User Stats

User-Level Statistics

Scan all session JSONL files for the current user and output aggregated statistics by model and tool.

Output Format

Usage by model:
    <ModelName>:  <calls> calls, <success_rate> success, <input> input, <output> output, <cache_read> cache read, <cache_write> cache write
Tool calls:
    <ToolName>:  <calls> calls, <errors> errors, <success_rate> success

• Token values: use k for thousands, M for millions (e.g. 11.8M, 45.8k, 0 if zero)
• Success rate: percentage with 1 decimal (e.g. 99.2%, 100.0%)
• Sort models by total calls descending; sort tools by total calls descending

Data Source

Session JSONL files are located at:

~/.claude/projects/<project-hash>/<sessionId>.jsonl

Claude Code 和 Trae 共用此目录存储会话历史。脚本默认扫描 ~/.claude/projects/。

也可以通过参数指定目录：node user-stats.js --dir <custom-path>

Each line is a JSON entry. Relevant entry types:

assistant entries (model stats)

{
  "type": "assistant",
  "message": {
    "model": "doubao-seed-code",
    "stop_reason": "tool_use",
    "usage": {
      "input_tokens": 31610,
      "output_tokens": 226,
      "cache_creation_input_tokens": 0,
      "cache_read_input_tokens": 0
    }
  }
}

• message.model → model name (use as-is for grouping)
• message.usage.input_tokens → input tokens
• message.usage.output_tokens → output tokens
• message.usage.cache_creation_input_tokens → cache write tokens
• message.usage.cache_read_input_tokens → cache read tokens
• message.stop_reason → "error" means failed call; any other value is success

tool_use blocks (tool stats)

Inside assistant entries, message.content is an array. Each block with type: "tool_use" represents a tool invocation:

{"type": "tool_use", "id": "call_xxx", "name": "Bash"}

• name → tool name. For Agent tool calls, the name may be a long string containing the full prompt — normalize it to just "Agent".
• id → tool call ID for matching with tool_result

tool_result blocks (tool success/failure)

Inside user entries, message.content is an array. Each block with type: "tool_result":

{"type": "tool_result", "tool_use_id": "call_xxx", "is_error": true}

• tool_use_id → matches the id from tool_use block
• is_error → true means the tool call failed

Workflow

1. Find all JSONL files:
   find ~/.claude/projects -name '*.jsonl' -type f

2. Write the aggregation script to /tmp/user-stats.js and run it:
   - Parse every line of every JSONL file
   - For each assistant entry:
     a. Extract model name from message.model
     b. Accumulate: call count, input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens
     c. If message.stop_reason === "error", increment model failure count
     d. For each tool_use block in message.content:
        - Normalize tool name: if name contains newline or is longer than 30 chars, use "Agent"
        - Record tool name, increment call count
        - Store mapping: tool_use_id → {tool_name}
   - For each user entry:
     a. For each tool_result block in message.content:
        - Look up tool_use_id in the stored mapping
        - If is_error === true, increment tool error count

3. Calculate success rates:
   - Model success rate = (total_calls - failures) / total_calls * 100
   - Tool success rate = (total_calls - errors) / total_calls * 100

4. Format and print output using the format above
5. Clean up temp file

Execution

Run the aggregation script directly:

node /Users/bytedance/code/ai-code/.agents/skills/user-stats/scripts/user-stats.js

Edge Cases

• No JSONL files found: Print "No session data found." and stop.
• Empty or corrupted JSONL lines: Skip silently (the script uses try/catch).
• Unknown model names: Use as-is from message.model.
• Long/multiline tool names (e.g. Agent with inline prompt): Normalize to "Agent".
• Tool calls without matching tool_result: Count the call but no error.
• tool_result without matching tool_use: Skip (orphaned result).