279 lines
9.2 KiB
Python
279 lines
9.2 KiB
Python
"""Working/coding prompt language routing and format-string contracts."""
|
|
|
|
import json
|
|
from string import Formatter
|
|
|
|
import pytest
|
|
|
|
from hy_memory.coding.extractor import (
|
|
EXTRACT_PROMPT,
|
|
EXTRACT_PROMPT_CODING,
|
|
CodingMemoryExtractor,
|
|
)
|
|
from hy_memory.coding.fact_reconciler import (
|
|
CHAIN_SUMMARY_PROMPT,
|
|
FACT_RECONCILE_PROMPT,
|
|
build_chain_summary_prompt,
|
|
build_fact_reconcile_prompt,
|
|
WorkingFactReconciler,
|
|
)
|
|
from hy_memory.coding.reconciler import (
|
|
RECONCILE_PROMPT,
|
|
CodingMemoryReconciler,
|
|
build_experience_reconcile_prompt,
|
|
)
|
|
from hy_memory.coding.types import CodingMemory, CodingMemoryDraft, WorkingFact, WorkingFactDraft
|
|
from hy_memory.pipelines.base import ChatMessage
|
|
|
|
from tests.coding._helpers import MockLLMProvider
|
|
|
|
|
|
def _fields(template: str) -> set[str]:
|
|
return {
|
|
field_name
|
|
for _, field_name, _, _ in Formatter().parse(template)
|
|
if field_name
|
|
}
|
|
|
|
|
|
@pytest.fixture
|
|
def zh_prompts():
|
|
from hy_memory.coding.prompts_zh import (
|
|
CHAIN_SUMMARY_PROMPT_ZH,
|
|
EXTRACT_PROMPT_CODING_ZH,
|
|
EXTRACT_PROMPT_ZH,
|
|
FACT_RECONCILE_PROMPT_ZH,
|
|
RECONCILE_PROMPT_ZH,
|
|
)
|
|
|
|
return {
|
|
"extract": EXTRACT_PROMPT_ZH,
|
|
"extract_coding": EXTRACT_PROMPT_CODING_ZH,
|
|
"reconcile": RECONCILE_PROMPT_ZH,
|
|
"fact_reconcile": FACT_RECONCILE_PROMPT_ZH,
|
|
"chain_summary": CHAIN_SUMMARY_PROMPT_ZH,
|
|
}
|
|
|
|
|
|
def test_chinese_templates_match_english_placeholder_sets(zh_prompts):
|
|
pairs = {
|
|
"extract": EXTRACT_PROMPT,
|
|
"extract_coding": EXTRACT_PROMPT_CODING,
|
|
"reconcile": RECONCILE_PROMPT,
|
|
"fact_reconcile": FACT_RECONCILE_PROMPT,
|
|
"chain_summary": CHAIN_SUMMARY_PROMPT,
|
|
}
|
|
for name, english in pairs.items():
|
|
assert _fields(zh_prompts[name]) == _fields(english), name
|
|
|
|
|
|
def test_all_chinese_templates_format_without_errors(zh_prompts):
|
|
rendered = {
|
|
"extract": zh_prompts["extract"].format(
|
|
workspace_present="true",
|
|
branch_present="false",
|
|
conversation_block="用户：请整理报告",
|
|
),
|
|
"extract_coding": zh_prompts["extract_coding"].format(
|
|
workspace_present="true",
|
|
branch_present="true",
|
|
conversation_block="用户：请修复 api.py",
|
|
),
|
|
"reconcile": zh_prompts["reconcile"].format(
|
|
drafts_block="draft[0]",
|
|
candidates_block="（无候选）",
|
|
),
|
|
"fact_reconcile": zh_prompts["fact_reconcile"].format(
|
|
drafts_block="fact[0]",
|
|
candidates_block="（无候选）",
|
|
),
|
|
"chain_summary": zh_prompts["chain_summary"].format(
|
|
chain_block="第一阶段\n第二阶段",
|
|
),
|
|
}
|
|
assert all(rendered.values())
|
|
|
|
|
|
def test_chinese_templates_preserve_core_behavior_contracts(zh_prompts):
|
|
working = zh_prompts["extract"]
|
|
coding = zh_prompts["extract_coding"]
|
|
experience_reconcile = zh_prompts["reconcile"]
|
|
fact_reconcile = zh_prompts["fact_reconcile"]
