toolpack-sdk 1.4.0 → 2.0.0

package/README.md

@@ -1,6 +1,6 @@
 # Toolpack SDK
 
-A unified TypeScript/Node.js SDK for building AI-powered applications with multiple providers, 90 built-in tools, a workflow engine, and a flexible mode system — all through a single API.
+A unified TypeScript/Node.js SDK for building AI-powered applications with multiple providers, 100+ built-in tools, a workflow engine, and a flexible mode system — all through a single API.
 
 [![npm version](https://img.shields.io/npm/v/toolpack-sdk.svg)](https://www.npmjs.com/package/toolpack-sdk)
 [![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
@@ -9,31 +9,34 @@ A unified TypeScript/Node.js SDK for building AI-powered applications with multi
 
 ## Features
 
-- **Unified API** — Single interface for OpenAI, Anthropic, Google Gemini, Ollama, and custom providers
+- **Unified API** — Single interface for OpenAI, Anthropic, Google Gemini, Ollama, OpenRouter, and custom providers
 - **Streaming** — Real-time response streaming across all providers
 - **Type-Safe** — Comprehensive TypeScript types throughout
 - **Multimodal** — Text and image inputs (vision) across all providers
 - **Embeddings** — Vector generation for RAG applications (OpenAI, Gemini, Ollama)
-- **Workflow Engine** — AI-driven planning and step-by-step task execution with progress events
+- **Workflow Engine** — AI-driven planning with plan-direct execution and parallel tool orchestration
 - **Mode System** — Built-in Agent and Chat modes, plus `createMode()` for custom modes with tool filtering
 - **HITL Confirmation** — Human-in-the-loop approval for high-risk operations with configurable bypass rules
 - **Custom Providers** — Bring your own provider by implementing the `ProviderAdapter` interface
-- **90 Built-in Tools** across 11 categories:
+- **101 Built-in Tools** across 14 categories (including 4 skill-tools and 2 mcp-tools):
 - **MCP Tool Server Integration** — dynamically bridge external Model Context Protocol servers into Toolpack as first-class tools via `createMcpToolProject()` and `disconnectMcpToolProject()`.
 
 | Category | Tools | Description |
 |----------|-------|-------------|
 | **`fs-tools`** | 18 | File system operations — read, write, search, tree, glob, batch read/write, etc. |
 | **`coding-tools`** | 12 | Code analysis — AST parsing, go to definition, find references, rename symbols, extract function |
-| **`git-tools`** | 9 | Version control — status, diff, log, blame, branch, commit, checkout |
+| **`git-tools`** | 10 | Version control — status, diff, log, blame, branch, commit, checkout, clone |
 | **`db-tools`** | 7 | Database operations — query, schema, tables, count, insert, update, delete (SQLite, PostgreSQL, MySQL) |
 | **`exec-tools`** | 6 | Command execution — run, run shell, background processes, kill, read output |
 | **`http-tools`** | 5 | HTTP requests — GET, POST, PUT, DELETE, download |
 | **`web-tools`** | 9 | Web interaction — fetch, search (Tavily/Brave/DuckDuckGo), scrape, extract links, map, metadata, sitemap, feed, screenshot |
 | **`system-tools`** | 5 | System info — env vars, cwd, disk usage, system info, set env |
+| **`github-tools`** | 9 | GitHub operations — PR reviews, review threads, file diffs, issue comments, GraphQL, repo contents |
+| **`slack-tools`** | 6 | Slack messaging — post messages, ephemeral messages, channel history, thread replies, reactions |
 | **`diff-tools`** | 3 | Patch operations — create, apply, and preview diffs |
 | **`cloud-tools`** | 3 | Deployments — deploy, status, list (via Netlify) |
 | **`k8s-tools`** | 11 | Kubernetes cluster inspection and management via kubectl |
+| **`skill-tools`** | 4 | Skill management — skill.create, skill.read, skill.update, skill.list |
 | **`mcp-tools`** | 2 | MCP integration — createMcpToolProject, disconnectMcpToolProject |
 
 ## Quick Start
@@ -60,7 +63,7 @@ const sdk = await Toolpack.init({
     anthropic: {},   // Reads ANTHROPIC_API_KEY from env
   },
   defaultProvider: 'openai',
-  tools: true,         // Load all 90 built-in tools
+  tools: true,         // Load all 100+ built-in tools
   defaultMode: 'agent', // Agent mode with workflow engine
 });
 
