956d98652d
代码审查(新鲜上下文,只给 diff 与验收标准)报了五条影响正确性的,逐条核实全部成立:
1. spec_for() 交出去的 parameters 就是注册表内部那份真字典。docstring 承诺的快照只挡住了
「调用方改自己那份」,没挡住「从注册表取出来往里伸一层改」——而后者一下同时改掉模型
看见的 schema 和校验用的 schema。改成逐层冻成只读视图,schema_for_model 出口再化回
普通字典与列表。
2. {type: integer} 拒掉 3.0。JSON Schema draft-06 起小数部分为零的浮点数是合法整数,
模型写 1e2 时 json.loads 给的就是 float。这是我自己在注释里点名最怕的那种假阳性。
3. additionalProperties: false 撞上 patternProperties 时拒掉一切匹配 pattern 的键。
那些正是这份 schema 专门要收的键,模型改名也绕不过去。patternProperties 在场就跳过。
4. 工具名不校验类型、纯空白名放行。名字要落进发给模型的 schema,不是字符串会让整个请求
被网关拒掉,报错指向请求体不指向注册表。
5. _by_name 是可变 dict,两条查询路径能被就地改到分岔。换成只读视图。
不可哈希那条不修,改在 docstring 里写明(参数 schema 是映射,注册表放不进 set)。
scope.md 的行号引用换成条目名——行号是最容易漂的一种参数,插一行就静默指错。
0008 按一轮硕士生冷读重写:字段位置那段原来自相矛盾(一边说位置是公共承诺、插在中间会
静默改掉后面字段的含义,一边就插在中间,且没讨论追加在末尾这个同一判据下的显然选项),
改成追加在末尾并说明规则;补上四个名字的就地解释(重放策略、两条完成通路、动作结果五个
字段、restrict_to);决策四那张表原来只有三列却被正文说成填五个字段,恒定的两个单列出来
并各自给了理由;补上 validate 不通过为什么算未执行、executor() 为什么全查、为什么必须是
具体类而不是闭包、异常栈去哪了。
408 lines
16 KiB
Python
408 lines
16 KiB
Python
"""工具规格与注册表的行为。
|
||
|
||
这里断言的不是「注册表有哪些方法」,是它对外的行为——名字与签名本身由
|
||
`research-wiki/design/0006-public-names-and-signatures.md` 决策六承诺,改了是破坏性变更。
|
||
|
||
最要紧的一条是**四者同源**:模型看见的 schema、存在性校验、参数校验、分发都由同一个实例
|
||
驱动。收窄之后模型看不见的工具,校验也必须拒绝它——这一条单独写在下面,因为它正是这个
|
||
模块存在的理由。
|
||
"""
|
||
|
||
import json
|
||
|
||
import pytest
|
||
|
||
from polyloop.ports import ToolCall
|
||
from polyloop.tools import ToolRegistry, ToolSpec, ToolValidationError
|
||
from polyloop.types import ReplayPolicy
|
||
|
||
pytestmark = pytest.mark.unit
|
||
|
||
|
||
def _spec(name: str, **kwargs: object) -> ToolSpec:
|
||
"""造一个规格,只有 `name` 是必须自己填的。"""
|
||
parameters = kwargs.pop("parameters", {})
|
||
return ToolSpec(name=name, description=f"{name} 的说明", parameters=parameters, **kwargs) # type: ignore[arg-type]
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# ToolSpec
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_spec_rejects_an_empty_name() -> None:
|
||
"""空名字在提示词里不可见,模型永远调不到它。"""
|
||
with pytest.raises(ValueError, match="工具名"):
|
||
ToolSpec(name="", description="", parameters={})
|
||
|
||
|
||
def test_spec_rejects_a_blank_name() -> None:
|
||
"""纯空白的名字和空串一样,在提示词里不可见。"""
|
||
with pytest.raises(ValueError, match="工具名"):
|
||
ToolSpec(name=" ", description="", parameters={})
|
||
|
||
|
||
def test_spec_rejects_a_name_that_is_not_a_string() -> None:
|
||
"""名字要落进发给模型的那份 schema,不是字符串的话整个请求会被网关拒掉。
|
||
|
||
那时报错指向请求体、不指向注册表,而两者隔着好几层。
|
||
"""
|
||
with pytest.raises(TypeError, match="工具名"):
|
||
ToolSpec(name=123, description="", parameters={}) # type: ignore[arg-type]
|
||
|
||
|
||
def test_spec_rejects_a_description_that_is_not_a_string() -> None:
|
||
with pytest.raises(TypeError, match="工具说明"):
|
||
ToolSpec(name="search", description=object(), parameters={}) # type: ignore[arg-type]
|
||
