Function build_effective_system_prompt

src/query/agent_loop_compat.py:176–257

Assemble the cold-start system prompt for the headless+TUI cutover. Returns the FULL system prompt as a **block list** (``list[dict]``) so the ``query()`` loop still engages prompt caching: the canonical base sections (``build_full_system_prompt_blocks`` — intro / # Doing tasks / # Exec

(
    style_prompt: str,
    tool_context: ToolContext,
    *,
    provider: Any | None = None,
    mcp_servers: list[Any] | None = None,
    query_source: str = "main",
)

def build_effective_system_prompt(
    style_prompt: str,
    tool_context: ToolContext,
    *,
    provider: Any | None = None,
    mcp_servers: list[Any] | None = None,
    query_source: str = "main",
) -> list[dict[str, Any]]:
    """Assemble the cold-start system prompt for the headless+TUI cutover.

    Returns the FULL system prompt as a **block list** (``list[dict]``) so the
    ``query()`` loop still engages prompt caching: the canonical base sections
    (``build_full_system_prompt_blocks`` — intro / # Doing tasks / # Executing
    actions / # Using your tools / # Tone / output-efficiency / env / memory /
    skills / …) with the resolved output-style prompt **appended**, then the
    existing workspace + git + CLAUDE.md context preserved as a trailing
    (uncached) block.

    Why this exists: the TUI (``tui/agent_bridge.py``) and headless
    (``entrypoints/headless.py``) cutover routes through ``query()``, which
    passes ``params.system_prompt`` **verbatim** to the model — it has no base
    build of its own. The engine/REPL path (``engine.py:124-188``) does the
    canonical build when ``system_prompt`` is unset, but the cutover pre-sets
    it, so before this fix the live TUI/headless agent received **no base
    instructions** at all (only the style line + context). This helper restores
    them, mirroring the engine's ``build_full_system_prompt_blocks`` +
    ``append_system_prompt`` shape.

    CLAUDE.md note: ``build_full_system_prompt_blocks``' memory section is
    *auto-memory* (``MEMORY.md`` via ``load_memory_prompt``), **not** CLAUDE.md.
    On the engine path CLAUDE.md is injected into the *messages* via
    ``prepend_user_context``; the cutover does not do that, so we keep
    ``build_context_prompt`` (which emits ``## Project Instructions``) to
    preserve CLAUDE.md — option (b) in
    ``my-docs/get-parity-by-folder/live-base-system-prompt-gap-analysis.md``.
    This overlaps the base ``# Environment`` section on CWD/date (a benign,
    documented duplication).

    ``tools``/``tool_registry`` are deliberately NOT passed to
    ``build_full_system_prompt_blocks`` (matching ``engine.py:167``): tool
    schemas reach the model via the API ``tools=`` param, so emitting a prose
    tool-docs section here would double-send them. ``provider``/``mcp_servers``
    feed only the global cache-scope gate; ``None`` is safe (disables the
    cross-user global scope, which TUI/headless should not use).
    """
    # Local imports — context_system is a heavier dep; only the cutover
    # callers need it, no need to drag it into agent_loop_compat's import time.
    from ..context_system import build_context_prompt
    from ..context_system.prompt_assembly import build_full_system_prompt_blocks

    cwd = str(tool_context.cwd or tool_context.workspace_root)

    # Skills listing (best-effort; mirrors engine.py:183).
    try:
        from ..command_system import get_skill_tool_commands
        skills = get_skill_tool_commands(cwd)
    except Exception:
        skills = None