mirror of https://github.com/Darkatse/TauriTavern.git synced 2026-10-19 07:35:14 +00:00

Darkatse f76139f961 feat(agent): Agent Phase 1 runtime foundation

- Introduce the Agent domain model, run/workspace/checkpoint repositories,
and a file-backed workspace store with event journaling.

- Wire the runtime service into Tauri commands and expose the frontend Agent
API for starting, observing, cancelling, reading, and committing runs.

- Derive workspaces from stable chat identity while keeping each run isolated,
and make unsupported Phase 1 paths fail fast.

- Document the Agent architecture, API contract, workspace model, LLM gateway,
tool/MCP boundaries, and implementation plan.

2026-04-26 14:28:08 -04:00

3.1 KiB

Raw Blame History

`window.TAURITAVERN.api.mcp` — MCP API Draft

本文档是 MCP Host ABI 草案。MCP 是独立平台能力，Agent 只是它的消费者之一。

状态：规划中，尚未实现。

1. 入口

await (window.__TAURITAVERN__?.ready ?? window.__TAURITAVERN_MAIN_READY__);

const mcp = window.__TAURITAVERN__.api.mcp;

2. 方法概览

type TauriTavernMcpApi = {
  listServers(): Promise<McpServerSummary[]>;
  connectServer(input: McpConnectServerInput): Promise<McpServerStatus>;
  disconnectServer(serverId: string): Promise<void>;
  listTools(serverId: string): Promise<McpToolSummary[]>;
  callTool(input: McpCallToolInput): Promise<McpToolResult>;
  listResources(serverId: string): Promise<McpResourceSummary[]>;
  readResource(input: McpReadResourceInput): Promise<McpResourceContent>;
  listPrompts(serverId: string): Promise<McpPromptSummary[]>;
  getPrompt(input: McpGetPromptInput): Promise<McpPromptContent>;
};

3. Server

type McpServerSummary = {
  id: string;
  displayName: string;
  transport: 'stdio' | 'http' | 'sse';
  enabled: boolean;
  connected: boolean;
  capabilities: {
    tools?: boolean;
    resources?: boolean;
    prompts?: boolean;
    sampling?: boolean;
  };
};

connectServer() 不应接受任意 command 字符串作为扩展/Agent 可传入参数。stdio server 配置必须来自用户设置或 allowlist。

4. Tools

type McpToolSummary = {
  serverId: string;
  name: string;
  title?: string;
  description?: string;
  inputSchema?: unknown;
  annotations?: unknown;
  approvalRequired?: boolean;
};

type McpCallToolInput = {
  serverId: string;
  name: string;
  arguments?: unknown;
  approvalToken?: string;
};

语义：

危险工具默认需要审批。
call result 可以被 Agent 映射为 ToolResult。
call 必须有超时与取消策略。

5. Resources

type McpResourceSummary = {
  serverId: string;
  uri: string;
  name?: string;
  description?: string;
  mimeType?: string;
};

type McpReadResourceInput = {
  serverId: string;
  uri: string;
};

MCP resource 不应自动进入 prompt。宿主、用户、profile 或 preset 决定是否纳入 ContextFrame。

6. Prompts

type McpPromptSummary = {
  serverId: string;
  name: string;
  title?: string;
  description?: string;
  arguments?: unknown;
};

type McpGetPromptInput = {
  serverId: string;
  name: string;
  arguments?: unknown;
};

MCP prompt 可以成为 PromptComponent，但不能覆盖 TauriTavern preset policy。

7. Security Contract

禁止：

Agent/Preset/角色卡/世界书直接写 MCP stdio command。
从远端 config 自动创建本地 stdio server。
初期启用 MCP Sampling 自动模型调用。
未经审批调用 destructive/high-cost tool。

要求：

server/command/args 对用户可见。
per-server capability allowlist。
per-tool approval。
MCP call 写入 Agent journal 或 MCP audit log。

8. Agent Integration

Agent 看到的是：

mcp.<server>.<tool> -> ToolRegistry ToolSpec
mcp resource        -> WorkspaceResource / ContextFrame
mcp prompt          -> PromptComponent

Agent 不直接操作 MCP config。

3.1 KiB Raw Blame History Unescape Escape

window.__TAURITAVERN__.api.mcp — MCP API Draft