Coordinator Load-outs

A load-out is a named profile that pins three things together:

| Field | What it does | |-------|--------------| | model | Which LLM backend to route the session to (e.g. qwen-coder, claude-sonnet) | | prompt | A system-prompt fragment appended to the default, scoping the agent's persona and rules | | tools | The exact subset of MCP tools the agent is allowed to call |

Together they form a reproducible, narrow context for a specific kind of task. The launch.py script in this directory reads a load-out YAML and assembles a claude CLI invocation that enforces all three constraints.

Current load-outs

coder.yaml

The first load-out. Targets a coding-focused sub-mind running on qwen-coder.

The agent is restricted to the codebean MCP tools — a code-as-model API that conceals git and the filesystem entirely. The agent builds packages, classes, and functions as structured objects, then calls codebean_save(message) to persist them; a real git commit is created in a hidden workspace with no further effort from the agent.

Allowed tools: coupler_status, bootstrap_orient, bootstrap_status, bootstrap_memory, bootstrap_checklist, codebean_new_module, codebean_add_class, codebean_add_function, codebean_set_body, codebean_render, codebean_tree, codebean_save.

Adding a new load-out

1. Create loadouts/<name>.yaml with name, model, prompt, and tools fields. 2. Verify the tool names against the coupler's mounted namespace ({service}_{tool} — e.g. codebean_save). 3. Run python loadouts/launch.py <name>.yaml (dry-run by default) to confirm the assembled command looks right. 4. Pass --go to actually launch a session.