Agent Review

Get feedback from an external AI agent. Useful for code review, architecture decisions, or getting a second opinion.

Running Commands

echo "code or prompt" | node .claude/skills/agent-review/query.mjs [options] "Your question"
# or
node .claude/skills/agent-review/query.mjs --file <path> [options] "Your question"

Options

Flag	Short	Description
--model <model>	-m	Model or alias (default: gemini)
--file <path>	-f	Read input from file instead of stdin
--lines <start-end>	-l	Extract specific lines from file (e.g., 50-100)
--context <path>	-c	Additional context file (can repeat)
--system <prompt>	-s	Custom system prompt
--temperature <n>	-t	Temperature 0-1 (default: 0.7)
--quiet	-q	Suppress status messages and usage stats
--list		List available models
--json		Output raw JSON response

Available Models

Model ID	Aliases	Provider	Notes
google/gemini-3-pro-preview	gemini, g3	OpenRouter	Default - good for external perspective
openai/gpt-5.1-codex	gpt, codex, gpt5	OpenRouter	Strong at code analysis
anthropic/claude-opus-4.5	opus, claude-opus	Agent SDK	Uses local subscription
anthropic/claude-sonnet-4.5	sonnet, claude	Agent SDK	Uses local subscription

Anthropic models route through Claude Agent SDK (uses your Claude subscription).
Other models route through OpenRouter API (requires OPENROUTER_API_KEY).

Environment Variables

Variable	Required	Description
OPENROUTER_API_KEY	For non-Anthropic models	OpenRouter API key
AGENT_REVIEW_DEFAULT_MODEL	No	Override default model

Examples

Code Review

# Review a file for security issues
cat src/server/auth.ts | node .claude/skills/agent-review/query.mjs \
  "Review this authentication code for security vulnerabilities"

# Review with a specific model (using alias)
node .claude/skills/agent-review/query.mjs -m gpt -f src/utils/parser.ts \
  "Review this parser for edge cases and error handling"

Reviewing Specific Lines

# Review a specific function (lines 50-100)
node .claude/skills/agent-review/query.mjs \
  -f src/server/auth.ts -l 50-100 \
  "Review this authentication function for security issues"

# Review a single line
node .claude/skills/agent-review/query.mjs \
  -f src/utils/parser.ts -l 42 \
  "Is this regex safe from ReDoS attacks?"

With Context Files

# Review code with type definitions as context
node .claude/skills/agent-review/query.mjs \
  -f src/api/routes.ts \
  -c src/types/api.ts \
  -c src/types/models.ts \
  "Review this API implementation"

# Review component with its hooks as context
node .claude/skills/agent-review/query.mjs \
  -f src/components/UserProfile.tsx \
  -c src/hooks/useUser.ts \
  "Review this React component for performance issues"

Architecture Feedback

# Get feedback on a proposed design
cat docs/design-proposal.md | node .claude/skills/agent-review/query.mjs \
  "What are the potential issues with this architecture?"

Custom System Prompt

node .claude/skills/agent-review/query.mjs \
  -f src/api/routes.ts \
  -s "You are a security expert specializing in API design" \
  "Audit this API for OWASP top 10 vulnerabilities"

Temperature Control

# Lower temperature for more deterministic analysis
node .claude/skills/agent-review/query.mjs -t 0.2 -f src/algo.ts \
  "Analyze the time complexity"

# Higher temperature for creative suggestions
node .claude/skills/agent-review/query.mjs -t 0.9 -f src/ui.tsx \
  "Suggest ways to improve the user experience"

Using Anthropic Models (Local Subscription)

# Uses Claude Agent SDK - no API credits consumed
node .claude/skills/agent-review/query.mjs -m opus -f complex-algorithm.ts \
  "Analyze the time complexity and suggest optimizations"

Quiet Mode (for scripting)

# Suppress all status messages, only output the response
REVIEW=$(node .claude/skills/agent-review/query.mjs -q -f src/auth.ts \
  "List security issues as JSON array")
echo "$REVIEW" | jq .

Output

The response text is written to stdout. After the response, usage stats are shown on stderr:

Tokens: 189 in / 871 out (1060 total)
Cost: $0.0108

This helps track token consumption and costs for OpenRouter requests. Use --quiet to suppress.

When to Use

• Code Review: Get a second opinion on code quality, security, or performance
• Architecture Decisions: Validate design choices with another perspective
• Bug Analysis: Share error context and get debugging suggestions
• Documentation Review: Check if docs are clear and complete
• Test Coverage: Identify missing test cases

Tips

• Default model (Gemini 3 Pro) is recommended for most external reviews
• Use Anthropic models when you want consistency with Claude's style
• Use -l to review specific functions without sending the entire file (saves tokens)
• Use -c to include type definitions or related files for better context
• Use -t 0.2 for more focused/deterministic responses, -t 0.8 for creative suggestions
• Pipe code directly for quick reviews: cat file.ts | node ... "review this"