@@ -78,7 +81,7 @@ for await (const chunk of sdk.stream({
 
 // Switch providers on the fly
 const anthropicResponse = await sdk.generate({
-  model: 'claude-sonnet-4-20250514',
+  model: 'your-model',
   messages: [{ role: 'user', content: 'Hello from Anthropic!' }],
 }, 'anthropic');
 ```
@@ -104,7 +107,7 @@ const sdk = await Toolpack.init({
 });
 
 const podsResponse = await sdk.generate({
-  model: 'gpt-4o',
+  model: 'your-model',
   messages: [
     {
       role: 'user',
@@ -115,7 +118,7 @@ const podsResponse = await sdk.generate({
 console.log(podsResponse.content);
 
 const applyResponse = await sdk.generate({
-  model: 'gpt-4o',
+  model: 'your-model',
   messages: [
     {
       role: 'user',
@@ -140,19 +143,20 @@ See `packages/toolpack-sdk/docs/examples/kubernetes-usage.ts` for a complete exa
 | **Anthropic** | Claude Sonnet 4, Claude 3.5 Haiku, Claude 3 Opus | No embeddings support |
 | **Google Gemini** | Gemini 2.0 Flash, Gemini 1.5 Pro, Gemini 1.5 Flash | Synthetic tool call IDs |
 | **Ollama** | Auto-discovered from locally pulled models | Capability detection via probing |
+| **OpenRouter** | All models at openrouter.ai (auto-discovered) | Access to 300+ models via OpenAI-compatible API |
 
 ### Provider Comparison
 
-| Capability | OpenAI | Anthropic | Gemini | Ollama |
-|------------|--------|-----------|--------|--------|
-| Chat completions | ✅ | ✅ | ✅ | ✅ |
-| Streaming | ✅ | ✅ | ✅ | ✅ |
-| Tool/function calling | ✅ | ✅ | ✅ | ✅ |
-| Multi-round tool loop | ✅ | ✅ | ✅ | ✅ |
-| Embeddings | ✅ | ❌ | ✅ | ✅ |
-| Vision/images | ✅ | ✅ | ✅ | ✅ (model-dependent) |
-| Tool name sanitization | ✅ (auto) | ✅ (auto) | ✅ (auto) | ✅ (auto) |
-| Model discovery | Static list | Static list | Static list | Dynamic (`/api/tags` + `/api/show`) |
+| Capability | OpenAI | Anthropic | Gemini | Ollama | OpenRouter |
+|------------|--------|-----------|--------|--------|------------|
+| Chat completions | ✅ | ✅ | ✅ | ✅ | ✅ |
+| Streaming | ✅ | ✅ | ✅ | ✅ | ✅ |
+| Tool/function calling | ✅ | ✅ | ✅ | ✅ | ✅ |
+| Multi-round tool loop | ✅ | ✅ | ✅ | ✅ | ✅ |
+| Embeddings | ✅ | ❌ | ✅ | ✅ | ❌ |
+| Vision/images | ✅ | ✅ | ✅ | ✅ (model-dependent) | ✅ (model-dependent) |
+| Tool name sanitization | ✅ (auto) | ✅ (auto) | ✅ (auto) | ✅ (auto) | ✅ (auto) |
+| Model discovery | Static list | Static list | Static list | Dynamic (`/api/tags` + `/api/show`) | Dynamic (`/models` endpoint) |
 
 #### Provider-Specific Notes
 
@@ -201,6 +205,7 @@ const sdk = await Toolpack.init({
 See `docs/MCP_INTEGRATION.md` and `docs/examples/mcp-integration-example.ts` for full instructions and best practices.
 - **Gemini**: Uses synthetic tool call IDs (`gemini_<timestamp>_<random>`) since the Gemini API doesn't return tool call IDs natively. Tool results are converted to `functionResponse` parts in chat history automatically. API key read from `GOOGLE_GENERATIVE_AI_KEY` or `TOOLPACK_GEMINI_KEY`.
 - **Ollama**: Auto-discovers all locally pulled models when registered as `{ ollama: {} }`. Uses `/api/show` and tool probing to detect capabilities (tool calling, vision, embeddings) per model. Models without tool support are automatically stripped of tools and given a system instruction to prevent hallucinated tool usage. Uses synthetic tool call IDs (`ollama_<timestamp>_<random>`). Embeddings use the modern `/api/embed` batch endpoint. Legacy per-model registration (`{ 'ollama-llama3': {} }`) is also supported.
+- **OpenRouter**: Routes requests to any of the 300+ models available on [openrouter.ai](https://openrouter.ai) via an OpenAI-compatible API. Models are discovered dynamically from the `/models` endpoint. Tool calling is fully supported; models that reject `tool_choice: 'none'` have tools stripped gracefully instead. No embeddings support. Optional `siteUrl` and `siteName` config for OpenRouter's attribution leaderboard. API key read from `OPENROUTER_API_KEY` or `TOOLPACK_OPENROUTER_KEY`.
 
 ### Custom Providers
 
@@ -276,8 +281,8 @@ Modes control AI behavior by setting a system prompt, filtering available tools,
 
 | Mode | Tools | Workflow | Description |
 |------|-------|----------|-------------|
-| **Agent** | All tools | Planning + step execution + dynamic steps | Full autonomous access — read, write, execute, browse |
-| **Coding** | All tools | Concise planning + step execution | Optimized for coding tasks — minimal text, file operations |
+| **Agent** | All tools | Plan-direct execution | Full autonomous access — read, write, execute, browse |
+| **Coding** | All tools | Plan-direct execution | Optimized for coding tasks — minimal text, file operations |
 | **Chat** | Web/HTTP only | Direct execution (no planning) | Conversational assistant with web access |
 
 ### Custom Modes
@@ -298,7 +303,6 @@ const reviewMode = createMode({
   },
   workflow: {
     planning: { enabled: true },
-    steps: { enabled: true, retryOnFailure: true },
     progress: { enabled: true },
   },
 });
@@ -338,19 +342,17 @@ sdk.cycleMode(); // Cycles through all registered modes
 | `blockedTools` | string[] | `[]` | Specific tools to block. Overrides allowed |
 | `blockAllTools` | boolean | `false` | If `true`, disables all tools (pure conversation) |
 | `baseContext` | object/false | `undefined` | Controls working directory and tool category injection |
-| `workflow` | WorkflowConfig | `undefined` | Planning, step execution, and progress configuration |
+| `workflow` | WorkflowConfig | `undefined` | Planning, execution mode, and progress configuration |
 
 ## Workflow Engine
 
-The workflow engine enables AI agents to plan and execute complex tasks step-by-step, with progress tracking, retries, and dynamic step additions.
+The workflow engine enables AI agents to plan and execute complex tasks with parallel tool orchestration.
 
 ### How It Works
 
-1. **Planning** — The AI generates a structured step-by-step plan from the user's request
-2. **Execution** — Each step is executed sequentially with tool access
-3. **Dynamic Steps** — New steps can be added during execution based on results
-4. **Retries** — Failed steps are retried automatically (configurable)
-5. **Progress** — Events are emitted at each stage for UI integration
+1. **Planning** — The AI generates a structured plan from the user's request
+2. **Execution** — The plan is injected as context and executed in a single call with parallel tool orchestration
+3. **Progress** — Events are emitted at each stage for UI integration
 
 ### Using the Workflow
 
@@ -361,7 +363,7 @@ const sdk = await Toolpack.init({
   defaultMode: 'agent', // Agent mode has workflow enabled
 });
 
-// Complex tasks are automatically planned and executed step-by-step
+// Complex tasks are automatically planned (plan-direct) with parallel tool execution
 const result = await sdk.generate('Build me a REST API with user authentication');
 
 // Or stream the response
@@ -383,35 +385,18 @@ const executor = sdk.getWorkflowExecutor();
 // Progress updates (ideal for status bars / shimmer text)
 executor.on('workflow:progress', (progress) => {
   // progress.status: 'planning' | 'awaiting_approval' | 'executing' | 'completed' | 'failed'
-  // progress.currentStep, progress.totalSteps, progress.percentage
-  // progress.currentStepDescription — includes retry info if retrying
-  console.log(`[${progress.percentage}%] Step ${progress.currentStep}/${progress.totalSteps}: ${progress.currentStepDescription}`);
+  // progress.percentage, progress.currentStepDescription
+  console.log(`[${progress.percentage}%] ${progress.currentStepDescription}`);
 });
 
-// Step lifecycle
-executor.on('workflow:step_start', (step, plan) => {
-  console.log(`Starting: ${step.description}`);
-});
-
-executor.on('workflow:step_complete', (step, plan) => {
-  console.log(`Completed: ${step.description}`);
-});
-
-executor.on('workflow:step_failed', (step, error, plan) => {
-  console.log(`Failed: ${step.description} — ${error.message}`);
-});
-
-executor.on('workflow:step_retry', (step, attempt, plan) => {
-  console.log(`Retrying: ${step.description} (attempt ${attempt})`);
-});
-
-executor.on('workflow:step_added', (step, plan) => {
-  console.log(`Dynamic step added: ${step.description}`);
+// Plan created
+executor.on('workflow:plan_created', (plan) => {
+  console.log('Plan:', plan.steps.map(s => s.description));
 });
 
 // Workflow completion
 executor.on('workflow:completed', (plan, result) => {
-  console.log(`Done! ${result.metrics.stepsCompleted} steps in ${result.metrics.totalDuration}ms`);
+  console.log(`Done in ${result.metrics.totalDuration}ms`);
 });
 
 executor.on('workflow:failed', (plan, error) => {
@@ -430,22 +415,10 @@ interface WorkflowConfig {
     maxSteps?: number;          // Max steps in a plan (default: 20)
   };
 
-  steps?: {
-    enabled: boolean;           // Enable step-by-step execution
-    retryOnFailure?: boolean;   // Retry failed steps (default: true)
-    maxRetries?: number;        // Max retries per step (default: 3)
-    allowDynamicSteps?: boolean; // Allow adding steps during execution
-    maxTotalSteps?: number;     // Max total steps including dynamic (default: 50)
-  };
-
   progress?: {
     enabled: boolean;           // Emit progress events (default: true)
     reportPercentage?: boolean; // Include completion percentage
   };
-
-  onFailure?: {
-    strategy: 'abort' | 'skip' | 'ask_user';
-  };
 }
 ```
 
