Function _split_system_prompt_blocks

src/query/query.py:419–457

Flatten system-prompt blocks for an OpenAI-compatible provider. Returns ``(system_text, volatile_tail_text)``. The ``__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__`` marker is always dropped (it is an Anthropic cache-only signal that would be unintelligible prose to other models). When ``

(
    blocks: list[dict[str, Any]], *, relocate_request_scope: bool
)

def _split_system_prompt_blocks(
    blocks: list[dict[str, Any]], *, relocate_request_scope: bool
) -> tuple[str, str]:
    """Flatten system-prompt blocks for an OpenAI-compatible provider.

    Returns ``(system_text, volatile_tail_text)``.

    The ``__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__`` marker is always dropped (it is
    an Anthropic cache-only signal that would be unintelligible prose to other
    models).

    When ``relocate_request_scope`` is True (DeepSeek only), blocks tagged
    ``_cache_scope == "request"`` — the env section, the auto-memory section
    (which embeds the mutable ``MEMORY.md`` body), plan-mode / non-interactive
    / tool-restriction sections — are routed into ``volatile_tail_text`` so the
    caller can place them AFTER the conversation history. That keeps the
    ``system + tools + history`` prefix byte-stable across turns, so DeepSeek's
    automatic prefix cache covers it even when memory or the environment
    changes mid-session.

    When False (every other provider), the tail is empty and all non-boundary
    text is concatenated into ``system_text`` — byte-for-byte the prior
    behaviour.
    """
    from ..context_system.cache_boundary import SYSTEM_PROMPT_DYNAMIC_BOUNDARY

    stable: list[str] = []
    volatile: list[str] = []
    for blk in blocks:
        if not isinstance(blk, dict):
            continue
        text = blk.get("text")
        if not text or text == SYSTEM_PROMPT_DYNAMIC_BOUNDARY:
            continue
        if relocate_request_scope and blk.get("_cache_scope") == "request":
            volatile.append(str(text))
        else:
            stable.append(str(text))
    return "\n\n".join(stable), "\n\n".join(volatile)


def _append_session_context_tail(