232 lines
7.0 KiB
Python
232 lines
7.0 KiB
Python
"""T4 交流文本 Workspace 单测（封闭，纯逻辑）。"""
|
||
import json
|
||
|
||
import pytest
|
||
|
||
from router_system.workspace import (
|
||
LIMITS,
|
||
STATUS_FLOW,
|
||
Workspace,
|
||
build_anchor,
|
||
estimate_tokens,
|
||
parse_anchor,
|
||
validate,
|
||
)
|
||
|
||
ALLOWED_BRIEF = {
|
||
"goal": "用 Python 实现快速排序并解释复杂度",
|
||
"constraints": ["必须用标准库", "时间复杂度 O(n log n)"],
|
||
"tags": ["code"],
|
||
"acceptance": [{"id": "a1", "check": "排序结果正确", "machine_checkable": True}],
|
||
"plan": [
|
||
{"id": "s1", "task": "实现快排主函数", "deps": [], "done_criteria": "可运行"},
|
||
{"id": "s2", "task": "补充单元测试", "deps": ["s1"], "done_criteria": "3/3 通过"},
|
||
],
|
||
}
|
||
|
||
|
||
def _ws(**kw):
|
||
return Workspace.new(request_id=kw.get("request_id", "abc123def456"),
|
||
query=kw.get("query", "写个快排"),
|
||
api_token_cap=kw.get("api_token_cap", 8000),
|
||
rounds_cap=kw.get("rounds_cap", 6))
|
||
|
||
|
||
# ---------- 结构/校验 ----------
|
||
def test_new_has_valid_defaults():
|
||
ws = _ws()
|
||
assert ws.status == "draft"
|
||
assert validate(ws.data) == []
|
||
b = ws.budget()
|
||
assert b["api_token_cap"] == 8000 and b["rounds_cap"] == 6
|
||
|
||
|
||
def test_brief_write_once():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
assert ws.status == "in_progress"
|
||
with pytest.raises(ValueError):
|
||
ws.apply_brief(ALLOWED_BRIEF) # 二次写入被拒
|
||
|
||
|
||
def test_apply_brief_missing_field_rejected():
|
||
ws = _ws()
|
||
bad = dict(ALLOWED_BRIEF)
|
||
del bad["plan"]
|
||
with pytest.raises(ValueError):
|
||
ws.apply_brief(bad)
|
||
|
||
|
||
def test_apply_brief_invalid_tag_rejected():
|
||
ws = _ws()
|
||
bad = dict(ALLOWED_BRIEF, tags=["hacker"])
|
||
with pytest.raises(ValueError):
|
||
ws.apply_brief(bad)
|
||
|
||
|
||
def test_apply_brief_overlong_goal_rejected():
|
||
ws = _ws()
|
||
bad = dict(ALLOWED_BRIEF, goal="长" * (LIMITS["goal"] + 1))
|
||
with pytest.raises(ValueError):
|
||
ws.apply_brief(bad)
|
||
|
||
|
||
def test_apply_brief_too_many_plan_steps_rejected():
|
||
ws = _ws()
|
||
plan = [{"id": f"s{i}", "task": "t", "deps": [], "done_criteria": "c"}
|
||
for i in range(LIMITS["plan_steps"] + 1)]
|
||
bad = dict(ALLOWED_BRIEF, plan=plan)
|
||
with pytest.raises(ValueError):
|
||
ws.apply_brief(bad)
|
||
|
||
|
||
# ---------- 状态机 ----------
|
||
def test_status_transitions():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
ws.transition("reviewing")
|
||
ws.transition("done")
|
||
assert ws.status == "done"
|
||
with pytest.raises(ValueError):
|
||
ws.transition("in_progress") # done 之后不允许再回
|
||
|
||
|
||
def test_illegal_transition():
|
||
ws = _ws()
|
||
with pytest.raises(ValueError):
|
||
ws.transition("done") # draft 不能直接到 done
|
||
|
||
|
||
# ---------- 写入 + 预算 ----------
|
||
def test_progress_and_issues_and_decisions():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
ws.add_progress("s1", "done", "实现完成", artifact="a://s1_main.py")
|
||
iid = ws.add_issue("s2", "a://s2_tests.py#L12-18",
|
||
"测试失败", "3/3 通过", "修了两次", "请裁决")
|
||
assert iid == "i1"
|
||
ws.add_decision(iid, "改用排序后断言", [{"id": "s2", "task": "修测试"}])
|
||
assert len(ws["progress"]) == 1
|
||
assert len(ws["issues"]) == 1
|
||
assert len(ws["decisions"]) == 1
|
||
|
||
|
||
def test_budget_accumulation_and_exhaustion():
|
||
ws = _ws(api_token_cap=10)
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
ws.add_budget(input_tokens=6, output_tokens=2)
|
||
assert ws.exhausted() is False
|
||
ws.add_budget(input_tokens=3) # 累计 11 > 10
|
||
assert ws.exhausted() is True
|
||
|
||
|
||
