1735 Claude Data Structure E8fc087c

Claude Code Data Structure Documentation

Overview

This document provides comprehensive information about the data structures and formats found in the .claude directory, used by Claude Code and the Analytics Dashboard.

Directory Structure

~/.claude/
├── projects/                    # Project-specific conversations
│   ├── -Users-user-Project1/
│   │   ├── conversation.jsonl   # Main conversation file
│   │   └── settings.json        # Project settings
│   └── -Users-user-Project2/
│       └── conversation.jsonl
├── desktop/                     # Claude Desktop app data
├── statsig/                     # Analytics and feature flags
│   ├── logs/
│   └── user_overrides.json
└── settings.json               # Global Claude Code settings

JSONL Conversation Format

File Structure

Each conversation is stored in a JSONL (JSON Lines) file where each line represents a single message or event.

Message Types

1. User Messages

{
  "parentUuid": "previous-message-uuid",
  "isSidechain": false,
  "userType": "external",
  "cwd": "/Users/user/project-path",
  "sessionId": "ae93d7b5-1c54-4578-b208-603b48a88c5e",
  "version": "1.0.35",
  "type": "user",
  "message": {
    "role": "user",
    "content": "User's message text here"
  },
  "uuid": "6a8f4604-6fdd-406f-87d6-436e6cf26bd1",
  "timestamp": "2025-07-01T19:06:05.237Z"
}

2. Assistant Messages

{
  "parentUuid": "6a8f4604-6fdd-406f-87d6-436e6cf26bd1",
  "isSidechain": false,
  "userType": "external",
  "cwd": "/Users/user/project-path",
  "sessionId": "ae93d7b5-1c54-4578-b208-603b48a88c5e",
  "version": "1.0.35",
  "message": {
    "id": "msg_016xDLMzLsNRmD5PsdEjPu3N",
    "type": "message",
    "role": "assistant",
    "model": "claude-sonnet-4-20250514",
    "content": [
      {
        "type": "text",
        "text": "Assistant's response text"
      }
    ],
    "stop_reason": null,
    "stop_sequence": null,
    "usage": {
      "input_tokens": 4,
      "cache_creation_input_tokens": 15116,
      "cache_read_input_tokens": 0,
      "output_tokens": 1,
      "service_tier": "standard"
    }
  },
  "requestId": "req_011CQgpcgetL2WTXxNz8FxVs",
  "type": "assistant",
  "uuid": "2f7d6c65-27a6-40b4-aa52-0fb8cad8f9a6",
  "timestamp": "2025-07-01T19:06:09.724Z"
}

3. Tool Use Messages

{
  "message": {
    "role": "assistant",
    "content": [
      {
        "type": "text",
        "text": "I'll help you with that task."
      },
      {
        "type": "tool_use",
        "id": "toolu_01A2B3C4D5E6F7G8H9I0J1K2",
        "name": "bash",
        "input": {
          "command": "ls -la"
        }
      }
    ]
  }
}

4. Tool Result Messages

{
  "message": {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "toolu_01A2B3C4D5E6F7G8H9I0J1K2",
        "content": "drwxr-xr-x  5 user  staff   160 Jul 20 10:30 .\ndrwxr-xr-x  8 user  staff   256 Jul 20 10:29 .."
      }
    ]
  }
}

Field Definitions

Root Level Fields

Message Object Fields

Content Formats

User Content (String)

"content": "Simple text message from user"

Assistant Content (Array of Blocks)

"content": [
  {
    "type": "text",
    "text": "Text response from assistant"
  },
  {
    "type": "tool_use",
    "id": "toolu_01A2B3C4D5E6F7G8H9I0J1K2",
    "name": "bash",
    "input": {
      "command": "ls -la",
      "description": "List directory contents"
    }
  }
]

Tool Result Content (Array)

"content": [
  {
    "type": "tool_result",
    "tool_use_id": "toolu_01A2B3C4D5E6F7G8H9I0J1K2",
    "content": "Command output here...",
    "is_error": false
  }
]

Usage Object (Token Information)

"usage": {
  "input_tokens": 156,                    // Tokens in the input
  "output_tokens": 45,                    // Tokens in the output
  "cache_creation_input_tokens": 2048,    // Tokens used to create cache
  "cache_read_input_tokens": 1024,        // Tokens read from cache
  "service_tier": "standard"              // Service tier: "standard" or "premium"
}

Tool Types and Input Formats

Available Tools

1. bash - Execute shell commands
2. read - Read file contents
3. write - Write/create files
4. edit - Edit existing files
5. glob - File pattern matching
6. grep - Search within files
7. ls - List directory contents

