Token count discrepancy - subagents/sub-tasks not fully captured

[rz1989s]

Summary

While building a custom cost tracking solution, I noticed a significant discrepancy between token counts from direct JSONL parsing vs ccusage output. After investigation, this appears related to #313.

Environment

• ccusage version: 17.0.3 (via bunx)
• Claude Code: 2.1.7
• macOS

Observed Behavior

For the same project directory, comparing token counts:

Source	Total Tokens	Cost
Direct JSONL parsing	1.18B	~$886
ccusage session	527M	~$372

That's a 2.25x difference in token count for the same project.

Investigation

1. Subagent files exist in ~/.claude/projects/<project>/subagents/ directory
2. Direct JSONL parsing includes these files
3. ccusage appears to not fully capture tokens from these subagent conversations

Example structure:

~/.claude/projects/-Users-me-myproject/
├── abc123.jsonl              # Main conversation
└── subagents/
    ├── agent-a1b2c3.jsonl    # Sub-task 1
    ├── agent-d4e5f6.jsonl    # Sub-task 2
    └── ...

Verification

Calculated costs excluding subagents:

• Main files only: 1.17B tokens (~$877)
• Subagent files: ~$126

Even excluding subagents, there's still a ~2x discrepancy, suggesting the issue may be broader than just subagent directories.

Pricing Verification

Confirmed ccusage uses LiteLLM pricing which has correct Opus 4.5 rates ($5/$25 per MTok). The discrepancy is in token counting, not pricing.

Related

This may be related to #313 (sub-task token tracking). Happy to provide more debug info if helpful.

---

Using Claude Code Max subscription, so accurate tracking is important for usage benchmarking.