code-dispatch

Code Dispatch (ACP-style Coding Agent Delegation)

Delegate coding, file-system, and research tasks to a real CLI agent running on the user's machine. The orchestrator stays clean — no token overhead, no context window bloat. The specialist executes.

When to use

• Complex coding tasks that require multi-file edits, running tests, or reading the repo
• File-system operations beyond simple read/write (refactors, scaffolding, migrations)
• Research tasks that benefit from live CLI tools (git, grep, compilers, package managers)
• Any task where Claude Code's local tools + MCP servers give better results than inline agent calls

Rules

Agent ID: Always agent: "claude" (not "claude-code"). Use "gemini" or "codex" only when explicitly requested.

Mode selection by channel:

• Discord: mode: "session" + thread: true for multi-step work; mode: "run" for one-shot
• Telegram / WhatsApp / Slack / Teams / BlueBubbles / WebChat: Always mode: "run" — async thread binding is not available on these channels

Permission mode:

• permission_mode: "approve-reads" (default): allows reads, prompts for writes/exec — safe for most tasks
• permission_mode: "approve-all": auto-approves all tool calls — only use for trusted, well-defined tasks
• permission_mode: "deny-all": read-only (Read, Glob, Grep, LS only) — use for audits or safe exploration

Timeout: Always set timeout_seconds (default 120, max 300). For complex refactors use 300.

Never nest: Do not call sessions_spawn from within a spawned session. The orchestrator dispatches; the specialist does the work.

Working directory: Always specify cwd pointing to the relevant project directory.

Example — Discord multi-step session

{
  "tool": "sessions_spawn",
  "agent": "claude",
  "mode": "session",
  "thread": true,
  "permission_mode": "approve-all",
  "cwd": "/path/to/project",
  "task": "Refactor the auth module to use JWT. Run tests after each file change.",
  "timeout_seconds": 300,
  "label": "Auth JWT refactor"
}