Initial commit

references/mcp-servers-guide.md

@@ -0,0 +1,387 @@
+# MCP Servers Guide
+
+Complete guide to creating and using Model Context Protocol (MCP) servers with Claude Agent SDK.
+
+---
+
+## What Are MCP Servers?
+
+MCP servers extend agent capabilities with custom tools. Think of them as plugins that give your agent new abilities.
+
+**Use Cases**:
+- Database access
+- API integrations
+- Custom calculations
+- External service interactions
+- File system operations
+
+---
+
+## Creating In-Process MCP Servers
+
+### Basic Server
+
+```typescript
+import { createSdkMcpServer, tool } from "@anthropic-ai/claude-agent-sdk";
+import { z } from "zod";
+
+const myServer = createSdkMcpServer({
+  name: "my-service",
+  version: "1.0.0",
+  tools: [
+    tool(
+      "tool_name",
+      "Tool description",
+      { /* Zod schema */ },
+      async (args) => {
+        // Implementation
+        return {
+          content: [{ type: "text", text: "Result" }]
+        };
+      }
+    )
+  ]
+});
+```
+
+### Tool Definition Pattern
+
+```typescript
+tool(
+  name: string,           // Tool identifier
+  description: string,    // What it does
+  inputSchema: ZodSchema, // Input validation
+  handler: Handler        // Implementation
+)
+```
+
+---
+
+## Zod Schemas
+
+### Common Patterns
+
+```typescript
+import { z } from "zod";
+
+// String
+z.string()
+z.string().email()
+z.string().url()
+z.string().min(5).max(100)
+z.string().describe("Description for AI")
+
+// Number
+z.number()
+z.number().int()
+z.number().positive()
+z.number().min(0).max(100)
+
+// Boolean
+z.boolean()
+z.boolean().default(false)
+
+// Enum
+z.enum(["option1", "option2", "option3"])
+z.union([z.literal("a"), z.literal("b")])
+
+// Optional
+z.string().optional()
+z.number().default(10)
+
+// Object
+z.object({
+  name: z.string(),
+  age: z.number(),
+  email: z.string().email().optional()
+})
+
+// Array
+z.array(z.string())
+z.array(z.number()).min(1).max(10)
+
+// Complex nested
+z.object({
+  user: z.object({
+    id: z.string().uuid(),
+    name: z.string(),
+    roles: z.array(z.enum(["admin", "user", "guest"]))
+  }),
+  metadata: z.record(z.any()).optional()
+})
+```
+
+### Best Practices
+
+```typescript
+// ✅ Good: Clear descriptions
+{
+  location: z.string().describe("City name or coordinates (e.g., 'San Francisco, CA')"),
+  radius: z.number().min(1).max(100).describe("Search radius in kilometers")
+}
+
+// ❌ Bad: No descriptions
+{
+  location: z.string(),
+  radius: z.number()
+}
+
+// ✅ Good: Validation constraints
+{
+  email: z.string().email(),
+  age: z.number().int().min(0).max(120),
+  role: z.enum(["admin", "user", "guest"])
+}
+
+// ❌ Bad: No validation
+{
+  email: z.string(),
+  age: z.number(),
+  role: z.string()
+}
+```
+
+---
+
+## Handler Implementation
+
+### Success Response
+
+```typescript
+async (args) => {
+  const result = await performOperation(args);
+  return {
+    content: [{
+      type: "text",
+      text: JSON.stringify(result, null, 2)
+    }]
+  };
+}
+```
+
+### Error Response
+
+```typescript
+async (args) => {
+  try {
+    const result = await riskyOperation(args);
+    return {
+      content: [{ type: "text", text: result }]
+    };
+  } catch (error) {
+    return {
+      content: [{
+        type: "text",
+        text: `Error: ${error.message}`
+      }],
+      isError: true  // Mark as error
+    };
+  }
+}
+```
+
+---
+
+## Complete Examples
+
+### Weather Service
+
+```typescript
+const weatherServer = createSdkMcpServer({
+  name: "weather",
+  version: "1.0.0",
+  tools: [
+    tool(
+      "get_weather",
+      "Get current weather for a location",
+      {
+        location: z.string().describe("City name"),
+        units: z.enum(["celsius", "fahrenheit"]).default("celsius")
+      },
+      async (args) => {
+        const response = await fetch(
+          `https://api.weather.com/v1/current?location=${args.location}&units=${args.units}`
+        );
+        const data = await response.json();
+        return {
+          content: [{
+            type: "text",
+            text: `Temp: ${data.temp}° ${args.units}\nConditions: ${data.conditions}`
+          }]
+        };
+      }
+    ),
+    tool(
+      "get_forecast",
+      "Get 7-day forecast",
+      {
+        location: z.string(),
+        days: z.number().min(1).max(7).default(7)
+      },
+      async (args) => {
+        const forecast = await fetchForecast(args.location, args.days);
+        return {
+          content: [{ type: "text", text: JSON.stringify(forecast, null, 2) }]
+        };
+      }
+    )
+  ]
+});
+```
+
+### Database Service
+
+```typescript
+const databaseServer = createSdkMcpServer({
+  name: "database",
+  version: "1.0.0",
+  tools: [
+    tool(
+      "query",
+      "Execute SQL query",
+      {
+        sql: z.string().describe("SQL query to execute"),
+        params: z.array(z.any()).optional().describe("Query parameters")
+      },
+      async (args) => {
+        try {
+          const results = await db.query(args.sql, args.params);
+          return {
+            content: [{ type: "text", text: JSON.stringify(results, null, 2) }]
+          };
+        } catch (error) {
+          return {
+            content: [{ type: "text", text: `SQL Error: ${error.message}` }],
+            isError: true
+          };
+        }
+      }
+    )
+  ]
+});
+```
+
+---
+
+## Using MCP Servers
+
+### In Query Options
+
+```typescript
+const response = query({
+  prompt: "What's the weather in NYC?",
+  options: {
+    mcpServers: {
+      "weather": weatherServer,
+      "database": databaseServer
+    },
+    allowedTools: [
+      "mcp__weather__get_weather",
+      "mcp__database__query"
+    ]
+  }
+});
+```
+
+### Tool Naming Convention
+
+**Format**: `mcp__<server-name>__<tool-name>`
+
+Examples:
+- `mcp__weather__get_weather`
+- `mcp__database__query`
+- `mcp__filesystem__read_file`
+
+**CRITICAL**: Server and tool names must match exactly.
+
+---
+
+## External MCP Servers
+
+### Stdio Servers
+
+```typescript
+options: {
+  mcpServers: {
+    "filesystem": {
+      command: "npx",
+      args: ["@modelcontextprotocol/server-filesystem"],
+      env: {
+        ALLOWED_PATHS: "/path/to/allowed/dir"
+      }
+    }
+  }
+}
+```
+
+### HTTP/SSE Servers
+
+```typescript
+options: {
+  mcpServers: {
+    "remote": {
+      url: "https://api.example.com/mcp",
+      headers: {
+        "Authorization": "Bearer token",
+        "Content-Type": "application/json"
+      }
+    }
+  }
+}
+```
+
+---
+
+## Best Practices
+
+### ✅ Do
+
+- Use clear tool names and descriptions
+- Add `.describe()` to all Zod fields
+- Implement error handling in handlers
+- Validate inputs with Zod constraints
+- Return clear, formatted responses
+- Test tools independently before integration
+- Use unique tool names across servers
+- Version your servers
+
+### ❌ Don't
+
+- Use generic names like "process" or "run"
+- Skip input validation
+- Return raw error objects
+- Forget `isError: true` on errors
+- Use duplicate tool names
+- Expose sensitive operations without checks
+- Skip testing in isolation
+
+---
+
+## Troubleshooting
+
+### Tool Not Found
+
+**Problem**: `"Tool mcp__server__tool not found"`
+
+**Solution**:
+1. Check server name matches
+2. Check tool name matches
+3. Include in `allowedTools` array
+4. Verify server added to `mcpServers`
+
+### Tool Name Collision
+
+**Problem**: Two tools with same name
+
+**Solution**: Use unique names or prefix with server name
+
+### Validation Errors
+
+**Problem**: Invalid input to tool
+
+**Solution**: Add descriptive Zod schemas with constraints
+
+---
+
+**For more details**: See SKILL.md
+**Official MCP docs**: https://modelcontextprotocol.io/

references/permissions-guide.md

@@ -0,0 +1,429 @@
+# Permissions Guide
+
+Complete guide to permission control in Claude Agent SDK.