@@ -457,12 +430,12 @@ The SDK provides built-in workflow presets for common use cases:
 import { DEFAULT_WORKFLOW, AGENT_WORKFLOW, CODING_WORKFLOW, CHAT_WORKFLOW } from 'toolpack-sdk';
 ```
 
-| Preset | Planning | Steps | Description |
-|--------|----------|-------|-------------|
-| `DEFAULT_WORKFLOW` | Disabled | Disabled | Direct execution, no planning |
-| `AGENT_WORKFLOW` | Enabled (detailed) | Enabled | Full autonomous agent with 11 planning rules |
-| `CODING_WORKFLOW` | Enabled (concise) | Enabled | Minimal prompts optimized for coding tasks |
-| `CHAT_WORKFLOW` | Disabled | Disabled | Simple conversational mode |
+| Preset | Planning | Description |
+|--------|----------|-------------|
+| `DEFAULT_WORKFLOW` | Disabled | Direct execution, no planning |
+| `AGENT_WORKFLOW` | Enabled (detailed) | Full autonomous agent with plan-direct execution |
+| `CODING_WORKFLOW` | Enabled (concise) | Minimal prompts optimized for coding tasks |
+| `CHAT_WORKFLOW` | Disabled | Simple conversational mode |
 
 ### Creating Custom Workflows
 
@@ -485,15 +458,6 @@ Rules:
 3. Generate docs in consistent format
 4. Output JSON: {"summary": "...", "steps": [...]}`,
   },
-  steps: {
-    ...AGENT_WORKFLOW.steps,
-    stepPrompt: `Execute step {stepNumber}: {stepDescription}
-
-Analyze code and write clear documentation.
-Focus on: purpose, parameters, return values, examples.
-
-Previous: {previousStepsResults}`,
-  },
 };
 
 // Use in a custom mode
@@ -508,23 +472,11 @@ const docMode = createMode({
 });
 ```
 