|
||
|
||
def test_spec_rejects_non_mapping_parameters() -> None:
|
||
with pytest.raises(TypeError, match="parameters"):
|
||
ToolSpec(name="search", description="", parameters=["query"]) # type: ignore[arg-type]
|
||
|
||
|
||
def test_spec_snapshots_the_parameters_it_was_given() -> None:
|
||
"""注册之后改调用方那份字典,注册表看见的仍是注册那一刻的形状。
|
||
|
||
不拷贝的话,「模型看见的 schema」会随调用方在别处的一次修改一起变,而那次修改没有任何
|
||
地方记录得到——事后翻轨迹,模型当时到底看见的是哪一份,查不出来。
|
||
"""
|
||
schema: dict[str, object] = {"properties": {"query": {"type": "string"}}}
|
||
spec = _spec("search", parameters=schema)
|
||
|
||
schema["properties"] = {"query": {"type": "integer"}} # type: ignore[index]
|
||
|
||
assert spec.parameters["properties"] == {"query": {"type": "string"}}
|
||
|
||
|
||
def test_a_spec_taken_out_of_the_registry_cannot_be_edited_in_place() -> None:
|
||
"""从注册表里取出规格、往里伸一层去改,改不动。
|
||
|
||
只冻最外面一层挡不住这种改法,而它一下同时改掉模型看见的 schema 和校验用的 schema——
|
||
第 1 步模型看到的和第 5 步校验用的就分了岔,而这次修改没有任何地方记录得到。
|
||
"""
|
||
registry = ToolRegistry(
|
||
[
|
||
_spec(
|
||
"read",
|
||
parameters={"properties": {"path": {"type": "string"}}, "required": ["path"]},
|
||
)
|
||
]
|
||
)
|
||
taken = registry.spec_for("read")
|
||
assert taken is not None
|
||
|
||
with pytest.raises(TypeError):
|
||
taken.parameters["required"] = ["path", "mode"] # type: ignore[index]
|
||
with pytest.raises(TypeError):
|
||
taken.parameters["properties"]["path"]["type"] = "integer" # type: ignore[index]
|
||
|
||
registry.validate(ToolCall(name="read", arguments={"path": "a.txt"}))
|
||
|
||
|
||
def test_spec_defaults_are_the_conservative_ones() -> None:
|
||
"""重放策略默认「绝不重放」、完成标记默认「不完成」。
|
||
|
||
两个默认值的方向都是「漏了声明只会多停一次,有人会看见」,反过来则是静默重复副作用、
|
||
或者一次运行永远停不下来。
|
||
"""
|
||
spec = _spec("write_file")
|
||
|
||
assert spec.replay_policy is ReplayPolicy.NEVER
|
||
assert spec.completes_run is False
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# 注册与查询
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_registry_rejects_duplicate_names() -> None:
|
||
"""重名会让「模型看见的那一份」和「分发时取到的那一份」取决于注册顺序。"""
|
||
with pytest.raises(ValueError, match="重复"):
|
||
ToolRegistry([_spec("search"), _spec("search")])
|
||
|
||
|
||
def test_registry_rejects_things_that_are_not_specs() -> None:
|
||
with pytest.raises(TypeError):
|
||
ToolRegistry([{"name": "search"}]) # type: ignore[list-item]
|
||
|
||
|
||
def test_an_empty_registry_is_constructible() -> None:
|
||
"""不注册任何工具是一种正常装配:模型输出的是一整段代码,不是工具调用。"""
|
||
registry = ToolRegistry()
|
||
|
||
assert registry.names() == ()
|
||
assert list(registry.schema_for_model()) == []
|
||
|
||
|
||
def test_registration_order_is_preserved() -> None:
|
||
"""工具清单要贴进提示词,顺序必须是确定的。"""
|
||
registry = ToolRegistry([_spec("read"), _spec("write"), _spec("search")])
|
||
|
||
assert registry.names() == ("read", "write", "search")
