cook. - Workflow loops for Claude Code, Codex, and OpenCode

Source
claude-codeagentic-workflowscode-reviewiteration-loopsparallel-agentsmulti-agentdeveloper-tools

Summary

Cook is a composable workflow DSL for AI coding agents (Claude Code, Codex, OpenCode) that adds review loops, sequential passes, parallel racing, and task-list progression on top of single-shot agent calls. It enables multi-pass iteration, quality gates, and branch-and-pick patterns without manual orchestration.

Key Insight

The key value is in the composition model - three categories of tokens that wrap left to right:

  1. Loop operators for iteration quality:
  • xN - run N sequential passes, each seeing previous output (refinement)
  • review - adds a review-then-gate loop with customizable prompts and max iterations
  • ralph - outer gate for task-list progression (reads project state, advances through a plan)
  1. Composition operators for exploring solution space:
  • vN / race N - run N identical cooks in parallel git worktrees
  • vs - run two different approaches in parallel worktrees
  • Resolvers: pick (choose winner), merge (synthesize), compare (write comparison doc)
  1. Per-step agent/model control - different agents for different steps (e.g., Codex for work, Claude for review), configured via CLI flags or .cook/config.json.

The ralph pattern is particularly powerful for autonomous task execution: the agent reads a plan file, works on the next task, passes through review, then ralph decides whether to advance or stop. This is a structured version of the “loop until done” pattern that keeps agents on track.

Sandbox modes: agent-native (default, no Docker) or Docker with restricted network. OpenCode only works in Docker mode.