-### Step Prompt Template Variables
-
-When using custom `stepPrompt`, these variables are automatically substituted:
-
-| Variable | Description |
-|----------|-------------|
-| `{stepNumber}` | Current step number (1-indexed) |
-| `{planSummary}` | Summary of the overall plan |
-| `{stepDescription}` | Description of the current step |
-| `{previousStepsResults}` | Output from completed steps (truncated to 2000 chars) |
-
 ### Workflow Prompt Tips
 
 - **Keep planning prompts concise** — LLMs perform better with 5-7 clear rules
 - **Use JSON schema examples** — Include the exact expected output format
-- **Avoid meta-commentary in step prompts** — The AI should just execute, not discuss
-- **Leverage previous results** — The `{previousStepsResults}` variable provides context
+- **Keep prompts task-oriented** — The AI should execute, not discuss
 
 ## Tool Call Events
 
@@ -548,7 +500,7 @@ client.on('tool:failed', (event) => { /* ... */ });
 
 ## Custom Tools
 
-In addition to the 90 built-in tools, you can create and register your own custom tool projects using `createToolProject()`:
+In addition to the 100+ built-in tools, you can create and register your own custom tool projects using `createToolProject()`:
 
 ```typescript
 import { Toolpack, createToolProject } from 'toolpack-sdk';
@@ -644,7 +596,424 @@ const response = await toolpack.chat('How do I configure authentication?');
 - **Progress Events**: Track embedding progress with `onEmbeddingProgress`
 - **Metadata Filtering**: Query with filters like `{ hasCode: true, category: 'api' }`
 
-See the [Knowledge package README](../toolpack-knowledge/README.md) for full documentation.
+See the [Knowledge package README](./packages/toolpack-knowledge/README.md) for full documentation.
+
+## Skills
+
+The skills system lets you define **reusable behavioral instructions** in `.skill.md` files and automatically inject them into requests based on message relevance — no agent code changes required.
+
+### Quick Start
+
+```typescript
+import { Toolpack, createSkillInterceptor, createSkillTools } from 'toolpack-sdk';
+
+const toolpack = await Toolpack.init({
+  provider: 'anthropic',
+  interceptors: [
+    createSkillInterceptor({ dir: '.toolpack/skills', maxSkills: 3, minScore: 0.3 }),
+  ],
+  customTools: [
+    createSkillTools({ dir: '.toolpack/skills' }),
+  ],
+});
+```
+
+Create a skill file at `.toolpack/skills/code-review.skill.md`:
+
+```markdown
+---
+name: code-review
+title: Code Review
+version: 1.0.0
+tags: ["coding", "quality"]
+updated: 2026-01-15T10:00:00.000Z
+---
+
+## Description
+
+Guides the agent through a structured code review process.
+
+## Triggers
+
+- "review this code"
+- "check my pull request"
+- "code review"
+
+## Instructions
+
+When reviewing code:
+1. Check for security vulnerabilities first
+2. Verify test coverage exists
+3. Flag naming inconsistencies
+4. Be constructive — suggest improvements, not just problems
+```
+
+When a user sends "review this PR", the interceptor automatically injects the `## Instructions` block before the LLM sees the message.
+
+### How It Works
+
+- **`createSkillInterceptor`** — An SDK interceptor that runs BM25 search on every user message and prepends matching skill instructions as a `<skill-instructions>` block. Validates all files at `Toolpack.init()` time.
+- **`createSkillTools`** — Four LLM-callable tools (`skill.create`, `skill.read`, `skill.update`, `skill.list`) for managing the skill library at runtime.
+
+### `createSkillInterceptor` Options
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `dir` | string | `.toolpack/skills` | Path to the skill files directory |
+| `maxSkills` | number | `3` | Maximum number of skills injected per message |
+| `minScore` | number | `0.3` | BM25 relevance threshold |
+| `onValidationError` | `'fail'` \| `'warn'` | `'fail'` | How to handle invalid skill files at startup |
+
+See the [Skills guide](https://toolpacksdk.com/guides/skills) and [Skill Tools reference](https://toolpacksdk.com/tools/skills) for full documentation.
+
+## AI Agents (@toolpack-sdk/agents)
+
+Build production-ready AI agents with channels, workflows, and event-driven architecture using the companion `@toolpack-sdk/agents` package:
+
+```bash
+npm install @toolpack-sdk/agents
+```
+
+### What are Agents?
+
+Agents are autonomous AI systems that:
+- **Listen** for events from channels (Slack, webhooks, schedules, etc.)
+- **Process** messages using the Toolpack SDK
+- **Execute** tasks with full tool access
+- **Respond** back through the same or different channels
+- **Remember** conversations using knowledge bases
+
+### Quick Start
+
+```typescript
+import { Toolpack } from 'toolpack-sdk';
+import { BaseAgent, AgentRegistry, SlackChannel } from '@toolpack-sdk/agents';
+
+// 1. Create a custom agent
+class SupportAgent extends BaseAgent {
+  name = 'support-agent';
+  description = 'Customer support agent that answers questions';
+  mode = 'chat';
+
+  async invokeAgent(input) {
+    const result = await this.run(input.message);
+    await this.sendTo('slack-support', result.output);
+    return result;
+  }
+}
+
+// 2. Set up channels
+const slackChannel = new SlackChannel({
+  name: 'slack-support',
+  token: process.env.SLACK_BOT_TOKEN,
+  signingSecret: process.env.SLACK_SIGNING_SECRET,
+});
+
+// 3. Register agent and channels
+const registry = new AgentRegistry([
+  { agent: SupportAgent, channels: [slackChannel] },
+]);
+
+// 4. Initialize Toolpack with agents
+const sdk = await Toolpack.init({
+  provider: 'openai',
+  tools: true,
+  agents: registry,
+});
+
+// Agents now listen and respond automatically!
+```
+
+### Built-in Agents
+
+The package includes 4 production-ready agents you can use directly or extend:
+
+#### ResearchAgent
+```typescript
+import { ResearchAgent } from '@toolpack-sdk/agents';
+
+const agent = new ResearchAgent(sdk);
+const result = await agent.invokeAgent({
+  message: 'Summarize recent developments in edge AI',
+});
+```
+- **Mode:** `agent`
+- **Tools:** web.search, web.fetch, web.scrape
+- **Use Cases:** Market research, competitive analysis, trend monitoring
+
+#### CodingAgent
+```typescript
+import { CodingAgent } from '@toolpack-sdk/agents';
+
+const agent = new CodingAgent(sdk);
+const result = await agent.invokeAgent({
+  message: 'Refactor the auth module to use the new SDK pattern',
+});
+```
+- **Mode:** `coding`
+- **Tools:** fs.*, coding.*, git.*, exec.*
+- **Use Cases:** Code generation, refactoring, debugging, test writing
+
+#### DataAgent
+```typescript
+import { DataAgent } from '@toolpack-sdk/agents';
+
+const agent = new DataAgent(sdk);
+const result = await agent.invokeAgent({
+  message: 'Generate a weekly summary of signups by region',
+});
+```
+- **Mode:** `agent`
+- **Tools:** db.*, fs.*, http.*
+- **Use Cases:** Database queries, reporting, data analysis, CSV generation
+
+#### BrowserAgent
+```typescript
+import { BrowserAgent } from '@toolpack-sdk/agents';
+
+const agent = new BrowserAgent(sdk);
+const result = await agent.invokeAgent({
+  message: 'Extract all product prices from acme.com/products',
+});
+```
+- **Mode:** `chat`
+- **Tools:** web.fetch, web.screenshot, web.extract_links
+- **Use Cases:** Web scraping, form filling, content extraction
+
+### Channels
+
+Channels connect agents to the outside world. The package includes 7 built-in channels:
+
+#### SlackChannel (Two-way)
+```typescript
+import { SlackChannel } from '@toolpack-sdk/agents';
+
+const slack = new SlackChannel({
+  name: 'slack-support',
+  token: process.env.SLACK_BOT_TOKEN,
+  signingSecret: process.env.SLACK_SIGNING_SECRET,
+});
+```
+- ✅ Receives messages from Slack
+- ✅ Replies in threads
+- ✅ Supports `ask()` for human input
+
+#### TelegramChannel (Two-way)
+```typescript
+import { TelegramChannel } from '@toolpack-sdk/agents';
+
+const telegram = new TelegramChannel({
+  name: 'telegram-bot',
+  token: process.env.TELEGRAM_BOT_TOKEN,
+});
+```
+- ✅ Receives messages from Telegram
+- ✅ Replies to users
+- ✅ Supports `ask()` for human input
+
+#### WebhookChannel (Two-way)
+```typescript
+import { WebhookChannel } from '@toolpack-sdk/agents';
+
+const webhook = new WebhookChannel({
+  name: 'github-webhook',
+  path: '/webhook/github',
+  port: 3000,
+  secret: process.env.WEBHOOK_SECRET,
+});
+```
+- ✅ Receives HTTP POST webhooks
+- ✅ Signature verification
+- ✅ Supports `ask()` for human input
+
+#### ScheduledChannel (Trigger-only)
+```typescript
+import { ScheduledChannel } from '@toolpack-sdk/agents';
+
+const scheduler = new ScheduledChannel({
+  name: 'daily-report',
+  cron: '0 9 * * 1-5', // 9am weekdays
+  notify: 'webhook:https://hooks.example.com/daily-report',
+  message: 'Generate the daily sales report',
+});
+// For Slack delivery, attach a named SlackChannel to the same agent and
+// call `this.sendTo('<slackChannelName>', output)` from within `run()`.
+```
+- ⏰ Triggers agents on cron schedules
+- ✅ Full cron expression support (ranges, steps, lists, combinations)
+- ❌ No `ask()` support (no human recipient)
+
+#### DiscordChannel (Two-way)
+```typescript
+import { DiscordChannel } from '@toolpack-sdk/agents';
+
+const discord = new DiscordChannel({
+  name: 'discord-bot',
+  token: process.env.DISCORD_BOT_TOKEN,
+  guildId: 'your-guild-id',
+  channelId: 'your-channel-id',
+});
+```
+- ✅ Receives messages from Discord
+- ✅ Replies in threads
+- ✅ Supports `ask()` for human input
+
+#### EmailChannel (Outbound-only)
+```typescript
+import { EmailChannel } from '@toolpack-sdk/agents';
+
+const email = new EmailChannel({
+  name: 'email-alerts',
+  from: 'bot@acme.com',
+  to: 'team@acme.com',
+  smtp: {
+    host: 'smtp.gmail.com',
+    port: 587,
+    auth: { user: 'bot@acme.com', pass: process.env.SMTP_PASSWORD },
+  },
+});
+```
+- 📧 Sends emails via SMTP
+- ❌ No `ask()` support (outbound-only)
+
+#### SMSChannel (Configurable)
+```typescript
+import { SMSChannel } from '@toolpack-sdk/agents';
+
+// Two-way with webhook
+const sms = new SMSChannel({
+  name: 'sms-alerts',
+  accountSid: process.env.TWILIO_ACCOUNT_SID,
+  authToken: process.env.TWILIO_AUTH_TOKEN,
+  from: '+1234567890',
+  webhookPath: '/sms/webhook', // Enables two-way
+  port: 3000,
+});
+
+// Outbound-only
+const smsOutbound = new SMSChannel({
+  name: 'sms-notifications',
+  accountSid: process.env.TWILIO_ACCOUNT_SID,
+  authToken: process.env.TWILIO_AUTH_TOKEN,
+  from: '+1234567890',
+  to: '+0987654321', // Fixed recipient
+});
+```
+- 📱 Twilio SMS integration
+- ✅ Two-way when `webhookPath` is set
+- ❌ Outbound-only without webhook
+
+### Agent Lifecycle & Events
+
+Agents emit events at each stage of execution:
+
+```typescript
+const agent = new MyAgent(sdk);
+
+agent.on('agent:start', (input) => {
+  console.log('Agent started:', input.message);
+});
+
+agent.on('agent:complete', (result) => {
+  console.log('Agent completed:', result.output);
+});
+
+agent.on('agent:error', (error) => {
+  console.error('Agent error:', error);
+});
+```
+
+### Knowledge Integration
+
+Agents can use knowledge bases for conversation memory and RAG:
+
+```typescript
+import { Knowledge, MemoryProvider, OllamaEmbedder } from '@toolpack-sdk/knowledge';
+import { BaseAgent } from '@toolpack-sdk/agents';
+
+class SmartAgent extends BaseAgent {
+  name = 'smart-agent';
+  description = 'Agent with memory';
+  mode = 'chat';
+  
+  constructor(toolpack) {
+    super(toolpack);
+    // Set up knowledge base
+    this.knowledge = await Knowledge.create({
+      provider: new MemoryProvider(),
+      embedder: new OllamaEmbedder({ model: 'nomic-embed-text' }),
+    });
+  }
+
+  async invokeAgent(input) {
+    // Conversation history is automatically loaded from knowledge
+    const result = await this.run(input.message);
+    return result;
+  }
+}
+```
+
+### Multi-Channel Routing
+
+Agents can send output to different channels:
+
+```typescript
+class MultiChannelAgent extends BaseAgent {
+  name = 'multi-agent';
+  description = 'Routes to multiple channels';
+  mode = 'agent';
+
+  async invokeAgent(input) {
+    const result = await this.run(input.message);
+    
+    // Send to multiple channels
+    await this.sendTo('slack:#general', result.output);
+    await this.sendTo('email-team', result.output);
+    await this.sendTo('sms-alerts', 'Task completed!');
+    
+    return result;
+  }
+}
+```
+
+### Extending Built-in Agents
+
+```typescript
+import { ResearchAgent } from '@toolpack-sdk/agents';
+
+class FintechResearchAgent extends ResearchAgent {
+  systemPrompt = `You are a research agent focused on fintech.
+                  Always cite sources and flag regulatory implications.`;
+  provider = 'anthropic';
+  model = 'your-model';
+
+  async onComplete(result) {
+    // Store research in knowledge base
+    if (this.knowledge) {
+      await this.knowledge.add(result.output, { 
+        category: 'research',
+        topic: 'fintech',
+      });
+    }
+    
+    // Send to Slack
+    await this.sendTo('slack-research', result.output);
+  }
+}
+```
+
+### Features
+
+- ✅ **7 Built-in Channels** — Slack, Telegram, Discord, Email, SMS, Webhook, Scheduled
+- ✅ **4 Built-in Agents** — Research, Coding, Data, Browser
+- ✅ **Event-Driven** — Full lifecycle events for monitoring
+- ✅ **Knowledge Integration** — Conversation memory and RAG
+- ✅ **Multi-Channel Routing** — Send to any registered channel
+- ✅ **Human-in-the-Loop** — `ask()` support for two-way channels
+- ✅ **Type-Safe** — Full TypeScript support
+- ✅ **199 Tests Passing** — Production-ready
+
+See the [Agents package README](./packages/toolpack-agents/README.md) for full documentation.
 
 ## Multimodal Support
 
@@ -704,6 +1073,7 @@ const response = await sdk.generate({
 export OPENAI_API_KEY="sk-..."
 export ANTHROPIC_API_KEY="sk-ant-..."
 export GOOGLE_GENERATIVE_AI_KEY="AIza..."
+export OPENROUTER_API_KEY="sk-or-..."
 
 # SDK logging (override — prefer toolpack.config.json instead)
 export TOOLPACK_SDK_LOG_FILE="./toolpack.log"    # Log file path (also enables logging)
@@ -926,6 +1296,7 @@ interface CompletionRequest {
   temperature?: number;
   max_tokens?: number;
   tools?: ToolCallRequest[];
+  requestTools?: RequestToolDefinition[];  // Request-scoped tools
   tool_choice?: 'auto' | 'none' | 'required';
 }
 
@@ -956,6 +1327,115 @@ interface ProviderModelInfo {
 }
 ```
 