|
|
chain_summary = zh_prompts["chain_summary"]
|
|
|
|
assert "等待用户" in working
|
|
assert '"experiences": []' in working
|
|
assert "不要为了满足本指令而杜撰" in working
|
|
assert "证据" in working and "不要" in working
|
|
assert "subject" in working and "不同" in working
|
|
|
|
assert "HARD CONSTRAINT" in coding
|
|
assert "BRANCH / WORKING LOCATION" in coding
|
|
assert "WHICH VERSION WON" in coding
|
|
assert "EXACT COORDINATES, NOT PASTED CODE" in coding
|
|
assert "VERIFIED vs MERELY WRITTEN" in coding
|
|
|
|
for action in ("ADD", "UPDATE", "DELETE", "SKIP"):
|
|
assert action in experience_reconcile
|
|
assert "明确证据" in experience_reconcile
|
|
assert "绝不要" in experience_reconcile
|
|
|
|
assert "SUPERSEDE" in fact_reconcile
|
|
assert "subject" in fact_reconcile
|
|
assert "同一个实体" in fact_reconcile
|
|
assert "保守" in fact_reconcile
|
|
|
|
assert "只能报告" in chain_summary
|
|
assert "证据状态" in chain_summary
|
|
assert "绝不要" in chain_summary
|
|
assert "当前状态" in chain_summary
|
|
|
|
|
|
@pytest.mark.parametrize("scene", ["working", "coding"])
|
|
def test_extractor_routes_chinese_new_user_message_to_chinese_prompt(scene, zh_prompts):
|
|
extractor = CodingMemoryExtractor(MockLLMProvider("{}"), scene=scene)
|
|
template, language = extractor._select_prompt_template([
|
|
ChatMessage(role="user", content="请修复 api_client.py 的 timeout 问题"),
|
|
ChatMessage(role="tool", content="Traceback: TypeError at line 100"),
|
|
])
|
|
expected = zh_prompts["extract_coding" if scene == "coding" else "extract"]
|
|
assert language == "zh"
|
|
assert template is expected
|
|
|
|
|
|
@pytest.mark.parametrize("scene", ["working", "coding"])
|
|
def test_extractor_keeps_english_prompt_when_chinese_term_appears_later(scene):
|
|
extractor = CodingMemoryExtractor(MockLLMProvider("{}"), scene=scene)
|
|
template, language = extractor._select_prompt_template([
|
|
ChatMessage(role="user", content="Please update the 中文 README file."),
|
|
])
|
|
expected = EXTRACT_PROMPT_CODING if scene == "coding" else EXTRACT_PROMPT
|
|
assert language == "en"
|
|
assert template is expected
|
|
|
|
|
|
def test_experience_reconcile_routes_from_new_drafts_not_candidates():
|
|
drafts = [CodingMemoryDraft(
|
|
task="修复搜索超时",
|
|
solution="修改 search_client.py 的 timeout 处理",
|
|
user_id="u1",
|
|
)]
|
|
candidates = [[CodingMemory(
|
|
experience_id="old",
|
|
user_id="u1",
|
|
task="Maintain the search client",
|
|
solution="Use the legacy retry helper",
|
|
)]]
|
|
prompt, language = build_experience_reconcile_prompt(drafts, candidates)
|
|
assert language == "zh"
|
|
assert prompt.startswith("你")
|
|
|
|
|
|
def test_experience_reconcile_keeps_english_for_english_leading_draft():
|
|
drafts = [CodingMemoryDraft(
|
|
task="Please update the 中文 README file",
|
|
solution="Keep the existing structure",
|
|
user_id="u1",
|
|
)]
|
|
prompt, language = build_experience_reconcile_prompt(drafts, [[]])
|
|
assert language == "en"
|
|
assert prompt.startswith("You")
|
|
|
|
|
|
def test_fact_reconcile_routes_from_new_fact_statement():
|
|
drafts = [WorkingFactDraft(
|
|
subject="api_client.py",
|
|
statement="请修复 timeout 问题；已定位到 retry 逻辑。",
|
|
)]
|
|
candidates = [[WorkingFact(
|
|
fact_id="old",
|
|
user_id="u1",
|
|
statement="The previous session inspected retry behavior.",
|
|
)]]
|
|
prompt, language = build_fact_reconcile_prompt(drafts, candidates)
|
|
assert language == "zh"
|
|
assert prompt.startswith("你")
|
|
|
|
|
|
def test_chain_summary_routes_from_current_statement():
|
|
from hy_memory.coding.prompts_zh import CHAIN_SUMMARY_PROMPT_ZH
|
|
|
|
prompt, language = build_chain_summary_prompt(
|
|
"(2026-10-11) The previous session inspected timeout behavior.\n"
|
|
"(now) 用户继续修复 timeout，并完成验证。",
|
|
current_statement="用户继续修复 timeout，并完成验证。",
|
|
)
|
|
assert language == "zh"
|
|
assert prompt == CHAIN_SUMMARY_PROMPT_ZH.format(
|
|
chain_block=(
|
|
"(2026-10-11) The previous session inspected timeout behavior.\n"
|
|
"(now) 用户继续修复 timeout，并完成验证。"
|
|
),
|
|
)
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
@pytest.mark.parametrize("scene", ["working", "coding"])
|
|
async def test_chinese_extractor_prompt_runs_and_parses_json(scene):
|
|
response = json.dumps({
|
|
"experiences": [{
|
|
"task": "修复搜索超时",
|
|
"search_keys": ["search timeout"],
|
|
"solution": "修改 api_client.py 并运行测试",
|
|
"boundary_envs": "",
|
|
"confidence": 0.9,
|
|
}],
|
|
"fact": {
|
|
"subject": "api_client.py",
|
|
"statement": "修复搜索超时并完成测试验证。",
|
|
"files": ["api_client.py"],
|
|
"details": "pytest passed",
|
|
},
|
|
}, ensure_ascii=False)
|
|
llm = MockLLMProvider(response)
|
|
extractor = CodingMemoryExtractor(llm, scene=scene)
|
|
|
|
drafts, facts = await extractor.extract(
|
|
[ChatMessage(role="user", content="请修复 api_client.py 的 timeout 问题")],
|
|
user_id="u1",
|
|
)
|
|
|
|
assert len(drafts) == 1
|
|
assert drafts[0].task == "修复搜索超时"
|
|
assert len(facts) == 1
|
|
assert facts[0].statement == "修复搜索超时并完成测试验证。"
|
|
assert "你" in llm.calls[0]["prompt"]
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_experience_reconciler_sends_chinese_prompt_to_llm():
|
|
llm = MockLLMProvider('{"ops": []}')
|
|
reconciler = CodingMemoryReconciler(None, None, llm)
|
|
drafts = [CodingMemoryDraft(
|
|
task="修复搜索超时",
|
|
solution="修改 api_client.py",
|
|
user_id="u1",
|
|
)]
|
|
|
|
await reconciler._llm_decide(drafts, [[]])
|
|
|
|
assert llm.calls[0]["prompt"].startswith("你")
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_fact_reconciler_sends_chinese_prompt_to_llm():
|
|
llm = MockLLMProvider('{"ops": []}')
|
|
reconciler = WorkingFactReconciler(None, None, llm)
|
|
drafts = [WorkingFactDraft(
|
|
subject="api_client.py",
|
|
statement="请修复搜索超时并完成验证。",
|
|
)]
|
|
|
|
await reconciler._llm_decide(drafts, [[WorkingFact(
|
|
fact_id="old",
|
|
user_id="u1",
|
|
statement="Previously inspected timeout behavior.",
|
|
)]])
|
|
|
|
assert llm.calls[0]["prompt"].startswith("你")
|