+
+---
+
+## Permission Modes
+
+### Overview
+
+Three built-in modes:
+
+| Mode | Behavior | Use Case |
+|------|----------|----------|
+| `default` | Standard checks | General use, production |
+| `acceptEdits` | Auto-approve file edits | Trusted refactoring |
+| `bypassPermissions` | Skip ALL checks | CI/CD, sandboxed envs |
+
+---
+
+## Default Mode
+
+Standard permission checks.
+
+```typescript
+options: {
+  permissionMode: "default"
+}
+```
+
+**Prompts user for**:
+- File writes/edits
+- Potentially dangerous bash commands
+- Sensitive operations
+
+**Auto-allows**:
+- Read operations (Read, Grep, Glob)
+- Safe bash commands
+
+---
+
+## Accept Edits Mode
+
+Automatically approves file modifications.
+
+```typescript
+options: {
+  permissionMode: "acceptEdits"
+}
+```
+
+**Auto-approves**:
+- File edits (Edit)
+- File writes (Write)
+
+**Still prompts for**:
+- Dangerous bash commands
+- Sensitive operations
+
+**Use when**: Refactoring, code generation workflows
+
+---
+
+## Bypass Permissions Mode
+
+⚠️ **DANGER**: Skips ALL permission checks.
+
+```typescript
+options: {
+  permissionMode: "bypassPermissions"
+}
+```
+
+**Auto-approves EVERYTHING**:
+- All file operations
+- All bash commands
+- All tools
+
+**ONLY use in**:
+- CI/CD pipelines
+- Sandboxed containers
+- Docker environments
+- Trusted, isolated contexts
+
+**NEVER use in**:
+- Production systems
+- User-facing environments
+- Untrusted inputs
+
+---
+
+## Custom Permission Logic
+
+### canUseTool Callback
+
+```typescript
+type CanUseTool = (
+  toolName: string,
+  input: any
+) => Promise<PermissionDecision>;
+
+type PermissionDecision =
+  | { behavior: "allow" }
+  | { behavior: "deny"; message?: string }
+  | { behavior: "ask"; message?: string };
+```
+
+### Basic Example
+
+```typescript
+options: {
+  canUseTool: async (toolName, input) => {
+    // Allow read-only
+    if (['Read', 'Grep', 'Glob'].includes(toolName)) {
+      return { behavior: "allow" };
+    }
+
+    // Deny dangerous bash
+    if (toolName === 'Bash' && input.command.includes('rm -rf')) {
+      return {
+        behavior: "deny",
+        message: "Destructive command blocked"
+      };
+    }
+
+    // Ask for confirmation
+    if (toolName === 'Write') {
+      return {
+        behavior: "ask",
+        message: `Create ${input.file_path}?`
+      };
+    }
+
+    return { behavior: "allow" };
+  }
+}
+```
+
+---
+
+## Common Patterns
+
+### Pattern 1: Block Destructive Commands
+
+```typescript
+canUseTool: async (toolName, input) => {
+  if (toolName === 'Bash') {
+    const dangerous = [
+      'rm -rf',
+      'dd if=',
+      'mkfs',
+      '> /dev/',
+      'shutdown',
+      'reboot',
+      'kill -9',
+      'pkill'
+    ];
+
+    for (const pattern of dangerous) {
+      if (input.command.includes(pattern)) {
+        return {
+          behavior: "deny",
+          message: `Blocked dangerous command: ${pattern}`
+        };
+      }
+    }
+  }
+
+  return { behavior: "allow" };
+}
+```
+
+### Pattern 2: Protect Sensitive Files
+
+```typescript
+canUseTool: async (toolName, input) => {
+  if (toolName === 'Write' || toolName === 'Edit') {
+    const sensitivePaths = [
+      '/etc/',
+      '/root/',
+      '.env',
+      'credentials',
+      'secrets',
+      'config/production',
+      '.ssh/',
+      'private_key'
+    ];
+
+    for (const path of sensitivePaths) {
+      if (input.file_path?.includes(path)) {
+        return {
+          behavior: "ask",
+          message: `⚠️  Modify sensitive file: ${input.file_path}?`
+        };
+      }
+    }
+  }
+
+  return { behavior: "allow" };
+}
+```
+
+### Pattern 3: Environment-Based Permissions
+
+```typescript
+const environment = process.env.NODE_ENV; // 'development' | 'staging' | 'production'
+
+canUseTool: async (toolName, input) => {
+  // Production: require approval for everything
+  if (environment === 'production') {
+    if (toolName === 'Bash' || toolName === 'Write' || toolName === 'Edit') {
+      return {
+        behavior: "ask",
+        message: `PRODUCTION: Approve ${toolName}?`
+      };
+    }
+  }
+
+  // Staging: auto-approve edits
+  if (environment === 'staging') {
+    if (toolName === 'Write' || toolName === 'Edit') {
+      return { behavior: "allow" };
+    }
+  }
+
+  // Development: allow most things
+  if (environment === 'development') {
+    return { behavior: "allow" };
+  }
+
+  return { behavior: "allow" };
+}
+```
+
+### Pattern 4: Deployment Confirmation
+
+```typescript
+canUseTool: async (toolName, input) => {
+  if (toolName === 'Bash') {
+    const deploymentPatterns = [
+      'deploy',
+      'kubectl apply',
+      'terraform apply',
+      'helm install',
+      'docker push',
+      'npm publish'
+    ];
+
+    for (const pattern of deploymentPatterns) {
+      if (input.command.includes(pattern)) {
+        return {
+          behavior: "ask",
+          message: `🚀 DEPLOYMENT: ${input.command}\n\nProceed?`
+        };
+      }
+    }
+  }
+
+  return { behavior: "allow" };
+}
+```
+
+### Pattern 5: Audit Logging
+
+```typescript
+const auditLog: Array<{
+  tool: string;
+  input: any;
+  decision: string;
+  timestamp: Date;
+}> = [];
+
+canUseTool: async (toolName, input) => {
+  // Log everything
+  console.log(`[${new Date().toISOString()}] ${toolName}`);
+
+  const decision = { behavior: "allow" as const };
+
+  // Store audit log
+  auditLog.push({
+    tool: toolName,
+    input,
+    decision: decision.behavior,
+    timestamp: new Date()
+  });
+
+  // Could also send to external service
+  // await logToDatadog(toolName, input, decision);
+
+  return decision;
+}
+```
+
+---
+
+## Combining Modes with Custom Logic
+
+```typescript
+options: {
+  permissionMode: "acceptEdits",  // Auto-approve file edits
+  canUseTool: async (toolName, input) => {
+    // But still block dangerous bash
+    if (toolName === 'Bash' && input.command.includes('rm -rf')) {
+      return { behavior: "deny", message: "Blocked" };
+    }
+
+    // Custom logic runs AFTER permission mode
+    return { behavior: "allow" };
+  }
+}
+```
+
+---
+
+## Security Best Practices
+
+### ✅ Do
+
+- Start with `"default"` mode
+- Use `canUseTool` for fine-grained control
+- Block known dangerous patterns
+- Require confirmation for sensitive ops
+- Log all tool usage for auditing
+- Test permission logic thoroughly
+- Use environment-based rules
+- Implement rate limiting if needed
+
+### ❌ Don't
+
+- Use `"bypassPermissions"` in production
+- Skip permission checks for "trusted" inputs
+- Allow arbitrary bash without filtering
+- Trust file paths without validation
+- Ignore audit logging
+- Assume AI won't make mistakes
+- Give blanket approvals
+
+---
+
+## Testing Permissions
+
+```typescript
+async function testPermissions() {
+  const tests = [
+    { tool: "Read", input: { file_path: "/etc/passwd" }, expectAllow: true },
+    { tool: "Bash", input: { command: "rm -rf /" }, expectDeny: true },
+    { tool: "Write", input: { file_path: ".env" }, expectAsk: true }
+  ];
+
+  for (const test of tests) {
+    const decision = await canUseTool(test.tool, test.input);
+
+    if (test.expectAllow && decision.behavior !== 'allow') {
+      console.error(`FAIL: Expected allow for ${test.tool}`);
+    }
+    if (test.expectDeny && decision.behavior !== 'deny') {
+      console.error(`FAIL: Expected deny for ${test.tool}`);
+    }
+    if (test.expectAsk && decision.behavior !== 'ask') {
+      console.error(`FAIL: Expected ask for ${test.tool}`);
+    }
+  }
+}
+```
+
+---
+
+## Common Scenarios
+
+### Allow Read-Only
+
+```typescript
+canUseTool: async (toolName, input) => {
+  if (['Read', 'Grep', 'Glob'].includes(toolName)) {
+    return { behavior: "allow" };
+  }
+  return {
+    behavior: "deny",
+    message: "Read-only mode"
+  };
+}
+```
+
+### Require Approval for All
+
+```typescript
+canUseTool: async (toolName, input) => {
+  return {
+    behavior: "ask",
+    message: `Approve ${toolName}?`
+  };
+}
+```
+
+### Block Bash Entirely
+
+```typescript
+canUseTool: async (toolName, input) => {
+  if (toolName === 'Bash') {
+    return {
+      behavior: "deny",
+      message: "Bash execution disabled"
+    };
+  }
+  return { behavior: "allow" };
+}
+```
+
+---
+
+## Error Handling
+
+Handle permission errors gracefully:
+
+```typescript
+for await (const message of response) {
+  if (message.type === 'error') {
+    if (message.error.type === 'permission_denied') {
+      console.log('Permission denied for:', message.error.tool);
+      // Continue with fallback or skip
+    }
+  }
+}
+```
+
+---
+
+**For more details**: See SKILL.md
+**Template**: templates/permission-control.ts

references/query-api-reference.md

@@ -0,0 +1,437 @@
+# Query API Reference
+
+Complete reference for the `query()` function - the primary interface for Claude Agent SDK.
+
+---
+
+## Function Signature
+
+```typescript
+function query(config: {
+  prompt: string | AsyncIterable<SDKUserMessage>;
+  options?: Options;
+}): AsyncGenerator<SDKMessage, void>;
+```
+
+---
+
+## Parameters
+
+### prompt
+
+**Type**: `string | AsyncIterable<SDKUserMessage>`
+**Required**: Yes
+
+The task or question for the agent.
+
+```typescript
+// Simple string prompt
+query({ prompt: "Analyze the codebase" })
+
+// Streaming prompt (advanced)
+query({ prompt: streamingUserMessages() })
+```
+
+### options
+
+**Type**: `Options`
+**Required**: No
+
+Configuration options for the query.
+
+---
+
+## Options Reference
+
+### model
+
+**Type**: `"sonnet" | "haiku" | "opus" | "claude-sonnet-4-5" | "inherit"`
+**Default**: `"sonnet"`
+
+Model to use for the agent.
+
+```typescript
+options: {
+  model: "claude-sonnet-4-5"  // Specific version
+  model: "haiku"               // Fast
+  model: "opus"                // Maximum capability
+  model: "inherit"             // Use parent model (subagents)
+}
+```
+
+### workingDirectory
+
+**Type**: `string`
+**Default**: Current working directory
+
+Directory where agent operates.
+
+```typescript
+options: {
+  workingDirectory: "/path/to/project"
+}
+```
+
+### systemPrompt
+
+**Type**: `string | { type: 'preset', preset: 'claude_code' }`
+**Default**: None
+
+System prompt that defines agent behavior.
+
+```typescript
+// Custom prompt
+options: {
+  systemPrompt: "You are a security-focused code reviewer."
+}
+
+// Use CLAUDE.md from project
+options: {
+  systemPrompt: { type: 'preset', preset: 'claude_code' },
+  settingSources: ["project"]  // Required to load CLAUDE.md
+}
+```
+
+### allowedTools
+
+**Type**: `string[]`
+**Default**: All tools
+
+Whitelist of tools agent can use.
+
+```typescript
+options: {
+  allowedTools: ["Read", "Grep", "Glob"]  // Read-only
+}
+```
+
+### disallowedTools
+
+**Type**: `string[]`
+**Default**: None
+
+Blacklist of tools agent cannot use.
+
+```typescript
+options: {
+  disallowedTools: ["Bash", "Write", "Edit"]  // No modifications
+}
+```
+
+**Note**: If both specified, `allowedTools` wins.
+
+### permissionMode
+
+**Type**: `"default" | "acceptEdits" | "bypassPermissions"`
+**Default**: `"default"`
+
+Permission strategy.
+
+```typescript
+options: {
+  permissionMode: "default"            // Standard checks
+  permissionMode: "acceptEdits"        // Auto-approve edits
+  permissionMode: "bypassPermissions"  // Skip all checks (caution!)
+}
+```
+
+### canUseTool
+
+**Type**: `(toolName: string, input: any) => Promise<PermissionDecision>`
+**Default**: None
+
+Custom permission logic.
+
+```typescript
+options: {
+  canUseTool: async (toolName, input) => {
+    if (toolName === 'Bash' && input.command.includes('rm -rf')) {
+      return { behavior: "deny", message: "Blocked" };
+    }
+    return { behavior: "allow" };
+  }
+}
+```
+
+**PermissionDecision**:
+- `{ behavior: "allow" }` - Allow execution
+- `{ behavior: "deny", message?: string }` - Block execution
+- `{ behavior: "ask", message?: string }` - Prompt user
+
+### agents
+
+**Type**: `Record<string, AgentDefinition>`
+**Default**: None
+
+Subagent definitions.
+
+```typescript
+options: {
+  agents: {
+    "test-runner": {
+      description: "Run test suites",
+      prompt: "You run tests. Fail if any test fails.",
+      tools: ["Bash", "Read"],
+      model: "haiku"
+    }
+  }
+}
+```
+
+**AgentDefinition**:
+- `description` (string, required) - When to use agent
+- `prompt` (string, required) - Agent's system prompt
+- `tools` (string[], optional) - Allowed tools
+- `model` (string, optional) - Model override
+
+### mcpServers
+
+**Type**: `Record<string, McpServerConfig>`
+**Default**: None
+
+MCP server configurations.
+
+```typescript
+options: {
+  mcpServers: {
+    "custom-server": customServer,  // In-process
+    "filesystem": {                 // External (stdio)
+      command: "npx",
+      args: ["@modelcontextprotocol/server-filesystem"]
+    },
+    "remote": {                     // External (HTTP)
+      url: "https://api.example.com/mcp",
+      headers: { "Authorization": "Bearer token" }
+    }
+  }
+}
+```
+
+### settingSources
+
+**Type**: `("user" | "project" | "local")[]`
+**Default**: `[]` (no filesystem settings)
+
+Filesystem settings to load.
+
+```typescript
+options: {
+  settingSources: ["project"]                   // Project only (CI/CD)
+  settingSources: ["user", "project", "local"]  // All sources
+  settingSources: []                            // Isolated (no files)
+}
+```
+
+**Files**:
+- `user` = `~/.claude/settings.json`
+- `project` = `.claude/settings.json`
+- `local` = `.claude/settings.local.json`
+
+**Priority**: Programmatic > Local > Project > User
+
+### resume
+
+**Type**: `string`
+**Default**: None
+
+Session ID to resume.
+
+```typescript
+options: {
+  resume: "session-id-here"
+}
+```
+
+### forkSession
+
+**Type**: `boolean`
+**Default**: `false`
+
+Create new branch from resumed session.
+
+```typescript
+options: {
+  resume: "session-id-here",
+  forkSession: true  // New branch, original unchanged
+}
+```
+
+---
+
+## Return Value
+
+**Type**: `AsyncGenerator<SDKMessage, void>`
+
+Asynchronous generator yielding messages.
+
+```typescript
+const response = query({ prompt: "..." });
+
+for await (const message of response) {
+  // Process message
+}
+```
+
+---
+
+## Message Types
+
+See full details in SKILL.md. Summary:
+
+| Type | When | Data |
+|------|------|------|
+| `system` | Session events | `session_id`, `model`, `tools` |
+| `assistant` | Agent response | `content` (string or blocks) |
+| `tool_call` | Tool requested | `tool_name`, `input` |
+| `tool_result` | Tool completed | `tool_name`, `result` |
+| `error` | Error occurred | `error` object |
+
+---
+
+## Usage Patterns
+
+### Basic Query
+
+```typescript
+const response = query({
+  prompt: "Analyze code",
+  options: { model: "sonnet" }
+});
+
+for await (const message of response) {
+  if (message.type === 'assistant') {
+    console.log(message.content);
+  }
+}
+```
+
+### With Tools
+
+```typescript
+const response = query({
+  prompt: "Review and fix bugs",
+  options: {
+    model: "sonnet",
+    allowedTools: ["Read", "Grep", "Edit"]
+  }
+});
+```
+
+### With Subagents
+
+```typescript
+const response = query({
+  prompt: "Deploy to production",
+  options: {
+    agents: {
+      "tester": { /* ... */ },
+      "deployer": { /* ... */ }
+    }
+  }
+});
+```
+
+### With Custom Tools
+
+```typescript
+const response = query({
+  prompt: "Get weather and send notification",
+  options: {
+    mcpServers: { "weather": weatherServer },
+    allowedTools: ["mcp__weather__get_weather"]
+  }
+});
+```
+
+### With Session Management
+
+```typescript
+// Start
+let session = await startSession("Build API");
+
+// Resume
+await resumeSession(session, "Add auth");
+
+// Fork
+await forkSession(session, "Try GraphQL instead");
+```
+
+---
+
+## Best Practices
+
+### ✅ Do
+
+- Set specific `allowedTools` for security
+- Use `canUseTool` for fine-grained control
+- Implement error handling for all queries
+- Capture `session_id` for resuming
+- Use `workingDirectory` for clarity
+- Test MCP servers independently
+- Monitor tool execution with `tool_call` messages
+
+### ❌ Don't
+
+- Use `bypassPermissions` in production (unless sandboxed)
+- Ignore error messages
+- Skip session ID capture if planning to resume
+- Allow unrestricted Bash without `canUseTool`
+- Load user settings in CI/CD
+- Use duplicate tool names
+
+---
+
+## Error Handling
+
+```typescript
+try {
+  const response = query({ prompt: "..." });
+  for await (const message of response) {
+    if (message.type === 'error') {
+      console.error('Agent error:', message.error);
+    }
+  }
+} catch (error) {
+  if (error.code === 'CLI_NOT_FOUND') {
+    console.error('Install Claude Code CLI');
+  }
+}
+```
+
+---
+
+## TypeScript Types
+
+```typescript
+type Options = {
+  model?: "sonnet" | "haiku" | "opus" | string;
+  workingDirectory?: string;
+  systemPrompt?: string | { type: 'preset', preset: 'claude_code' };
+  allowedTools?: string[];
+  disallowedTools?: string[];
+  permissionMode?: "default" | "acceptEdits" | "bypassPermissions";
+  canUseTool?: (toolName: string, input: any) => Promise<PermissionDecision>;
+  agents?: Record<string, AgentDefinition>;
+  mcpServers?: Record<string, McpServerConfig>;
+  settingSources?: ("user" | "project" | "local")[];
+  resume?: string;
+  forkSession?: boolean;
+};
+
+type AgentDefinition = {
+  description: string;
+  prompt: string;
+  tools?: string[];
+  model?: "sonnet" | "opus" | "haiku" | "inherit";
+};
+
+type PermissionDecision =
+  | { behavior: "allow" }
+  | { behavior: "deny"; message?: string }
+  | { behavior: "ask"; message?: string };
+```
+
+---
+
+**For more details**: See SKILL.md
+**Official docs**: https://docs.claude.com/en/api/agent-sdk/typescript

references/session-management.md

@@ -0,0 +1,419 @@
+# Session Management Guide
+
+Complete guide to sessions, resuming, and forking in Claude Agent SDK.
+
+---
+
+## What Are Sessions?
+
+Sessions enable:
+- **Persistent conversations** - Resume where you left off
+- **Context preservation** - Agent remembers everything
+- **Alternative paths** - Fork to explore different approaches
+
+---
+
+## Session Lifecycle
+
+```
+Start → Capture Session ID → Resume → Resume → ... → End
+                                  ↓
+                                Fork (alternative path)
+```
+
+---
+
+## Starting a Session
+
+Every `query()` call creates a session.
+
+```typescript
+let sessionId: string | undefined;
+
+const response = query({
+  prompt: "Build a REST API",
+  options: { model: "sonnet" }
+});
+
+for await (const message of response) {
+  if (message.type === 'system' && message.subtype === 'init') {
+    sessionId = message.session_id;
+    console.log(`Session started: ${sessionId}`);
+  }
+}
+```
+
+**CRITICAL**: Capture `session_id` from `system` init message.
+
+---
+
+## Resuming a Session
+
+Continue a previous conversation.
+
+```typescript
+const resumed = query({
+  prompt: "Now add authentication",
+  options: {
+    resume: sessionId,  // Resume previous session
+    model: "sonnet"
+  }
+});
+```
+
+**What's preserved**:
+- All previous messages
+- Agent's understanding of context
+- Files created/modified
+- Decisions made
+
+**What's NOT preserved**:
+- Environment variables
+- Tool availability (specify again)
+- Permission settings (specify again)
+
+---
+
+## Forking a Session
+
+Create alternative path without modifying original.
+
+```typescript
+const forked = query({
+  prompt: "Actually, make it GraphQL instead",
+  options: {
+    resume: sessionId,
+    forkSession: true,  // Creates new branch
+    model: "sonnet"
+  }
+});
+```
+
+**Result**:
+- New session created
+- Starts from same point as original
+- Original session unchanged
+- Can compare approaches
+
+---
+
+## Use Case Patterns
+
+### Pattern 1: Sequential Development
+
+Step-by-step feature building.
+
+```typescript
+// Step 1: Initial implementation
+let session = await startSession("Create user authentication");
+
+// Step 2: Add feature
+session = await resumeSession(session, "Add OAuth support");
+
+// Step 3: Add tests
+session = await resumeSession(session, "Write integration tests");
+
+// Step 4: Deploy
+session = await resumeSession(session, "Deploy to production");
+```
+
+### Pattern 2: Exploration & Decision
+
+Try multiple approaches, choose best.
+
+```typescript
+// Start main conversation
+let mainSession = await startSession("Design payment system");
+
+// Explore option A
+let optionA = await forkSession(mainSession, "Use Stripe");
+
+// Explore option B
+let optionB = await forkSession(mainSession, "Use PayPal");
+
+// Explore option C
+let optionC = await forkSession(mainSession, "Use Square");
+
+// Choose winner
+let chosenSession = optionA;  // Decision made
+await resumeSession(chosenSession, "Implement chosen approach");
+```
+
+### Pattern 3: Multi-User Collaboration
+
+Multiple developers, independent work.
+
+```typescript
+// Developer A starts work
+let sessionA = await startSession("Implement user profile page");
+
+// Developer B forks for different feature
+let sessionB = await forkSession(sessionA, "Add avatar upload");
+
+// Developer C forks for another feature
+let sessionC = await forkSession(sessionA, "Implement search");
+
+// All can work independently
+```
+
+### Pattern 4: Error Recovery
+
+Backup and restore points.
+
+```typescript
+// Save checkpoint before risky operation
+let checkpoint = sessionId;
+
+try {
+  await resumeSession(checkpoint, "Refactor entire auth system");
+} catch (error) {
+  console.log("Refactor failed, restoring from checkpoint");
+  await forkSession(checkpoint, "Try safer incremental refactor");
+}
+```
+
+### Pattern 5: A/B Testing
+
+Test different implementations.
+
+```typescript
+let baseline = await startSession("Implement feature X");
+
+// Approach A
+let approachA = await forkSession(baseline, "Use algorithm A");
+const metricsA = await measurePerformance(approachA);
+
+// Approach B
+let approachB = await forkSession(baseline, "Use algorithm B");
+const metricsB = await measurePerformance(approachB);
+
+// Compare and choose
+const winner = metricsA.better(metricsB) ? approachA : approachB;
+```
+
+---
+
+## Session Best Practices
+
+### ✅ Do
+
+- Always capture `session_id` from init message
+- Store session IDs for later use
+- Use descriptive prompts when resuming
+- Fork for alternative approaches
+- Test resuming before deploying
+- Consider session lifetime limits
+
+### ❌ Don't
+
+- Forget to capture session ID
+- Assume sessions last forever
+- Resume with completely unrelated prompts
+- Fork excessively (creates many branches)
+- Rely on sessions for critical state
+- Skip testing resume functionality
+
+---
+
+## Helper Functions
+
+```typescript
+async function startSession(prompt: string): Promise<string> {
+  let sessionId: string | undefined;
+
+  const response = query({ prompt, options: { model: "sonnet" } });
+
+  for await (const message of response) {
+    if (message.type === 'system' && message.subtype === 'init') {
+      sessionId = message.session_id;
+    }
+  }
+
+  if (!sessionId) throw new Error('Failed to start session');
+  return sessionId;
+}
+
+async function resumeSession(
+  sessionId: string,
+  prompt: string
+): Promise<void> {
+  const response = query({
+    prompt,
+    options: { resume: sessionId, model: "sonnet" }
+  });
+
+  for await (const message of response) {
+    if (message.type === 'assistant') {
+      console.log(message.content);
+    }
+  }
+}
+
+async function forkSession(
+  sessionId: string,
+  prompt: string
+): Promise<string> {
+  let newSessionId: string | undefined;
+
+  const response = query({
+    prompt,
+    options: {
+      resume: sessionId,
+      forkSession: true,
+      model: "sonnet"
+    }
+  });
+
+  for await (const message of response) {
+    if (message.type === 'system' && message.subtype === 'init') {
+      newSessionId = message.session_id;
+    } else if (message.type === 'assistant') {
+      console.log(message.content);
+    }
+  }
+
+  if (!newSessionId) throw new Error('Failed to fork session');
+  return newSessionId;
+}
+```
+
+---
+
+## Session Storage
+
+Store sessions for later use:
+
+```typescript
+// In-memory storage
+const sessions = new Map<string, { id: string; created: Date }>();
+
+async function saveSession(prompt: string) {
+  const id = await startSession(prompt);
+  sessions.set(id, { id, created: new Date() });
+  return id;
+}
+
+function getSession(id: string) {
+  return sessions.get(id);
+}
+
+// Database storage
+async function saveSessionToDb(sessionId: string, metadata: any) {
+  await db.insert('sessions', {
+    id: sessionId,
+    created_at: new Date(),
+    metadata
+  });
+}
+```
+
+---
+
+## Session Limits
+
+### Context Window
+
+Sessions have context window limits (200k tokens for Sonnet).
+
+**Strategies**:
+- SDK auto-compacts context
+- Fork to start fresh from a point
+- Summarize and start new session
+
+### Lifetime
+
+Sessions may expire after inactivity.
+
+**Strategies**:
+- Don't rely on sessions lasting indefinitely
+- Store important state separately
+- Test resume functionality
+
+---
+
+## Troubleshooting
+
+### Session Not Found
+
+**Problem**: `"Invalid session ID"`
+
+**Causes**:
+- Session expired
+- Invalid session ID
+- Session from different CLI instance
+
+**Solution**: Start new session
+
+### Context Preserved Incorrectly
+
+**Problem**: Agent doesn't remember previous work
+
+**Causes**:
+- Different settings/tools specified
+- Context window exceeded
+- Fork instead of resume
+
+**Solution**: Verify using `resume` not `forkSession`
+
+### Too Many Forks
+
+**Problem**: Hard to track branches
+
+**Solution**: Limit forking, clean up unused branches
+
+---
+
+## Complete Example
+
+```typescript
+class SessionManager {
+  private sessions = new Map<string, string>();
+
+  async start(name: string, prompt: string): Promise<string> {
+    const id = await startSession(prompt);
+    this.sessions.set(name, id);
+    console.log(`✨ Started: ${name}`);
+    return id;
+  }
+
+  async resume(name: string, prompt: string): Promise<void> {
+    const id = this.sessions.get(name);
+    if (!id) throw new Error(`Session ${name} not found`);
+    console.log(`↪️  Resuming: ${name}`);
+    await resumeSession(id, prompt);
+  }
+
+  async fork(
+    fromName: string,
+    newName: string,
+    prompt: string
+  ): Promise<string> {
+    const fromId = this.sessions.get(fromName);
+    if (!fromId) throw new Error(`Session ${fromName} not found`);
+    console.log(`🔀 Forking: ${fromName} → ${newName}`);
+    const newId = await forkSession(fromId, prompt);
+    this.sessions.set(newName, newId);
+    return newId;
+  }
+
+  list(): string[] {
+    return Array.from(this.sessions.keys());
+  }
+}
+
+// Usage
+const manager = new SessionManager();
+
+await manager.start("main", "Build a web app");
+await manager.resume("main", "Add authentication");
+await manager.fork("main", "option-a", "Use JWT tokens");
+await manager.fork("main", "option-b", "Use sessions");
+
+console.log("Sessions:", manager.list());
+// ["main", "option-a", "option-b"]
+```
+
+---
+
+**For more details**: See SKILL.md
+**Template**: templates/session-management.ts

references/subagents-patterns.md

@@ -0,0 +1,464 @@
+# Subagents Patterns
+
+Guide to designing and orchestrating specialized subagents.
+
+---
+
+## What Are Subagents?
+
+Specialized agents with:
+- **Specific expertise** - Focused domain knowledge
+- **Custom tools** - Only tools they need
+- **Different models** - Match capability to task cost
+- **Dedicated prompts** - Tailored instructions
+
+---
+
+## When to Use Subagents
+
+### ✅ Use Subagents When:
+
+- Task requires different expertise areas
+- Some subtasks need different models (cost optimization)
+- Tool access should be restricted per role
+- Clear separation of concerns needed
+- Multiple steps with specialized knowledge
+
+### ❌ Don't Use Subagents When:
+
+- Single straightforward task
+- All work can be done by one agent
+- Overhead of orchestration > benefit
+- Tools/permissions don't vary
+
+---
+
+## AgentDefinition Structure
+
+```typescript
+type AgentDefinition = {
+  description: string;        // When to use this agent
+  prompt: string;             // System prompt for agent
+  tools?: string[];           // Allowed tools (optional)
+  model?: 'sonnet' | 'opus' | 'haiku' | 'inherit';  // Model (optional)
+}
+```
+
+### Field Guidelines
+
+**description**:
+- Clear, action-oriented
+- When to invoke this agent
+- 1-2 sentences
+- Examples: "Run test suites", "Deploy to production"
+
+**prompt**:
+- Agent's role and behavior
+- Instructions and constraints
+- What to do and what not to do
+- Can be detailed (100-500 tokens)
+
+**tools**:
+- If omitted, inherits all tools from main agent
+- Use to restrict agent to specific tools
+- Examples: `["Read", "Grep"]` for read-only
+
+**model**:
+- `"haiku"` - Fast, cost-effective ($0.25/$1.25 per MTok)
+- `"sonnet"` - Balanced ($3/$15 per MTok)
+- `"opus"` - Maximum capability ($15/$75 per MTok)
+- `"inherit"` - Use main agent's model
+- If omitted, inherits main agent's model
+
+---
+
+## Design Patterns
+
+### Pattern 1: DevOps Pipeline
+
+```typescript
+agents: {
+  "test-runner": {
+    description: "Run automated test suites and verify coverage",
+    prompt: `You run tests.
+
+Execute:
+- Unit tests
+- Integration tests
+- End-to-end tests
+
+FAIL if any tests fail. Report clear errors.`,
+    tools: ["Bash", "Read"],
+    model: "haiku"  // Fast, cost-effective
+  },
+
+  "security-checker": {
+    description: "Security audits and vulnerability scanning",
+    prompt: `You check security.
+
+Scan for:
+- Exposed secrets
+- Dependency vulnerabilities
+- Permission issues
+- OWASP compliance
+
+Block deployment if critical issues found.`,
+    tools: ["Read", "Grep", "Bash"],
+    model: "sonnet"  // Balance for analysis
+  },
+
+  "deployer": {
+    description: "Application deployment and rollbacks",
+    prompt: `You deploy applications.
+
+Process:
+1. Deploy to staging
+2. Verify health checks
+3. Deploy to production
+4. Create rollback plan
+
+ALWAYS have rollback ready.`,
+    tools: ["Bash", "Read"],
+    model: "sonnet"  // Reliable for critical ops
+  }
+}
+```
+
+### Pattern 2: Code Review Workflow
+
+```typescript
+agents: {
+  "syntax-checker": {
+    description: "Check syntax, formatting, and linting",
+    prompt: `You check syntax and formatting.
+
+Run:
+- ESLint
+- Prettier
+- TypeScript type checking
+
+Report all violations clearly.`,
+    tools: ["Bash", "Read"],
+    model: "haiku"  // Fast checks
+  },
+
+  "logic-reviewer": {
+    description: "Review logic, algorithms, and architecture",
+    prompt: `You review code logic.
+
+Check:
+- Algorithmic correctness
+- Edge cases
+- Performance issues
+- Design patterns
+
+Suggest improvements.`,
+    tools: ["Read", "Grep"],
+    model: "sonnet"  // Complex analysis
+  },
+
+  "security-reviewer": {
+    description: "Review for security vulnerabilities",
+    prompt: `You review security.
+
+Check:
+- SQL injection
+- XSS vulnerabilities
+- Authentication bypass
+- Data exposure
+
+FAIL on critical issues.`,
+    tools: ["Read", "Grep"],
+    model: "sonnet"  // Security expertise
+  }
+}
+```
+
+### Pattern 3: Content Generation
+
+```typescript
+agents: {
+  "researcher": {
+    description: "Research topics and gather information",
+    prompt: `You research topics.
+
+Gather:
+- Relevant information
+- Latest trends
+- Best practices
+- Examples
+
+Provide comprehensive research.`,
+    tools: ["WebSearch", "WebFetch", "Read"],
+    model: "haiku"  // Fast research
+  },
+
+  "writer": {
+    description: "Write content based on research",
+    prompt: `You write content.
+
+Write:
+- Clear, engaging prose
+- Well-structured
+- Audience-appropriate
+- Fact-based
+
+Use research provided.`,
+    tools: ["Write", "Read"],
+    model: "sonnet"  // Quality writing
+  },
+
+  "editor": {
+    description: "Edit and polish content",
+    prompt: `You edit content.
+
+Check:
+- Grammar and spelling
+- Clarity and flow
+- Consistency
+- Formatting
+
+Polish to perfection.`,
+    tools: ["Read", "Edit"],
+    model: "sonnet"  // Quality editing
+  }
+}
+```
+
+### Pattern 4: Incident Response
+
+```typescript
+agents: {
+  "incident-detector": {
+    description: "Detect and triage incidents",
+    prompt: `You detect incidents.
+
+Monitor:
+- Error rates
+- Response times
+- System health
+- Alerts
+
+Assess severity and impact.`,
+    tools: ["Bash", "Read"],
+    model: "haiku"  // Fast detection
+  },
+
+  "root-cause-analyzer": {
+    description: "Analyze root cause of incidents",
+    prompt: `You analyze root causes.
+
+Investigate:
+- Logs
+- Metrics
+- Recent changes
+- Dependencies
+
+Identify exact cause.`,
+    tools: ["Bash", "Read", "Grep"],
+    model: "sonnet"  // Deep analysis
+  },
+
+  "fix-implementer": {
+    description: "Implement fixes for incidents",
+    prompt: `You implement fixes.