+### Request-Scoped Tools
+
+Request-scoped tools are dynamic tools attached to a single completion request. Unlike globally registered tools in the ToolRegistry, they:
+
+- **Don't pollute the shared registry** — Each request can have its own tools
+- **Can close over request-specific state** — e.g., `conversationId`, user context
+- **Are safe for multi-agent/multi-request usage** — No cross-request contamination
+- **Execute through the same SDK orchestration** — Events, logging, HITL all work
+
+#### Built-in Request-Scoped Tools
+
+**Knowledge Tools** (when `knowledge` is configured):
+- `knowledge_search` — Search the knowledge base for relevant information
+- `knowledge_add` — Add new content to the knowledge base at runtime
+
+**Conversation Tools** (when using `ConversationHistory`):
+- `conversation_search` — Search conversation history for past messages
+
+#### Creating Custom Request Tools
+
+```typescript
+import { RequestToolDefinition, ConversationHistory } from 'toolpack-sdk';
+
+// Example: Session-specific calculator
+const createCalculatorTool = (sessionId: string): RequestToolDefinition => ({
+  name: 'calculate',
+  displayName: 'Calculator',
+  description: 'Perform mathematical calculations',
+  category: 'math',
+  parameters: {
+    type: 'object',
+    properties: {
+      expression: { type: 'string', description: 'Math expression to evaluate' },
+    },
+    required: ['expression'],
+  },
+  execute: async (args) => {
+    // Can safely close over sessionId
+    console.log(`Session ${sessionId}: calculating ${args.expression}`);
+    
+    // Simple eval (use a proper math library in production)
+    const result = eval(args.expression);
+    return { result, sessionId };
+  },
+});
+
+// Use in a request
+const result = await sdk.generate({
+  messages: [{ role: 'user', content: 'What is 15 * 23?' }],
+  model: 'gpt-4',
+  requestTools: [createCalculatorTool('user-123')],
+});
+```
+
+#### Using ConversationHistory with Request Tools
+
+```typescript
+import { ConversationHistory } from 'toolpack-sdk';
+
+const history = new ConversationHistory('./chat.db');
+
+// Add some messages
+await history.addUserMessage('conv-1', 'What is the API rate limit?');
+await history.addAssistantMessage('conv-1', 'The rate limit is 100 requests per minute.');
+
+// Use conversation search in a request
+const result = await sdk.generate({
+  messages: [
+    { role: 'user', content: 'What did we discuss about rate limits?' }
+  ],
+  model: 'gpt-4',
+  requestTools: [
+    history.toTool('conv-1'),  // Scoped to conversation 'conv-1'
+  ],
+});
+
+// AI can now call conversation_search to find the earlier discussion
+```
+
+#### Request Tools vs Registry Tools
+
+| Feature | Request Tools | Registry Tools |
+|---------|---------------|----------------|
+| **Scope** | Single request | All requests |
+| **State** | Can close over request state | Stateless |
+| **Registration** | Per-request via `requestTools` | Global via `ToolRegistry` |
+| **Use Case** | Dynamic, stateful tools | Reusable, static tools |
+| **Priority** | Higher (checked first) | Lower |
+| **Examples** | `conversation_search`, `knowledge_add` | `fs.read_file`, `web.search` |
+
+#### Automatic Guidance Injection
+
+When request-scoped tools are present, the SDK automatically injects usage guidance into the system prompt:
+
+```
+Knowledge Base:
+- Use `knowledge_search` when you need factual or domain-specific information.
+- Use `knowledge_add` when you learn durable information that should be saved.
+
+Conversation History:
+- Only recent messages may be present in context.
+- Use `conversation_search` to find details from earlier in this conversation.
+```
+
+This guidance is:
+- **Per-request** — Only injected when tools are actually present
+- **Derived from effective tool set** — Reflects the actual tools available
+- **Idempotent** — Won't duplicate if already present
+
 ## Error Handling
 
