incorrect model is used despite configuration on startup

[RussBrown00]

I have multiple agents: Promethus, Sisyphus.... plus a couple a custom agents, that on startup are all using the incorrect configuration. The big one is Promethus is configured for Opus 4.8 and Sisyphus Sonnet 4.6, but they both use Sonos 4.6 anyway. I've also got custom ones configured to use haiku 4.6 for example and they startup as Opus 4.8... WTH?

I'm really hoping that I'm missing something, but it makes it hard to trust that any sub-agents are using the models that I want them too. Changing them in Opencode and then restarting changes nothing. I'm at a loss for what to do.

[RussBrown00]

If I update the actual agent YAML file with the model instead of using the oh-my-openagent.json file for my custom agents it seems to stick to the correct one.

I am using Meridian with a plugin... I'm having some issues with Meridian using the wrong models on passthrough, but that should effect how things are selected in Opencode

[VectorPeak]

This usually happens when there are two sources of truth for the same agent: the generated/cached agent YAML that opencode actually loads, and the oh-my-openagent.json config that oh-my-openagent uses to generate/update those files. If the YAML already exists, changing the model in opencode UI or only in oh-my-openagent.json may not affect the file being loaded at startup.

Check the resolved agent files directly and confirm the model field there:

grep -R "model:" ~/.config/opencode/agent ~/.config/opencode/agents 2>/dev/null
grep -R "Promethus\|Sisyphus" ~/.config/opencode ~/.local/share/opencode 2>/dev/null

If Promethus/Sisyphus show the wrong model in the YAML, edit that YAML or delete/regenerate the agent files from oh-my-openagent. If the YAML is correct but startup still uses the wrong model, then not sure about Meridian passthrough/model aliasing, worth checking whether Meridian is rewriting opus/sonnet/haiku aliases or applying a default model per provider.

[UnknownHawkins]

We'll produce maybe 2-3 short paragraphs. When oh-my-openagent starts it (re)generates those YAML files from the JSON, overwriting any manual changes you made in the UI.

[VectorPeak]

The quickest way to split this is to check the generated agent files after startup, not the Opencode UI state.

If the generated YAML for Prometheus/Sisyphus already contains the wrong model immediately after startup, the issue is in the oh-my-openagent.json → YAML generation path, or a default/model inheritance rule is winning over the per-agent value. In that case, look for a shared/default model field that applies to all agents, and make sure each custom agent has its own explicit model entry rather than relying on a partially inherited config.

If the generated YAML contains the correct model but Opencode still runs the wrong one, then oh-my-openagent is probably not the layer making the final choice anymore. That points more toward Opencode/provider/passthrough resolution - especially with Meridian in the middle - where model aliases or fallback behavior may be collapsing different requested models into the same backend model.

One practical debug step is to add a deliberately invalid model to one agent in oh-my-openagent.json, restart, and see where the failure happens. If startup/generation accepts it but the agent later falls back silently, that’s provider/runtime behavior. If the generated YAML normalizes or overwrites it before Opencode sees it, that’s on the config generation side.

Also worth checking that the generated YAML filenames match the agents you think are being loaded. A stale generated agent with the same display name can make it look like the right config is being edited while Opencode is loading a different file.

---

If my answer solved your problem, you can click answered the question. I'm really here to help, and along the way I'm also collecting Galaxy Brain badges haha 😆