Problem
I’ve been working on opencode/deepseek for several months now, and I tried out a coding session on deepseek-tui.
Specifically, I found the UI to be somewhat hard to read. That might seem like a minor issue if the work is done well, but I personally like to follow the process, and in this case, it’s pretty tough. Well, maybe it’s just a matter of getting used to it.
On the plus side, even if it’s hard to quantify, I really felt like the job was done, and that it was done faster than usual.
On the other hand, what is perfectly quantifiable is the cache hit rate. And there, when comparing a Deepseek API-oriented agent to a general-purpose agent, the problem is clearly visible.
Classic opencode session : 95.43% cache hit
| Total |
74,995,786 tokens |
| Input (Cache hit) |
71,571,072 tokens |
| Input (Cache miss) |
3,254,817 tokens |
| Output |
169,897 tokens |
deepseek-tui session : 71.20% cache hit
| Total |
29,992,833 tokens |
| Input (Cache hit) |
21,356,928 tokens |
| Input (Cache miss) |
8,470,281 tokens |
| Output |
165,624 tokens |
opencode session on same project : 94.36% cache hit
| Total |
55,300,522 tokens |
| Input (Cache hit) |
52,185,216 tokens |
| Input (Cache miss) |
2,924,970 tokens |
| Output |
190,336 tokens |
I don’t know exactly what’s breaking the KV cache, but it certainly seems like something is.
Edit: Just to clarify, I use only the deepseek-v4-flash API on both agents, they don't call anything else. (@superzmy )
Problem
I’ve been working on opencode/deepseek for several months now, and I tried out a coding session on deepseek-tui.
Specifically, I found the UI to be somewhat hard to read. That might seem like a minor issue if the work is done well, but I personally like to follow the process, and in this case, it’s pretty tough. Well, maybe it’s just a matter of getting used to it.
On the plus side, even if it’s hard to quantify, I really felt like the job was done, and that it was done faster than usual.
On the other hand, what is perfectly quantifiable is the cache hit rate. And there, when comparing a Deepseek API-oriented agent to a general-purpose agent, the problem is clearly visible.
Classic opencode session : 95.43% cache hit
deepseek-tui session : 71.20% cache hit
opencode session on same project : 94.36% cache hit
I don’t know exactly what’s breaking the KV cache, but it certainly seems like something is.
Edit: Just to clarify, I use only the deepseek-v4-flash API on both agents, they don't call anything else. (@superzmy )