_______ __ _______
| | |.---.-..----.| |--..-----..----. | | |.-----..--.--.--..-----.
| || _ || __|| < | -__|| _| | || -__|| | | ||__ --|
|___|___||___._||____||__|__||_____||__| |__|____||_____||________||_____|
on Gopher (inofficial)
(HTM) Visit Hacker News on the Web
COMMENT PAGE FOR:
(HTM) DeepSeek reasonix, DeepSeek native coding agent with high caching and low cost
perseusai wrote 9 hours 13 min ago:
This is a nice companion to the token saving context app I made. Even
has the same Claude Design site, which I think looks awesome! Even
though something is cheap, the concepts that make using Deepseek more
efficiently can surely be applied elsewhere. Cool stuff!
danborn26 wrote 13 hours 22 min ago:
The caching strategy here looks really solid for keeping API costs
down. Curious how it handles state invalidation when the agent context
gets too large though.
ElenaDaibunny wrote 14 hours 49 min ago:
The caching strategy is doing most of the heavy lifting here cost-wise.
mkrd wrote 16 hours 24 min ago:
God, I whish there was a code harness I donât have to install a
JavaScript runtime for
tylerdurden91 wrote 16 hours 58 min ago:
Given the number of supply chain attacks via npm, maybe the recommended
approach to use should be pnpm instead of npx.
naaqq wrote 17 hours 40 min ago:
I don't think it's helpful, you can already get a 99+% cache hit on
claude code, just change the api settings to deepseek. I would like to
use a agent built by deepseek itself using deepseek models. Deepseek
should make their own agent based on their model, just like OpenAI and
Anthropic.
m00dy wrote 17 hours 37 min ago:
same here, using claude code on deepseekv4. just burnt 24.1M input
hit and 170k cache miss.
JSR_FDED wrote 17 hours 55 min ago:
Maybe the first problem this tool can tackle is creating a better web
page? Content continually shifting, super annoying.
tw1984 wrote 19 hours 15 min ago:
deepseek is building an official coding harness, why would anyone waste
time on such 3rd party toy when official one is coming probably in
weeks?
edg5000 wrote 19 hours 30 min ago:
Side note: In DeepSeek API docs they mention that coding clients
automatically are assigned the highest thinking effort, despite any
settings. This is what I suspected when using OpenCode with V4; it
keeps reasoning in very long cycles, this felt like a flaw in the
model. May just be a weird API thing.
Overall I find their API design and docs so messy. It's a shame, since
it's the main entrypoint to using their service.
yanhangyhy wrote 20 hours 31 min ago:
In the open-source contributors section, when you see a lot of anime or
cartoon avatars, you know most of the devs are Chinese.
treexs wrote 1 day ago:
codex generated sites are so easy to spot lmao
cloudengineer94 wrote 1 day ago:
Quite interesting being Terminal based and the AI skills staying within
a file of it's own.
Will give a go and see how cache behaves
trollbridge wrote 1 day ago:
Well folks here we have it: DeepSeekâs brand is now strong enough
people want to jump on their brand recognition.
nikolay wrote 1 day ago:
This is not an agent by DeepSeek, so the title is misleading.
jedisct1 wrote 1 day ago:
It's probably good, and the best for Deepseek models, but do we really
need one harness per model?
mark_l_watson wrote 1 day ago:
I tried it and the text input area was black with a dark font. I
checked the documentation, and asked DeepSeek v4, Claude, and Gemini
for help with the fonts/style and nothing works except to run in a
terminal with a dark theme. Crazy. None of the devs on the project use
a light theme?
miav wrote 1 day ago:
I agree that this is an issue, but..
no, they probably donât. Light themes are very rarely used.
jofzar wrote 22 hours 15 min ago:
I understand why, but I didn't even think of light themed terminals
till now..
.
jbellis wrote 1 day ago:
As someone who has been writing harnesses for a year: the people at
opencode etc aren't stupid, when they decide to break the prefix cache
[usually partially] it's always because they've tested it and it gives
better results overall.
