halt the loop, TaskCreate a blocker task, surface the failure to the user, and ask via AskUserQuestion to Retry / Skip / Abort.

Core rules: never silently skip a failure; completed means verified; use deleted (with a reason) for explicit skips.

See references/failure-resumption.md for the full protocol (when to trigger, task-state transitions, anti-patterns).

Step 4.5: Next-Pass Gate

After fixes land, do NOT auto-run Step 5. Ask the user first whether to spend another review pass. A second pass often surfaces new gaps (regressions, issues revealed by the fix, adjacent code the fix touched) — but it's not always wanted. Let the user decide.

Invoke the AskUserQuestion tool (verbatim — must appear as a structured panel, not a prose question):

{
  "questions": [
    {
      "question": "Run another review pass to check for new or remaining gaps?",
      "header": "Next pass",
      "multiSelect": false,
      "options": [
        {
          "label": "Run next pass",
          "description": "Re-review the changed code (Step 5) and surface any new gaps"
        },
        {
          "label": "Stop — I'm done",
          "description": "Exit the review with a summary of what was fixed"
        }
      ]
    }
  ]
}

Routing:

• Run next pass → proceed to Step 5 (Re-review)
• Stop — I'm done → proceed to Step 6 (Exit) with summary
• Free-form in "Other" — parse intent: "once more" / "yes" → Run; "enough" / "no" / "that's fine" → Stop; anything ambiguous → confirm with a short follow-up

Also skip the gate and go straight to Step 6 when:

• No fixes were applied in Step 4 (user picked "Done" earlier, or subset resolved to empty) — there's nothing new to re-review
• All gaps were deleted via the Failure & Resumption Skip option and nothing actually landed

Step 5: Re-review After Fixes

Task list handling across iterations:

• Leave tasks from the previous iteration intact as an audit trail (completed, deleted, and any in_progress blockers). Do not delete completed tasks between iterations.
• For each new gap found in this iteration, create a new task via TaskCreate. Do not reuse or relabel previous tasks.
• If a new iteration re-surfaces a gap that was deleted (skipped) earlier, treat it as a new task and let the user decide again.

After applying fixes, re-review based on the mode:

PR Review Mode:

1. Run git diff "${BASE}...HEAD" again (using the base detected in Step 1) to get the updated diff
2. Re-analyze the diff for:
   • Remaining unfixed gaps
   • New issues introduced by fixes
   • Regression in previously clean code
3. Present the updated gap list

Holistic Review Mode:

1. Re-read the files that were modified during fixes
2. Re-analyze those files plus any files they interact with. Delegation follows $ARGUMENTS (see "Subagent Mode"): subagent → Explore for re-scan; no-subagent → read directly; auto → Explore only when the re-scan is large, direct reads otherwise
3. Check if fixes introduced new issues or broke related code
4. Re-validate any agent-reported gaps (Step 2g) before presenting
5. Present the updated gap list for the affected areas

Step 6: Iterate or Exit

The review loop continues until:

• User says "done": Respect the user's decision to stop
• No gaps remain: Congratulate and summarize changes made
• User explicitly cancels: Stop immediately

When exiting, provide a summary:

## Review Complete

Iterations: 3
Gaps fixed: 7
Gaps skipped: 2 (user choice)

Changes made:
- src/auth.ts: Fixed SQL injection, added input validation
- src/api.ts: Added error handling
- src/utils.ts: Fixed off-by-one error

Output Behavior

• Do not narrate workflow steps to the user. Avoid "Step 1: ...", "Step 2: ..." announcements. Work like a colleague reviewing code, not a robot reading a checklist.
• First user-visible output should be the gap list, not chatter about fetching the diff, resolving the base branch, or detecting the stack. Keep internal plumbing silent.
• Only surface results, decisions, and questions that need user input.