|
||
assert [entry["name"] for entry in registry.schema_for_model()] == ["read", "write", "search"]
|
||
|
||
|
||
def test_spec_for_returns_none_for_an_unregistered_name() -> None:
|
||
"""问一个不在注册表里的名字是正常情形,不是错误。
|
||
|
||
没有工具的动作(一整段代码)就问不出规格,那时调用方取「绝不重放」。抛异常的话这条
|
||
正常路径要用捕获异常来走。
|
||
"""
|
||
registry = ToolRegistry([_spec("read")])
|
||
|
||
assert registry.spec_for("read") is not None
|
||
assert registry.spec_for("write") is None
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# 收窄
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_restrict_to_keeps_registration_order_not_the_caller_order() -> None:
|
||
"""收窄结果按注册顺序排,不按传进来的那个集合的迭代顺序。
|
||
|
||
收窄清单常常是一个 `set`,而 Python 的字符串哈希每进程随机——照集合顺序输出会让同一份
|
||
配置在不同进程里渲染出不同的提示词。
|
||
"""
|
||
registry = ToolRegistry([_spec("read"), _spec("write"), _spec("search")])
|
||
|
||
narrowed = registry.restrict_to({"search", "read"})
|
||
|
||
assert narrowed.names() == ("read", "search")
|
||
|
||
|
||
def test_restrict_to_leaves_the_original_alone() -> None:
|
||
"""注册表是不可变值对象,取子集返回新实例。"""
|
||
registry = ToolRegistry([_spec("read"), _spec("write")])
|
||
|
||
narrowed = registry.restrict_to(["read"])
|
||
|
||
assert narrowed.names() == ("read",)
|
||
assert registry.names() == ("read", "write")
|
||
|
||
|
||
def test_restrict_to_rejects_a_name_it_does_not_have() -> None:
|
||
"""静默丢弃的话,模型看不见调用方以为已经开放的那个工具,而表现是模型在绕圈。"""
|
||
registry = ToolRegistry([_spec("read")])
|
||
|
||
with pytest.raises(ValueError, match="不在注册表里"):
|
||
registry.restrict_to(["read", "reed"])
|
||
|
||
|
||
def test_two_registries_with_the_same_specs_are_equal() -> None:
|
||
"""相等按「注册了哪些规格、什么顺序」判,不按对象身份判。
|
||
|
||
构造运行请求时要比对「执行器持有的注册表」和「本次可见的注册表」,两份内容相同的注册表
|
||
在模型可见 schema 与实际分发上完全一致,没有可失败的地方。
|
||
"""
|
||
specs = [_spec("read"), _spec("write")]
|
||
|
||
assert ToolRegistry(specs) == ToolRegistry(specs)
|
||
assert ToolRegistry(specs) != ToolRegistry(list(reversed(specs)))
|
||
assert ToolRegistry(specs) != ToolRegistry([_spec("read")])
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# 模型可见的 schema
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_schema_for_model_is_json_serialisable() -> None:
|
||
"""这份清单要贴进提示词或者发给模型 API,必须能直接序列化。"""
|
||
registry = ToolRegistry(
|
||
[_spec("read", parameters={"properties": {"path": {"type": "string"}}})]
|
||
)
|
||
|
||
json.dumps(list(registry.schema_for_model()))
|
||
|
||
|
||
def test_mutating_the_returned_schema_does_not_touch_the_registry() -> None:
|
||
"""每次调用现造一份新的普通字典,调用方改它不会影响后面几步看见的东西。"""
|
||
registry = ToolRegistry(
|
||
[_spec("read", parameters={"properties": {"path": {"type": "string"}}})]
|
||
)
|
||
|
||
entry = registry.schema_for_model()[0]
|
||
entry["parameters"]["properties"]["path"]["type"] = "integer" # type: ignore[index]
|
||
|
||