If you think that dsv4 behaves differently enough from the aggregate of
other models, submit a PR with a patch to special case that to your
harness of choice with evidence. Just blindly assuming "append only all
the time because cache" is a waste of everyone's time.
schaefer wrote 11 hours 16 min ago:
> As someone who has been writing harnesses for a yearâ¦
Your agent harness, brokk, looks great. Iâm going to try it this
morning.
phrotoma wrote 13 hours 26 min ago:
Is "harness" in this context ~= "agent"?
furyofantares wrote 4 hours 38 min ago:
I think agent = harness + model.
abustamam wrote 11 hours 45 min ago:
I've understood harness to be the software that runs the agent
(open code, pi, Claude code)
anon373839 wrote 19 hours 8 min ago:
Are there any learning resources you'd recommend on writing
harnesses? I'm interested in doing a non-coding one, but not really
sure where to start.
sams99 wrote 2 hours 1 min ago:
My agent wrote a pile of very interesting articles at wasnotwas.com
I have been a bit quiet there for a bit, but it covers lots of
areas that are very interesting to harness builders (albeit less
interesting to the general public)
jbellis wrote 15 hours 54 min ago:
Generically, I would say, just start building it and ask your
favorite coding agent for advice when you get stuck. This is the
first technology that can teach you how to use it! (But do ask a
model with a recent knowledge cutoff, i.e. not gemini.)
stiray wrote 1 day ago:
If only author would understand, that some people want single, self
sustained binary that doesnt take half of computer memory and would
rather write it in rust or golang.
Defenestresque wrote 1 day ago:
github.com/charmracelet/crush
The company that had that acrimonious split from OpenCode. Still,
fully written in Go and compared to node-based harnesses, uses 1/5th
the RAM. (At least for me.)
Works with any provider (including OpenRouter free ones).
No conflict of interest here, just a happy "customer" of this
excellent resource.
_joel wrote 8 hours 13 min ago:
it's a 404
mikodin wrote 7 hours 59 min ago:
[1] (Haven't used it, but also hit the 404 and wanted to see it)
(HTM) [1]: https://github.com/charmbracelet/crush
wg0 wrote 1 day ago:
Can someone explain that was use of AI (and all the claims) that a
coding agent cannot be written in plain go for example? Given there
are tons of good terminal libraries for golang?
xlii wrote 1 day ago:
It can be written in Golang but interaction libraries are very
limited and with sharp edges.
There's Google's genkit, charmbracelet's fantasy and LangChainGo.
Each has ugly hacks and omissions. Then handling slice streaming of
data into Elm architecture (bubbletea) is also complex.
So in theory nothing stand against but in practice one has to get
quite low to the ground to get anything done.
Also: Golang agent exist! It's called crush and is developed by
charmbracelet people. It's so-so though I prefer Pi myself.
crystal_revenge wrote 1 day ago:
If this is what you want, especially in the age of coding agents, why
not just build it yourself?
Xeoncross wrote 1 day ago:
I'm really happy to see a lot of new software come out in Rust, Go,
or Zig.
The value and ease of development that slow interpreted languages
used to offer is disappearing. New languages have all the nice things
built in, or rather, our 1am pager alarms are starting to make us
mad.
zozbot234 wrote 1 day ago:
If you want to try a single self-contained binary that does take half
of your computer memory or more, there's always ds4-agent.
pancsta wrote 1 day ago:
Having a coding bot but skimming on coding? That should tell us
something.
sunaookami wrote 1 day ago:
Another day, another vibeslopped "product" on the front page of hacker
news with over 200 points. When will you guys learn?
adi_kurian wrote 1 day ago:
There is an uncanny valley effect to websites where FE is created in
full via an AI.
These sites have the immediate scent of 'high design', with errors that
no 'high designer' would dare make.
The italics give me nausea.
Text promoted with orange fill is seemingly random.
There is no thought behind the combination of art and copy.
Random smattering of Title Case and Sentence case and lower case.
A lack of commitment to a full stop
Widowed H1s.
H1s with random spaces .
At the same time, if I hammer CMD - to 25%, it looks fancy.
Perhaps nobody gives a fuck.