def test_round_exhaustion():
|
||
ws = _ws(rounds_cap=2)
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
ws.mark_round()
|
||
ws.mark_round()
|
||
assert ws.exhausted() is True
|
||
|
||
|
||
def test_issue_text_overlong_rejected():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
with pytest.raises(ValueError):
|
||
ws.add_issue("s1", "a://f.py", "观" * (LIMITS["issue_text"] + 1), "e", "t", "q")
|
||
|
||
|
||
# ---------- rollup ----------
|
||
def test_rollup_folds_done_progress_and_resolved_issues():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
ws.add_progress("s1", "done", "快排实现完成", "a://s1_main.py")
|
||
ws.add_progress("s2", "failed", "测试未过")
|
||
iid = ws.add_issue("s2", "a://s2_tests.py#L1", "失败", "通过", "试过", "问")
|
||
ws.add_decision(iid, "已修复")
|
||
n = ws.rollup()
|
||
assert n == 1 # 折叠 1 条 done progress
|
||
assert "s1" in "".join(ws["archive"])
|
||
# s2 failed 保留；已 resolve 的 issue 从 issues 移除
|
||
steps = [p["step"] for p in ws["progress"]]
|
||
assert steps == ["s2"]
|
||
assert len(ws["issues"]) == 0
|
||
|
||
|
||
def test_rollup_keeps_unresolved_issue():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
ws.add_progress("s1", "done", "ok", "a://s1.py")
|
||
ws.add_issue("s2", "a://s2.py#L1", "obs", "exp", "try", "ask") # 无 decision
|
||
ws.rollup()
|
||
assert len(ws["issues"]) == 1
|
||
|
||
|
||
# ---------- 渲染 ----------
|
||
def test_render_for_architect_within_budget():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
for i in range(6):
|
||
ws.add_progress(f"s{i}", "done", f"步骤{i}完成" * 3)
|
||
ws.add_issue(f"s{i}", f"a://f{i}.py#L1", "obs", "exp", "try", "ask" * 10)
|
||
ws.add_decision(f"i{i+1}", "裁决" * 20)
|
||
out = ws.render_for_architect()
|
||
assert estimate_tokens(out) <= 1200
|
||
assert "== meta ==" in out
|
||
|
||
|
||
def test_render_for_worker_has_step_and_criteria():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
out = ws.render_for_worker("s2", artifact_text="def test(): pass")
|
||
assert "当前步" in out
|
||
assert "s2" in out
|
||
assert "def test(): pass" in out
|
||
assert "验收标准" in out and "a1" in out
|
||
|
||
|
||
def test_render_for_worker_marks_current_step():
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
out = ws.render_for_worker("s1")
|
||
assert "<-- 当前步" in out
|
||
|
||
|
||
# ---------- 锚点 ----------
|
||
def test_anchor_build_and_parse():
|
||
a = build_anchor("s1_main.py", 12, 18)
|
||
assert a == "a://s1_main.py#L12-18"
|
||
p = parse_anchor(a)
|
||
assert p == {"file": "s1_main.py", "start": 12, "end": 18}
|
||
# 无行号
|
||
assert parse_anchor("a://f.py") == {"file": "f.py", "start": 1, "end": 1}
|
||
assert parse_anchor("not-an-anchor") is None
|
||
|
||
|
||
# ---------- 持久化 ----------
|
||
def test_save_load_roundtrip(tmp_path):
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
ws.add_progress("s1", "done", "ok")
|
||
path = tmp_path / "ws.json"
|
||
ws.save(path)
|
||
loaded = Workspace.load(path)
|
||
assert loaded.to_dict() == ws.to_dict()
|
||
assert loaded.request_id == ws.request_id
|
||
|
||
|
||
def test_json_serializable(tmp_path):
|
||
ws = _ws()
|
||
ws.apply_brief(ALLOWED_BRIEF)
|
||
ws.add_issue("s1", "a://f.py#L1", "obs", "exp", "try", "ask")
|
||
# 应能被 json 直接序列化（中文 ensure_ascii=False）
|
||
s = json.dumps(ws.to_dict(), ensure_ascii=False)
|
||
assert json.loads(s)["request_id"] == ws.request_id
|
||
|
||
|
||
def test_validate_rejects_non_object():
|
||
assert validate("nope") != []
|
||
|
||
|
||
def test_status_flow_enumeration():
|
||
assert set(STATUS_FLOW.keys()) == {"draft", "in_progress", "reviewing", "escalated", "done", "failed"}
|