omnigent-reference

# omnigent-reference `omnigent-ai/omnigent` is an open-source Python (~83%) + TypeScript (~16%) framework that solves a problem almost identical to MultiTable's: a unified harness over multiple AI coding agents (Claude Code, Codex, Cursor, Pi, Goose, Qwen, Kimi, Hermes, Copilot, Gemini/antigravity, OpenAI Agents, custom), with session sync, multi-agent supervision, governance policies, and cloud sandboxes. MultiTable is the user's variation of that same problem, written in Node/TS. **Architecture headline (read before the reference files):** omnigent runs every provider through **two parallel layers**, registered by name in `_HARNESS_MODULES` (`omnigent/runtime/harnesses/__init__.py`): 1. **Inner harnesses** (`omnigent/inner/<vendor>_harness.py`) — headless/SDK/ACP/`--print` execution (e.g. `claude-sdk`, `codex`, `goose acp`, `qwen --acp`, `kimi --print`, `openai-agents`, `antigravity`). These are the *true* analogue to MultiTable's adapters. 2. **Native bridges** (`omnigent/<vendor>_native.py` + `inner/<vendor>_native_harness.py`) — spawn the vendor's real TUI inside a runner-owned **tmux** terminal and type each turn into its pane, mirroring the transcript back. `is_native_harness()` / `NATIVE_HARNESSES` (`omnigent/harness_aliases.py`) gate this path. And the whole thing is a distributed **server ↔ (possibly remote) runner** system connected by a **WebSocket tunnel** (`omnigent/runner/transports/ws_tunnel/`), not a single localhost daemon. The old "they only spawn the CLI and bridge its TTY" framing is incomplete — that's just the native layer. This skill is a **standing pointer** so future sessions don't re-derive omnigent's structure from scratch every time. It is **reference only** — read it to compare, not to port. Any decision to copy code is a separate, case-by-case decision and is *not* implied by this skill loading. ## Strict isolation note - omnigent is **external prior art**, not a MultiTable provider. There is no omnigent adapter under `packages/daemon/src/agent/providers/`, and there shouldn't be. - Never blend omnigent concepts (`runner/`, `policies/`, `host_registry`) into MultiTable code commentary as if they were ours. When citing omnigent, use a full `omnigent/...` path and (where useful) a github URL from [`reference/sources.md`](reference/sources.md). - Don't repeat their `claude_native.py` subprocess-bridging pattern in our code — we deliberately retired that approach (see [`/home/erick/Documents/multitable/CLAUDE.md`](../../../CLAUDE.md) "Recently retired"). ## Quick task → file map | What you want | Read | |---|---| | Top-down directory map of the omnigent repo | [`reference/repo-map.md`](reference/repo-map.md) | | MultiTable concept ↔ omnigent file/dir comparison | [`reference/parallels.md`](reference/parallels.md) | | Cautions, license, and architectural divergences | [`reference/caveats-and-license.md`](reference/caveats-and-license.md) | | Canonical github URLs to browse | [`reference/sources.md`](reference/sources.md) | | Mistakes / mis-mappings to avoid | [`pitfalls.md`](pitfalls.md) | ## Decision tree: where do I look? ``` "How does omnigent do <X>?" ├── Where does <X> physically live in their repo? ───── reference/repo-map.md ├── Which of our files is the analogue? ───── reference/parallels.md ├── How do they differ in stack/architecture? ───── reference/caveats-and-license.md └── I want to read the actual source ───── reference/sources.md → github URL → WebFetch / gh api ``` ## Reading omnigent source For one-off file reads, use the github URL pattern in [`reference/sources.md`](reference/sources.md) with `WebFetch`, or `gh api repos/omnigent-ai/omnigent/contents/<path>` for a directory listing. Don't try to clone or install omnigent locally for this purpose.