That said, I'm excited to try this tool!
arikrahman wrote 1 day ago:
Saw nix suffixed and was excited a new dotfiles was about to hit the
market.
carterschonwald wrote 1 day ago:
i cant find anything substantiated in the code that actually
differentiates it from any other harness.
my fork of oh my pi that i have a lot of experiments in, is lterally
designed to only work well with models that have decent reasoning
levels, like deep seek models. check it out! [1] â thats the install
script for after clone
fair warning: tis my dog food test bed as i build even fancier stuff
(HTM) [1]: https://github.com/cartazio/oh-punkin-pi/blob/main/scripts/bui...
agrippanux wrote 1 day ago:
This website seems to have been generated by Codex - I asked Codex to
create an HTML overview of a feature for my team and it made an overly
produced monstrosity - complete with the same large stat boxes that
were for the most part devoid of meaningful information - using the
same font, colors, layout, hero section, etc. It was also terrible on
mobile just like this is.
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.
easygenes wrote 17 hours 19 min ago:
Claude Opus 4.7 defaults to exactly this design language for a lot of
"just make me a rich html presentation page" requests without further
specification.
ritonlajoie wrote 23 hours 9 min ago:
strange, I got the same design with claude design, same fonts, same
title designs with the strange character etc...
port11 wrote 1 day ago:
A lot of LLM-driven design now looks like this. I donât understand
how people donât find ugly the pairings with an heavily italicised
serif. You also canât read much of the page on mobile, because the
code example keeps shifting the content around.
Now, that is overly critical, Iâm sure their heart is in the right
place. But a simpler website would do :)
gizajob wrote 16 hours 52 min ago:
Yeah such amazing tech used to produce a tediously unreadable
website with great flair.
krm01 wrote 1 day ago:
Itâs sad to see companies not spending a bit more on design.
Sure, ai will help you get something decent out fast. But thereâs
a threshold where design becomes an indicator of trust. Especially
for b2b software that tailor to large corps. Good design,
character, adds directly to the bottom line.
schaefer wrote 11 hours 37 min ago:
> Itâs sad to see companiesâ¦
The article is about an open source agent harness, Reasonix, that
is built to leverage the DeepSeek native api.
Thereâs no company here. No design budget. These people are
graciously sharing a project they made in their free time.
port11 wrote 2 hours 53 min ago:
I agree. I didnât mean to be too critical. But if theyâd
made something simpler, I think it would save them tokens and
end up more likely to convince their target audience of
developers.
(The series of âmotherfucking websitesâ comes to mind, they
were all very readable and simple, even if satire.)
darkmatriarch wrote 7 hours 44 min ago:
You're right, but I find as a solo engineer it's still
important to check the frontends I create on mobile
locknitpicker wrote 1 day ago:
> 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.
That doesn't say much about any model though. For starters, any
software engineer can tell you that leaving out features can
drastically simplify any project.
mmarcant wrote 1 day ago:
"byte-stable prefix cache" -- give us your codebase in a way that's
even EASIER for us to train on.
wg0 wrote 1 day ago:
Performance is horrible when you type but caching is magical.
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.
snqb wrote 1 day ago:
Deepseek API caches very efficiently itself. I use it heavily via pi
agent, and a lot of times I get 99%+ caching for longer sessions.
Have you tried using Deepseek API via other agents? This project tbh
looks like a S-tier slop
wg0 wrote 1 day ago:
I have used it with OpenCode and was good enough.
nextaccountic wrote 1 day ago:
> Tool-call repair
> 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?
WhereIsTheTruth wrote 1 day ago:
Y'all should not be writing js/ts/slop/npm based clis anymore
It's the agentic era, pick a better option
Just stop
fHr wrote 1 day ago:
yep codex opensource rust cli clears this night and day long
Alifatisk wrote 1 day ago:
Whats that option?
singingtoday wrote 1 day ago:
That site does not render correctly on my android. Lots of text on the
right breaking the reactive layout.
storus wrote 1 day ago:
Can it instruct DeepSeek during an LLM call to start removing old tool
calls from the context instead of waiting for the LLM call to finish if
the context size approaches DeepSeek's dumb zone? Claude Code can't do
that, /compact can only happen after the LLM call; it's often
preferable to start cleaning up context during an LLM call, especially
when tool calls are huge like reading markdown files;
implementation-wise all that is needed is to start removing earliest
... and replacing them just with some log entry stating this tool call
was already performed, then re-running KV cache prefill (so the
"online" compaction would get 0.5s latency hit every time it's
performed). That way one can read 1000 files in one LLM call.
m101 wrote 1 day ago:
For those of you that use deepseek v4 occasionally, what harness do you
use it with? Iâm only familiar with claude code and codex.
