Function decide_advisor_mode

src/utils/advisor.py:339–411

Pick activation mode for the upcoming turn. Returns one of: * ``ADVISOR_MODE_INACTIVE`` — no advisor on this request. * ``ADVISOR_MODE_SERVER_SIDE`` — Anthropic 1P beta path. * ``ADVISOR_MODE_CLIENT_SIDE`` — separate provider call from the tool dispatcher. Decision tree:

(
    provider: "BaseProvider | None",
    main_loop_model: str | None,
    advisor_model: str | None,
    *,
    force_client_mode: bool = False,
    advisor_provider: str | None = None,
    advisor_enabled: bool = True,
)

def decide_advisor_mode(
    provider: "BaseProvider | None",
    main_loop_model: str | None,
    advisor_model: str | None,
    *,
    force_client_mode: bool = False,
    advisor_provider: str | None = None,
    advisor_enabled: bool = True,
) -> str:
    """Pick activation mode for the upcoming turn.

    Returns one of:
    * ``ADVISOR_MODE_INACTIVE`` — no advisor on this request.
    * ``ADVISOR_MODE_SERVER_SIDE`` — Anthropic 1P beta path.
    * ``ADVISOR_MODE_CLIENT_SIDE`` — separate provider call from the
      tool dispatcher.

    Decision tree:

    1. ``advisor_model`` empty / env-disabled → INACTIVE.
    2. ``advisor_provider`` empty → INACTIVE (the multi-provider
       rewrite requires explicit provider; name-based inference was
       removed because the same model name can sit behind multiple
       providers).
    3. ``force_client_mode`` set → CLIENT_SIDE iff the advisor
       provider is a configured key; else INACTIVE.
    4. 1P + main_loop_model supports server advisor + advisor_model is
       a valid server target + advisor_provider == "anthropic" →
       SERVER_SIDE (the optimized path; one roundtrip, prompt-cache
       friendly). Server-side only makes sense when the advisor call
       lands on the same Anthropic API as the main loop.
    5. Otherwise, if the advisor provider is configured → CLIENT_SIDE.
    6. Else INACTIVE — the configured advisor can't be reached.

    ``advisor_enabled`` is the master switch (settings ``advisor_enabled``,
    default False in production): when False the advisor is INACTIVE regardless
    of model/provider. The parameter defaults True so direct callers (the
    activation truth-table tests) keep their behavior; production call sites pass
    ``get_settings().advisor_enabled``.
    """
    if not advisor_enabled:
        return ADVISOR_MODE_INACTIVE
    if _env_truthy(_DISABLE_ENV):
        return ADVISOR_MODE_INACTIVE
    if not advisor_model:
        return ADVISOR_MODE_INACTIVE
    if not advisor_provider:
        return ADVISOR_MODE_INACTIVE

    # Provider must be configured in ~/.clawcodex/config.json. Use the
    # provider class registry as the lightweight check (a key with no
    # class registered can't be instantiated anyway).
    advisor_routes = False
    try:
        from src.providers import get_provider_class
        get_provider_class(advisor_provider)
        advisor_routes = True
    except Exception: