Claude Code Metadata

Claude Code Metadata & Data Formats

Complete technical reference for Claude Code's local data storage architecture. All data is stored locally in ~/.claude/.

Directory Structure Overview

~/.claude/
├── CLAUDE.md                    # User memory file (personal instructions)
├── settings.json                # User settings (global)
├── settings.json.backup         # Backup of settings
├── stats-cache.json             # Token usage statistics cache
├── history.jsonl                # Session index/metadata
│
├── projects/                    # Session transcripts by project path
│   └── -Users-name-project/     # Path encoded as directory name
│       ├── session-uuid.jsonl   # Main session files
│       └── agent-xyz.jsonl      # Agent/subagent sessions
│
├── session-env/                 # Session environment snapshots
│   └── session-uuid/            # Per-session env directory
│
├── todos/                       # Todo lists per session/agent
│   └── session-uuid-agent-uuid.json
│
├── plans/                       # Plan mode files (markdown)
│   └── adjective-verb-noun.md   # Random generated names
│
├── debug/                       # Debug logs per session
│   └── session-uuid.txt
│   └── latest                   # Symlink to most recent
│
├── file-history/                # File backup history
│   └── session-uuid/
│       └── content-hash@v1      # Versioned file backups
│
├── shell-snapshots/             # Shell state snapshots
│   └── snapshot-zsh-timestamp-random.sh
│
├── skills/                      # User skills (symlinks or dirs)
│   └── skill-name/
│       ├── SKILL.md
│       └── references/
│
├── agents/                      # Custom agent definitions
│   └── agent-name.md
│
├── plugins/                     # Plugin system
│   ├── config.json
│   ├── installed_plugins.json
│   ├── known_marketplaces.json
│   ├── cache/                   # Downloaded plugin cache
│   ├── marketplaces/            # Marketplace repos
│   └── repos/
│
├── config/                      # Additional config
│   └── notification_states.json
│
├── statsig/                     # Feature flags (Statsig)
│   ├── statsig.cached.evaluations.*
│   ├── statsig.failed_logs.*
│   ├── statsig.last_modified_time.*
│   ├── statsig.session_id.*
│   └── statsig.stable_id.*
│
├── chrome/                      # Chrome extension support
│   └── chrome-native-host
│
├── ide/                         # IDE integration (empty by default)
│
└── telemetry/                   # Telemetry data (empty by default)

Session JSONL Format

Sessions are stored as JSON Lines files. Each line is a complete JSON object representing a message or event.

Message Types

User Message Schema

{
  "parentUuid": "previous-message-uuid",
  "isSidechain": false,
  "userType": "external",
  "cwd": "/working/directory",
  "sessionId": "session-uuid",
  "version": "2.0.76",
  "gitBranch": "main",
  "type": "user",
  "message": {
    "role": "user",
    "content": "message text or array of content blocks"
  },
  "isMeta": false,
  "uuid": "unique-message-id",
  "timestamp": "2025-12-26T23:15:21.914Z"
}

Assistant Message Schema

{
  "parentUuid": "user-message-uuid",
  "isSidechain": false,
  "userType": "external",
  "cwd": "/working/directory",
  "sessionId": "session-uuid",
  "version": "2.0.76",
  "gitBranch": "main",
  "slug": "adjective-verb-noun",
  "message": {
    "model": "claude-opus-4-5-20251101",
    "id": "msg_01XYZ...",
    "type": "message",
    "role": "assistant",
    "content": [
      {
        "type": "thinking",
        "thinking": "internal reasoning...",
        "signature": "base64-signature..."
      },
      {
        "type": "text",
        "text": "response text"
      },
      {
        "type": "tool_use",
        "id": "toolu_01ABC...",
        "name": "ToolName",
        "input": { "param": "value" }
      }
    ],
    "stop_reason": "end_turn",
    "stop_sequence": null,
    "usage": {
      "input_tokens": 100,
      "cache_creation_input_tokens": 5000,
      "cache_read_input_tokens": 20000,
      "cache_creation": {
        "ephemeral_5m_input_tokens": 5000,
        "ephemeral_1h_input_tokens": 0
      },
      "output_tokens": 500,
      "service_tier": "standard"
    }
  },
  "requestId": "req_01XYZ...",
  "type": "assistant",
  "uuid": "message-uuid",
  "timestamp": "2025-12-26T23:16:04.923Z"
}

