# OMP delegation reference

Inspected on this machine: OMP 18.12.5, using its official
`omp agents unpack --dir /tmp/pi-native-omp-reference` command. The extracted bundle
contains scout, task, sonic, reviewer, security-reviewer and seance. It does not
contain this package's researcher role. No OMP binary was modified, and no OMP
implementation is bundled.

The installed scout advertises fast read-only discovery and compressed handoffs;
task/sonic emphasize assigned scope, minimal useful results and avoiding redundant
reads/writes. Reviewer emphasizes actionable, evidence-backed findings and checking
consumers across module boundaries.

Current upstream (separate from the installed-version evidence):
- [Model-facing task prompt](https://github.com/can1357/oh-my-pi/blob/main/packages/coding-agent/src/prompts/tools/task.md)
- [Task implementation reference](https://github.com/can1357/oh-my-pi/blob/main/docs/tools/task.md)
- [Subagent usage / when not to delegate](https://omp.sh/docs/subagents)

The task prompt chooses the most specific specialist, uses scout when files are
unknown, prefers one investigate+edit owner, assigns one integration owner for shared
edits, and coordinates validation after a batch. The docs keep small sequential
changes in the main session. Automatic role selection is guidance interpreted by
the parent model, not proof a deterministic classifier always makes a given choice.

Adaptation here is original prompt text: opportunistic automatic delegation under
the approved task scope; tiny known-file/sequential work stays in the parent; minimal
useful fan-out; no forced scout→worker→reviewer chain; compact evidence/result
handoffs; no nested agents. Model/thinking overrides remain explicit user choices.
We do not copy OMP's yield schemas, auto-thinking classifier, background lifecycle,
unlimited budgets, approval bypass, isolation machinery or model-role implementation.

This is not a claim that prompts alone guarantee lower latency. Each Pi child still
starts a fresh process/context, native server/model latency remains, all inherited
MCP schemas can add input tokens, and inheriting max thinking can be expensive.
The role editor exposes that inheritance and suggests effort appropriate to each
role without changing the user's model or thinking defaults automatically.