Agentsy

Production-ready LLM stream parsing and VS Code integration for multi-step agent workflows.

🎯 Overview

Agentsy is a Turbo monorepo with independent, composable packages for processing LLM responses in agent systems. Each package is published separately to npm under the @agentsy scope.

📦 Packages

• @agentsy/core (v0.1.0) — Stream parsing & orchestration (foundation layer). npm | docs
• @agentsy/vscode (v0.1.0) — VS Code integration utilities. npm

🚀 Coming Soon

• @agentsy/agents — Multi-step LLM agents with memory, planning, and reflection
• @agentsy/tools — Standardized tool definitions for code execution, web search, and file I/O
• @agentsy/cli — Command-line agent runner with live streaming and JSONL output
• @agentsy/ui — React components for agent progress, thinking visualization, and tool interaction

⚡ Quick Start

Stream Parsing

Parse LLM responses chunk-by-chunk with @agentsy/core:

npm install @agentsy/core

import { LLMStreamProcessor } from '@agentsy/core/processor';

const processor = new LLMStreamProcessor({
  parseThinkTags: true,
  knownTools: new Set(['search', 'edit_file']),
});

processor.on('thinking', delta => console.log(`[💭] ${delta}`));
processor.on('text', delta => process.stdout.write(delta));
processor.on('tool_call', call => executeTool(call));

for await (const chunk of llmStream) {
  processor.process({
    content: chunk.content,
    done: chunk.done,
  });
}

VS Code Provider

Build Language Model Chat Providers with @agentsy/vscode:

npm install @agentsy/vscode @agentsy/core vscode

import { BaseLanguageModelChatProvider } from '@agentsy/vscode';

export class MyProvider extends BaseLanguageModelChatProvider {
  constructor(context: ExtensionContext) {
    super(context, {
      providerId: 'my-provider',
      vendor: 'MyVendor',
      displayName: 'My Provider',
      maxInputTokens: 4096,
      supportedCapabilities: ['thinking', 'tool-calls'],
    });
  }

  protected async buildRequest(messages, request) {
    return { model: request.model.id, messages };
  }

  protected async callApi(request) {
    // Call your LLM API
    return response;
  }
}

export function activate(context: ExtensionContext) {
  context.subscriptions.push(languages.registerLanguageModelChatProvider('my-provider', new MyProvider(context)));
}

📚 Documentation

• Getting Started — Setup and basic patterns
• API Reference — Complete parser API docs
• Developer Guide — Local development and contribution
• Integration Examples — Anthropic, OpenAI, Gemini patterns
• VS Code Integration — Provider setup and hooks

🏗️ Monorepo Structure

agentsy/
├── packages/
│   ├── core/                    # @agentsy/core (v1.0.0 — foundation)
│   │   ├── src/
│   │   │   ├── thinking/        # <think> tag extraction
│   │   │   ├── xml-filter/      # XML context scrubbing
│   │   │   ├── tool-calls/      # Tool call extraction
│   │   │   ├── structured/      # JSON parsing & validation
│   │   │   ├── processor/       # Event-driven orchestrator
│   │   │   ├── adapters/        # Provider normalizers
│   │   │   ├── agent/           # Multi-step agent loops
│   │   │   └── ...more modules
│   │   └── package.json
│   └── vscode/                  # @agentsy/vscode (v0.1.0)
│       ├── src/
│       │   ├── extension/       # Extension hooks
│       │   ├── provider/        # BaseLanguageModelChatProvider
│       │   ├── api-key-manager/ # SecretStorage integration
│       │   ├── error-handling/  # Error mapping
│       │   └── ...more modules
│       └── package.json
├── docs/                        # Unified documentation
│   ├── index.md                 # Overview
│   ├── getting-started.md
│   ├── api.md
│   └── developers/
├── turbo.json                   # Monorepo orchestration
├── pnpm-workspace.yaml          # Workspace config
├── package.json                 # Root config
├── tsconfig.json                # Shared TypeScript config
├── .oxlintrc.json               # Linting config
├── .oxfmtrc.json                # Formatting config
└── pnpm-lock.yaml

🛠️ Development

Prerequisites

• Node.js 22+ (verified in CI)
• pnpm 10+ (workspace package manager)

Setup

# Install dependencies
pnpm install

# Build all packages
pnpm turbo run build

# Run tests
pnpm turbo run test

# Lint and format
pnpm turbo run lint
pnpm turbo run format

Turbo Tasks

All build, test, and lint tasks are orchestrated via Turbo with intelligent caching:

pnpm turbo run build        # Build all packages (cached)
pnpm turbo run test         # Run all tests
pnpm turbo run check-types  # TypeScript strict mode check
pnpm turbo run lint         # Lint all packages (oxlint)
pnpm turbo run format       # Format all packages (oxfmt)
pnpm turbo run precommit    # Pre-commit hook (types + lint + format)

Local Package Development

Both packages use workspace:* dependencies, so local changes are reflected immediately:

// packages/vscode/package.json
{
  "dependencies": {
    "@agentsy/core": "workspace:*"  // Always uses local packages/ version
  }
}

🧪 Testing

Each package has full test coverage with Vitest:

# Test single package
cd packages/parser && pnpm test

# Test all packages
pnpm turbo run test

# Test with coverage
pnpm turbo run test -- --coverage

📦 Publishing

Both packages are published independently to npm:

# Build for distribution
pnpm turbo run build

# Create release tags
git tag @agentsy/[email protected]
git tag @agentsy/[email protected]

# Push tags to trigger GitHub Actions release workflow
git push --tags

See .github/workflows/release.yml for CI/CD automation.

🔒 Security

• TypeScript strict mode — Zero any types; exact optional property types enforced
• Input validation — All LLM output is validated and bounded (depth limits, key counts, tool call sizes)
• Privacy by default — Privacy tags are always scrubbed from output
• Dependency scanning — Trivy security scans in CI/CD

📄 License

MIT

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

📞 Support

• GitHub Issues — Report bugs and request features
• GitHub Discussions — Ask questions and share ideas
• Email — [email protected]

🙏 Acknowledgments

Built with Turbo, pnpm, TypeScript, and Vitest.