Claude Code sub-agents¶

This repository is a Claude Code plugin. It adds sub-agents that hand a task to a craze session — on craze's own native provider, Grok or Cursor — and return the session's result to Claude Code, the way a native Claude Code sub-agent returns its own. Use it to put another model on a task, to get a second opinion, or to run several models side by side.

It needs craze 0.2.0 or newer (v0.1.x is too old) on your PATH, signed in to the providers you use (Quick Start). An older craze fails with this craze is too old; the plugin needs craze ≥ 0.2.0.

Install¶

In Claude Code:

/plugin marketplace add charliek/craze
/plugin install craze@craze

The agents¶

Dispatch one with the Agent tool: subagent_type craze:<name>, and the task as the prompt.

craze:delegate reads a first line naming the target, then the task:

target: native chatgpt/gpt-6.1-sol effort=high mode=plan
Review internal/cache/evict.go for races. Report each one with the line and a fix.

The provider is native, cursor or grok; the model, effort (low, medium, high, xhigh, max) and mode (plan or ask) are optional. The presets take the task alone, or an optional first line mode=plan or mode=ask.

The agents run on Haiku and only forward: each writes the task to a file under /tmp, runs one craze new --wait (reference) and returns what it printed. They never do the task themselves, and never retry on another model. The session works in Claude Code's current directory, with its own model's tools.

The task must stand alone: the session has not seen the conversation. Give it the goal, the paths and what to return, and put a large context in a file in the project that the task names.

The result¶

The agent returns the turn's last message and a footer:

The flaky test raced the cache's eviction; I made eviction wait for readers.
---
completed · 27s · 5 tool calls · $0.04 · context 16.6k/272k
files: stats/core.py
session 175bfe79 · craze attach --session 175bfe79

The status is completed, failed, cancelled, dropped, needs_input or running, and the command's exit code follows it (Waiting commands). A message over 20,000 characters is cut and the whole result written to a file the output names, so a long answer does not overflow Claude Code's Bash output.

Long tasks and follow-ups¶

A call waits up to nine minutes (--timeout 9m, under a Bash timeout of 600000 — Claude Code's default is two minutes). A longer task comes back running with a handle, and the agent follows it with craze wait <short> --from <handle>; it asks again at most three times, so about 36 minutes in all.

SendMessage to the agent continues the same session with its context: the agent runs craze send <short> --wait (reference). A session whose host has since exited is loaded again from the session index.

To watch or answer a session, run craze attach --session <short> (the footer prints it). A session that stops on a question or a permission request that nothing answers reports needs_input (exit 4).

Running several at once¶

Dispatch several agents in one message, each in the background. Each is its own session; their results arrive as each finishes. Give them different tasks or different models, and do not point two at the same files to edit.

Plan and ask modes¶

With mode=plan the plan comes back in the result (plan:), the status is completed, and nothing is implemented — on every provider. The session stays in plan mode, follow-ups and reloads included, so a follow-up asking it to implement gets another plan. To carry the plan out, start a new delegation without mode=plan and put the plan in its task. A task that says "read-only" or "do not edit" is a request to the model, not something craze enforces; only mode=ask removes the shell.

Permissions¶

Claude Code's permission rules and hooks apply to Claude Code's own tools. A craze session's tools run inside craze, so they do not apply to them. Only the mode above limits the session, and it asks nothing of you while it works: questions it asks are answered with their first option and a plan it presents is returned, not accepted. Do not delegate what you would not let that model do unattended in this directory.

In Claude Code's default permission mode the agents themselves run craze and write the task file, so each call prompts unless you allow them in your settings:

{
  "permissions": {
    "allow": [
      "Bash(craze new:*)",
      "Bash(craze send:*)",
      "Bash(craze wait:*)",
      "Bash(craze cancel:*)",
      "Edit(//tmp/craze-delegate-*)"
    ]
  }
}

Claude Code applies Edit rules to its Write tool, and // marks an absolute path (a single leading / is relative to the settings file's own directory).

Against a native Claude Code sub-agent¶

A session left running when Claude Code is interrupted keeps going until its turn ends; craze ps lists it and craze stop <short> ends it.