BREAKING CHANGE(api): Migrate public API to ai-sdk v6 and refactor core agent architecture: replace class-based DualAgent/Driver/Guardian with a single runAgent function; introduce ts_tools factories for tools, a compactMessages compaction subpath, and truncateOutput utility; simplify ToolRegistry to return ToolSet and remove legacy BaseToolWrapper/tool classes; update package exports and dependencies and bump major version.

This commit is contained in:
2026-03-06 11:39:01 +00:00
parent 903de44644
commit f9a9c9fb48
36 changed files with 3928 additions and 6586 deletions

View File

@@ -1,97 +1,51 @@
# Project Readme Hints
## Overview
`@push.rocks/smartagent` is a dual-agent agentic framework built on top of `@push.rocks/smartai`. It implements a Driver/Guardian architecture where the Driver proposes tool calls and the Guardian evaluates them against security policies.
`@push.rocks/smartagent` v2.0.0 is an agentic loop built on Vercel AI SDK v6 via `@push.rocks/smartai`. It wraps `streamText` with `stopWhen: stepCountIs(n)` for parallel multi-step tool execution.
## Architecture
- **DualAgentOrchestrator**: Main entry point, coordinates Driver and Guardian agents
- **DriverAgent**: Reasons about tasks, plans steps, proposes tool calls (supports both XML and native tool calling)
- **GuardianAgent**: Evaluates tool calls against configured policies
- **ToolRegistry**: Manages tool lifecycle, visibility, and discovery
- **BaseToolWrapper**: Base class for creating custom tools
- **plugins.ts**: Imports and re-exports smartai and other dependencies
## Architecture (v2)
- **`runAgent()`**: Pure async function — the core agentic loop. No class state.
- **`ToolRegistry`**: Lightweight helper for collecting tools into a `ToolSet`.
- **`truncateOutput()`**: Utility to prevent tool output from bloating context.
- **`compactMessages()`**: Context overflow handler (separate subpath export).
## Standard Tools (via registerStandardTools)
1. **FilesystemTool** - File operations with scoping and exclusion patterns
2. **HttpTool** - HTTP requests
3. **ShellTool** - Secure shell commands (no injection possible)
4. **BrowserTool** - Web page interaction via Puppeteer
5. **DenoTool** - Sandboxed TypeScript/JavaScript execution
## Additional Tools (must register manually)
6. **JsonValidatorTool** - JSON validation and formatting (NOT in registerStandardTools)
7. **ToolSearchTool** - AI-facing interface for tool discovery and activation
8. **ExpertTool** - Wraps a DualAgentOrchestrator as a specialized expert tool
## Tool Visibility System
Tools can be registered with visibility modes:
- **initial**: Always visible to Driver, included in system prompt (default)
- **on-demand**: Only discoverable via search, must be activated before use
```typescript
// Register with visibility options
orchestrator.registerTool(myTool, {
visibility: 'on-demand',
tags: ['database', 'sql'],
category: 'data'
});
## Source Layout
```
ts/ → core: runAgent, ToolRegistry, truncateOutput, interfaces
ts_tools/ → built-in tool factories (filesystem, shell, http, json)
ts_compaction/ → compactMessages helper for onContextOverflow
```
## Expert/SubAgent System
Experts are specialized agents wrapped as tools, enabling hierarchical agent architectures:
```typescript
orchestrator.registerExpert({
name: 'code_reviewer',
description: 'Reviews code for quality and best practices',
systemMessage: 'You are a code review expert...',
guardianPolicy: 'Allow read-only file access',
tools: [new FilesystemTool()],
visibility: 'on-demand',
tags: ['code', 'review']
});
```
## Tool Search
Enable tool discovery for the Driver:
```typescript
orchestrator.enableToolSearch();
// Driver can now use:
// - tools.search({"query": "database"})
// - tools.list({})
// - tools.activate({"name": "database_expert"})
// - tools.details({"name": "filesystem"})
```
## Key Features
- Token streaming support (`onToken` callback)
- Vision support (pass images as base64)
- Progress events (`onProgress` callback)
- Scoped filesystem with exclusion patterns
- Result truncation with configurable limits
- History windowing to manage token usage
- **Native tool calling mode** (`useNativeToolCalling: true`) for providers like Ollama
- **Tool visibility system** (initial vs on-demand)
- **Expert/SubAgent system** for hierarchical agents
- **Tool search and discovery** via ToolSearchTool
## Native Tool Calling
When `useNativeToolCalling` is enabled:
- Uses provider's built-in tool calling API instead of XML parsing
- Tool names become `toolName_actionName` (e.g., `json_validate`)
- Streaming includes `[THINKING]` and `[OUTPUT]` markers
- More efficient for models that support it
## Built-in Tools (ts_tools/)
Each exports a factory returning a flat `ToolSet` (Record<string, Tool>):
1. **filesystemTool()**`read_file`, `write_file`, `list_directory`, `delete_file`
2. **shellTool()**`run_command`
3. **httpTool()**`http_get`, `http_post`
4. **jsonTool()**`json_validate`, `json_transform`
## Key Dependencies
- `@push.rocks/smartai`: Multi-provider AI interface
- `@push.rocks/smartfs`: Filesystem operations
- `@push.rocks/smartshell`: Shell command execution
- `@push.rocks/smartbrowser`: Browser automation
- `@push.rocks/smartdeno`: Deno code execution
- `@push.rocks/smartrequest`: HTTP requests
- `minimatch`: Glob pattern matching for exclusions
- `@push.rocks/smartai` ^2.0.0 — provider registry, `getModel()`, re-exports `tool`, `jsonSchema`
- `ai` ^6.0.0 — Vercel AI SDK v6 (`streamText`, `stepCountIs`, `ModelMessage`, `ToolSet`)
- `zod` ^3.25.0 — tool input schema definitions
- `@push.rocks/smartfs`, `smartshell`, `smartrequest` — tool implementations
## AI SDK v6 Key APIs
- `streamText({ model, messages, tools, stopWhen: stepCountIs(20) })` — agentic loop
- `tool({ description, inputSchema: z.object({...}), execute })` — define tools
- `ModelMessage` — message type (replaces v4's `CoreMessage`)
- `LanguageModelV3` — model type from `@ai-sdk/provider`
- Result is `StreamTextResult` with PromiseLike properties (`await result.text`, etc.)
## Package Exports
- `.` → core (runAgent, ToolRegistry, truncateOutput, re-exports)
- `./tools` → built-in tool factories
- `./compaction` → compactMessages
## Build
- `pnpm build``tsbuild tsfolders --allowimplicitany`
- Cross-folder imports via each folder's `plugins.ts` (tsbuild unpack resolves them)
## Test Structure
- Tests use `@git.zone/tstest/tapbundle`
- Tests must end with `export default tap.start();`
- Tests must end with `export default tap.start()`
- `pnpm test``tstest test/ --verbose`