Skip to content

sensai

SensAIDeps dataclass

Python
SensAIDeps(
    job_status: JobStatus | None,
    column_manifest: ColumnManifest,
    current_plot_config: PlotConfig | None,
)

Runtime context available to the SensAI chat agent and every domain agent it delegates to.

Used by:

job_status instance-attribute

Python
job_status: JobStatus | None

The current job, or None if no job is loaded.

column_manifest instance-attribute

Python
column_manifest: ColumnManifest

Signal column names available in the current trial's telemetry data.

current_plot_config instance-attribute

Python
current_plot_config: PlotConfig | None

The plot config currently active in the trial viewer, or None if not set.

build_fallback_model

Python
build_fallback_model() -> FallbackModel

Builds the FallbackModel shared by every SensAI agent from SensAISettings.models, in the priority order the user configured there - the first (provider, model name) pair is tried first, and each later one is only used if every earlier one errors.

Each entry's Model comes from ModelProvider.build_model, which builds that entry's provider exactly as Agent('provider:model') would - reading whatever environment variable that provider looks for by default (see ModelProvider.env_var_name) - except for an OLLAMA entry, which gets SensAISettings.base_url explicitly: unlike most providers, pydantic_ai's OllamaProvider has no built-in localhost default and raises without one. Only OLLAMA gets this - passing it to a provider that also happens to accept base_url (e.g. Google) would silently redirect that provider's real endpoint to Ollama's URL instead.

Raises:

Type Description
ValueError

If SensAISettings.models is empty - FallbackModel needs at least one model to try.
Source code in src/mujoco_mojo/utils/layers/dojo/sensai/model.py
Python
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
def build_fallback_model() -> FallbackModel:
    """
    Builds the `FallbackModel` shared by every SensAI agent from `SensAISettings.models`, in the priority order the user configured there - the first `(provider, model name)` pair is tried first, and each later one is only used if every earlier one errors.

    Each entry's `Model` comes from `ModelProvider.build_model`, which builds that entry's provider exactly as `Agent('provider:model')` would - reading whatever environment variable that provider looks for by default (see `ModelProvider.env_var_name`) - except for an `OLLAMA` entry, which gets `SensAISettings.base_url` explicitly: unlike most providers, `pydantic_ai`'s `OllamaProvider` has no built-in `localhost` default and raises without one. Only `OLLAMA` gets this - passing it to a provider that also happens to accept `base_url` (e.g. Google) would silently redirect that provider's real endpoint to Ollama's URL instead.

    Raises:
        ValueError: If `SensAISettings.models` is empty - `FallbackModel` needs at least one model to try.

    """
    settings = MujocoMojoSettings().dojo.sensai
    if not settings.models:
        msg = "dojo.sensai.models is empty; configure at least one (provider, model name) pair."
        raise ValueError(msg)

    models = [
        entry.provider.build_model(
            entry.model_name,
            base_url=settings.base_url
            if entry.provider is ModelProvider.OLLAMA
            else None,
        )
        for entry in settings.models
    ]

    return FallbackModel(
        models[0],
        *models[1:],
        # pydantic-ai's Google model only wraps errors.APIError as ModelAPIError when
        # opening the connection, not when the first streamed chunk raises it (the actual
        # 503 case). Include the raw SDK error type too so streaming failures still fall back.
        # This only helps entries that are actually Google models - a non-Google-only
        # fallback chain gets no equivalent coverage for its own provider's raw SDK errors.
        fallback_on=(ModelAPIError, google_errors.APIError),
    )