feat(server): add workbench API, DSH planner worker and RAG pipeline

2026-09-16 19:06:20 +08:00
parent 72fc5b47c7
commit d4b5b3f43a
122 changed files with 31906 additions and 0 deletions
@@ -0,0 +1,149 @@
+"""冻结并校验运行配置，避免历史任务被后续 Agent/Provider 编辑污染。"""
+from __future__ import annotations
+
+import hashlib
+import json
+
+from sqlmodel import Session, select
+
+from ..models import Agent, AgentPromptTemplate, AppUser, ModelCatalog, ModelProvider, now
+from .coordinator import _provider_model, provider_fingerprint
+from .skills import agent_skills
+from .memory import memory_candidates
+from ..digital_persona import freeze_persona
+from .default_composer import resolve_composer_agent
+
+
+def _digest(value: object) -> str:
+    raw = json.dumps(value, ensure_ascii=False, sort_keys=True, separators=(",", ":"), default=str)
+    return hashlib.sha256(raw.encode("utf-8")).hexdigest()
+
+
+def freeze_agent_snapshot(session: Session, agent: Agent, visited: set[int] | None = None,
+                          user: AppUser | None = None, allow_agent_scope: bool = False,
+                          allowed_agent_ids: set[int] | None = None) -> dict:
+    """返回无密钥的不可变 Agent 配置，嵌套融合 Agent 也一并冻结。"""
+    visited = visited or set()
+    if agent.id in visited:
+        return {"id": agent.id, "name": agent.name, "cycle": True}
+    visited.add(agent.id or 0)
+    provider: ModelProvider | None = None
+    model = agent.model_name
+    catalog: ModelCatalog | None = session.get(ModelCatalog, agent.model_catalog_id) if agent.model_catalog_id else None
+    try:
+        provider, model = _provider_model(session, agent)
+    except Exception:
+        provider = session.get(ModelProvider, agent.model_provider_id) if agent.model_provider_id else None
+    skills = [{"id": skill.id, "name": skill.name, "version": skill.version,
+               "instructions": skill.instructions} for skill in agent_skills(session, agent.id or 0)]
+    memories = memory_candidates(session, agent, user, allow_agent_scope=allow_agent_scope,
+                                 allowed_agent_ids=allowed_agent_ids)
+    # 人物表达统一由已绑定的数字人提供。旧模板仅保留身份元数据用于历史
+    # 审计和兼容快照，正文不会注入运行时，避免两套表达规则互相覆盖。
+    template = session.get(AgentPromptTemplate, agent.prompt_template_id) if agent.prompt_template_id else None
+    template_snapshot = {}
+    if template and template.enabled:
+        # 保留完整历史快照供审计/旧任务回放；当前 runner 会忽略其表达字段，
+        # 新任务的文风统一从数字人快照注入。
+        template_snapshot = {
+            "id": template.id, "name": template.name, "version": template.version,
+            "role_prompt": template.role_prompt, "task_template": template.task_template,
+            "constraints": list(template.constraints or []), "examples": list(template.examples or []),
+            "output_schema": dict(template.output_schema or {}), "style": template.style,
+        }
+    fusions = []
+    for fusion_id in agent.fusion_agent_ids or []:
+        fused = session.get(Agent, fusion_id)
+        if fused:
+            fusions.append(freeze_agent_snapshot(session, fused, visited.copy(), user=user,
+                                                 allow_agent_scope=allow_agent_scope,
+                                                 allowed_agent_ids=allowed_agent_ids))
+    snapshot = {
+        "id": agent.id, "name": agent.name, "role": agent.role,
+        "organization_id": agent.organization_id, "owner_user_id": agent.owner_user_id,
+        **freeze_persona(session, agent, user),
+        "rules": agent.rules, "rules_version": agent.rules_version,
+        "rules_hash": hashlib.sha256((agent.rules or "").encode("utf-8")).hexdigest(),
+        "kb_ids": list(agent.kb_ids or []),
+        "kb_category_ids": dict(getattr(agent, "kb_category_ids", {}) or {}),
