script-build-and-framework-tmp/script_build_host/tests/test_prompt_quality.py

from __future__ import annotations

from agent import get_preset

from script_build_host.agents.presets import register_script_presets
from script_build_host.agents.prompts import (
    COMPARISON_WORKER_PROMPT,
    COMPOSE_WORKER_PROMPT,
    DECODE_RETRIEVAL_WORKER_PROMPT,
    DIRECTION_WORKER_PROMPT,
    ELEMENT_SET_WORKER_PROMPT,
    EXTERNAL_RETRIEVAL_WORKER_PROMPT,
    KNOWLEDGE_RETRIEVAL_WORKER_PROMPT,
    PARAGRAPH_WORKER_PROMPT,
    PATTERN_RETRIEVAL_WORKER_PROMPT,
    PLANNER_PROMPT,
    ROOT_VALIDATOR_PROMPT,
    SCRIPT_CANDIDATE_VALIDATOR_PROMPT,
    STRUCTURE_WORKER_PROMPT,
    script_build_prompt_manifest,
)


def test_retrieval_roles_freeze_distinct_source_specific_prompts() -> None:
    retrieval_presets = {
        "script_pattern_retrieval_worker": (
            "pattern_retrieval_worker.md",
            PATTERN_RETRIEVAL_WORKER_PROMPT,
            "query_pattern_qa",
        ),
        "script_decode_retrieval_worker": (
            "decode_retrieval_worker.md",
            DECODE_RETRIEVAL_WORKER_PROMPT,
            "search_script_decode_case",
        ),
        "script_external_retrieval_worker": (
            "external_retrieval_worker.md",
            EXTERNAL_RETRIEVAL_WORKER_PROMPT,
            "external_search_case",
        ),
        "script_knowledge_retrieval_worker": (
            "knowledge_retrieval_worker.md",
            KNOWLEDGE_RETRIEVAL_WORKER_PROMPT,
            "search_knowledge",
        ),
    }
    manifest = {str(item["preset"]): item for item in script_build_prompt_manifest()}

    assert len({content for _, content, _ in retrieval_presets.values()}) == 4
    for preset, (filename, content, tool_name) in retrieval_presets.items():
        assert manifest[preset]["source_id"] == filename
        assert manifest[preset]["content"] == content
        assert tool_name in content
        assert "且仅再调用一次" in content


def test_registered_retrieval_presets_use_their_own_prompt() -> None:
    register_script_presets()
    expected = {
        "script_pattern_retrieval_worker": PATTERN_RETRIEVAL_WORKER_PROMPT,
        "script_decode_retrieval_worker": DECODE_RETRIEVAL_WORKER_PROMPT,
        "script_external_retrieval_worker": EXTERNAL_RETRIEVAL_WORKER_PROMPT,
        "script_knowledge_retrieval_worker": KNOWLEDGE_RETRIEVAL_WORKER_PROMPT,
    }
    for preset, prompt in expected.items():
        assert get_preset(preset).system_prompt == prompt


def test_creative_prompts_preserve_quality_and_convergence_contracts() -> None:
    assert "前期可探索不同的真实路径" in PLANNER_PROMPT
    assert "一个最主要的缺陷簇" in PLANNER_PROMPT
    assert "受众可感知结果" in DIRECTION_WORKER_PROMPT
    assert "不预写 Phase 2" in DIRECTION_WORKER_PROMPT
    assert "段落方案" in DIRECTION_WORKER_PROMPT
    assert "不是完成的正文" in STRUCTURE_WORKER_PROMPT
    assert "不可随意互换" in STRUCTURE_WORKER_PROMPT
    assert "受众最终能直接看到" in PARAGRAPH_WORKER_PROMPT
    assert "不是提纲、创作说明" in PARAGRAPH_WORKER_PROMPT
    assert "只是抽象评价" in ELEMENT_SET_WORKER_PROMPT
    assert "先按完全相同的 criteria 独立评价" in COMPARISON_WORKER_PROMPT
    assert "不在 Compose 阶段补写或润色" in COMPOSE_WORKER_PROMPT


def test_validator_prompts_require_audience_visible_semantic_evidence() -> None:
    dimensions = (
        "D1 Goal 实现",
        "D2 具体可执行",
        "D3 结构叙事",
        "D4 元素一致",
        "D5 主题与 persona",
        "D6 信息与事实",
    )
    for dimension in dimensions:
        assert dimension in SCRIPT_CANDIDATE_VALIDATOR_PROMPT
    assert "标签、自述和 Artifact URI 本身都不能证明内容质量" in (
        SCRIPT_CANDIDATE_VALIDATOR_PROMPT
    )
    assert "真实受众的阅读顺序" in ROOT_VALIDATOR_PROMPT
    assert "空 evidence、空正文、零独立 Element 或零 Link" in ROOT_VALIDATOR_PROMPT