+
+Fix:
+- Apply patches
+- Rollback changes
+- Update config
+- Deploy hotfixes
+
+Verify fix resolves issue.`,
+    tools: ["Read", "Edit", "Bash"],
+    model: "sonnet"  // Careful fixes
+  }
+}
+```
+
+---
+
+## Orchestration Strategies
+
+### Sequential Execution
+
+Main agent coordinates agents in order:
+
+```
+test-runner → security-checker → deployer
+```
+
+**Use when**: Steps must happen in order
+
+### Parallel Execution
+
+Main agent delegates to multiple agents at once:
+
+```
+test-runner
+security-checker  } in parallel
+code-reviewer
+```
+
+**Use when**: Steps are independent
+
+### Conditional Execution
+
+Main agent decides which agents to use based on context:
+
+```
+if (tests fail) → test-fixer
+if (security issue) → security-fixer
+if (all pass) → deployer
+```
+
+**Use when**: Different paths needed
+
+---
+
+## Model Selection Strategy
+
+### Cost vs Capability
+
+| Model | Cost (in/out) | Speed | Use For |
+|-------|---------------|-------|---------|
+| Haiku | $0.25/$1.25 | Fastest | Monitoring, simple checks |
+| Sonnet | $3/$15 | Medium | Code review, analysis |
+| Opus | $15/$75 | Slowest | Complex reasoning |
+
+### Optimization Tips
+
+```typescript
+// ✅ Good: Match model to task
+agents: {
+  "monitor": { model: "haiku" },      // Simple checks
+  "reviewer": { model: "sonnet" },     // Analysis
+  "architect": { model: "opus" }       // Complex design
+}
+
+// ❌ Bad: Opus for everything
+agents: {
+  "monitor": { model: "opus" },   // Wasteful
+  "reviewer": { model: "opus" },  // Wasteful
+  "simple-task": { model: "opus" } // Wasteful
+}
+```
+
+---
+
+## Tool Restriction Patterns
+
+### Read-Only Agent
+
+```typescript
+{
+  description: "Analyze code without modifications",
+  tools: ["Read", "Grep", "Glob"],
+  model: "haiku"
+}
+```
+
+### Write-Only Agent
+
+```typescript
+{
+  description: "Generate new files",
+  tools: ["Write"],
+  model: "sonnet"
+}
+```
+
+### Modify Agent
+
+```typescript
+{
+  description: "Edit existing files",
+  tools: ["Read", "Edit"],
+  model: "sonnet"
+}
+```
+
+### Full Access Agent
+
+```typescript
+{
+  description: "Complete control",
+  tools: ["Read", "Write", "Edit", "Bash", "Grep", "Glob"],
+  model: "sonnet"
+}
+```
+
+---
+
+## Communication Between Agents
+
+Agents communicate through the main agent:
+
+```
+Main Agent: "Run tests"
+  ↓
+Test Runner: "Tests passed"
+  ↓
+Main Agent: "Deploy to staging"
+  ↓
+Deployer: "Deployed to staging"
+  ↓
+Main Agent: "Verify health"
+  ↓
+Monitor: "Health checks passing"
+  ↓
+Main Agent: "Deploy to production"
+```
+
+**No direct agent-to-agent communication**.
+
+---
+
+## Best Practices
+
+### ✅ Do
+
+- Give agents clear, specific roles
+- Match model to task complexity
+- Restrict tools per agent's needs
+- Write detailed prompts with constraints
+- Use descriptive agent names
+- Test agents independently
+- Monitor which agents are invoked
+
+### ❌ Don't
+
+- Create overlapping responsibilities
+- Use Opus for simple tasks
+- Give all agents all tools
+- Write vague prompts
+- Use generic names like "agent1"
+- Skip testing in isolation
+- Assume agents will coordinate perfectly
+
+---
+
+## Troubleshooting
+
+### Agent Not Invoked
+
+**Problem**: Main agent doesn't call subagent
+
+**Solution**: Improve `description` to be more specific
+
+### Wrong Agent Invoked
+
+**Problem**: Main agent calls incorrect subagent
+
+**Solution**: Make descriptions more distinct
+
+### Agent Lacks Capability
+
+**Problem**: Agent can't complete task
+
+**Solution**: Add required tools or upgrade model
+
+---
+
+**For more details**: See SKILL.md
+**Examples**: templates/subagents-orchestration.ts

references/top-errors.md

@@ -0,0 +1,503 @@
+# Top Errors & Solutions
+
+Complete reference for common Claude Agent SDK errors and how to fix them.
+
+---
+
+## Error #1: CLI Not Found
+
+### Error Message
+```
+"Claude Code CLI not installed"
+```
+
+### Why It Happens
+The SDK requires Claude Code CLI to be installed globally, but it's not found in PATH.
+
+### Solution
+```bash
+npm install -g @anthropic-ai/claude-code
+```
+
+Verify installation:
+```bash
+which claude-code
+# Should output: /usr/local/bin/claude-code or similar
+```
+
+### Prevention
+- Install CLI before using SDK
+- Add to project setup documentation
+- Check CLI availability in CI/CD
+
+---
+
+## Error #2: Authentication Failed
+
+### Error Message
+```
+"Invalid API key"
+"Authentication failed"
+```
+
+### Why It Happens
+- ANTHROPIC_API_KEY environment variable not set
+- API key is invalid or expired
+- API key has wrong format
+
+### Solution
+```bash
+# Set API key
+export ANTHROPIC_API_KEY="sk-ant-..."
+
+# Verify it's set
+echo $ANTHROPIC_API_KEY
+```
+
+Get API key:
+1. Visit https://console.anthropic.com/
+2. Navigate to API Keys section
+3. Create new key
+4. Copy and save securely
+
+### Prevention
+- Use environment variables (never hardcode)
+- Check key before running
+- Add to .env.example template
+- Document setup process
+
+---
+
+## Error #3: Permission Denied
+
+### Error Message
+```
+"Tool use blocked"
+"Permission denied for tool: Bash"
+```
+
+### Why It Happens
+- Tool not in `allowedTools` array
+- `permissionMode` is too restrictive
+- Custom `canUseTool` callback denied execution
+
+### Solution
+```typescript
+// Add tool to allowedTools
+options: {
+  allowedTools: ["Read", "Write", "Edit", "Bash"]  // Add needed tools
+}
+
+// Or use less restrictive permission mode
+options: {
+  permissionMode: "acceptEdits"  // Auto-approve edits
+}
+
+// Or check canUseTool logic
+options: {
+  canUseTool: async (toolName, input) => {
+    console.log("Tool requested:", toolName);  // Debug
+    return { behavior: "allow" };
+  }
+}
+```
+
+### Prevention
+- Set appropriate `allowedTools` from start
+- Test permission logic thoroughly
+- Use `permissionMode: "bypassPermissions"` in CI/CD
+
+---
+
+## Error #4: Context Length Exceeded
+
+### Error Message
+```
+"Prompt too long"
+"Context length exceeded"
+"Too many tokens"
+```
+
+### Why It Happens
+- Input prompt exceeds model's context window (200k tokens)
+- Long conversation without pruning
+- Large files in context
+
+### Solution
+SDK auto-compacts context, but you can:
+
+```typescript
+// Fork session to start fresh from a point
+const forked = query({
+  prompt: "Continue with fresh context",
+  options: {
+    resume: sessionId,
+    forkSession: true  // Start fresh
+  }
+});
+
+// Or reduce context
+options: {
+  allowedTools: ["Read", "Grep"],  // Limit tools
+  systemPrompt: "Keep responses concise."
