delegating-to-omp-rpc

Delegating to omp-rpc

Overview

omp-rpc is a persistent second agent you can drive from the shell. You (the driving agent) send it scoped tasks with omp-rpc send, read the streamed result, and iterate — its session remembers earlier turns, so a delegation is a conversation, not a series of cold prompts. Mechanics of each command live in the using-omp-rpc skill; this skill is about when and how to delegate well.

When to use

A sub-task is self-contained and you'd rather not spend your own context on it (bulk edits, a focused investigation, boilerplate, a first-draft).
You want a second opinion or a different model's take in parallel.
You want work done under a separate accumulating context (e.g. iterating on one file over several turns) without polluting yours.

Not for: tasks needing your live conversation context (the sub-agent can't see it — you must put everything it needs in the task text), or trivial one-liners you can just do.

The delegation loop

omp-rpc status                                   # running? busy?
omp-rpc models --json                            # discover exact selectors (never guess them)
omp-rpc start --models "<sel>,<sel>" --model <sel> --cwd /path/to/repo  # scope + active
omp-rpc send "Task 1 — <clear, self-contained instruction>"
omp-rpc send "Task 2 — builds on task 1"         # context carries over
# ...review each streamed reply, correct course as needed...
omp-rpc stop                                     # when the whole job is done

Doing it well

Point --cwd at the repo the sub-agent should work in, so its file tools operate on the right tree.
Make each task self-contained. The sub-agent has no access to your conversation — restate the goal, name the files, include constraints.
One task at a time. The session is single-flight; wait for a send to return before the next. omp-rpc status shows busy.
Steer instead of waiting when a running turn is going the wrong way: omp-rpc steer "actually, skip the tests dir" injects mid-turn; omp-rpc abort cancels it outright.
Discover models, don't recall them. Selectors live in omp's catalog and change over time — run omp-rpc models --json and copy exact selectors; never invent a selector from memory.
Match the model to the task, live. Scope the session to a few models at start (--models), then switch among them per turn — a big-context model for a sprawling session, a coding-tuned one for a focused turn — with omp-rpc model <sel>. No restart; the accumulated context is preserved. Switches are limited to the session's scope.
Capture cleanly when you'll parse the output: omp-rpc send -q "..." sends the answer to stdout and thoughts/tools to stderr. --json gives {stopReason, usage}.
Review before trusting. It runs a different (often smaller) model with auto-approved tools; read its diffs/output rather than assuming success.

Safety when delegating

Destructive shell commands are auto-denied and stop won't discard in-flight work — but the guard is a net, not a sandbox. For untrusted or high-risk tasks, start the session with --cwd inside a throwaway checkout or container. Details in the using-omp-rpc safety reference.

Common mistakes

Assuming shared context. It sees only the task text you send. Missing context → wrong output.
Firing a second task while busy. Returns a busy error; serialize your sends (or steer into the running one).
Leaving a session running on the wrong repo. Its --cwd is fixed at start; to work elsewhere, stop and start again there.