assert registry.schema_for_model()[0]["parameters"] == {
|
||
"properties": {"path": {"type": "string"}}
|
||
}
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# 校验
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_validate_accepts_a_well_formed_call() -> None:
|
||
registry = ToolRegistry(
|
||
[
|
||
_spec(
|
||
"read",
|
||
parameters={
|
||
"properties": {"path": {"type": "string"}},
|
||
"required": ["path"],
|
||
"additionalProperties": False,
|
||
},
|
||
)
|
||
]
|
||
)
|
||
|
||
registry.validate(ToolCall(name="read", arguments={"path": "a.txt"}))
|
||
|
||
|
||
def test_validate_rejects_a_tool_that_is_not_registered() -> None:
|
||
registry = ToolRegistry([_spec("read")])
|
||
|
||
with pytest.raises(ToolValidationError, match="不存在"):
|
||
registry.validate(ToolCall(name="write", arguments={}))
|
||
|
||
|
||
def test_validate_rejects_a_missing_required_parameter() -> None:
|
||
registry = ToolRegistry(
|
||
[
|
||
_spec(
|
||
"read",
|
||
parameters={"properties": {"path": {"type": "string"}}, "required": ["path"]},
|
||
)
|
||
]
|
||
)
|
||
|
||
with pytest.raises(ToolValidationError, match="必填"):
|
||
registry.validate(ToolCall(name="read", arguments={}))
|
||
|
||
|
||
def test_validate_rejects_an_undeclared_parameter_only_when_the_schema_says_so() -> None:
|
||
"""多出来的键只在 schema 显式写了 `additionalProperties: false` 时才拒绝。
|
||
|
||
JSON Schema 里这一项默认为真,跟着它走;自作主张收紧的话,一份合法 schema 在库里和在别处
|
||
的含义就不一样了,而不一样的地方没有任何标记。
|
||
"""
|
||
lenient = ToolRegistry([_spec("read", parameters={"properties": {"path": {"type": "string"}}})])
|
||
strict = ToolRegistry(
|
||
[
|
||
_spec(
|
||
"read",
|
||
parameters={
|
||
"properties": {"path": {"type": "string"}},
|
||
"additionalProperties": False,
|
||
},
|
||
)
|
||
]
|
||
)
|
||
call = ToolCall(name="read", arguments={"path": "a.txt", "encoding": "utf-8"})
|
||
|
||
lenient.validate(call)
|
||
with pytest.raises(ToolValidationError, match="未声明"):
|
||
strict.validate(call)
|
||
|
||
|
||
def test_validate_rejects_a_wrong_top_level_type() -> None:
|
||
registry = ToolRegistry(
|
||
[_spec("read", parameters={"properties": {"path": {"type": "string"}}})]
|
||
)
|
||
|
||
with pytest.raises(ToolValidationError, match="类型"):
|
||
registry.validate(ToolCall(name="read", arguments={"path": 7}))
|
||
|
||
|
||
def test_a_boolean_is_not_an_integer() -> None:
|
||
"""布尔在 Python 里是整数的子类,在 JSON 里不是。
|
||
|
||
不排掉的话,一个声明收整数的工具会收下 `True`,然后在工具内部被当成 1 用。
|
||
"""
|
||
registry = ToolRegistry([_spec("head", parameters={"properties": {"n": {"type": "integer"}}})])
|
||
|
||
with pytest.raises(ToolValidationError, match="类型"):
|
||
registry.validate(ToolCall(name="head", arguments={"n": True}))
|
||
|
||
|
||
def test_a_float_with_no_fractional_part_counts_as_an_integer() -> None:
|
||
"""JSON Schema draft-06 起,小数部分为零的浮点数是合法的整数。
