Skip to content

AGENTS.md

Latest commit

 

History

History
306 lines (242 loc) · 9.43 KB

AGENTS.md

File metadata and controls

306 lines (242 loc) · 9.43 KB

agents.md — Autonomous Agent Execution Contract (v3)

This document defines mandatory rules for any automated agent (LLM, code generator, bot) operating in this repository.

Violation of these rules produces incorrect output.

0. SOURCE OF TRUTH HIERARCHY

PRD.md > agents.md > existing code > your assumptions

Before writing, modifying, or suggesting ANY code:

  1. Read PRD.md — Product and architecture specification
  2. Read agents.md — Runtime and implementation contract
  3. Cross-reference the Vercel AI SDK docs — Authoritative API reference

Rules:

  • Do NOT reimplement functionality that Vercel AI SDK provides (see Section 2)
  • Do NOT invent custom streaming protocols (use AI SDK UI message stream via toUIMessageStreamResponse())
  • Do NOT create custom message types (use AI SDK UIMessage/ModelMessage)
  • Do NOT build custom LLM↔tool loops (use streamText())
  • Cite specific doc sections when referencing requirements

If conflict exists: Quote the exact text and stop.

1. RUNTIME CONTRACT: Bun-Only Architecture

Bun is the exclusive runtime. Node.js compatibility is explicitly deprioritized.

Mandatory Rules

✅ Use: Bun.serve(), Bun.file(), Bun.spawn()
✅ Use: bun:sqlite (local file mode, sync API)
❌ Never: Node fs, child_process, Express, Fastify
✅ Default: bun --watch engine/index.ts serve
❌ Never: npm start, nodemon, ts-node

2. VERCEL AI SDK CONTRACT (Critical)

PocketAgents is infrastructure for Vercel AI SDK, not a competing framework.

What Vercel AI SDK Owns (NEVER Reimplement)

streamText()                    — Agentic loop (LLM ↔ tool execution ↔ streaming)
toUIMessageStreamResponse()     — Standardized UI message stream response
useChat()                       — React hook: message state, streaming, tool approval
UIToolInvocation               — Client-side tool call state
ModelMessage / UIMessage        — Standard message formats
AI SDK UI Message Stream        — The wire format

What PocketAgents Owns

resolveAgent()                  — DB lookup: agent → model → provider key → tools → RAG
Tool HTTP proxy                 — executeHttpTool() with provider keys injection
Tool internal handlers          — search_memory
RAG pipeline                    — chunk → embed → store → search → inject
Event logging                   — Store run events to events table
Auth middleware                 — Admin/client key validation (Bearer token)
CRUD API routes                 — agents, models, provider keys, tools, collections
Admin console                   — React SPA using useChat() + TanStack Query
Single binary                   — bun build --compile

The Chat Pipeline (Entire Server-Side Streaming)

const resolved = await resolveAgent(agentId);     // 1. Resolve from DB (agent, model, provider key, tools)
const result = streamText({ ...resolved });        // 2. Vercel AI SDK does everything
return result.toUIMessageStreamResponse();         // 3. Return Response

Any code that reimplements what streamText() already does is wrong.

3. LANGUAGE: TypeScript Everywhere (Strict)

All code must be TypeScript. .js files are prohibited.

Type Requirements

✅ Explicit types for public APIs
✅ Use AI SDK types: UIMessage, ModelMessage, UIToolInvocation
✅ Interfaces for module boundaries
✅ No `any` unless documented exception
❌ Custom ChatMessage / ChatResponse types (use AI SDK's)
❌ Implicit `any` → wrong

File Structure

engine/
├── index.ts              # Single entrypoint (bun build --compile target)
├── agent-engine.ts       # resolveAdapter() — provider adapter factory
├── config.ts             # CLI flags → runtime config
├── api/                  # Service modules (CRUD + business logic)
│   ├── agents.ts
│   ├── chat.ts           # handleChat() — resolveAgent + streamText() + toDSR
│   ├── collections.ts
│   ├── keys.ts
│   ├── models.ts
│   ├── provider-keys.ts
│   ├── runs.ts           # handleRun() — logAndForward() async generator
│   └── tools.ts
├── server/
│   ├── routes.ts         # Bun.serve route handlers
│   ├── auth.ts           # Auth middleware
│   └── helpers.ts        # Response helpers
├── runtime/
│   ├── resolve.ts        # resolveAgent() — shared resolution pipeline
│   ├── tools.ts          # rowToServerTool() — DB → AI SDK tool definition
│   ├── rag.ts            # buildRagContext() — passive RAG injection
│   ├── embeddings.ts     # Embedding client
│   └── ingestion.ts      # Document chunking pipeline
├── db/
│   └── index.ts          # bun:sqlite connection + migrations
├── shared/
│   └── types.ts          # DB row types (AgentRow, ModelRow, ToolRow, etc.)
└── utils/
    ├── logger.ts
    ├── crypto.ts
    ├── provider-keys.ts
    └── validation.ts     # JSON Schema validation for tool args

