deep-review

Deep analysis of a focused area. You are Opus analyzing directly — no delegation, no team. The value is YOUR cross-domain reasoning across all related files in one context.

ultrathink

Pre-flight

1. Validate argument — $ARGUMENTS is REQUIRED. If empty, STOP: "Please specify a target area to review. Example: /deep-review settings page"
2. Verify Linear MCP — Call mcp__linear__list_teams. If unavailable, STOP: "Linear MCP is not connected. Run /mcp to reconnect, then re-run."
3. Read CLAUDE.md — Load project rules, conventions, and accepted patterns. Extract the team name from the LINEAR INTEGRATION section (look for "Team:" line). If CLAUDE.md doesn't have a LINEAR INTEGRATION section, use the team name from mcp__linear__list_teams (there should only be one team in most cases).
4. Query existing Backlog issues — mcp__linear__list_issues with discovered team name, state "Backlog". Record titles and file paths to avoid creating duplicates.

Scope Discovery

Trace the full dependency graph for $ARGUMENTS. Find EVERY file that participates in the target feature.

Step 1: Find entry points

Read CLAUDE.md's STRUCTURE section to discover file patterns. Use Glob to locate primary files matching $ARGUMENTS based on the discovered patterns. Common patterns include:

• Pages/routes (e.g., src/app/**/page.tsx, src/routes/**/*.ts)
• Components (e.g., src/components/**/*.tsx, src/ui/**/*.jsx)
• Hooks (e.g., src/hooks/**/*.ts)
• API routes (e.g., src/app/api/**/*.ts, api/**/*.ts)

If the argument is ambiguous, use Grep to search for the feature name across the codebase.

Step 2: Trace imports

Read each entry point. For every import:

• Component imports → add the component file
• Hook imports → add the hook file
• Lib/utility imports → add the lib module
• Type imports → add the type file
• Database imports → flag if imported from unexpected locations (check CLAUDE.md for DB access conventions)

Follow imports recursively until the full dependency tree is mapped.

Step 3: Map data flows

For any API calls, data store access, or external service interactions:

• Extract the data operation (API path, query, service call)
• Find the corresponding handler or data source (consult CLAUDE.md for patterns)
• Add the handler/source AND its dependencies
• Trace the data flow from source through processing layers to consumer

Step 4: Find related files

For each source file:

• Test file: check for colocated __tests__/ or *.test.ts patterns (consult CLAUDE.md)
• Loading/skeleton files: check for related UI state files

Step 5: Output file manifest

List all discovered files grouped by role:

• Pages/Layouts — entry point routes
• Components — UI components
• Hooks — custom hooks
• API Routes — server-side handlers
• Lib Modules — business logic
• Types — shared type definitions
• Tests — test files
• Config — loading, middleware, etc.

Output the manifest before proceeding so the user can verify scope.

Read All Files

Read EVERY file in the manifest. Cross-domain reasoning requires holding all related code in context simultaneously.

Read in dependency order: types → lib → hooks → API routes → components → pages → tests.

Deep Analysis

Read references/deep-review-checklist.md for the comprehensive cross-domain checklist.

Critical instruction: Do NOT analyze each file in isolation. Reason about how files INTERACT. For each finding, trace the impact across the full stack.

Analysis approach

Walk through the user's journey for this feature:

1. Arrival — User navigates to the screen. What loads? What can fail during load?
2. Interaction — User takes actions. What state changes? What API calls fire? What race conditions exist?
3. Server processing — What validates input? What can fail? How do errors propagate back?
4. AI integration — If the feature involves Claude API: trace input preparation → API call → tool loop → response validation → client rendering. Check tool definitions, system prompts, agentic loop correctness, and response validation (see checklist section 9).
5. UI update — Does the component handle all response shapes? All error types? Loading states?
6. Edge cases — Empty data, concurrent actions, stale cache, network failure, large payloads, back/forward navigation

Findings format

For each finding, record:

• Severity: [critical] | [high] | [medium] | [low]
• Domain: code | security | ux | performance
• Location: file:line (and related files if cross-cutting)
• Description: What's wrong
• Impact: Who is affected and how
• Cross-domain note: How this interacts with other parts of the system (if applicable)

Create Linear Issues

For each finding, check against existing Backlog issues. Skip if a matching issue already exists.

Use mcp__linear__create_issue:

team: "<discovered-team-name>"
state: "Backlog"
title: "[Brief description]"
priority: [1-4] (critical=1, high=2, medium=3, low=4)
labels: [mapped label]
description: (format below)

Issue description format:

**Problem:**
[1-2 sentence problem statement]

**Context:**
[File paths with line numbers — include ALL related files, not just the primary location]

**Impact:**
[Who is affected and how — trace the user-facing consequence]

**Acceptance Criteria:**
- [ ] [Verifiable criterion]
- [ ] [Another criterion]

Label mapping:

| Finding domain | Linear label | |---------------|-------------| | code (bugs, logic, async, edge cases) | Bug | | security (auth, validation, XSS) | Security | | ux (loading, feedback, accessibility, responsive) | Improvement | | performance (re-renders, bundle, images, CWV) | Performance | | convention (CLAUDE.md violation) | Convention |

Termination

Output this report and STOP:

## Deep Review: $ARGUMENTS

**Files analyzed:** N
**Scope:** [list of file groups with counts]

### Findings (ordered by severity)

| # | ID | Severity | Domain | Title |
|---|-----|----------|--------|-------|
| 1 | PROJ-XX | Critical | code | Brief title |
| 2 | PROJ-XX | High | ux | Brief title |
| ... | ... | ... | ... | ... |

X issues created | Duplicates skipped: N

### Cross-cutting Observations

[Systemic patterns observed across the feature — e.g., "error handling is inconsistent between the API route and client component" or "loading states exist but don't cover all async paths"]

Do not ask follow-up questions. Do not offer to fix issues.