Agent Go

agent_go

[](https://www.python.org/) [](tests/) [](LICENSE) []()

**Plan Mode orchestration tool** — wraps Claude Code with a structured `Plan -> Decompose -> Execute` workflow. LLM generates an execution plan, each step runs as an isolated subtask in its own git worktree with Claude Code. Subtasks execute concurrently with topological wave scheduling.

Why agent_go?

Give Claude Code a complex task — refactoring auth, upgrading dependencies, adding a feature — and it can drift. agent_go breaks the work into **2–5 independently executable subtasks**, each with its own isolated worktree, verification command, agent role, and skill injection. Results flow downstream via git merge.

• Structured execution — LLM generates the plan, not ad-hoc decisions
• Isolated worktrees — shared .git object db, each subtask on its own branch
• Concurrent execution — topological wave scheduling with --parallel N
• Smart role/skill routing — config-driven rules match subtasks to Agent types and Skills
• Interrupt & resume — SIGINT pauses, agent_go resume picks up where you left off
• Remote push — push worktree branches to a remote for CI/CD integration
• Zero dependencies — pure Python stdlib
• Plan cache — SHA256 cache key + 24h TTL reduces API costs
• Interrupt & crash recovery — SIGINT pauses with resume; SIGKILL/abnormal exits rebuild meta.json via recover
• Result review — review --task aggregates per-file diffs with approve/reject decisions; --deep adds independent-model analysis
• MCP server & client — expose agent_go as an MCP server (stdio or HTTP/SSE, 6 tools); subtasks can consume tools from external MCP servers (mcp__{server}__{tool})
• **