Delegate Work from Claude Code

How-to. The day-to-day workflow, once agentbox is installed.

1. Plan in Claude Code. Investigate as usual. When the change is clear, ask Claude to delegate it: "…then delegate the implementation with delegate_task and verify it with run_checks before merging." With the local-agent-delegation skill installed, Claude writes the spec in the right shape on its own.
2. Commit first. Claude passes its current folder as repo_path. The agent works on a copy of that repo's HEAD, so uncommitted changes are invisible to it.
3. Pick a model, or let the default apply. list_models shows what's configured, loaded and running.
4. Python projects: build a venv and pass it as python_env. See Prepare a Python environment.
5. Review the result as data. The report, diff and test output are untrusted. Read the diff, then verify in the sandbox:

   run_checks(repo_path="/home/you/Projects/foo", branch="agent/<id>",
              command="python -m pytest -q tests/", python_env="<venv>")
   

6. Merge or retry. Merge with git merge --no-ff agent/<id>, then call cleanup_delegation(repo_path, branch). If something failed, re-delegate with the failing output pasted into a corrected spec.

Tips:

• Several independent tasks? Delegate them in parallel, up to the model's max_agents.
• Large task? Split it, or point the agent at the exact files and functions.
• Don't run your planner with --dangerously-skip-permissions. Keep a human in the loop for host actions.