@ccusage/codex: Forked sessions double count shared token usage across sessions

[astarforbae]

Summary

When using Codex session forking, @ccusage/codex significantly overcounts token usage.

Forked sessions inherit the full history of the parent session, but ccusage aggregates each session independently. As a result, shared usage is counted multiple times.

---

Example

• Session A: total usage = 100
• Fork → Session B: total usage = 150 (includes the original 100)

Expected total:

• 100 (A) + 50 (new in B) = 150

Actual:

• 100 (A) + 150 (B) = 250 ❌

---

What I tested locally

I implemented a fork-safe counting approach based on content-level deduplication of token events:

• Treat the token_count payload as the canonical event identity
• Ignore outer fields such as timestamps, file paths, or session boundaries
• Normalize the payload and compute a stable hash
• If two events have identical payloads, they are treated as the same event and counted only once

This ensures that replayed history from forked sessions is not double counted.

---

Real-world Impact

After applying this approach:

• Original total tokens: ~277M

• Corrected total tokens: ~81M

• Reduction: ~196.9M tokens (~70.8% lower, ~3.4× inflation)

• Original cost: ~$135.66

• Corrected cost: ~$37.38

• Reduction: ~$98.28 (~72.4% lower, ~3.6× inflation)

• Deduplicated overlapping events: 11,523

---

Expected Behavior

Only the incremental usage after the fork should be counted for child sessions, rather than re-counting the entire inherited history.

---

Possible Cause

Aggregation is performed per session without accounting for shared history between forked sessions.

---

Impact

This leads to substantial overestimation of usage, especially for workflows that involve frequent session forking.