Any comments on what you can or cannot rely on it for relative to cc
and codex would be appreciated too!
eikenberry wrote 1 day ago:
Maybe check out Goose. It is the standard agent harness being
developed by The Linux Foundation under the AAIF. Under active
development and the implementation seems to have a good leg up on the
other popular agents. [1]
(HTM) [1]: https://github.com/aaif-goose/goose
(HTM) [2]: https://goose-docs.ai/
nsonha wrote 1 day ago:
I see their name mentiod everywere along with Aider, presumably for
being among the first agents, but I've never met anyone that
actually uses them.
droidjj wrote 1 day ago:
Check out pi.dev. OpenCode is a nice batteries-included Claude Code
replacement, but Iâm in love with the extensibility of Pi.
chuckadams wrote 1 day ago:
Any Pi extensions you'd specifically recommend? I'm just starting
out with Pi, but I've had mixed results with extensions. I'm using
Pi with gemma4 26b locally, so anything that's friendly to small
local models would be appreciated. I think the only extension I'm
using right now is pi-total-recall.
gck1 wrote 1 day ago:
I think pi wants you to write your own extensions, adapted to
your meeds.
I haven't had a need for any extensions though. Maybe subagents,
but I solved that with tmux. For all the rest, I just use
"skills".
ankitwarbhe wrote 1 day ago:
you created it yourself ?
Alifatisk wrote 1 day ago:
No.
andai wrote 1 day ago:
But Claude made the website?
Alifatisk wrote 1 day ago:
What conclusion are you drawing from that?
andai wrote 1 day ago:
If Deepseek can't even make a static site, why would I want to use
it for anything else? (Not saying it can't, just that it's a weird
choice to present your Deepseek-oriented product.)
Alifatisk wrote 1 day ago:
I see your point, but as we know, devs from Google and OpenAI
regularly use Claude Code because of its edge in frontend. I
think using another model to build your own thing is a pragmatic
engineering decision, not a sign of failure.
Hfuffzehn wrote 1 day ago:
This is really tickling the conspiracy theorist part of my brain.
"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?
danborn26 wrote 1 day ago:
High caching rates for coding agents can drastically reduce latency and
API costs. I am curious to see how the caching strategy handles context
invalidation across multiple files.
fouric wrote 1 day ago:
I don't think it's particularly effective to create a new coding agent
when there's existing open-source agents (especially extremely
extensible ones like Pi) that already optimize for cache hits, have far
larger communities, and work for providers other than Deepseek.
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.
hmokiguess wrote 1 day ago:
Click on the download page, it's hilarious. It has a lot of information
about the "smart probe" on the download and it's a realtime probe you
can rerun.
That's the pinnacle of AI slop over engineered garbage in my opinion.
All of that information is noise.
ricardobeat wrote 1 day ago:
> The loop is append-only, engineered around DeepSeek's byte-stable
prefix cache â long sessions hold 90%+ cache hit and input-token cost
collapses to ~1/5. Terminal-first, leave it running.
AI marketing slop. This is how all models and coding harnesses work,
isn't it?
The author claims (in another AI-written post):
> LangChain â along with every generic agent framework I checked â
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.
am17an wrote 1 day ago:
This Claude front end skill is now soon to be slop.
auggierose wrote 1 day ago:
Oh, I was wondering why all new websites look shitty in the same way.
aratahikaru5 wrote 9 hours 19 min ago:
Not a maintainer, but I've fixed some of the really jarring issues
on desktop (mobile needs a complete overhaul though). IMO It's not
that bad, and it gets the job done.
