Starting model and effort¶

ExOShell can pick the model and reasoning effort an agent starts with. Model IDs and supported effort levels depend on the agent, model, and provider; ExOShell does not maintain a model catalog or translate effort levels into a universal scale.

Set a starting model¶

Set optional model and effort strings in each [agents.codex], [agents.claude], or [agents.opencode] section. Only the selected agent's values apply. Omitting either setting preserves that agent's native defaults, including defaults supplied by your image.

[agents.codex]
# model = "example-model"
# effort = "high" # Codex: model_reasoning_effort

[agents.claude]
# model = "sonnet"
# effort = "high" # Claude Code: --effort

[agents.opencode]
# model = "openrouter/example-model"
# effort = "high" # OpenCode: model variant (provider/model#variant); requires model

Override for one launch¶

Override a configured value for one launch with --model VALUE or --effort VALUE before the launcher separator. --no-model and --no-effort ignore the corresponding ExOShell setting for that launch; they do not clear the agent's own image, project, or user configuration. Omitting a launcher option uses the configured value. For example:

./run-exoshell-agent.sh --effort low /path/to/project
./run-exoshell-agent.sh --no-model --no-effort /path/to/project

Explicit native options forwarded after -- take precedence over ExOShell's starting values. Codex accepts repeatable -c/--config overrides for model and model_reasoning_effort; native -m/--model also overrides the model. Claude accepts native --model and --effort. OpenCode accepts native -m/--model for run and mini (the full-screen TUI accepts neither; use ExOShell model/effort settings for it). Values after a native literal -- remain positional arguments. Verbose output reports ExOShell's selected values, not native overrides or the agent's eventual model selection.

Per-agent notes¶

Codex, Claude, and OpenCode run selections use native startup options. Users can change models and effort during the session.

OpenCode effort requires an explicit model from the config or launcher option (native --model cannot combine with a configured effort for run; express the variant natively as --model provider/model#variant). Interactive launches merge the model (with #variant when effort is set) into the built-in build and plan agents through disposable OPENCODE_CONFIG_CONTENT, preserving unrelated configuration and MCP settings. Custom OpenCode agents keep their own model configuration. opencode run receives --model after the subcommand instead.

Rebuild the base image and dependent images before using interactive OpenCode model/effort: its startup helper consumes the launcher's model/variant pair. See ADR-0036 and ADR-0040.