Besides being even better at the caching, I'm not sure what benefits you'd get compared to just firing up OpenCode with the DeepSeek API yourself, it'll similarly do caching for sure and also "talks directly to api.deepseek.com" if that matters, and you'll get a much more mature harness.
> tool call pruning breaks cache and people will tell you this is horrible and expensive
> except i looked at some anthropic data and real user behavior ends up with better cache hits and 30% less spend
> even this is needs to be analyzed further, it's just not simple
> for openai data it's inverted! cache hit ratio is actually better [sic: I think he meant worse based on the screenshot] with tool call pruning turned on
> but the net $ saved is only 5%
> kimi is a funny one - it has better cache hits with pruning on...but is also more expensive!
There was also another thread recently where he discussed that pruning improves user experience (models are smarter with less context) but I can't find it.
This can also be disabled in the config: https://opencode.ai/docs/config/#compaction
Ah, reminds me of good old "There are only 2 hard problems in computer science: cache invalidation, naming things, and off-by-1 errors."
You quip, but LLM KV caching (from the harness side) is quite easy: You get a cache hit on stable prompt prefixes, period. That means you want to keep the prefix stable, and only append at the end of the conversation. Made up example: Don't put the git branch name into the system prompt part (that comes first), as whenever the branch name changes, that'd trigger a cache invalidation of the entire prompt.
Getting this right requires some care to not by accident modify the prefix, basically, and some design on communicating the things that can change (user configuration, working dir, git information, ...).
They explain some of the the reasons why they have a better solution and why they are very opinionated
>Automatic prefix caching activates only when the exact byte prefix of the previous request matches. Most agent loops reorder, rewrite, or inject fresh timestamps each turn — cache hit rate in practice: <20%.
So they optimize on this plus other techniques to improve cache hits, making it cheaper.
Can you share the bridge. DeepSeek v4 is awesome paired with claude-code or opencode. I found that claude code costs me less than opencode and I am presuming this is due to a better engineered harness.
I only used it for a few hours to play around with stuff before the quota issue was fixed and I could resume using GPT models, and the bridge was coded by DeepSeek-V4-Flash-IQ2XXS + DwarfStar4 locally, I take no responsibility for what might happen with your computer or you, during usage or just reading the code.
Edit: heh, like don't look at line 117 for example where seemingly it likes to handle misspellings in the .env file which totally wasn't my fault for typo'ing the API key in that file... I'm sure there are tons of sharp edges and dumb stuff in there.
Obviously, if you do deal with any sort of secrets, then using local LLMs over OpenAI, Anthropic, DeepSeek or whoever is obviously preferred, and in the case of personal data of users, probably a requirement.
Getting the source code of facebook or instagram doesn't mean you could compete with them.
I work for a company that has built relationship with event organizers over the past 10 years. The code I maintain could be written from scratch in maybe 2-3 months even though it was built over the past 10 years but besides that you have frontend / DB / hardware / logistics etc
Same with codex? codex-rs at least, is a TUI as well, it does run a "app-server" in the background, that the TUI actually interacts with, but that's just an implementation detail. Also makes it easy to hook in your own programs to fire of codex "headless" sessions even without the TUI.
In the end I had Claude produce a one-page html file that was 95% of the way there and it took minor editing to clearly explain the intent of the feature.
Extremely pro consumer tool. I have been hammering it hard with 97% cache utilization and barely $0.03 dollar spent for me constantly exploring a codebase.
> Tool arguments the model produces occasionally have JSON typos, unclosed quotes, or shape mismatches. Reasonix runs a schema-aware repair pass before dispatch so malformed args still execute.
So Deepseek API doesn't have a structured output option where you give a grammar and the model promises the output will follow this grammar?
Or it does, but it's buggy?
I'm concerned since i really want SOTA reasoning, but DeepSeek still has me interested.
I think you should give other models a try and see how much they differ from SOTA models. I did this and realized, even Qwen-2.5-Max was enough. I am sure even Claude Sonnet 3.5 is enough for things I play around with. I am not really striving for fields medal in Mathematics.
After about 6 hours, both ultimately failed to fully RE, however, there were some drastic differences:
DS stopped every 30 minutes or so, saying it did full RE and it should all work now, while in fact, it didn't complete even 1% of it. It also looked for shortcuts again and again, despite me prompting heavily that the specific shortcut may not be used. It was a complete and utter failure.
GPT-5.5, on the other hand, blew me away. It just did the right things, didn't jump to next steps until it was sure it completed the initial layers and had a full understanding of what's required. The only time I prompted it during the 6 hours was when I saw it going in the right direction and I could nudge it slightly towards an even better way. I never felt I was fighting it. Okay, maybe a little bit - after compaction, it sometimes would go on a "no I'm not helping you with reverse engineering" tangent, but it would resolve in a clean session.
I cancelled my Claude subscription a month ago, so I haven't tested that, but DeepSeek has reminded me a lot of how I worked with Opus 4.6/4.7. Which perhaps could be a positive sign to some, but GPT-5.5 showed me that the way claude/ds work is just way too annoying.