+        "allowed_tools": list(agent.allowed_tools or []),
+        "provider_id": provider.id if provider else agent.model_provider_id,
+        "provider_name": provider.name if provider else "",
+        "provider_base_url": provider.base_url if provider else "",
+        "provider_fingerprint": provider_fingerprint(provider, model) if provider else "",
+        "model": model, "model_catalog_id": catalog.id if catalog else agent.model_catalog_id,
+        "model_catalog": {
+            "id": catalog.id, "name": catalog.model_name,
+            "display_name": catalog.display_name, "capabilities": list(catalog.capabilities or []),
+            "capability_status": catalog.capability_status, "rate_version": catalog.rate_version,
+        } if catalog else {},
+        "skills": skills, "memories": memories, "fusions": fusions,
+        "prompt_template": template_snapshot,
+    }
+    snapshot["snapshot_hash"] = _digest(snapshot)
+    return snapshot
+
+
+def freeze_stage_snapshot(session: Session) -> dict:
+    """一次运行中 QC 和组合阶段使用的完整快照。"""
+    qc: dict[str, dict] = {}
+    for gate, name in {
+        "structure": "结构质检", "content": "内容质检",
+        "style": "风格尺度质检", "cross": "跨稿质检",
+    }.items():
+        agent = session.exec(select(Agent).where(Agent.name == name)).first()
+        if agent:
+            qc[gate] = freeze_agent_snapshot(session, agent)
+    composer = resolve_composer_agent(session)
+    payload = {
+        "schema_version": 1,
+        "captured_at": now(),
+        "qc": qc,
+        "composer": freeze_agent_snapshot(session, composer),
+    }
+    payload["snapshot_hash"] = _digest(payload)
+    return payload
+
+
+def agent_from_snapshot(snapshot: dict) -> Agent:
+    """把冻结的数据转为运行时 Agent，不再读取当前的规则或技能绑定。"""
+    agent = Agent(
+        id=int(snapshot.get("id") or 0), name=str(snapshot.get("name") or "冻结 Agent"),
+        role=str(snapshot.get("role") or "service"), rules=str(snapshot.get("rules") or ""),
+        rules_version=int(snapshot.get("rules_version") or 1),
+        organization_id=int(snapshot.get("organization_id") or 1),
+        owner_user_id=int(snapshot.get("owner_user_id") or 0),
+        digital_human_id=snapshot.get("digital_human_id"),
+        persona_json=dict(snapshot.get("persona_json") or {}),
+        runtime_snapshot={"primary": snapshot},
+        kb_ids=list(snapshot.get("kb_ids") or []),
+        kb_category_ids=dict(snapshot.get("kb_category_ids") or {}), status="active",
+        allowed_tools=list(snapshot.get("allowed_tools") or []),
+    )
+    object.__setattr__(agent, "_frozen_persona_snapshot", snapshot)
+    return agent
+
+
+def skills_prompt_from_snapshot(snapshot: dict) -> str:
+    skills = list(snapshot.get("skills") or [])
+    if not skills:
+        return ""
+    lines = ["## 已冻结技能（按下列规则执行）"]
+    for skill in skills:
+        lines.append(f"### {skill.get('name') or '未命名技能'} v{skill.get('version') or 1}")
+        lines.append(str(skill.get("instructions") or ""))
+    return "\n".join(lines)
+
+
+def resolve_snapshot_provider(session: Session, snapshot: dict) -> tuple[ModelProvider, str]:
+    """只允许以原 Provider 和模型续跑；配置漂移必须由人工创建新运行。"""
+    provider_id = int(snapshot.get("provider_id") or 0)
+    model = str(snapshot.get("model") or "")
+    provider = session.get(ModelProvider, provider_id)
+    if not provider or not provider.enabled or not model:
+        raise ValueError("冻结的模型配置已不可用，请创建新运行后重试")
+    expected = str(snapshot.get("provider_fingerprint") or "")
+    actual = provider_fingerprint(provider, model)
+    if expected and expected != actual:
+        raise ValueError("Provider 配置已变更，已阻止使用新配置覆盖历史运行")
+    return provider, model