M5：百工与司命——能力注册表、授契行契请命、禁域裁决与观察标证
- 百工域 domains/baigong：capabilities（统一能力契约：执行类型 builtin/qiankun/mcp/http + 四维风险标注 + 来源版本）、mandates（授契 grant / 行契 charter 预算时间窗，撤销立即生效）、tool_calls（净化摘要 + 观察标证）；请命复用 pending_confirmations 加 expires_at/brief_json（超时=拒绝执行 HJ-1302，六要素说明） - 司命裁决器 siming.py：四维取最高档（能力/数据/后果/情境）；禁域为代码常量不可配置（递归删根/格式化/写裸设备/fork 炸弹、file:// 协议、密钥材料外发）——专项测试断言恶意标注与篡改对象仍拒 HJ-1301 - 执行链：禁域→裁决（HJ-1303 未授契拒且双留痕）→请命→执行（乾坤经 SandboxClient HMAC；MCP 经 call_tool；内置无副作用）→净化标证（观察非事实，敏感遮蔽+体量截断）→行动账 - MCP 登记/启停/刷新自动派生与撤下 capabilities（博闻登记接入点）；内置能力种子（echo/datetime/qiankun.shell/web.fetch） - 端点：/capabilities /mandates /petitions /tool-calls + /tools 兼容投影；前端百工页四区（能力目录四维标注/授契行契/请命批准拒绝/行动账观察标证） - 退役：modules/safety|skills|tool_runtime|agent、application/tool_runtime、services/skills|safety、tool_policies/skills/model_endpoint_overrides 表（迁移 006）、前端 SkillManager/ToolsConsole/旧 tools 页 - 验收（test_baigong_siming 四用例）：未授契被拒留痕、请命超时不执行、禁域不可关闭、结果为观察不入忆；后端 454、前端 220、BFF 4、神识 5 包全绿（2 个预存环境失败不变）
This commit is contained in:
chenxin committed
1 parent
144415247e
commit
684261f4b4
97 files changed
+3075
-5554
No files matched your search
@@ -1,6 +1,5 @@
|
||||
"""运行时：Run / Event / Trace 与小周天执行状态（见后端架构计划 5.2）。"""
|
||||
"""运行时事件契约（Event/EventStore）：一炁事件流的数据结构。"""
|
||||
|
||||
from app.modules.runtime.events import Event, EventStore, InMemoryEventStore
|
||||
from app.modules.runtime.goals import build_goals_context
|
||||
|
||||
__all__ = ["Event", "EventStore", "InMemoryEventStore", "build_goals_context"]
|
||||
__all__ = ["Event", "EventStore", "InMemoryEventStore"]
|
||||
@@ -1,41 +0,0 @@
|
||||
"""运行时契约。"""
|
||||
|
||||
from typing import Any
|
||||
|
||||
from pydantic import BaseModel, ConfigDict, Field
|
||||
|
||||
from app.modules.agent import AgentDefinition
|
||||
|
||||
|
||||
class RuntimeContract(BaseModel):
|
||||
"""从业务上下文传递给 LLM 基础设施的结构化契约。"""
|
||||
|
||||
model_config = ConfigDict(arbitrary_types_allowed=True)
|
||||
|
||||
character_id: str
|
||||
session_id: str
|
||||
run_id: str
|
||||
model_hint: str
|
||||
agent: AgentDefinition | None = None
|
||||
system_instructions: list[str] = Field(default_factory=list)
|
||||
# 卡的 system_prompt（酒馆）：作道体增补层注入，绝不覆盖恒定层（§6 安全边界）
|
||||
system_prompt: str | None = None
|
||||
session_goal: str | None = None
|
||||
personality_summary: str | None = None
|
||||
# 道体核心使命（设计理念 §2.1）：对应太一 highest goal 的恒定快照，组装时不可被覆盖。
|
||||
mission: str | None = None
|
||||
# 道体安全边界：人物自身基础行为约束（区别于司命 safety_constraints 的动态治理）
|
||||
safety_guidelines: str | None = None
|
||||
# 世谱（Lorebook）注入：before/after 相对 personality_summary 的位置（酒馆 before_char/after_char）
|
||||
lorebook_before: str | None = None
|
