Function _extract_routing_features

uncommon_route/proxy.py:1688–1836

Extract routing features from the current request step.

(
    body: dict,
    *,
    step_type: str,
    tool_names: list[str] | tuple[str, ...] | None = None,
    prompt: str = "",
    max_output_tokens: int = 4096,
    session_id: str | None = None,
)

def _extract_routing_features(
    body: dict,
    *,
    step_type: str,
    tool_names: list[str] | tuple[str, ...] | None = None,
    prompt: str = "",
    max_output_tokens: int = 4096,
    session_id: str | None = None,
) -> RoutingFeatures:
    """Extract routing features from the current request step."""
    messages = body.get("messages", [])
    raw_tools = body.get("tools") or body.get("customTools") or []
    normalized_tool_names = tuple(tool_names or ())
    if not normalized_tool_names:
        derived_tool_names: list[str] = []
        for tool in raw_tools:
            fn = tool.get("function") or tool.get("definition") or {}
            name = str(fn.get("name") or tool.get("name") or "").strip()
            if name:
                derived_tool_names.append(name)
        normalized_tool_names = tuple(derived_tool_names)

    has_vision = any(_has_vision_content(msg.get("content")) for msg in messages if isinstance(msg, dict))
    needs_tool_calling = bool(raw_tools)
    has_tool_results = step_type == "tool-result-followup"
    suggestion_mode = _is_suggestion_mode_prompt(prompt)
    title_generation_sidechannel = _body_has_title_generation_sidechannel(messages)

    response_format = body.get("response_format")
    response_format_name: str | None = None
    wants_structured_output = False
    if isinstance(response_format, dict):
        response_format_name = str(response_format.get("type") or "json_schema").strip().lower() or "json_schema"
        wants_structured_output = True
    elif isinstance(response_format, str):
        response_format_name = response_format.strip().lower() or None
        wants_structured_output = response_format_name in {"json", "json_schema"}
    if (
        not title_generation_sidechannel
        and not wants_structured_output
        and _body_has_structured_output_directive(messages)
    ):
        response_format_name = "system"
        wants_structured_output = True

    step_risk = (
        "low"
        if title_generation_sidechannel
        else _estimate_step_risk(
            messages=messages,
            step_type=step_type,
            tool_names=normalized_tool_names,
            prompt=prompt,
            needs_tool_calling=needs_tool_calling,
            wants_structured_output=wants_structured_output,
        )
    )
    contextual_floor = (