+}
+```
+
+### Prevention
+- Use session forking for long tasks
+- Keep prompts focused
+- Don't load unnecessary files
+- Monitor context usage
+
+---
+
+## Error #5: Tool Execution Timeout
+
+### Error Message
+```
+"Tool did not respond"
+"Tool execution timeout"
+```
+
+### Why It Happens
+- Tool takes too long (>5 minutes default)
+- Infinite loop in tool implementation
+- Network timeout for external tools
+
+### Solution
+
+For custom tools:
+```typescript
+tool("long_task", "Description", schema, async (args) => {
+  // Add timeout
+  const timeout = 60000; // 1 minute
+  const promise = performLongTask(args);
+  const result = await Promise.race([
+    promise,
+    new Promise((_, reject) =>
+      setTimeout(() => reject(new Error('Timeout')), timeout)
+    )
+  ]);
+  return { content: [{ type: "text", text: result }] };
+})
+```
+
+For bash commands:
+```typescript
+// Add timeout to bash command
+command: "timeout 60s long-running-command"
+```
+
+### Prevention
+- Set reasonable timeouts
+- Optimize tool implementations
+- Use background jobs for long tasks
+- Test tools independently
+
+---
+
+## Error #6: Session Not Found
+
+### Error Message
+```
+"Invalid session ID"
+"Session not found"
+```
+
+### Why It Happens
+- Session ID is incorrect
+- Session expired (old)
+- Session from different CLI instance
+
+### Solution
+```typescript
+// Ensure session ID captured correctly
+let sessionId: string | undefined;
+for await (const message of response) {
+  if (message.type === 'system' && message.subtype === 'init') {
+    sessionId = message.session_id;  // Capture here
+    console.log("Session:", sessionId);  // Verify
+  }
+}
+
+// Use correct session ID
+query({
+  prompt: "...",
+  options: { resume: sessionId }  // Must match exactly
+});
+```
+
+### Prevention
+- Always capture session_id from system init
+- Store session IDs reliably
+- Don't rely on sessions lasting indefinitely
+- Handle session errors gracefully
+
+---
+
+## Error #7: MCP Server Connection Failed
+
+### Error Message
+```
+"Server connection error"
+"MCP server not responding"
+"Failed to connect to MCP server"
+```
+
+### Why It Happens
+- Server command/URL incorrect
+- Server crashed or not running
+- Network issues (HTTP/SSE servers)
+- Missing dependencies
+
+### Solution
+
+For stdio servers:
+```typescript
+// Verify command works independently
+// Test: npx @modelcontextprotocol/server-filesystem
+options: {
+  mcpServers: {
+    "filesystem": {
+      command: "npx",  // Verify npx is available
+      args: ["@modelcontextprotocol/server-filesystem"],
+      env: {
+        ALLOWED_PATHS: "/path"  // Verify path exists
+      }
+    }
+  }
+}
+```
+
+For HTTP servers:
+```typescript
+// Test URL separately
+const testResponse = await fetch("https://api.example.com/mcp");
+console.log(testResponse.status);  // Should be 200
+```
+
+### Prevention
+- Test MCP servers independently before integration
+- Verify command/URL works
+- Add error handling for server failures
+- Use health checks
+
+---
+
+## Error #8: Subagent Definition Error
+
+### Error Message
+```
+"Invalid AgentDefinition"
+"Agent configuration error"
+```
+
+### Why It Happens
+- Missing required fields (`description` or `prompt`)
+- Invalid `model` value
+- Invalid `tools` array
+
+### Solution
+```typescript
+agents: {
+  "my-agent": {
+    description: "Clear description of when to use",  // Required
+    prompt: "Detailed system prompt",                  // Required
+    tools: ["Read", "Write"],                          // Optional
+    model: "sonnet"                                    // Optional
+  }
+}
+```
+
+### Prevention
+- Always include `description` and `prompt`
+- Use TypeScript types
+- Test agent definitions
+- Follow examples in templates
+
+---
+
+## Error #9: Settings File Not Found
+
+### Error Message
+```
+"Cannot read settings"
+"Settings file not found"
+```
+
+### Why It Happens
+- `settingSources` includes non-existent file
+- File path incorrect
+- File permissions deny read
+
+### Solution
+```typescript
+// Check file exists before loading
+import fs from 'fs';
+
+const projectSettingsPath = '.claude/settings.json';
+const settingSources = [];
+
+if (fs.existsSync(projectSettingsPath)) {
+  settingSources.push('project');
+}
+
+options: {
+  settingSources  // Only existing files
+}
+```
+
+### Prevention
+- Check file exists before including in sources
+- Use empty array for isolated execution
+- Handle missing files gracefully
+
+---
+
+## Error #10: Tool Name Collision
+
+### Error Message
+```
+"Duplicate tool name"
+"Tool already defined"
+```
+
+### Why It Happens
+- Two MCP servers define same tool name
+- Tool name conflicts with built-in tool
+
+### Solution
+```typescript
+// Use unique tool names
+const server1 = createSdkMcpServer({
+  name: "service-a",
+  tools: [
+    tool("service_a_process", ...)  // Prefix with server name
+  ]
+});
+
+const server2 = createSdkMcpServer({
+  name: "service-b",
+  tools: [
+    tool("service_b_process", ...)  // Different name
+  ]
+});
+```
+
+### Prevention
+- Use unique tool names
+- Prefix tools with server name
+- Test integration before deployment
+
+---
+
+## Error #11: Zod Schema Validation Error
+
+### Error Message
+```
+"Invalid tool input"
+"Schema validation failed"
+```
+
+### Why It Happens
+- Agent provided data that doesn't match Zod schema
+- Schema too restrictive
+- Missing `.describe()` on fields
+
+### Solution
+```typescript
+// Add descriptive schemas
+{
+  email: z.string().email().describe("User email address"),
+  age: z.number().int().min(0).max(120).describe("Age in years"),
+  role: z.enum(["admin", "user"]).describe("User role")
+}
+
+// Make fields optional if appropriate
+{
+  email: z.string().email(),
+  phoneOptional: z.string().optional()  // Not required
+}
+```
+
+### Prevention
+- Use `.describe()` on all fields
+- Add validation constraints
+- Test with various inputs
+- Make optional fields explicit
+
+---
+
+## Error #12: Filesystem Permission Denied
+
+### Error Message
+```
+"Access denied"
+"Cannot access path"
+"EACCES: permission denied"
+```
+
+### Why It Happens
+- Path outside `workingDirectory`
+- No read/write permissions
+- Protected system directory
+
+### Solution
+```typescript
+// Set correct working directory
+options: {
+  workingDirectory: "/path/to/accessible/dir"
+}
+
+// Or fix permissions
+// chmod +r file.txt  (add read)
+// chmod +w file.txt  (add write)
+```
+
+### Prevention
+- Set appropriate `workingDirectory`
+- Verify file permissions
+- Don't access system directories
+- Use dedicated project directories
+
+---
+
+## General Error Handling Pattern
+
+```typescript
+try {
+  const response = query({ prompt: "...", options: { ... } });
+
+  for await (const message of response) {
+    if (message.type === 'error') {
+      console.error('Agent error:', message.error);
+      // Handle non-fatal errors
+    }
+  }
+} catch (error) {
+  console.error('Fatal error:', error);
+
+  // Handle specific errors
+  switch (error.code) {
+    case 'CLI_NOT_FOUND':
+      console.error('Install: npm install -g @anthropic-ai/claude-code');
+      break;
+    case 'AUTHENTICATION_FAILED':
+      console.error('Check ANTHROPIC_API_KEY');
+      break;
+    case 'RATE_LIMIT_EXCEEDED':
+      console.error('Rate limited. Retry with backoff.');
+      break;
+    case 'CONTEXT_LENGTH_EXCEEDED':
+      console.error('Reduce context or fork session');
+      break;
+    default:
+      console.error('Unexpected error:', error);
+  }
+}
+```
+
+---
+
+**For more details**: See SKILL.md
+**Template**: templates/error-handling.ts