Skip to content

ShamirSecret/auto-code-machine

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14,705 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Auto Code Machine logo

An AI coding agent controlled like a machine.

Auto Code Machine treats LLM code generation as a controlled process—not free-form writing. Inspired by mechanical control engineering, it uses design rules, small work orders, and checkpoints to keep output reliable, reviewable, and reversible.


How It Works

  • Design first — agree on scope and architecture before coding
  • Try a small slice first — prove the design with a minimal end-to-end path
  • Small work orders — split work into verifiable steps, generated as you go
  • Gates by risk — light checks for safe changes, full review for risky ones
  • Fix by feedback — let tests and linters guide repairs
  • Remember failures — don't repeat the same mistake

Work Orders

Each work order is one small, verifiable task—not one tiny code edit. It defines the goal, owner, allowed files, how to verify, and how to roll back.

Agent Taxonomy

Auto Code Machine (ACM) adds a role-based execution layer on top of native opencode agents.

Agent Type Role
plan (native) Primary Explore, plan, and review without making changes
build (native) Primary Full development work with all tools
chief-engineer (ACM) Primary Orchestrate work orders, route to subagents
planning-engineer (ACM) Subagent Decompose goals into small work orders
code-engineer (ACM) Subagent Implement code within scope using TDD
review-engineer (ACM) Subagent Read-only review of diffs, security, and permissions
  • Primary agents are full conversational agents you interact with directly.
  • Subagents are specialized assistants invoked by the chief-engineer or via @mention.
  • Native plan and build are not ACM agents — ACM starts only through chief-engineer.
  • Legacy role values (design, test, security, repair, memory) are accepted from historical data but are not active dispatch targets.

Native vs ACM

  • Native opencode plan and build are normal TUI primary agents.
  • Native plan uses plan_exit to switch to build.
  • Native build does not create .machine intake or work orders.
  • ACM starts only through chief-engineer, which is itself a primary agent.
  • chief-engineer dispatches work to ACM subagents: planning-engineer, code-engineer, and review-engineer.

Chief Commands

When using chief-engineer, the following slash commands are available:

Command Behavior
/session Manage sessions and child sessions
/model Override model for current session
/goal Set or update the current goal
/commit Auto-commit from diff, or /commit <message> to stage and commit directly
/distill Generate a work pattern report (report-only)
/dream Generate a memory consolidation report (report-only)
/engineers Configure subagent model overrides for planning/code/review
/new Start a new session
/export Export session data

Runtime

Tuned for cache-friendly models like DeepSeek and MiMo to keep cost down.

  • Tier routing — fast models for coding, stronger models for decisions
  • Stable context — structured prompts to improve cache hits
  • Tool call repair — auto-fix broken model tool calls
  • Cost tracking — spend per turn, order, and session

Contributing

See CONTRIBUTING.md.

About

The open source coding agent.

Resources

License

Contributing

Security policy

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages

  • TypeScript 70.7%
  • MDX 25.9%
  • CSS 2.9%
  • HTML 0.2%
  • JavaScript 0.1%
  • Astro 0.1%
  • Other 0.1%