✨ feat: WP-06 工具域——注册表、四维裁决、审批与 RunBudget

- 迁移 10:tool(注册表:风险级/审批声明/输入 Schema)、tool_call(六态含 pending_approval、幂等唯一约束)、run.max_tool_calls(RunBudget 硬上限)
- 内置工具三件:get-current-time/calculate(AST 白名单四则,只读循律)、record-note(真实副作用:写入记忆候选,必须请命)
- 四维裁决(能力/数据/后果/情境→autonomy_level):read_only 循律直接执行;side_effect/high_impact 进入 pending_approval 请命;禁用工具 forbidden
- 审批流(朱批):approve → 执行并落审计;deny → 拒绝且不可逆(重复审批 409);副作用工具产物走既有记忆审核通道
- 工具入参按不可信模型输出处理:JSON Schema 子集校验(required/type/未知字段拒绝);RunBudget 超限 422 不静默放行
- 幂等:同 (run, tool, Idempotency-Key) 重放返回 200 原调用
- 测试 67 项全绿(+7):循律即执行、请命→朱批→真实候选产物、拒绝不可逆与不执行、Schema 三类拒绝、未注册 404、幂等重放、预算硬上限
This commit is contained in:
chenxin committed 2026-09-11 11:29:43 +08:00
1 parent e333dc5ea5
commit dfce60f7db
7 files changed
+669

No files matched your search

+59
View File
@@ -240,6 +240,7 @@ class Run(Base):
ForeignKey("conversation.id", ondelete="cascade"), nullable=True, index=True
)
status: Mapped[str] = mapped_column(String, nullable=False, default="queued")
max_tool_calls: Mapped[int] = mapped_column(nullable=False, default=10) # RunBudget 硬上限
created_at: Mapped[datetime] = mapped_column(
DateTime(timezone=True), nullable=False, server_default=func.now()
)
@@ -614,3 +615,61 @@ class BigCycle(Base):
"status IN ('pending', 'running', 'completed', 'failed')", name="status_values"
),
)
class Tool(Base):
"""工具注册表:内置工具与后续插件贡献统一登记(WP-06)。"""
__tablename__ = "tool"
name: Mapped[str] = mapped_column(String, primary_key=True) # kebab-case
description: Mapped[str] = mapped_column(String, nullable=False, default="")
input_schema: Mapped[dict[str, Any]] = mapped_column(JSONB, nullable=False, default=dict)
risk_level: Mapped[str] = mapped_column(String, nullable=False, default="read_only")
requires_approval: Mapped[bool] = mapped_column(nullable=False, default=False)
enabled: Mapped[bool] = mapped_column(nullable=False, default=True)
created_at: Mapped[datetime] = mapped_column(
DateTime(timezone=True), nullable=False, server_default=func.now()
)
__table_args__ = (
CheckConstraint(
"risk_level IN ('read_only', 'side_effect', 'high_impact')",
name="risk_level_values",
),
)
class ToolCall(Base):
"""工具调用:Schema 校验与 Policy 判定后的执行记录(含审批态)。"""
__tablename__ = "tool_call"
id: Mapped[str] = mapped_column(
String, primary_key=True, default=lambda: new_id("tool_call_")
)
run_id: Mapped[str] = mapped_column(
ForeignKey("run.id", ondelete="cascade"), nullable=False, index=True
)
tool_name: Mapped[str] = mapped_column(String, nullable=False)
input: Mapped[dict[str, Any]] = mapped_column(JSONB, nullable=False, default=dict)
output: Mapped[dict[str, Any] | None] = mapped_column(JSONB, nullable=True)
status: Mapped[str] = mapped_column(String, nullable=False, default="pending_approval")
error: Mapped[str | None] = mapped_column(String, nullable=True)
idempotency_key: Mapped[str | None] = mapped_column(String, nullable=True)
created_at: Mapped[datetime] = mapped_column(
DateTime(timezone=True), nullable=False, server_default=func.now()
)
completed_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
__table_args__ = (
CheckConstraint(
"status IN ('pending_approval', 'approved', 'denied', "
"'executing', 'succeeded', 'failed')",
name="status_values",
),
Index(
"uq_tool_call_idempotency", "run_id", "tool_name", "idempotency_key",
unique=True,
),
)