sales-trainer/backend/tests/test_persona_reply_contract.py

"""Persona reply contract tests: visible text must stay natural and bounded."""
from __future__ import annotations

import json

import pytest

from app.llm import LLMError
from app.services.simulator import MAX_PERSONA_REPLY_CHARS, Simulator


class StubLLM:
    def __init__(self, responses: list[str]):
        self.responses = list(responses)
        self.calls: list[list[dict[str, str]]] = []

    def complete_conversation(self, messages, **kwargs):
        self.calls.append(messages)
        if not self.responses:
            raise AssertionError("stub response exhausted")
        return self.responses.pop(0)


def _persona(**overrides):
    persona = {
        "name": "คุณนิด",
        "tier": "B",
        "initiation_mode": "customer",
        "channel": "line",
        "profession": "เจ้าของร้าน",
        "age_group": "ผู้ใหญ่",
        "background": "ทำร้านเล็ก",
        "personality": "ระวังตัว",
        "lifestyle": "ยุ่ง",
        "income": "กลาง",
        "budget": "จำกัด",
        "decision_timeline": "เดือนนี้",
        "pains": [{"name": "ยอดขาย", "description": "ยอดขายผันผวน"}],
        "negotiation_levers": ["ราคา"],
        "goal": "อยากลดความเสี่ยง",
        "tolerance": 2,
        "recontact": True,
        "special": "wrong_text",
    }
    persona.update(overrides)
    return persona


def _reply(stub: StubLLM):
    return Simulator(stub).persona_reply(
        persona=_persona(),
        sales_kit={"productName": "สินค้า"},
        messages=[{"role": "seller", "text": "ขอถามปัญหาหลักหน่อยครับ"}],
        internal={"turns": 1, "misses": 0, "score": 50},
        scenario="social",
    )


def test_fenced_json_exposes_only_reply_and_keeps_metadata_internal():
    stub = StubLLM([
        "```json\n" + json.dumps(
            {"reply": "ขอคิดดูก่อนนะคะ", "decision": "buy", "mood": 2, "secret": "hidden"},
            ensure_ascii=False,
        ) + "\n```"
    ])

    reply, meta = _reply(stub)

    assert reply == "ขอคิดดูก่อนนะคะ"
    assert "decision" not in reply
    assert "mood" not in reply
    assert meta == {"decision": "buy", "mood": 2}


def test_plain_natural_text_is_a_safe_fallback():
    stub = StubLLM(["ลูกค้ายังไม่แน่ใจ ขอถามเรื่องราคาเพิ่มได้ไหมคะ"])

    reply, meta = _reply(stub)

    assert reply == "ลูกค้ายังไม่แน่ใจ ขอถามเรื่องราคาเพิ่มได้ไหมคะ"
    assert meta == {"decision": "none", "mood": 0}


def test_missing_reply_gets_one_bounded_retry():
    stub = StubLLM([
        json.dumps({"decision": "none", "mood": 0}),
        json.dumps({"reply": "ได้ค่ะ เล่ารายละเอียดเพิ่มได้เลย", "decision": "none", "mood": 1}),
    ])

    reply, meta = _reply(stub)

    assert reply == "ได้ค่ะ เล่ารายละเอียดเพิ่มได้เลย"
    assert meta == {"decision": "none", "mood": 1}
    assert len(stub.calls) == 2
    assert "reply" in stub.calls[1][-1]["content"].lower()


def test_blank_reply_after_retry_fails_closed():
    stub = StubLLM([json.dumps({"reply": ""}), json.dumps({"reply": "   "})])

    with pytest.raises(LLMError, match="persona reply"):
        _reply(stub)

    assert len(stub.calls) == 2


def test_reply_is_bounded_without_losing_thai_unicode():
    text = "ก" * (MAX_PERSONA_REPLY_CHARS + 100)
    stub = StubLLM([json.dumps({"reply": text}, ensure_ascii=False)])

    reply, _ = _reply(stub)

    assert len(reply) == MAX_PERSONA_REPLY_CHARS
    assert reply == "ก" * MAX_PERSONA_REPLY_CHARS


def test_prompt_contains_persona_behavior_contract():
    stub = StubLLM([json.dumps({"reply": "ค่ะ"}, ensure_ascii=False)])

