ollama - 💡(How to fix) Fix context utilization not accurate [1 pull requests]

What is the issue?

Since upgrading Ollama past version 0.24.0 (specifically experiencing this on 0.30.0), token usage tracking in external development tools (like VS Code extensions, OpenCode, and other CLI chat clients) is erratic.

Instead of showing a stable, incrementally growing context size as the session builds up, the numbers fluctuate completely at random from turn to turn.

What I am seeing in my editors/CLI tools: During a single, continuous chat session, the reported context/token usage jumps up and down wildly without any logic. For example, over a 5-turn conversation, the token usage numbers look like this:

Turn 1: 3,000 tokens

Turn 2: 19,000 tokens

Turn 3: 1,000 tokens (dropped drastically, even though the history grew)

Turn 4: 100 tokens

Turn 5: 9,000 tokens (spiked to an impossible number)

The numbers are quite random.

I run on CUDA. But I assume is irrelevant

Relevant log output

I don't have any.
I don't see anything unusual in the logs.

OS

Linux

GPU

Nvidia

CPU

Intel

Ollama version

0.30.0