Glm Delegate — Testing skill for Claude Code

glm-delegate

**Run implementation work as swarms of GLM agents in disposable git worktrees, while an Anthropic model does the scoping and the review.**

The expensive model earns its price on judgement — deciding what to build, reading a diff, catching the thing that pages you at 3am. It is a poor use of a rate limit for typing out a migration across forty files. This plugin moves that middle to GLM-5.2, keeps it inside a throwaway worktree, and hands back a diff nobody has merged yet.

[](LICENSE) [](.github/workflows/ci.yml) [](plugin/server/) [](plugin/server/glm-executor.mjs) [](plugin/)

┌───────────┐   ┌───────────┐   ┌────────────┐   ┌──────────┐   ┌──────────┐
│   SPEC    │──►│ WORKTREE  │──►│ EXECUTORS  │──►│  GUARD   │──►│  REVIEW  │──► apply
│ you scope │   │ throwaway │   │  n shards  │   │  blocks  │   │ you read │
│ and write │   │ + branch  │   │ in parallel│   │ escapes  │   │ the diff │
└───────────┘   └───────────┘   └────────────┘   └──────────┘   └──────────┘
      │                                               │
  the judgement                                 tripwire, not a
  is never delegated                            sandbox — docs/guard.md

Quick start

The plugin is a zero-dependency MCP server. The guard suite runs offline — no key, no CLI, no network:

git clone https://github.com/samuelowles/glm-delegate
cd glm-delegate
node plugin/server/test-guard-rules.mjs

write-outside → BLOCK
  ok  open(abs, "w")
  ok  pathlib write_text
  ok  %%bash redirect to abs path
in-tree writes → ALLOW
  ok  relative write
relative writes when cwd drifts from the workt