omp packages

@siriuslan/subagent-dispatch0.0.1

Pi extension that delegates a task to a separate, isolated agent session (pi or Claude Code) and returns its result

siriuslan Published 7 days ago MIT ~251 downloads/mo

Untested on omp

This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.

Did it work for you?

Result

One report per person. A new report replaces your old one.

Reports 0

No reports yet.

Readme

@siriuslan/subagent-dispatch

A pi extension that hands a self-contained task to a separate, isolated agent session, waits for it to finish and returns its answer. The work doesn't fill your current session's context window. Only the final answer comes back.

Backend Runs Model
pi (default) pi --mode json -p --no-session Your session's current model; override per call
claude claude -p --output-format json --dangerously-skip-permissions Claude Code's default

Install

pi install npm:@siriuslan/subagent-dispatch

The claude backend needs the Claude Code CLI on your PATH and uses your existing login.

Usage

Just ask. A prompt of this form is dispatched directly, without going through the model:

start a subagent to do <task>
start a subagent using claude to do <task>
start a subagent with claude code session to handle <task>

using / with / via choose the backend (pi, claude, claude code). do / complete / finish / handle are all accepted, and the task may span several lines. When the subagent finishes, its answer appears in the transcript as a message, and the main agent sees it on its next turn.

Or let the model decide. The start_subagent tool is available to the model:

Parameter Description
task The task, used as-is as the subagent's prompt
backend "pi" (default) or "claude"
cwd Working directory, relative to the current one (default: current)
model pi backend only: an id, provider/id or pattern, as pi --model takes. Default: your session's current model
provider pi backend only: the provider to look model up in. Ignored without model

The tool result shows the backend and model used, the answer (Ctrl+O expands the task, full output, duration and cost), or the error if the run failed. Pressing Esc while the tool runs stops the subagent.

Notes

  • The claude backend runs with --dangerously-skip-permissions. It can edit files and run commands in the working directory without asking. Only use it for tasks you'd let it do unsupervised.
  • The pi backend runs the same pi version as your session, with its default tools, and saves no session.
  • The pi backend uses the model your session is on when the subagent starts. To pick a cheaper model, ask for one ("start_subagent with model haiku") or switch models first. The natural-language trigger always uses the session's model. If the session has no model, pi's own default is used.

Development

npm test   # uses fake pi / claude CLIs; no agent is called