Model Context Mismatch after Image Read

[logar16]

Describe the bug

I had a bug where I was using the 1 million context LoRA opus 4.6. It was doing some work, and then it tried to read some images using PlayWrite. Suddenly it said the context report in lower-right corner went from like ~10% to 44%, so that's like hundreds of thousands of tokens! I then ran the /context command and it showed the context only using 9% of the total context. I switched models to a different model and then went back to the 1 million context, and now it shows 7%. So who can I trust? There's a disparity in the lower right context percentage and the context command output.

Something perhaps related to this is I've had it where the model reads in a couple of images and suddenly it starts compacting its context history, which doesn't make any sense because these images aren't actually taking up that much space. It seems to trigger the compaction, thinking that it has suddenly filled up its context. It's different from if I paste in the images, because if I paste in the images there's very little change to the context. I think there's a bug when the model itself is reading the images.

Affected version

1.0.35-4

Steps to reproduce the behavior

1. Have the model look for and "view" several pngs in a folder or via playwright mcp
2. Watch the context report in lower right corner and compare to /context
3. You may even "fill up" context and have it automatically start compacting

Switching models will mostly repair the discrepancy but it's no longer aligned.

Expected behavior

I expect the model context reporting to be the same always, and I also expect that reading images should not totally fill up the context, nor should it cause discrepancy in context reading.

Additional context

This is on Windows 11 running Windows Terminal

[examon]

This should be resolved as of v1.0.35. Binary tool-result content (such as images read by tools or MCP servers) was being counted as base64, which inflated local token counts and triggered spurious compaction; updating from a 1.0.35-prerelease to the v1.0.35 stable release should fix it. If you still see the footer-vs-/context mismatch after updating, please comment with a fresh repro and we'll take another look.