Agent - Team Mode Choosing Wrong Model

[styles01]

Working on MacOS - via Terminal/Tmux - Ollama - it works fine when doing regular work, but when in agent-team mode with sub-panes, it does not resolve to Ollama and ollama is not in the model list.

err: "There is an issue with the selected model (claude-sonnet-4-6). It may not exist or you may not have a access to it. Run /model to pick a different model. "

And of course ollama doens't show when you run /model

[jedwhite]

Hey thanks for reporting this. Unfortunately, that's currently a limitation when using the Claude Code harness with ollama. Using agent teams mode looks to be Claude-model specific. It looks like the Claude harness itself here is choosing the model for the subagents. When regular --model override is set, it then gets used for everything (including small/fast model), but for sub-agents Claude doesn't appear to honour it.

Some of the Claude specific flags only work with a Claude subscription sadly. Other examples where the Claude harness drops support if it isn't running on a Claude subscription include web search and the --chrome browser support.

Interestingly, that appears to be true even using Anthropic's own API instead of the Anthropic subscription.

With airun, some shortcut flags are shared, but the underlying harness still sets the boundaries for support of features.

Let me do some investigation. Ollama cloud itself has web search support. But I don't think they have it when ollama is used with claude code's harness directly either.

The trick may end up being to make available some shared skills/plugins that unify support for things like browser use across model providers and interpreters. We just added Codex support as the harness (ai --codex) and we're adding more shortly (Pi, OpenCode etc). So one of the issues we'll have to work through is the variable support for features.

The current approach is very simple:

• There are some shared shortcuts that map between models/providers/interpreters (like ai --high)
• When there is a flag or argument that we can map between interpreters we'll try to map (ai --skip)
• Interpreter/provider specific flags/args get passed through to the hardness, and silently dropped if they aren't supported for an interpreter, with a stderr warning.

We'll likely run into more of these, and I'll update the docs to flag them as we find them!