Any feedback on how to make it less "shitty"? I feel like doing
some vibe coding tonight.
ricardobeat wrote 1 day ago:
Already is. Every new website looks exactly the same.
imagetic wrote 1 day ago:
(HTM) [1]: https://shittycodingagent.ai
peheje wrote 1 day ago:
having issues with truncated output from deepseek v4 pro through
openrouter via pi-harness on ptyxis-terminal using ubuntu
trying reasonix with direct api..
peheje wrote 1 day ago:
first impression: the tui flickers a lot, unpleasent. very laggy to
write in.
mi_lk wrote 1 day ago:
Not sure about the story but it would be funny if pi folks actually
own this domain.
chuckadams wrote 1 day ago:
They do. That's Pi's old name.
chabes wrote 1 day ago:
Aka pi.dev
pkulak wrote 1 day ago:
Doesn't Pi Agent do exactly this? Assuming "append only" means they do
some kind of compaction as well.
yalogin wrote 1 day ago:
Can someone give me a eli5 version of what this is? It really sounds
useful to Claude subscribers.
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?
timcobb wrote 1 day ago:
You can also ask Claude and get an immediate answer, the power is
yours
Salgat wrote 1 day ago:
Certainly you realize that these comments exist for more than a
single person right? You expect potentially hundreds of viewers to
each burn through AI tokens instead of just getting a direct and
relevant answer here? This has the same vibe as the old forum posts
where the only response was a "google it".
quotemstr wrote 1 day ago:
> no reordering, no marker-based compaction
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.
hebetude wrote 1 day ago:
Wow the UI looks exactly what I vibe coded yesterday. What a
coincidence
huqedato wrote 1 day ago:
It's obvious why...
singiamtel wrote 1 day ago:
I would've liked benchmarks against other harnesses showing the caching
performance
Havoc wrote 1 day ago:
Just checked the stats on my opencode/DS usage...looks like 70%ish
hit rate.
Pretty shaky datapoint though...don't use it as primary model
Alifatisk wrote 1 day ago:
Is there benchmarks and measurements that offers comparisons between
different harnesses?
unshavedyak wrote 1 day ago:
It's pretty funny, i'm a $200/m Claude subscriber and i've had little
need to use anything else. However the more Claude has been restricting
my workflow (notably around the recent IDE/-p usage change) the more
i've been wanting to go elsehwere.
I'm concerned since i really want SOTA reasoning, but DeepSeek still
has me interested.
gck1 wrote 1 day ago:
I gave a fairly complex reverse engineering task to DS-4 xhigh and
GPT-5.5 xhigh today.
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.
Aurornis wrote 1 day ago:
> 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.
This is my experience with non-SOTA models across the board. When
you try them on little tasks and they work it feels amazing, but
then you go deeper and you're back to going in loops and fighting
the model for hours.
Switching back to a SOTA model immediately yields progress again.
When I read all of the comments from people saying they can't tell
a difference between Opus and I don't know if they haven't really
used it much yet, or if they're just not doing anything
complicated.
am17an wrote 12 hours 58 min ago:
Did you read the OP when he's exactly chiding the model you're
glazing?
Aurornis wrote 9 hours 54 min ago:
Did you intentionally miss the point of my comment? Substitute
Opus for GPT-5.5 if you will. I use both as well as locally
hosted models using some of your branches, even.
am17an wrote 7 hours 58 min ago:
Fair enough. I agree with you - although DS4 Pro is a GPT 5
class model which scores 46% on ARC-AGI-2[^1]. It's behind by
maybe 9 months, I think it's still good enough for a lot of
complex tasks as well. They definitely need to work on a
"just fucking works" harness like CC/Codex. Also thanks!
[^1]
(HTM) [1]: https://www.nist.gov/news-events/news/2026/05/caisi-...
cmrdporcupine wrote 1 day ago:
The GPT models are heavily biased to a more incremental, empirical,
evidence based approach. Sometimes to a fault. I prefer them for
this reason, but it requires coaxing or strategic use of /goal to
break it out if its highly staged, one piece at a time, approach..
if you don't like it.
I suspect for people doing more... website ... type development,
the more "yeet this into existence" style of Opus feels preferable.
