Skill

deep_review

One-shot adversarial review - three parallel subagents analyze independently, orchestrator synthesizes. No debate between agents. Fast and cost-effective.

Invocation

How this skill is triggered — by the user, by Claude, or both

Slash command

/deep-review:deep_review

User invocable

Model invocable

Inline context

Default effort

Context Preview

The summary Claude sees in its skill listing — used to decide when to auto-load this skill

Perform thorough code review using three parallel agents with opposing mindsets.

SKILL.md

303 lines · ~2.3k tokens

Stats

LanguageJavaScript

Parent stars0

MaintenanceGood

Last CommitMar 12, 2026

Actions

View Source View Plugin View on GitHub View README

Stats

Actions

Deep Review

Perform thorough code review using three parallel agents with opposing mindsets.

Why This Works

Single-reviewer blind spots miss issues. Three perspectives create productive tension:

Advocate - "Why is this correct?" Trust boundaries, design rationale, false-positive defense
Skeptic - "How can I break this?" Bugs, edge cases, code smells that indicate bugs
Architect - "Is this the right direction?" System impact, scope vs correctness, structural smells

Their disagreement surfaces issues. Their agreement signals confidence.

Usage

/deep_review                                        # Review uncommitted local changes
/deep_review <PR-ID>                                # Review a pull request
/deep_review <commit>                               # Review a specific commit
/deep_review <file1> <file2>                        # Review specific files
/deep_review --base=main --head=feat/PROJ-123       # Review branch diff

Workflow

Phase 1: GATHER CONTEXT (you, the orchestrator)
    ├── Collect change information
    ├── Fetch file contents
    └── Write to /tmp/deep_review-${CLAUDE_SESSION_ID}-context.yaml

Phase 2: PARALLEL ANALYSIS (three agents)
    ├── Verify context file exists
    ├── Spawn three agents in parallel
    └── Each reads context file independently

Phase 3: SYNTHESIS (you, the orchestrator)
    ├── Reconcile into final review
    └── Delete context file (always)

Phase 1: Gather Context

1.1 Identify the Changes

Determine what's being reviewed based on arguments:

--base and --head flags → Branch diff: git diff {base}...{head} and git log {base}...{head}
PR/MR identifier → Remote pull request
Commit hash → Specific commit
File paths → Named files
No arguments → Local uncommitted changes

Use project-specific tools if available (check project CLAUDE.md). Fall back to git commands: git diff, git show, git log.

1.2 Fetch Content

For each changed file, get the new version (what's being reviewed).

Local files: Read directly
Remote PRs: Use project-specific fetch scripts or APIs

1.3 Write Context File

Write the context file to: /tmp/deep_review-${CLAUDE_SESSION_ID}-context.yaml

Use /tmp in bash - it maps to the system temp directory on all platforms. Do NOT create temp files in the source directory being reviewed.

review:
  type: pr | local | commit | files | branch_diff
  id: <identifier if applicable>
  base: <base branch, if branch_diff>
  head: <head branch, if branch_diff>
  title: <summary>
  description: <details>

changed_files:
  - path: <relative path>
    action: add | edit | delete | rename
    content: |
      <full file content - new version>

existing_feedback:               # optional
  - author: <who>
    location: <file:line or general>
    comment: <text>

observations:                    # your notes
  - <anything notable>

Large changes: If total content exceeds ~50KB, summarize less-critical files or split into multiple reviews.

1.4 Verify Context File

Before Phase 2, confirm the file exists:

test -f "$CONTEXT_FILE" && echo "Ready" || echo "ERROR: Context file missing"

Do NOT spawn agents if the context file doesn't exist.

Phase 2: Parallel Analysis

Spawn THREE agents in a SINGLE message using the Task tool. Agents do NOT inherit conversation history - the context file is their only input.

2.1 Agent Prompts

Agent instructions: agents/

Find the plugin path first: **/deep_review/agents/advocate.agent.md

Advocate:

subagent_type: deep-review:Advocate
prompt: |
  You are the ADVOCATE in an adversarial code review.
  Read your instructions: {plugin-path}/agents/advocate.agent.md
  Read the context: {context-file-path}

Skeptic:

subagent_type: deep-review:Skeptic
prompt: |
  You are the SKEPTIC in an adversarial code review.
  Read your instructions: {plugin-path}/agents/skeptic.agent.md
  Read the context: {context-file-path}

Architect:

subagent_type: deep-review:Architect
prompt: |
  You are the ARCHITECT in an adversarial code review.
  Read your instructions: {plugin-path}/agents/architect.agent.md
  Read the context: {context-file-path}

2.2 Handle Failures

If an agent fails or times out, offer user options:

Re-trigger - spawn just that agent again
Proceed without - continue with available results
Abort - if critical perspective is missing

Phase 3: Synthesis

3.1 Present Raw Perspectives

Show each agent's analysis in full before synthesizing. This lets the user see raw reasoning and overrule if needed.