||||
lorebook_after: str | None = None
|
||||
emotion_state: str | None = None
|
||||
memory_context: str | None = None
|
||||
temporal_context: str | None = None
|
||||
active_goals: str | None = Field(
|
||||
default=None,
|
||||
description="格式化后的活跃目标上下文（由 build_goals_context 生成）。",
|
||||
)
|
||||
recent_messages: list[Any] = Field(default_factory=list)
|
||||
safety_constraints: list[str] = Field(default_factory=list)
|
||||
trace_id: str
|
||||
@@ -1,36 +0,0 @@
|
||||
"""运行时目标上下文构建（纯函数，无 IO）。"""
|
||||
|
||||
from typing import Any
|
||||
|
||||
_LEVEL_LABEL: dict[str, str] = {
|
||||
"highest": "最高目标",
|
||||
"long_term": "长期目标",
|
||||
"session": "会话目标",
|
||||
"task": "任务目标",
|
||||
"action": "行动",
|
||||
}
|
||||
|
||||
|
||||
def build_goals_context(goals: list[dict[str, Any]]) -> str:
|
||||
"""把活跃目标列表格式化为 system prompt 片段。
|
||||
|
||||
按 priority 降序排列；空列表返回空串。
|
||||
"""
|
||||
if not goals:
|
||||
return ""
|
||||
|
||||
sorted_goals = sorted(goals, key=lambda g: g.get("priority", 0), reverse=True)
|
||||
lines: list[str] = ["当前活跃目标："]
|
||||
for g in sorted_goals:
|
||||
level_label = _LEVEL_LABEL.get(g.get("level", ""), g.get("level", ""))
|
||||
progress = g.get("progress", 0.0)
|
||||
bar = _progress_bar(progress)
|
||||
line = f"- [{level_label}] {g['title']} [{bar} {int(progress * 100)}%]"
|
||||
lines.append(line)
|
||||
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
def _progress_bar(value: float, width: int = 8) -> str:
|
||||
filled = int(value * width)
|
||||
return "■" * filled + "□" * (width - filled)
|
||||
@@ -1,30 +0,0 @@
|
||||
"""道体完整性校验。
|
||||
|
||||
组装运行时契约后，确认恒定层（mission / safety_guidelines）仍等于 Character 的值，
|
||||
未被任何动态注入（system_prompt / 世谱 / 眉批）在数据层覆盖或稀释。
|
||||
这是魂镜相对酒馆的核心安全增量——酒馆的 system_prompt 设计上完全替换前端 prompt，
|
||||
魂镜则保证道体恒定层不可被覆盖（设计理念 §2.2、融合设计 §6）。
|
||||
|
||||
注：字段层校验防组装错误与数据层覆盖；prompt 层的注入诱导（「忽略你的使命」）
|
||||
属另一类问题，不在此校验范围。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from app.modules.runtime.contract import RuntimeContract
|
||||
|
||||
|
||||
def validate_daoti_integrity(
|
||||
contract: RuntimeContract,
|
||||
*,
|
||||
mission: str | None = None,
|
||||
safety_guidelines: str | None = None,
|
||||
) -> None:
|
||||
"""校验契约道体字段仍等于 Character 恒定层；失配抛 ValueError。
|
||||
|
||||
``None`` 的道体字段跳过校验（未设定时不强求）。
|
||||
"""
|
||||
if mission is not None and contract.mission != mission:
|
||||
raise ValueError("道体完整性校验失败：mission 与 Character 恒定层不一致")
|
||||
if safety_guidelines is not None and contract.safety_guidelines != safety_guidelines:
|
||||
raise ValueError("道体完整性校验失败：safety_guidelines 与 Character 恒定层不一致")
|
||||
Reference in new issue
Block a user