|
||
|
||
模型写出 `1e2` 或者 `3.0`,`json.loads` 给的就是 float。照「必须是 int」判会拒掉一次
|
||
合法调用,而模型怎么改都过不去——它写的东西按标准就是对的。
|
||
"""
|
||
registry = ToolRegistry([_spec("head", parameters={"properties": {"n": {"type": "integer"}}})])
|
||
|
||
registry.validate(ToolCall(name="head", arguments={"n": 3.0}))
|
||
with pytest.raises(ToolValidationError, match="类型"):
|
||
registry.validate(ToolCall(name="head", arguments={"n": 3.5}))
|
||
|
||
|
||
def test_pattern_properties_switches_off_the_unknown_key_check() -> None:
|
||
"""`patternProperties` 在场时,「哪些键被声明过」要靠正则才答得出,这里不答。
|
||
|
||
照 `properties` 的键去判的话,每一个匹配到 pattern 的键都会被拒——而那些正是这份 schema
|
||
专门要收的键,模型改名字也绕不过去。
|
||
"""
|
||
registry = ToolRegistry(
|
||
[
|
||
_spec(
|
||
"put",
|
||
parameters={
|
||
"patternProperties": {"^x_": {"type": "string"}},
|
||
"additionalProperties": False,
|
||
},
|
||
)
|
||
]
|
||
)
|
||
|
||
registry.validate(ToolCall(name="put", arguments={"x_1": "a"}))
|
||
|
||
|
||
def test_validate_rejects_arguments_that_are_not_a_mapping() -> None:
|
||
"""`ToolCall` 自己不校验字段,一个列表参数能从下游的解释器直接产出。
|
||
|
||
不在这里拦住的话,后面那句 `.items()` 会抛 `AttributeError` 打断整次运行;拦住了它就
|
||
只是一条正常观察,模型收到说明可以自己纠正。
|
||
"""
|
||
registry = ToolRegistry([_spec("read")])
|
||
|
||
with pytest.raises(ToolValidationError, match="映射"):
|
||
registry.validate(ToolCall(name="read", arguments=["a.txt"])) # type: ignore[arg-type]
|
||
|
||
|
||
def test_a_type_may_be_declared_as_a_list_of_alternatives() -> None:
|
||
registry = ToolRegistry(
|
||
[_spec("head", parameters={"properties": {"n": {"type": ["integer", "null"]}}})]
|
||
)
|
||
|
||
registry.validate(ToolCall(name="head", arguments={"n": 3}))
|
||
registry.validate(ToolCall(name="head", arguments={"n": None}))
|
||
with pytest.raises(ToolValidationError, match="类型"):
|
||
registry.validate(ToolCall(name="head", arguments={"n": "3"}))
|
||
|
||
|
||
def test_an_empty_schema_accepts_anything() -> None:
|
||
"""没声明参数就等于不约束参数,不等于「不许带参数」。"""
|
||
registry = ToolRegistry([_spec("ping")])
|
||
|
||
registry.validate(ToolCall(name="ping", arguments={"anything": [1, 2, 3]}))
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# 四者同源
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_narrowing_hides_a_tool_from_the_model_and_from_validation_together() -> None:
|
||
"""收窄之后模型看不见的工具,校验也拒绝它。
|
||
|
||
这是这个模块存在的理由:模型看见的 schema、存在性校验、参数校验、分发四者同源。不同源
|
||
的表现是「模型调了一个它看得见的工具却说不存在」,或者反过来——校验放行了一个分发时
|
||
找不到的名字。
|
||
"""
|
||
registry = ToolRegistry([_spec("read"), _spec("write")])
|
||
|
||
narrowed = registry.restrict_to(["read"])
|
||
|
||
assert [entry["name"] for entry in narrowed.schema_for_model()] == ["read"]
|
||
assert narrowed.spec_for("write") is None
|
||
with pytest.raises(ToolValidationError):
|
||
narrowed.validate(ToolCall(name="write", arguments={}))
|