[Bug]: 升级后缓存命中率从之前99%降到78%，字 小得看不清，比雷布斯还小。不针对deepseek优化估计慢慢就没人用了。

[qiyuanlicn]

Version line

v2 — Go rewrite (1.x), main-v2 (active development)

Exact version

1.1.0

What happened?

升级后缓存命中率从之前99%降到78%，字 小得看不清，比雷布斯还小。不针对deepseek优化估计慢慢就没人用了。

Steps to reproduce

0

OS / platform

windows

Relevant logs or output

[esengine]

Thanks for the report — and sorry the 1.1.0 upgrade landed worse for you on all three fronts. Let me take them one at a time, since they're really separate issues bundled into one.

1. Cache hit rate 99% → 78% (this is the one I care about most)

A drop this size is not expected and not something I'd shrug off — cache-first is the whole point of this project, so a regression here is a real bug, not a tuning preference. To track it down I need a bit more from you:

• The output of the in-session cache stats (the per-turn cache hit line, or the summary at session end). Even a screenshot is fine.
• Roughly what your workflow looks like: long single sessions, lots of /clear or branching, frequent file edits between turns? Anything that rewrites the prefix of the prompt will cost you hits.
• Which model/provider endpoint you're on (official DeepSeek API vs. a proxy), since cache behavior is endpoint-side.

My first suspicion is a prefix-stability regression — something in 1.1.0 that perturbs the immutable prefix of the request (system block, tool list ordering, or per-turn metadata) and silently invalidates the upstream cache. If you can drop a cache hit log line from a 1.0.x run and a 1.1.0 run side by side, that alone will probably point straight at it.

2. Font "too small to read"

The TUI doesn't ship its own font — it renders into your terminal at the terminal's own font size, so the size itself is controlled by your terminal (e.g. Windows Terminal: Ctrl + + to zoom, or set the font size in the profile). That said, if there's a specific element that got smaller or dimmer in 1.1.0 (dimmed/secondary text, the status line, a particular panel), that would be a real styling regression and I'd like to fix it. Can you share a screenshot pointing at the text you mean?

3. "Not optimizing for DeepSeek"

For what it's worth, DeepSeek is the primary target — the caching, pricing, and token accounting are all built around its context-cache pricing model first. If you've hit a place where the behavior is clearly tuned against some other provider's assumptions, point me at it and I'll treat it as a bug.

If you can grab the two cache-stat lines for #1, that's the fastest path to a fix. I'd also suggest we split the font issue into its own report once you've got a screenshot, so neither gets lost.