After each turn where the agent edits files, a language-aware review checklist is injected into the system prompt. The agent self-reviews before proceeding, catching type safety issues, error handling gaps, security concerns, and naming problems.

Supports: TypeScript, Python, Go, Rust, Java, PHP.

On-demand review (`/review`)

Run a thorough code review with structured findings:

/review                    # review all uncommitted changes
/review --staged           # only staged changes
/review --ref=main         # diff against main
/review src/foo.ts         # specific files

When an Anthropic API key is available, /review uses a direct Haiku call for structured output with severity-leveled findings (CRITICAL / HIGH / MEDIUM / INFO). Without an API key, it falls back to a prompt-based review via the session agent.

How it works

Edit tracking: hooks into tool_execution_end to collect files modified by Write/Edit tools during each turn
Turn batching: at turn_end, snapshots the accumulated edits (no per-edit overhead)
Prompt injection: at before_agent_start, injects a brief language-specific review checklist into the system prompt
On-demand: /review reads file contents, calls Haiku for structured analysis, and formats findings with severity, line numbers, and suggestions