Tool Input Examples

Bash Tool

{
  "type": "tool_use",
  "name": "bash",
  "input": {
    "command": "npm install --save express",
    "description": "Install Express.js package"
  }
}

Read Tool

{
  "type": "tool_use",
  "name": "read",
  "input": {
    "file_path": "/path/to/file.js",
    "limit": 100,
    "offset": 0
  }
}

Write Tool

{
  "type": "tool_use",
  "name": "write",
  "input": {
    "file_path": "/path/to/new-file.js",
    "content": "console.log('Hello World');"
  }
}

Edit Tool

{
  "type": "tool_use",
  "name": "edit",
  "input": {
    "file_path": "/path/to/file.js",
    "old_string": "const oldCode = 'old';",
    "new_string": "const newCode = 'new';",
    "replace_all": false
  }
}

Settings Files

Message Flow and Relationships

Conversation Threading

Messages are linked through parentUuid fields:

Message 1 (uuid: A, parentUuid: null)          # First message
├── Message 2 (uuid: B, parentUuid: A)         # Response to Message 1
    ├── Message 3 (uuid: C, parentUuid: B)     # Response to Message 2
    └── Message 4 (uuid: D, parentUuid: B)     # Alternative response

Tool Use Flow

User Message → Assistant with Tool Use → Tool Result → Assistant Response
     ↓                    ↓                    ↓               ↓
   uuid: A             uuid: B              uuid: C        uuid: D
parentUuid: null    parentUuid: A        parentUuid: B   parentUuid: C

Analytics Data Extraction

Extractable Metrics

1. Conversation Metrics

   • Total messages per conversation
   • Message frequency over time
   • User vs Assistant message ratio
   • Tool usage patterns

2. Token Usage

   • Input/Output token consumption
   • Cache hit rates
   • Model usage patterns
   • Cost analysis

3. Tool Analytics

   • Most used tools
   • Tool success/error rates
   • Tool execution time (estimated)
   • File operation patterns

4. Project Analytics

   • Active projects
   • Project-specific usage patterns
   • File types being worked on
   • Development patterns

Common Patterns and Edge Cases

User Confirmation Messages

User responses to Claude's prompts often appear as simple strings:

{
  "message": {
    "role": "user",
    "content": "[ok]"  // or "yes", "1", "y", etc.
  }
}

Large Tool Results

Long command outputs are stored as complete strings:

{
  "type": "tool_result",
  "content": "Very long output that could be thousands of characters..."
}

Error Messages

Tool errors are marked with is_error: true:

{
  "type": "tool_result",
  "tool_use_id": "toolu_xxx",
  "content": "bash: command not found: invalidcommand",
  "is_error": true
}

Data Processing Notes

Performance Considerations

1. File Sizes: JSONL files can become large (100MB+) for long conversations
2. Parsing: Each line must be parsed individually as valid JSON
3. Memory: Large conversations should be streamed or paginated
4. Caching: Parsed conversations should be cached to avoid re-parsing

Data Validation

1. Required Fields: Always check for required fields before processing
2. Timestamps: Parse ISO strings carefully, handle timezone differences
3. Content Arrays: Assistant messages may have mixed content types
4. UUIDs: Validate UUID format for consistency checks

Usage in Analytics Dashboard

Current Implementation

The analytics dashboard extracts the following data:

// Simplified message object after parsing
{
  id: item.message.id || item.uuid,           // Message identifier
  role: item.message.role,                    // "user" or "assistant"
  timestamp: new Date(item.timestamp),        // Parsed timestamp
  content: item.message.content,              // Raw content
  model: item.message.model || null,          // AI model used
  usage: item.message.usage || null,          // Token usage
}

Available Extensions

With this data structure, the dashboard could be extended to show:

• Conversation threading/branching
• Tool usage analytics
• Project-specific insights
• Cost tracking per project
• Development velocity metrics
• Error rate analysis

File System Integration

Directory Monitoring

Watch for changes in:

• ~/.claude/projects/*/conversation.jsonl - New messages
• ~/.claude/projects/ - New projects
• ~/.claude/settings.json - Setting changes

File Reading Strategies

1. Tail Reading: Read only new lines from JSONL files
2. Full Parse: Parse entire file for complete analysis
3. Chunk Processing: Process large files in smaller chunks
4. Incremental Updates: Track file modification times

This documentation provides a complete reference for working with Claude Code data structures and can be updated as new formats or features are discovered.