 The SDK provides typed error classes for common failure scenarios:
@@ -1008,21 +1488,25 @@ toolpack-sdk/
 │   │   ├── openai/        # OpenAI adapter
 │   │   ├── anthropic/     # Anthropic adapter
 │   │   ├── gemini/        # Google Gemini adapter
+│   │   ├── openrouter/    # OpenRouter adapter (OpenAI-compatible, dynamic model discovery)
 │   │   └── ollama/        # Ollama adapter + provider (auto-discovery)
 │   ├── modes/             # Mode system (Agent, Chat, createMode)
-│   ├── workflows/         # Workflow engine (planner, step executor, progress)
-│   ├── tools/             # 90 built-in tools + registry + router + BM25 search
+│   ├── workflows/         # Workflow engine (planner, executor, progress)
+│   ├── tools/             # 100+ built-in tools + registry + router + BM25 search
 │   │   ├── fs-tools/      # File system (18 tools)
 │   │   ├── coding-tools/  # Code analysis (12 tools)
-│   │   ├── git-tools/     # Git operations (9 tools)
-│   │   ├── db-tools/      # Database operations (6 tools)
+│   │   ├── git-tools/     # Git operations (10 tools)
+│   │   ├── db-tools/      # Database operations (7 tools)
 │   │   ├── exec-tools/    # Command execution (6 tools)
 │   │   ├── http-tools/    # HTTP requests (5 tools)
-│   │   ├── web-tools/     # Web interaction (5 tools)
+│   │   ├── web-tools/     # Web interaction (9 tools)
 │   │   ├── system-tools/  # System info (5 tools)
+│   │   ├── github-tools/  # GitHub API (9 tools)
+│   │   ├── slack-tools/   # Slack messaging (6 tools)
 │   │   ├── diff-tools/    # Patch operations (3 tools)
 │   │   ├── cloud-tools/   # Deployments (3 tools)
 │   │   ├── k8s-tools/     # Kubernetes management (11 tools)
+│   │   ├── skill-tools/   # Skill management (4 tools)
 │   │   ├── registry.ts    # Tool registry and loading
 │   │   ├── router.ts      # Tool routing and filtering
 │   │   └── search/        # BM25 tool discovery engine (internal)
@@ -1037,9 +1521,9 @@ toolpack-sdk/
 
 **Current Version:** 0.1.0
 
-- ✓ **4 Built-in Providers** — OpenAI, Anthropic, Gemini, Ollama (+ custom provider API)
-- ✓ **90 Built-in Tools** — fs, exec, git, diff, web, coding, db, cloud, http, system, Kubernetes
-- ✓ **Workflow Engine** — AI-driven planning, step execution, retries, dynamic steps, progress events
+- ✓ **5 Built-in Providers** — OpenAI, Anthropic, Gemini, Ollama, OpenRouter (+ custom provider API)
+- ✓ **100+ Built-in Tools** — fs, exec, git, diff, web, coding, db, cloud, http, system, Kubernetes, GitHub, Slack, Skills
+- ✓ **Workflow Engine** — AI-driven planning, plan-direct execution, parallel tool orchestration, progress events
 - ✓ **Mode System** — Agent, Coding, Chat, and custom modes via `createMode()` with `blockAllTools` support
 - ✓ **Tool Search** — BM25-based on-demand tool discovery for large tool libraries
 - ✓ **545 Tests** passing across 81 test files