With Claude I was constantly jamming my finger on the escape key
"wait, you did what?! based on what proof?!"
beering wrote 1 day ago:
You make it sound as if Codex is for people who know what they
want and Claude Code is for people who donât know what
theyâre doing.
cmrdporcupine wrote 1 day ago:
I was trying to not sound that biased, but ok ;-)
ttul wrote 1 day ago:
What youâre experiencing is the difference in model intelligence.
Most models can seem pretty good at simple stuff over short time
horizons. Complex work requires that more intelligence be stuffed
into those trillion-dimensional spaces.
KronisLV wrote 1 day ago:
> i've been wanting to go elsehwere.
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 ( [1] ) 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.
(HTM) [1]: https://ccode.kronis.dev
0xbadcafebee wrote 1 day ago:
You should definitely stick to the $200 plan, and not try the $10
coding plans with open weight models and higher limits. Anthropic
needs your money to stay solvent, and you'll sleep better knowing
you're using SOTA.
constantius wrote 1 day ago:
The world would be better long-term if we chise tonfund open models
instead however.
If you think short-term and only about yourself, paying for SOTA
regardless of how many military contracts the lab has is the best
thing, but paying for open models is both better ethically, and for
a future where AI belongs to everyone and not just to Altman et al.
port11 wrote 1 day ago:
(Zero reason to defend Anthropic.)
Iâve gone that route. I really wanted to stop using Claude, but
Deepseek v4 Pro and Kimi 2.6 didnât do the job. For a lot of
coding tasks or well-specced plans, maybe⦠but then thatâs a
plan made by Opus anyway.
Even Sonnet is sometimes not worth the trouble. Opus is very
thorough and reviews its own mistakes quite well. Catches a lot of
edge cases.
Iâm not saying we shouldnât try other things â I did! â,
but itâs more or less okay that people just like Claude Code
subscriptions? The back and forth I had with Kimi on a small
feature came out to ~1.8â¬, which is 10% of my Claude subscription
each month. And that was a single session. CC with Serena uses
tokens fairly well.
bazhand wrote 1 day ago:
/advisor is like the old /opusplan mode but for running tasks not
just pre-planning. It can work nicely with Sonnet as the main
agent and escalates to Opus as needed.
port11 wrote 2 hours 57 min ago:
Advisor-mode has been very helpful indeed, I can now plan with
Opus, have Haiku code, and escalate back to Opus for review.
Itâs a decent flow for Pro subscribers trying to max their
usage. But as Iâve said above, sometimes itâs not worth it:
Sonnet and Haiku can produce stuff thatâs not worth
reviewing.
Alifatisk wrote 1 day ago:
> I'm concerned since i really want SOTA reasoning
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.
unshavedyak wrote 1 day ago:
That's fair, neither am i - i do tend to work in large, complex,
full of legacy decision based codebases. Eg i have access to Sonnet
(of course), but i choose to solely work in Opus because i find its
output reads better, analyzes better, etc.
The "cost" is dumb models is just so high for me. Eg every bad
decision they make increases my frustration quite a bit. Despite
putting a lot of effort into my workflow to help reduce the number
of decisions they make, they always will. So my hedge is always
against that.. trying to reduce how insane they can be heh.
logicchains wrote 1 day ago:
If you want SOTA reasoning you should be using GPT 5.5 Pro.
unshavedyak wrote 1 day ago:
This is fair, but i've found the different models to have different
moods and require different interactions to get them to stick to
just the specific edits i ask for, etc.
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.
auggierose wrote 1 day ago:
Codex has only GPT 5.5
mmaunder wrote 1 day ago:
Unusable thanks to the top animation pushing the rest of the site down
repeatedly as youâre trying to read.
busymom0 wrote 1 day ago:
The layout of the entire page is horrendous on mobile too. Looks like
a huge wide site where content is only in a tiny column on left side.
schaefer wrote 1 day ago:
Okay, I'm curious.
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 â
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?
thomasfromcdnjs wrote 23 hours 21 min ago:
I would wonder that too, I'm only a novice openrouter user, but I do
notice it reroutes my same-model requests to different providers.
Maybe users reporting otherwise are just looking at their client
reports which wouldn't be able to tell the difference.