3.2 Agreement Analysis

What do multiple agents agree on? High-confidence findings - include in final review.

3.3 Conflict Resolution

When agents disagree:

Skeptic finds bug, Advocate defends → Does Advocate cite file:line that refutes? If not, Skeptic wins.
Advocate says intentional, Skeptic says bug → If Skeptic shows reproducible path, it's a bug regardless of intent.
Architect says blocking, Skeptic disagrees on priority → Use Skeptic's priority (they own correctness).
Architect says blocking, Advocate defends → Architect wins on architectural concerns (they own direction).
No evidence either way → Mark as "Disputed" for user to decide.

Evidence beats assertion - file:line wins over "probably."

3.4 Completeness Assessment

For large changes where agents could not examine everything, assess coverage honestly:

What was reviewed - files, areas, and patterns agents examined
What was not reviewed - files or areas agents skipped or only mentioned in passing
Confidence level - HIGH (all significant paths examined), MEDIUM (representative sample), LOW (spot-checked only)

If agents focused heavily on some areas and ignored others, note the gap. The user needs to know where the review is strong and where it's thin.

3.5 Final Review

## Deep Review: <title>

### Summary
<1-2 sentence overview>

### Perspectives

**Author's Intent** (Advocate)
<key defenses and rationale>

**Risk Analysis** (Skeptic)
<bugs found, edge cases, concerns>

**Architectural Impact** (Architect)
<patterns, debt, direction>

### Consolidated Findings

- <issue>
  - **Priority**: Critical/High/Medium/Low
  - **Advocate**: <view>
  - **Skeptic**: <view>
  - **Architect**: <view>

### Disputed (if any)

- <issue>
  - **Advocate**: <defense>
  - **Skeptic/Architect**: <concern>
  - **Resolution**: User to decide

### Recommendations
<prioritized actions>

### Follow-up Items
<non-blocking concerns worth tracking>

### Coverage (if partial)
<what was reviewed, what was skipped, confidence level>

3.6 Structured Findings Block

After the human-readable final review, append a machine-readable findings block. This enables downstream tools (e.g., code_review) to parse findings programmatically.

<!-- structured-findings
findings:
  - id: 1
    priority: critical | high | medium | low
    file: <relative path>
    line: <line number or null>
    summary: "<one-line description>"
    agents: [advocate | skeptic | architect]
  - id: 2
    ...
structured-findings -->

Each finding in the consolidated review MUST have a corresponding entry in this block. The agents field lists which agents flagged the finding. The priority field uses the same levels as the human-readable review.

Schema Contract

This schema is a contract between deep_review (producer) and its consumers: code_review and full-orchestration:implement_and_review. Changes must be synchronized across all three plugins.

Field	Type	Allowed Values
`id`	integer	Sequential from 1
`priority`	string	`critical`, `high`, `medium`, `low`
`file`	string	Relative path from repo root
`line`	integer or null	Line number, or null if not applicable
`summary`	string	One-line description of the finding
`agents`	list of strings	`advocate`, `skeptic`, `architect`

3.7 Cleanup (Always)

Delete the context file regardless of outcome:

rm -f "$CONTEXT_FILE"

Context file may contain sensitive code - always delete.

Notes

Cost: Spawns 3 parallel agents. Use for complex reviews.
Pre-existing bugs: Agents may find issues in surrounding code. Include them.
Contradictions: If agents find opposing evidence, include both views.
Failure recovery: This plugin uses human-in-the-loop for failure recovery - on agent failure, the user chooses whether to re-trigger, proceed without, or abort. There are no automatic retry limits by design.

deep_review

Invocation

Context Preview

SKILL.md

deep_review

Invocation

Context Preview

SKILL.md

Deep Review

Why This Works

Usage

Workflow

Phase 1: Gather Context

1.1 Identify the Changes

1.2 Fetch Content

1.3 Write Context File

1.4 Verify Context File

Phase 2: Parallel Analysis

2.1 Agent Prompts

2.2 Handle Failures

Phase 3: Synthesis

3.1 Present Raw Perspectives

3.2 Agreement Analysis

3.3 Conflict Resolution

3.4 Completeness Assessment

3.5 Final Review

3.6 Structured Findings Block

Schema Contract

3.7 Cleanup (Always)

Notes

Similar Skills

Deep Review

Why This Works

Usage

Workflow

Phase 1: Gather Context

1.1 Identify the Changes

1.2 Fetch Content

1.3 Write Context File

1.4 Verify Context File

Phase 2: Parallel Analysis

2.1 Agent Prompts

2.2 Handle Failures

Phase 3: Synthesis

3.1 Present Raw Perspectives

3.2 Agreement Analysis

3.3 Conflict Resolution

3.4 Completeness Assessment

3.5 Final Review

3.6 Structured Findings Block

Schema Contract

3.7 Cleanup (Always)

Notes

Similar Skills