Skip to content
Back to skills

Omni Agent Runtime

ASecurity

Internal contract for the subagent omni-agent:worker: exact command syntax, wait loop, project mode, provider choice. Do not call directly.

  • 3 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added November 13, 2026
ai-agentspythongobash

Security analysis

A100/100

Scanned November 13, 2026

npx -y skills add alexodexo/omni-agent --skill omni-agent-runtime --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Omni Agent Runtime?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Omni Agent Runtime
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/alexodexo-omni-agent-runtime/badge)](https://www.skillsdirectory.com/skills/alexodexo-omni-agent-runtime)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: omni-agent-runtime
description: Internal contract for the subagent omni-agent:worker: exact command syntax, wait loop, project mode, provider choice. Do not call directly.
user-invocable: false
---

# Omni-Agent runtime

Only for the subagent `omni-agent:worker`. It is a forwarder, not a planner.

## Command

`omni-agent` is on the PATH while the plugin is active. If it is missing, use `python3 "${CLAUDE_PLUGIN_ROOT}/bin/omni-agent"` in the same place.

```
omni-agent run --background --wait 540 [--account NAME | --provider ID] [--model M] [--effort low|medium|high] [--project] [--write] [--resume] --text "task text"
omni-agent wait RUN --max 540
```

| Part | Meaning |
|---|---|
| `--background --wait 540` | Start the run detached, wait at most 540 s for it to end, then report the state. Always set both. |
| `--text "…"` | Task text. Long? Write a file and pass the path as an argument instead of `--text`. **There is no `--task` flag.** |
| `--account NAME` | Claude account (fbs, felipe, llc, see `omni-agent account list`). |
| `--provider ID` | Third-party provider or `codex`. Only if the user named it. |
| `--model`, `--effort` | Only on explicit request. `--effort` applies to Codex only. |
| `--host TARGET` | Execute on a remote target (Mac mini or similar). Only if the user named a target; `--project` then means a path there, `--files` is not possible. |
| `--project` | The agent works in the project folder (current directory), read-only by default. Possible with any provider. |
| `--write` | With `--project`: files in the project may be changed. Only if the user explicitly wants changes. |
| `--resume` | Follow-up to the last agent (without a name: last run of this session). `--text` is the follow-up. |

Always call Bash with `timeout: 600000` (10 minutes, the maximum).

## Procedure

1. One `run` call. The output starts with `Started: <run>` and ends with a state message.
2. First state line `[still running]`: repeat `omni-agent wait <run> --max 540` until the line changes.
3. `[done]`, `[failed]` or `[cancelled]`: return the entire output unchanged. Do not summarize, do not comment.

## Choosing the provider

- The user gave nothing: neither `--account` nor `--provider`. Then the user's default applies.
- You **never** choose a third-party provider.
- Project mode: if the task refers to files or code of the project, set `--project`. This works with any provider; with a third-party provider the user named, the files read go to its data location, which is the user's decision.

## Do not

- No `review`, `status`, `result`, `cancel`, `account`, `setup` or `doctor`.
- No own work on the project, no reading files, no checking whether the run is still going (`wait` does that).
- Never print a key or ask for one.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

View all in ai-agents →

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.