Class QueryParams

src/query/query.py:86–134

@dataclass
class QueryParams:
    messages: list[Message]
    # WI-1.1: ``system_prompt`` accepts either the legacy ``str`` shape
    # (joined sections, no cache_control markers) OR the block-list shape
    # ``list[dict]`` produced by ``build_full_system_prompt_blocks``. The
    # block-list shape is what engages Anthropic's prompt cache via
    # ``cache_control: {type: 'ephemeral'}`` markers; the str shape is
    # retained for backward compat with callers that pass a custom prompt.
    system_prompt: str | list[dict[str, Any]]
    tools: Tools
    tool_registry: ToolRegistry
    tool_use_context: ToolContext
    provider: BaseProvider
    abort_controller: AbortController
    query_source: str = "repl_main_thread"
    # ch05 round-3 G2: the +500k turn budget. Deliberate deviation from
    # TS's ambient bootstrap-global design — params is the carrier; the
    # bootstrap globals remain the mechanism (snapshot at query() entry).
    token_budget: int | None = None
    max_output_tokens_override: int | None = None
    max_turns: int | None = None
    user_context: dict[str, str] | None = None
    system_context: dict[str, str] | None = None
    pipeline_config: PipelineConfig | None = None
    # Ch5/F-followup: live streaming text callback. When set, the
    # provider's chat_stream_response receives this callback so each
    # SSE text-delta drives the UI in real time. Critical for the
    # TUI/headless live-stream UX after the F.2/F.3 migration to this
    # loop — without it, callers see the entire response materialize
    # at once after the model turn completes. The callback can also
    # raise AbortError from inside the SDK's stream context to tear
    # down the HTTP socket on ESC.
    on_text_chunk: Callable[[str], None] | None = None
    # Live thinking deltas (separate channel from on_text_chunk) for the TUI's
    # streaming thinking view. None → thinking isn't surfaced live.
    on_thinking_chunk: Callable[[str], None] | None = None

    # Extended thinking ("adaptive" mode) — opt the model into a private
    # reasoning scratchpad before producing its visible answer. Defaults
    # to ``None`` which auto-enables on Anthropic Claude 4.x models
    # (the only family the API supports it on) and stays off elsewhere.
    # Pass ``False`` to force-disable (e.g. for determinism in tests).
    # Mirrors the TS reference which always passes
    # ``thinking: {type: "adaptive"}`` on these models.
    extended_thinking: bool | None = None
    # Output-effort hint forwarded as ``output_config.effort``. Anthropic
    # accepts ``"low" | "medium" | "high"``. Only sent when extended
    # thinking is active.
    thinking_effort: str = "medium"


@dataclass