Tool Result Message Schema

{
  "parentUuid": "tool-use-message-uuid",
  "isSidechain": false,
  "userType": "external",
  "cwd": "/working/directory",
  "sessionId": "session-uuid",
  "version": "2.0.76",
  "gitBranch": "main",
  "slug": "adjective-verb-noun",
  "type": "user",
  "message": {
    "role": "user",
    "content": [
      {
        "tool_use_id": "toolu_01ABC...",
        "type": "tool_result",
        "content": "tool output text"
      }
    ]
  },
  "uuid": "result-uuid",
  "timestamp": "2025-12-26T23:16:05.510Z",
  "toolUseResult": {
    "mode": "files_with_matches",
    "filenames": ["file1.go", "file2.go"],
    "numFiles": 2
  }
}

File History Snapshot Schema

{
  "type": "file-history-snapshot",
  "messageId": "associated-message-uuid",
  "snapshot": {
    "messageId": "associated-message-uuid",
    "trackedFileBackups": {
      "/path/to/file": "content-hash"
    },
    "timestamp": "2025-12-26T23:15:21.917Z"
  },
  "isSnapshotUpdate": false
}

Content Types in Messages

Messages can have content as:

• String: Plain text content
• Array: Multiple content blocks

Content block types:

• text: Plain text { "type": "text", "text": "..." }
• image: Base64 image { "type": "image", "source": { "type": "base64", "media_type": "image/png", "data": "..." } }
• thinking: Model reasoning { "type": "thinking", "thinking": "...", "signature": "..." }
• tool_use: Tool invocation { "type": "tool_use", "id": "...", "name": "...", "input": {...} }
• tool_result: Tool output { "type": "tool_result", "tool_use_id": "...", "content": "..." }

Stats Cache Format (stats-cache.json)

Aggregated usage statistics cached locally.

{
  "version": 1,
  "lastComputedDate": "2025-12-25",
  "totalSessions": 766,
  "totalMessages": 58731,
  "firstSessionDate": "2025-11-12T06:27:12.809Z",

  "dailyActivity": [
    {
      "date": "2025-12-24",
      "messageCount": 4800,
      "sessionCount": 16,
      "toolCallCount": 1464
    }
  ],

  "dailyModelTokens": [
    {
      "date": "2025-12-24",
      "tokensByModel": {
        "claude-opus-4-5-20251101": 33212,
        "claude-sonnet-4-5-20250929": 103056,
        "claude-haiku-4-5-20251001": 272249
      }
    }
  ],

  "modelUsage": {
    "claude-opus-4-5-20251101": {
      "inputTokens": 1640347,
      "outputTokens": 2475769,
      "cacheReadInputTokens": 1242320484,
      "cacheCreationInputTokens": 66151371,
      "webSearchRequests": 0,
      "costUSD": 0,
      "contextWindow": 0
    }
  },

  "hourCounts": {
    "9": 58,
    "10": 43,
    "17": 59,
    "18": 94
  },

  "longestSession": {
    "sessionId": "uuid",
    "duration": 96242264,
    "messageCount": 4,
    "timestamp": "2025-11-17T02:12:38.105Z"
  }
}

Settings Format (settings.json)

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "model": "opus",
  "alwaysThinkingEnabled": true,

  "hooks": {
    "PreCompact": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "echo 'hook output'"
          }
        ]
      }
    ],
    "SessionStart": []
  },

  "statusLine": {
    "type": "command",
    "command": "shell command for status line"
  },

  "enabledPlugins": {
    "plugin-name@marketplace": true
  },

  "feedbackSurveyState": {
    "lastShownTime": 1754025174206
  }
}

Hook Event Types

• SessionStart: When a new session begins
• PreCompact: Before context compaction
• SubagentStart: When a subagent is spawned

History Index (history.jsonl)

Each line contains session metadata:

{
  "display": "user message preview",
  "pastedContents": {},
  "timestamp": 1766792756159,
  "project": "/Users/name/code/project",
  "sessionId": "session-uuid"
}

Todo Format (todos/*.json)

[
  {
    "content": "Task description",
    "status": "in_progress",
    "activeForm": "Working on task"
  }
]

Status values: pending, in_progress, completed

Installed Plugins Format

{
  "version": 2,
  "plugins": {
    "plugin-name@marketplace-name": [
      {
        "scope": "user",
        "installPath": "/path/to/plugin",
        "version": "1.0.0",
        "installedAt": "2025-12-04T02:42:33.270Z",
        "lastUpdated": "2025-12-04T02:42:33.270Z",
        "gitCommitSha": "abc123...",
        "isLocal": true
      }
    ]
  }
}

Agent Definition Format

Markdown file with YAML frontmatter:

---
name: agent-name
description: When to use this agent
model: sonnet
color: green
---

Agent system prompt instructions here...