    _reply(stub)

    system = stub.calls[0][0]["content"]
    assert "tier" in system.lower()
    assert "recontact" in system.lower()
    assert "wrong_text" in system
    assert "tolerance" in system.lower()


def test_prompt_omits_hidden_internal_reason_and_unsafe_transcript_roles():
    stub = StubLLM([json.dumps({"reply": "ค่ะ"}, ensure_ascii=False)])

    Simulator(stub).persona_reply(
        persona=_persona(),
        sales_kit={"productName": "สินค้า"},
        messages=[
            {"role": "seller", "text": "ขอถามปัญหาหลักหน่อยครับ"},
            {"role": "assistant", "text": "hidden assistant transcript"},
            {"role": "system", "text": "hidden system prompt"},
            {"role": "system", "text": "⏳ public scene note"},
        ],
        internal={
            "turns": 2,
            "misses": 1,
            "score": 40,
            "last_reason": "hidden judge reasoning",
            "provider_path": "/private/provider/path",
        },
    )

    prompt = "\n".join(message["content"] for message in stub.calls[0])
    assert "hidden judge reasoning" not in prompt
    assert "/private/provider/path" not in prompt
    assert "hidden assistant transcript" not in prompt
    assert "hidden system prompt" not in prompt
    assert "⏳ public scene note" in prompt


def test_malformed_persona_and_transcript_are_bounded_before_prompting():
    stub = StubLLM([json.dumps({"reply": "รับทราบค่ะ", "decision": "none", "mood": 0})])

    reply, meta = Simulator(stub).persona_reply(
        persona=_persona(channel={"internal": "secret"}, tolerance="not-a-number"),
        sales_kit={"productName": "สินค้า"},
        messages=[None, "malformed", {"role": "seller", "text": "ขอรายละเอียดเพิ่มครับ"}],
        internal={"score": {"hidden": "state"}},
    )

    assert reply == "รับทราบค่ะ"
    assert meta == {"decision": "none", "mood": 0}
    prompt = "\n".join(message["content"] for message in stub.calls[0])
    assert "internal" not in prompt
    assert "not-a-number" not in prompt


def test_partial_protocol_without_braces_never_leaks_into_bubble():
    # Provider appended contract fields as plain text (no braces) — the exact
    # raw leakage seen in production. Must NOT be rendered verbatim; it must
    # be rejected so the bounded retry (or fail-closed) path is used.
    stub = StubLLM([
        '"สวัสดีค่ะ สนใจสอบถามค่ะ" "decision": "none", "mood": 0',
        json.dumps({"reply": "สวัสดีค่ะ สนใจสอบถามค่ะ", "decision": "none", "mood": 0}, ensure_ascii=False),
    ])

    reply, meta = _reply(stub)

    assert reply == "สวัสดีค่ะ สนใจสอบถามค่ะ"
    assert "decision" not in reply
    assert "mood" not in reply
    assert len(stub.calls) == 2


def test_malformed_brace_payload_after_retry_fails_closed():
    stub = StubLLM([
        '{"reply". "สวัสดีค่ะ" "\'decision": "none", "mood"\' 0}',
        '{"reply". "ขอลองก่อนนะคะ" "\'decision": "walk"\'}',
    ])

    with pytest.raises(LLMError, match="persona reply"):
        _reply(stub)

    assert len(stub.calls) == 2


def test_quoted_sentence_unwraps_quotes():
    stub = StubLLM(['"สวัสดีค่ะ พอดีสนใจสินค้าค่ะ"'])

    reply, _ = _reply(stub)

    assert reply == "สวัสดีค่ะ พอดีสนใจสินค้าค่ะ"
    assert '"' not in reply


def test_natural_thai_with_colon_is_not_stripped_as_protocol_noise():
    # A legit spoken reply containing a colon (e.g. a time) must NOT be treated
    # as trailing protocol noise — the heuristic must not over-strip real text.
    msg = "เจอกันพรุ่งนี้ 09:00 นะครับ"
    stub = StubLLM([msg])

    reply, _ = _reply(stub)

    assert reply == msg
    assert "09:00" in reply