subagent-orchestration

Orchestrating subagents with 09orche

You are the orchestrator. This project's MCP tools (ask_<alias>, agent_<alias>, and the profile tools below) give you access to other models as subagents — text in, text out for ask_*; real sandboxed file and shell access, scoped to one workspace directory, for agent_*. This skill is about when and how to use them well, not what the tools do — see the project README for that.

One-off delegation vs. a saved profile

For a single task, just call the tool directly with an inline system_prompt:

ask_ox_alpha(prompt="...", system_prompt="You are a terse code reviewer. Flag only real bugs.")

Reach for a profile when you'll want the same persona again — across several calls in this session, or in future sessions on the same project. Creating one is one call:

save_profile(
  name="reviewer",
  base_alias="ox_alpha",
  system_prompt="You are a terse code reviewer. Flag only real bugs, no style nits.",
  agent_tools="read",          # optional: only if this profile needs agent mode
)

After that, ask_profile("reviewer", prompt) or agent_profile("reviewer", prompt, workspace) reuses the persona without restating it. Check list_profiles() before creating one — it may already exist from an earlier session.

Don't create a profile for a task you'll only run once. A profile that never gets reused is just an extra file to maintain.

Verification is your judgment call, not a mandatory step

This project deliberately does not force every delegated task through an automatic verifier — that would be slower and often redundant with what you'd naturally check anyway. Pick the right level for the task:

Default: review it yourself. For most delegated work, read the subagent's output (or its agent_* diff/changes) the way you'd review a colleague's pull request. This is usually enough, and it's cheap. A worked example: two subagents built a small Conway's Game of Life implementation in a shared workspace — one wrote the core logic, the other wrote a CLI and tests on top of it. The second subagent's own tests initially looked fine to a skim, but running them (as the orchestrator, not trusting the subagent's self-report) found a real bug: a test assumed a class was iterable when it wasn't. Actually running the check, not just reading the code, is what caught it.