Claude Code

Claude Code records every assistant message — including token usage and tool calls — to JSONL files on disk. tokenuse reads these directly.

Status: implemented (src/tools/claude_code/).

Where the data lives

Subagent transcripts live in a subagents/ subdirectory under each project and are read in addition to the main *.jsonl files.

Env var override: CLAUDE_CONFIG_DIR replaces ~/.claude for the CLI projects path. Useful for sandboxed installs.

Claude entries include a top-level cwd field, and that is the authoritative project path for parsed calls. The project directory name is only a lossy fallback: names like -Users-me-Code-ai-commit-dev cannot distinguish path separators from real hyphens, so never treat the directory-derived value as canonical when cwd is present.

Discovery rules (src/tools/claude_code/discovery.rs):

• Enumerate immediate subdirectories of ~/.claude/projects/.
• Walk the Desktop sessions tree to depth 8 looking for any directory named projects; treat each child as a session source.
• Skip node_modules and .git while walking.

Record format

Each *.jsonl is one JSON object per line. Two entry types matter:

// User turn
{
  "type": "user",
  "timestamp": "2026-04-26T10:00:00Z",
  "sessionId": "session-uuid",
  "message": {
    "role": "user",
    "content": "refactor the parser"            // string OR array of {type:"text", text:"..."}
  }
}

// Assistant turn — the only entry type that produces a ParsedCall
{
  "type": "assistant",
  "timestamp": "2026-04-26T10:00:01Z",
  "sessionId": "session-uuid",
  "message": {
    "role": "assistant",
    "id": "msg_01ABC...",                       // dedup key
    "model": "claude-opus-4-7-20250514",
    "usage": {
      "input_tokens": 100,
      "output_tokens": 50,
      "cache_creation_input_tokens": 1000,
      "cache_read_input_tokens": 5000,
      "speed": "fast",                          // optional, "standard" | "fast"
      "server_tool_use": {
        "web_search_requests": 0
      }
    },
    "content": [
      { "type": "tool_use", "name": "Bash", "input": { "command": "ls -la | grep foo" } },
      { "type": "tool_use", "name": "Edit",  "input": { /* ... */ } },
      { "type": "text", "text": "Done." }
    ]
  }
}

Token & cost mapping

Anthropic-specific quirk: cache reads are billed at ~10% of the input rate, and cache_read_input_tokens is not included in input_tokens. The pricing formula handles this directly — do not sum the buckets together before pricing.

Deduplication

dedup_key = message.id if present, otherwise claude:<timestamp>.

Re-reading the same JSONL across runs is normal; the shared seen: &mut HashSet<String> ensures every assistant message contributes once per process.

Tools / bash extraction

Walk message.content[] and collect name from every { "type": "tool_use" } block.

• mcp__server__tool names are kept in tools and surface separately in the dashboard’s MCP servers panel (split on __).
• For Bash and BashOutput tool calls, parse input.command and split on unquoted ;, |, &&, ||. Each split is a separate command (tools::jsonl::split_bash_commands). The dashboard then groups by first word (first_word).

Known limitations

• The user message captured per call is the most recent user turn before the assistant response, truncated to 500 chars. If a user sends multiple messages in rapid succession before any assistant reply, only the last is recorded.
• Synthetic models (<synthetic>, used by Claude Code for placeholder rows) hit the pricing fallback — they cost $0 because their token counts are zero, but they still count toward call totals.
• No live file watching: press r or wait for the 15-minute background archive sync to pick up new sessions.