There's always the option of using Anthropic's models for some tasks like planning and then just hand over the implementation task to something like DeepSeek. Across different tools, a Markdown plan works pretty okay. That's what I'm planning to do if I go from the 5x Max subscription down to the Pro.
I am also writing a launcher that makes using 3rd party providers with Claude Code easy (https://ccode.kronis.dev) and I already have a local proxy up and running, just not dynamic model switching yet. Though it shouldn't be too hard to add, will probably be there within a week or two, depending on my schedule.
I don't think it's wise to leave Anthropic altogether because their models are great (and a subscription gives you features like Remote Control which I like), but switching tiers and maybe saving a bit of money seems viable! On the other hand, you do need a quality baseline, because I remember using Cerebras with GLM 4.6 way back and there was a bit too much slop.
I used to surf the three big players frequently and got really tired of the effort needed to steer some models. In the end i ended up sticking with Claude because it required less steering effort. While not strictly reasoning, a models ability to follow clear directions consistently is something i'd consider part of its SOTA capabilities.
Eventually i just tired of exploring. I just want stability.
Which ironically is why i'm thinking about moving from Claude. The very basic IDE/-p usage getting removed from my plan is a UX stability issue. I'm trying to progressively improve my workflows and efficiency, not have to establish a new foundation anytime something shifts. Quite frustrating.
I have been using it for a while, and I wholeheartedly agree. imo, it is as good as codex or claude which I also use. It is a winner in the cost-sensitive tier, and if some startup could put it together with data-retention in mind, it could be a great product sold to the enterprise, as data-retention and privacy are the main issues for the coding-assistant usecase.
> Independent open-source project · not affiliated with DeepSeek
Yes, but a lot of harnesses change previous context. E.g. the system prompt injects the current time/date, working directory, files in the working directory, etc. Compaction also changes the whole previous context. I _think_ changing the list of tools also invalidates cache, so invoking a subagent with different tools would invalidate the cache.
My vague impression is that it's in a similar vein to functional programming languages. It generally disallows doing things that lead to bugs (cache misses in this case), and presumably allows you to do those things in a way that makes it much clearer that this is likely to cause cache misses. I would guess that in this paradigm, you don't mutate your existing session, you derive a new session by mutating the prior context into a new context.
From the FAQ, I see:
>Can I point it at a self-hosted / private DeepSeek endpoint?
>Yes. Since 0.30 we accept non-standard key prefixes for self-hosted DeepSeek endpoints. Just point `baseUrl` at your internal address — the loop, cache strategy, and tool protocol are unchanged.
But my question is: If I use Reasonix to talk to a deepseek endpoint through openrouter, am I still getting the cache-hit benifits of this agent harness?
Any comments on what you can or cannot rely on it for relative to cc and codex would be appreciated too!
This is still art as much as science and the different harnesses take different approaches.
Is this improving the cache hit and hence overall efficiency of coding workflows?
Does it also let me host a local llm (deepseek)? What are model min requirements for this?
I specifically use multiple different models and providers, so this wouldn't be useful for me.
And it contributes to the problem of each person vibe-coding their own, incompatible, half-baked tool in a space, instead of contributing to a small set of tools and expanding them.
It'd be better to just extend an existing tool.
AI marketing slop. This is how all models and coding harnesses work, isn't it?
The author claims (in another AI-written post):
> LangChain — along with every generic agent framework I checked — rebuilds the prompt every turn. Timestamps get injected. History gets reordered. Tool schemas re-serialize with different whitespace.
I haven't touched LangChain in a long, long time, but don't think any of the current harnesses, Claude Code, Pi, Crush, OpenCode etc do that except if you change configuration? Keeping the context stable for caching is a very basic principle and not a wild innovation.
This posing as DeepSeek-specific is also a mystery.
That's the pinnacle of AI slop over engineered garbage in my opinion. All of that information is noise.
Is this really the behavior you want? Yes, doing tool-result clearing and such will blow your cache, but if you do it only occasionally, it's still likely a win. Yes, cache hits are good, but not so good that it's okay to be profligate with context to preserve those precious, precious KVs.
"Independent open-source project · not affiliated with DeepSeek" "Reasonix only targets DeepSeek because..." "Why DeepSeek only? Can I swap to Claude / GPT? It's a design choice, not a limitation"
The lady doth protest too much, methinks?
Nicely timed shortly after the making the rebate permanent anouncement.
Could just be Chinese devs trying to help western devs with some software and a western facing marketing campaign to raise awareness. Could be DeepSeek astroturfing. Could be "someone" in China trying to get more access to western data.
Who knows?
(that is, different places on the Pareto efficiency graph)
It's bad enough that I'm working on guardrails at the harness level because prompting appears to be useless.
Do you have the same issue?
Now that you mention it, though, I have seen it do a few things that weren't in the plan. The reviewer caught them, though, so they didn't cause a problem, and it's so cheap that overall it's a massive improvement.
It's the agentic era, pick a better option
Just stop
