src/case_zero/api/scripted.py

"""Deterministic scripted interrogation for golden / seed cases (no LLM).

Reads the golden's per-suspect answer/delta tables directly, so the golden case plays
instantly and identically every time - ideal for the demo, tests, and offline CI. The
live in-process llama.cpp engine replaces this for generated cases (same wire result).
The deltas never reach the client; only the suspect's spoken reply + the resulting
server-side suspicion do.
"""

from __future__ import annotations

RATTLED_DELTA = 14
CORNERED_DELTA = 24


def _suspect(golden: dict, sus_id: str) -> dict:
    for s in golden["suspects"]:
        if s["id"] == sus_id:
            return s
    raise KeyError(sus_id)


def scripted_turn(
    golden: dict,
    sus_id: str,
    *,
    question_id: str | None = None,
    present_evidence_id: str | None = None,
    free_text: str | None = None,
) -> tuple[str, int]:
    """Return (spoken_reply, suspicion_delta) for one scripted turn."""
    suspect = _suspect(golden, sus_id)
    default = suspect.get("default", "I've told you what I know, detective.")

    if present_evidence_id is not None:
        entry = (suspect.get("present") or {}).get(present_evidence_id)
        if entry:
            return entry["a"], int(entry["d"])
        return default, 1

    if question_id is not None:
        for q in suspect.get("questions", []):
            if q["id"] == question_id:
                return q["a"], int(q["d"])
        return default, 2

    # free text: a small model would answer live; the scripted engine deflects.
    return default, 2