feat: 添加ObjectAI实现文档，定义AI代理和RAG管道的实施规则

Author: hotlong

packages/spec/llms.txt

@@ -133,6 +133,21 @@ function GridView({ config }: { config: z.infer<typeof ViewSchema> }) {
 }
 ```
 
+### ObjectAI Implementation (Intelligence Layer)
+*Reference: `node_modules/@objectstack/spec/prompts/implement-objectai.md`*
+
+AI Agents and RAG pipelines must be schema-aware and permission-safe.
+
+```typescript
+import { AgentSchema, RAGPipelineSchema } from '@objectstack/spec/ai';
+
+// Agents are configured via Protocol, not arbitrary prompts
+const supportAgent = AgentSchema.parse({
+  role: 'support',
+  tools: ['search_knowledge_base', 'create_ticket']
+});
+```
+
 ### 1. Strict Type Compliance
 Use `z.infer` to derive types directly from the protocol. Do not manually re-declare interfaces.
 

packages/spec/prompts/implement-objectai.md

@@ -0,0 +1,39 @@
+# ObjectAI Implementation Agent
+
+**Role:** You are the Lead AI Engineer building the `objectai` intelligent layer.
+**Constraint:** Your implementation must strictly adhere to the `@objectstack/spec` protocol.
+
+## 1. Setup
+
+You are working in a repository that depends on `@objectstack/spec`.
+Your source of truth is `node_modules/@objectstack/spec`.
+
+## 2. Implementation Rules
+
+### Rule #1: Protocol-Driven Intelligence
+AI Agents must be defined using `ai/agent.zod.ts`.
+- Do not create ad-hoc agent configurations.
+- Agents must expose their capabilities via standard `tools` definitions.
+
+### Rule #2: Schema-Aware RAG
+The RAG pipeline (`ai/rag-pipeline.zod.ts`) must obey ObjectQL security rules.
+- When retrieving documents, the AI **MUST** enforce `PermissionSchema`.
+- Do not allow the LLM to access data the user cannot see (RLS Enforcement).
+
+### Rule #3: Orchestration & chaining
+Complex workflows are orchestrated using `ai/orchestration.zod.ts`.
+- Steps must be strongly typed (input/output validated by Zod).
+- Use `ModelRegistrySchema` to select LLM providers (OpenAI, Anthropic, Local).
+
+## 3. Workflow
+
+1.  **Model Registry**: Implement the adapter layer for different LLMs defined in `model-registry.zod.ts`.
+2.  **Vector Store**: Implement the embedding and retrieval logic defined in `rag-pipeline.zod.ts`.
+3.  **Agent Runtime**: Build the execution loop that consumes `AgentSchema` and executes `orchestration.zod.ts` plans.
+
+## 4. Key Files to Watch
+
+- `ai/agent.zod.ts`: The Agent definition.
+- `ai/model-registry.zod.ts`: LLM provider configuration.
+- `ai/rag-pipeline.zod.ts`: Vector search and retrieval.
+- `ai/orchestration.zod.ts`: Tool execution and planning.