OMP Agents

Hand a task to one of your OMP subagents by naming it in the thread. Type @ in the composer, pick a row under OMP agents, and the pill expands at send time into agent-only context naming the agent, its specialization, its tools and model, its definition file, and the task call that delegates to it.

What you get

• @ mentions for every OMP task agent visible to the workspace. Matching is substring, subsequence, and token-based, so jns6, jns65, and jns 65 all find jns65-agent.
• An OMP agents page in the sidebar listing each agent with its scope, tools, model, and definition path.
• bb omp-agents list for the same list from a shell, with --json for the full records and the directories that were searched.

Where agents come from

Discovery follows OMP's own precedence, so a name defined higher up shadows the same name below it.

• Project — the nearest .omp/agents/*.md, walking up from the workspace.
• User — $PI_CODING_AGENT_DIR/agents, then ~/.omp/agent/agents, or the active OMP_PROFILE's directory.
• Extra — absolute directories you list in the plugin's settings.

Every read goes through BB's own file API on the machine running BB, so a symlinked agent directory resolves to its real path. OMP's bundled agents have no file on disk, so run omp agents unpack to give them definitions and make them appear here.

Offered only where it can run

These subagents are spawned by the task tool inside OMP, so the menu appears only on an OMP-backed surface. On a Claude, Codex, or Pi thread it stays empty, and a message that still carries an OMP mention is refused with an error naming the provider rather than reaching the model as an instruction it cannot follow. Pi is a separate tool from OMP, despite the similar name.

Settings

• Extra agent directories — absolute paths, one per line, searched after the project and user scopes.
• Mention results — how many agents the menu lists at once.
• Providers that run OMP — which provider ids get the menu.