omp (oh-my-pi) — a Pi fork by Mario Zechner, rewritten for a real development workflow: sessions, subagents, slash commands, extensions. TypeScript + Rust on macOS/Linux/Windows without WSL.
https://omp.sh/docs/agents-and-roles
There are three layers:
- Model — a specific provider and model (Claude, GPT…).
- Role — a named slot:
defaultfor regular requests,smolfor cheap small jobs,slowfor deep thinking and review,plan,vision,commit,task, more on that later. - Agents and subagents — who exactly does the work (reviewer, scout, designer). The agent looks at the role, the role looks at the model. Out of the box:
scout(fast read-only repo overview),designer,reviewer,security-reviewer,librarian(checking external APIs),task,sonic(mechanical edits).
When you change the model in the slow role, all future reviewer agents automatically switch to it. The agent harness picks them itself depending on the prompt.
https://omp.sh/docs/advisor
The special Advisor agent can be assigned a separate model that reads every step of the main agent (prompts, responses, reasoning, tool calls) and adds notes directly into the session, and sometimes can stop the whole flow right away if it is clearly going the wrong way!
It runs on its own context and a completely different model (preferably with a matching context window size): the main stream can run on Gemini 3.8 Flash, while the advisor is, say, Meta Spark 1.3 — different models look at the same work from different angles. The Advisor does not approve actions: the main agent must weigh each note against the request, not blindly obey it.
Enabled with two lines in the config, or via /advisor. Three levels of advice: nit — an uninterrupted aside (“this could be simplified”); concern — a risk or wrong direction, may interrupt the run; blocker — continuing is clearly pointless, the result will be broken. Subagents have no advisor by default.
The advisor idea has already been ported back to pure Pi: the pasky/pi-omplike-advisor port https://github.com/pasky/pi-omplike-advisor — a permanent read-only model that reviews the main agent's work on every step.
#omp #piagent #subagents #advisor