Repository navigation
omo (oh-my-openagent) as a first-class provider #10376
ai-ops-agentic-coder
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Versions: Paseo 0.16.2 (tested on 0.12.0) · omo 5.1.79 (engine: senpi 2026.11.18-3) · Linux daemon host
What I'm trying to do
Run omo as an agent provider in Paseo — pick it in the provider list, get modes, models, streaming, permissions and session resume, the same way Pi or Oh My Pi work today.
How I do it today
omo's binary is a fork of Pi (its engine,
@code-yeongyu/senpi, ships bothomoandpibins), and--mode rpcspeaks Pi's JSONL protocol. So a custom provider gets it running with no core changes:{ "agents": { "providers": { "omo": { "extends": "pi", "label": "OmO", "command": ["~/.local/bin/omo", "--omo-task-usage-hint=false"], "enabled": true } } } }That works end to end: omo's full model catalog is discovered, streaming text/thinking/tool calls appear in the timeline, bash + edit tools run, multi-turn works, and sessions persist and resume after a daemon restart. The
--omo-task-usage-hint=falsepart is a workaround I'll explain below.Where Paseo gets in the way
It's undocumented guesswork. Nothing in Supported providers mentions omo, and the only way to find
extends: "pi"is to read the custom-providers docs and know omo is Pi-compatible. The provider shows up labelled "OmO" but the diagnostic panel still says "Pi" and checks~/.pi/agent/auth.json.Session import is broken, resume is not. Paseo's Pi provider reads sessions from
~/.pi/agent/sessions; omo writes~/.omo/agent/sessions/…. Resume works (Paseo passes the native session path back), butpaseo importcan't see terminal-started omo sessions.A boot-time custom message kills the turn. omo emits a
senpi-task.usagecustom message at session start. Paseo's Pi adapter treats any custommessage_endarriving beforeactiveTurnStartedas a completed turn (handleMessageEnd→if (!this.activeTurnStarted) this.completeTurn(...)), so the agent can go idle before its first token.--omo-task-usage-hint=falsesuppresses that one message and fixes it, but a user shouldn't have to know that flag exists.paseo runreturns before the agent finishes with the Pi adapter. Foregroundpaseo runreportscompletedwhile the agent is still running (verified: agent idle only ~40 s later). A 10 s-sleep prompt takes 14–16 s via OpenCode and ~5 s via omo. I believe this is await_for_finish/skipQueuerace rather than anything omo-specific —agent.create.updateresolves the CLI before the initial prompt dispatches — but it makes the Pi adapter path look broken from the CLI. Happy to file it separately as a bug if that's more useful.What a better workflow looks like
A first-class
omoentry inAGENT_PROVIDER_DEFINITIONS— idomo, label "OmO", no modes (omo has no permission modes, same as Pi) — that reuses the Pi adapter, plus asessionDir-style option so import can find~/.omo/agent/sessions.Minimum viable version, in order of value:
sessionDir/session-path option so import works.omoprovider id so it appears in the list without hand-written config.Protocol notes, in case they're useful
What I verified against omo 5.1.79 in
--mode rpc:get_state,get_commands,get_messages,get_available_models,set_model,set_thinking_level,compact,set_auto_compaction,abort,steer,clear_queue,get_session_stats,extension_ui_response,prompt(ack{"disposition":"started","turnId":"1"}).readyframe and nonegotiate_protocol(theompadapter waits on both), and it'sget_commands, notget_available_commands. Soextends: "omp"can't work.text,json,rpconly, soextends: "acp"is out too.~/.omo/agent/sessions/<slug>/<ts>_<uuid>.jsonl.All reactions