DeepSeek Harness, the logging and debugging tool designed for developers working with DeepSeek models, has a split personality when it comes to correctness. The append-only architecture that stores your session data works exactly as intended—but if you're using its token projection features to track usage or estimate costs, you're looking at numbers that are precisely double what they should be.

The Exact Problem

Testing conducted this week across two different providers on a nine-day-old codebase revealed the issue clearly: summing every usage record in a DeepSeek Harness session log yields exactly 2.000000× the correct total. This isn't approximately double—it's a mathematically precise factor of two, reproducible without remainder and confirmed independently across both routes.

What Gets Affected

The implications hit forked sessions particularly hard. In one documented case, a single forked session reported 263,790 tokens through the projection system when the actual usage was half that figure. For teams relying on these numbers for cost tracking or capacity planning, this isn't a rounding error—it's a systematic overcounting that compounds across longer projects.

Append-Only Works Fine

Here's where it gets interesting: DeepSeek Harness gets one critical piece right. The append-only logging mechanism itself is sound—it doesn't suffer from the same doubling issue. Your raw session logs are accurate; it's only when you run them through the token projection layer that the numbers go sideways.

Where Compaction Costs You

The root of the problem appears tied to how the system handles data compaction without properly adjusting its projections. When usage records get merged, compacted, or deduplicated in the append-only log, the projection layer doesn't account for the fact that those operations reduce the total token count. It keeps counting as if every original record still exists.

What This Means for Your Budget

If you're using DeepSeek Harness to track API spend across your team, you need to factor this bug into your calculations—or wait for a patch. The 2x overcounting means any budget projections based on these numbers are going to be significantly inflated.

Key Takeaways

  • Append-only logging is working correctly; don't distrust your raw session data
  • Token projection multiplies everything by exactly 2x, not approximately
  • Forked sessions amplify the problem with proportionally larger discrepancies
  • Compaction logic appears to trigger the bug when merging usage records

The Bottom Line

DeepSeek Harness shows promise on the infrastructure side—append-only semantics done right—but the token projection layer needs a serious rethink before I'd trust it for anything financial. Either patch that math or keep your own spreadsheets until this gets sorted.