4. DEPENDENCY POLICY: Approved Only

Approved

@ai-sdk/openai, @ai-sdk/anthropic, etc. (Model Providers)
 ai, @ai-sdk/react                           (AI SDK Core + React)
@tanstack/react-query, @tanstack/react-router         (UI data + routing)
bun:sqlite                                                  (database)
React + radix-ui + shadcn/ui + tailwind                (console UI)

Prohibited

ORMs (Drizzle, Prisma), test frameworks (Vitest), CSS frameworks (Chakra),
state managers (Zustand, Jotai), HTTP clients (axios), @libsql/client

5. AI SDK DATA STREAM PROTOCOL (Streaming Contract)

All chat/run endpoints MUST return the UI message stream via toUIMessageStreamResponse().

Wire Format

0: "Hello"\n
1: " world"\n
d: {"finishReason":"stop","usage":{"promptTokens":10,"completionTokens":5}}\n

SSE Wire Format

(Standard AI SDK UI message stream)

Endpoint Contract

POST /api/agents/:id/chat  → Content-Type: text/plain; charset=utf-8 (ephemeral)
POST /api/agents/:id/runs  → Content-Type: text/plain; charset=utf-8 (durable + logged)

6. MONITORING: Events Logging

The events table stores serialized run events. No translation layer.

events (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  run_id TEXT NOT NULL REFERENCES runs(id) ON DELETE CASCADE,
  type TEXT NOT NULL,           -- Platform event type
  payload_json TEXT NOT NULL,   -- Full event payload
  created_at INTEGER NOT NULL
);

Metrics are derived from event queries:

Tool success rate    → COUNT TOOL_CALL_END where result IS NOT NULL
Token usage          → SUM from RUN_FINISHED.usage
Agent success rate   → COUNT RUN_FINISHED / (COUNT RUN_FINISHED + COUNT RUN_ERROR)

7. UI CONTRACT

Chat = useChat() from @ai-sdk/react

const { messages, sendMessage } = useChat({ transport });

Messages = Message components

messages.map(message => (
  <MessageBubble key={message.id} message={message} />
))

CRUD = TanStack Query hooks (unchanged)

useQuery({ queryKey: ["agents"], queryFn: fetchAgents })

NEVER do:

❌ useState for chat messages (useChat manages state)
❌ Manual fetch to chat endpoints
❌ Custom streaming parsers (useChat handles streaming)
❌ Custom tool call state tracking (UIToolInvocation is authoritative)

8. TOOL EXECUTION CONTRACT

Bridge: DB → AI SDK Tool

const tool = {
  description: row.description,
  parameters: JSON.parse(row.parameters_json),
  execute: async (args) => executeHttpTool(row, args)
};

Schema Validation Before HTTP Execution

const error = validateToolArgs(args, schema);
if (error) return { error: `Schema validation failed: ${error}` };
// Only THEN call external URL

HITL Approval

  • Streaming: AI SDK native (addToolResult())
  • Async runs: Server-side (POST /api/runs/:id/approve)

9. EXECUTABLE REQUIREMENTS

✅ Single file: bun build --compile engine/index.ts → pocket-agents
✅ Startup: <2s cold start
✅ Size: <100MB
✅ Data: --data-dir → runtime.db + ./data/files/
✅ Ports: --port 4631 (default)

10. DEVELOPER WORKFLOW

Dev:    bun run dev          # bun --watch engine/index.ts serve
Check:  bun run check        # biome check --write
Format: bun run format       # biome format --write
Types:  bun run typecheck    # tsc --noEmit
Build:  bun run build        # bun build --compile → pocket-agents
Start:  bun run start        # production mode

11. STYLE GUIDE

✅ 2-space indent, semicolons, explicit returns
✅ Named functions for public APIs
✅ Comments explain WHY, not WHAT
✅ JSDoc for public module exports
✅ Max 400 LOC per file
✅ Single responsibility per module
❌ "utils.ts" catch-all files
❌ God files (routes.ts is the exception — route mappings are inherently flat)

SUMMARY (For LLMs)

1. PRD.md = architecture authority
2. Vercel AI SDK streamText() + toUIMessageStreamResponse() = streaming contract
3. PocketAgents = resolveAgent() + infrastructure (DB, RAG, auth, tools, monitoring)
4. NEVER reimplement: agentic loop, message state, streaming format, tool approval
5. Events table = run events (no translation)
6. UI chat = useChat() from @ai-sdk/react
7. Single binary, Bun-only, TypeScript strict

**Wrong output reimplements what AI SDK already provides.**