Anti-Patterns to Avoid

1. Blind diff review - Don't review the diff in isolation. Always gather context: check callers, verify dependencies, trace data flow
2. Ignoring related issues - Pre-existing issues that affect the changes ARE in scope. Flag callers missing new parameters, contracts that are now violated, etc.
3. Over-fixing - Do not exceed the scope the user approved — either all gaps (from "Apply all fixes") or the subset parsed from their free-form input. Never fold in adjacent cleanups, related refactors, or new issues noticed mid-fix unless the user approves them.
4. Skipping the Next-Pass Gate - Never auto-run Step 5 (Re-review) after fixes land. Always ask the user first via the Next-Pass Gate whether to spend another pass. If they say stop, stop.
5. Silent fixes - Always explain what was changed and why
6. Trusting agent claims blindly - When using Task agents for parallel analysis, ALWAYS validate their reported gaps by reading the actual code before presenting to user. Agents can hallucinate issues or misunderstand context.
7. Grounding against latest instead of pinned version - Never assume library behavior from the latest release or general knowledge. Always verify against the exact version resolved in the manifest/lockfile for the specific package being reviewed. Behavior in v5.x may not exist in v4.x.
8. Applying global assumptions in a monorepo - Different packages can pin different versions and enforce different conventions. Ground per-package, not per-repo.
9. Mixing unrelated work into review fixes - If git status is dirty with unrelated changes, resolve that first. Don't commit review fixes alongside in-flight work.
10. Silent batch fixes - For batch work (Tracked Fix Loop, or any selective run ≥3 gaps), always surface progress via TaskCreate/TaskUpdate. No invisible mass edits.
11. Marking a task completed before verification - completed requires a post-fix check confirming the change addresses the gap. Edit tool success alone is not completion.
12. Silently skipping a failed fix - Failures route through the Failure & Resumption Protocol. Never drop a gap without the user deciding Retry / Skip / Abort.
13. Starting fix work without the pre-gate checklist - Validation gate, file existence, clean working tree, and confirmed working context must all pass before any edit.
14. Paraphrasing AskUserQuestion as a prose question - Wherever this skill says to present choices (Step 3 fix menu, Step 4 dirty-tree pre-gate, Step 4.5 Next-Pass Gate, Failure & Resumption Retry/Skip/Abort), invoke the AskUserQuestion tool verbatim so the structured panel renders. A natural-language "Apply all fixes or done?" in chat is a bug — the user should see the panel, not plain text.

Example Session

User: review my PR

[Internal: resolve base branch, fetch diff, gather context, validate gaps]

Claude: ## PR Review - Iteration 1

Found 3 gaps in your PR:

1. [HIGH] src/auth.ts:45 - Security: Password stored in plain text
2. [MED] src/api.ts:89 - Error Handling: Unhandled promise rejection
3. [LOW] src/utils.ts:12 - Style: Unused import

[AskUserQuestion: "Apply fixes for all 3 validated gaps?"
 Options: Apply all fixes / Done — looks good]

User: Apply all fixes

[Internal pre-gate: ✓ validated, ✓ files exist, ✓ clean tree, ✓ branch confirmed]

[TaskCreate × 3]
  #1 pending  — Fix [HIGH] src/auth.ts:45 — Security
  #2 pending  — Fix [MED] src/api.ts:89 — Error Handling
  #3 pending  — Fix [LOW] src/utils.ts:12 — Style

[#1 in_progress → Read → Edit → verify (re-read confirms bcrypt hashing) → completed]
[#2 in_progress → Read → Edit → verify (try/catch wraps await) → completed]
[#3 in_progress → Read → Edit → verify (import removed) → completed]

Claude: Fixed 3/3 gaps.

[AskUserQuestion: "Run another review pass to check for new or remaining gaps?"
 Options: Run next pass / Stop — I'm done]

User: Run next pass

[Internal: re-run diff, re-validate]

Claude: ## PR Review - Iteration 2

Found 1 gap:

1. [LOW] src/auth.ts:48 - Style: Magic number should be a constant

[AskUserQuestion: "Apply the fix for the 1 validated gap?"
 Options: Apply all fixes / Done — looks good]

User: Done — looks good

Claude: ## Review Complete

Iterations: 2
Gaps fixed: 3
Gaps skipped: 1 (user choice)

Changes made:
- src/auth.ts: Added password hashing with bcrypt
- src/api.ts: Added try/catch with proper error handling
- src/utils.ts: Removed unused import

Variant — free-form subset via "Other":

[AskUserQuestion: "Apply fixes for all 3 validated gaps?"]

User (types in Other): skip #3

Claude: Interpreted as: fix gaps #1 and #2, skip #3.

[Selective fix — 2 gaps, below ≥3 threshold, lightweight loop]
[#1 Read → Edit → verify]
[#2 Read → Edit → verify]

Claude: Fixed 2/3 gaps (skipped #3 per your request). Re-reviewing...