Memory File Hierarchy

Priority order (highest to lowest):

Import Syntax

See @README for project overview.
@docs/instructions.md
@~/.claude/custom-rules.md

File History Backup Format

Files in file-history/session-uuid/:

• Filename: content-hash@vN (e.g., 89a91e02ba2f0439@v1)
• Content: Raw file contents at that version

Debug Log Format

Plain text with timestamps:

2025-11-25T19:05:43.547Z [DEBUG] Message here...

Shell Snapshot Format

Shell script capturing environment:

# Snapshot file
# Unset all aliases to avoid conflicts with functions
unalias -a 2>/dev/null || true
# Functions
function_name () { ... }
# Environment variables
export VAR=value

Notification States (config/notification_states.json)

{
  "switch_to_custom": {
    "triggered": false,
    "timestamp": null
  },
  "exceed_max_limit": {
    "triggered": true,
    "timestamp": "2025-09-27T19:00:38.269745"
  },
  "tokens_will_run_out": {
    "triggered": false,
    "timestamp": null
  },
  "cost_will_exceed": {
    "triggered": true,
    "timestamp": "2025-09-27T16:56:18.406988"
  }
}

Statsig Feature Flags

Files in statsig/:

• statsig.cached.evaluations.*: Feature flag cache
• statsig.failed_logs.*: Failed telemetry
• statsig.stable_id.*: User identifier
• statsig.session_id.*: Session identifier

Project Path Encoding

Project paths become directory names by replacing / with -:

• /Users/name/code/project -> -Users-name-code-project

Token Usage Calculation

IMPORTANT: Usage data is nested at message.usage, NOT at the top level of a JSONL line.

From assistant messages (type: "assistant"), extract from message.usage:

• message.usage.input_tokens: Direct input tokens
• message.usage.output_tokens: Response tokens
• message.usage.cache_creation_input_tokens: New cache entries
• message.usage.cache_read_input_tokens: Cache hits

Field names use snake_case (input_tokens, not inputTokens).

Cache efficiency = cache_read_input_tokens / (cache_read_input_tokens + input_tokens)

What dailyModelTokens tracks

The dailyModelTokens field in stats-cache.json tracks a proprietary aggregate computed by Claude Code. It does NOT equal the sum of per-message usage.input_tokens + usage.output_tokens from JSONL files. Empirically, dailyModelTokens values are ~0.04x–0.39x of the JSONL-derived input_tokens + output_tokens sum for the same day. The exact formula is internal to Claude Code.

Key implication: If your tool is calibrated against dailyModelTokens (e.g., by correlating it with scraped usage percentages), you MUST continue using dailyModelTokens for consistency. Switching to JSONL-summed tokens will produce values ~3-25x larger, breaking any calibration.

stats-cache.json Volatility Warning

stats-cache.json values can change retroactively. Claude Code rebuilds this file by scanning JSONL session files. Key issues:

1. Retroactive rewriting: When sessions are deleted (30-day retention cleanup), the cache is rebuilt and historical daily totals can drop.
2. Timezone uncertainty: The date bucketing in the cache may use a different timezone than expected.
3. Race conditions: If Claude Code is running when you read the cache, values may include in-progress data.

Direct JSONL Scanning

For per-message ground truth, scan JSONL session files directly. Filter .jsonl files by os.FileInfo.ModTime() for performance (skip files not modified in your date range), then parse message.usage from type: "assistant" lines. Note that the per-message totals will NOT match dailyModelTokens from stats-cache.

Session Retention

Default: 30 days. Configure with cleanupPeriodDays setting.

Key UUIDs

All UUIDs follow RFC 4122 format:

• Session IDs: xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx
• Message UUIDs: Same format, linked via parentUuid
• Request IDs: req_ prefix
• Tool IDs: toolu_ prefix
• Message IDs: msg_ prefix

Model Identifiers

Common model IDs:

• claude-opus-4-6 (latest)
• claude-opus-4-5-20251101
• claude-sonnet-4-5-20250929
• claude-haiku-4-5-20251001

Cost Calculation Reference

Per-million tokens (approximate):

• Haiku: $0.25 input / $1.25 output
• Sonnet: $3.00 input / $15.00 output
• Opus: $5.00 input / $25.00 output
• Cache read: 90% discount
• Batch: 50% discount

Useful Queries

Count total tokens per session

Parse JSONL, sum message.usage.input_tokens + message.usage.output_tokens for type: "assistant" lines.

Find all sessions for a project

List files in projects/-path-to-project/.

Get session timeline

Sort messages by timestamp field.

Reconstruct conversation tree

Follow parentUuid chains to build message tree.

Third-Party Tools

• ccusage: CLI for analyzing usage from JSONL files
• Claude Code Usage Monitor: Real-time terminal dashboard
• LiteLLM: Proxy for tracking enterprise usage

References

• Claude Code Docs - Memory
• Claude Code Docs - Costs
• Usage and Cost API