# Omnius Agent Guide This project uses Omnius. Search the installed discovery contract before guessing a command, route, tool, provider, state path, or source owner: ```bash omnius discover "your task" omnius show omnius docs omnius capabilities --json ``` Start with `workflow.choose-entrypoint`, then expand the relevant `layer.*`, `module.*`, `runtime.*`, `store.*`, or task-specific `workflow.*` entry. | Need | Use | | --- | --- | | Interactive chat and human-only slash commands | `omnius` | | One foreground coding task | `omnius ""` | | Stateful daemon conversation | `POST /v1/chat` | | Long task with ID, events, polling, and cancel | `POST /v1/run` | | OpenAI-compatible client | `POST /v1/chat/completions` | | One direct tool | Only the exact `rest-call` interface returned by tool metadata | For a running daemon, bootstrap from `GET /v1/discovery/bootstrap`, then check `GET /version`, `GET /health/ready`, `GET /openapi.json`, and the relevant live registry such as `GET /v1/tools`. Send `X-Omnius-Min-Version` on execution requests so a stale runtime fails before creating work. State is deliberately split: project sessions/tasks/context belong in `/.omnius`; daemon identity, credentials, managed runtimes, and shared model/media storage belong in `~/.omnius`. Resolve the active workspace before touching project state, and never emit secrets from global state. `web_search` is agent-bound: inspect `GET /v1/tools/web_search`, then use `/v1/run`, `/v1/chat`, or `/v1/chat/completions` with `agent_loop:true` and `include_daemon_tools:["read"]`. Never infer REST exposure from a slash command or direct-call exposure from a tool name. Treat queued/accepted as non-terminal; verify the observable result. Small-context agents should expand one workflow and its references. Maintainer agents should also expand the owning layer/module and run discovery freshness.