Lapel2742 wrote 20 hours 21 min ago:
Look into Openrouter's provider routing.
csunoser wrote 1 day ago:
Yes*. At least from my limited usage of deepseek-flash for a few
billion tokens on openrouter, the cache-hit rate is >95%. And I
simply used the claude code harness pointed at the openrouter
anthropic compatible endpoint with no fluff.
port11 wrote 1 day ago:
Did you get proper tool use? Some CC-driven models seem to get a
bit off when it comes to MCP usage. For example: I really struggled
to get Kimi to use Serena, which I think ended up costing too many
tokens.
schaefer wrote 1 day ago:
thank you!
sergiotapia wrote 1 day ago:
What AI model did you use for the website design? This is the second
one I see with the exact same font and color scheme. Just curious
because Claude models lean towards purples for example. Thank you!
FergusArgyll wrote 1 day ago:
Frontend design skill by Anthropic specifically says not to use
purple. I'd be surprised if it still uses purple. Have you seen that
recently?
pcwelder wrote 1 day ago:
Opus 4.7 selects such palette and motifs by default. Might even be
first iteration of claude design.
sheepscreek wrote 1 day ago:
DeepSeek v4 perhaps?
franga2000 wrote 1 day ago:
This design still screams Claude to me, but a newer version than what
you're thinking of. At some point they added a markdown file that
tells it to use obviously AI designs like lots of blue/purple and
gradients. Since then, this is its new style.
declan_roberts wrote 1 day ago:
I love the focus on cache hit efficiency. Hats off to the deekseek team
for creating a great product that maximizes cost efficiency for the
user.
nicce wrote 1 day ago:
Just in case, note that this project is someone's side project
> Independent open-source project · not affiliated with DeepSeek
Bombthecat wrote 1 day ago:
Adding already cheap API cost and you probably could let it run for
days and the same task..
stavros wrote 1 day ago:
How can you have cache hit efficiency? Isn't it just a matter of not
changing the previous context? I don't understand what knobs there
are to tweak on this.
everforward wrote 1 day ago:
> Isn't it just a matter of not changing the previous context?
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.
chillfox wrote 1 day ago:
changing between plan/build mode in some agents will change the
tools list, which breaks the cache.
brookst wrote 1 day ago:
Cache is always there, itâs just that it only caches up to
the point where an input token changes. So if the tools list is
early in the prompt, changing it would limit cache for most of
the prompt. If the tools list is the last thing, you could
still get 99% cache hits even if it changes every turn.
chillfox wrote 23 hours 17 min ago:
Depends upon the service and how the harness is built, Some
of the services allow for very few cache keys, so you won't
necessarily get any cache if you edit recent messages as the
cache is not per message, but big blocks of everything up to
a cache key.
This was actually surprising to me when I learned about it as
I have never worked with (or built) any cache working like
that before.
RevEng wrote 1 day ago:
After a couple of turns the system prompt is a small part of
the context. Not changing the system prompt at all is key so
that the rest of the history is itself part of the prefix.
bwfan123 wrote 1 day ago:
> Hats off to the deekseek team for creating a great product
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.
chillfox wrote 1 day ago:
Deepseek v4 pro is definitely my preferred cheap model, it's very
good, and I use it all the time for my personal projects (opencode
go plan), but I also use Claude Opus all the time at work and
Deepseek is not as good as that, but it does compete with Sonnet
for capability, and beats it on price.
spaceman_2020 wrote 18 hours 52 min ago:
I genuinely donât think you need Opus 4.7/GPT-5.5 tier models
for 95% of tasks in a normal workplace
People are out there using frontier intelligence to make
responsive headers and weekly work reports. Absolutely donât
need the latest and greatest models for this stuff
HDBaseT wrote 1 day ago:
Deepseek V4 Pro is an amazing model, even without the unreal cost
factored in.
It is my default model at the moment. I'm not doing anything too
complex though. I honestly found more expensive models like Qwen
3.6 to fail in tasks Deepseek nails.
I'm interested in knowing what people are using for tasks which
require a bit more thinking. Kimi 2.6? Qwen 3.7? GLM 5.1?
Akamant wrote 13 hours 25 min ago:
17 GoLang microservices for a serious project were written
perfectly using the latest version of QWEN(3.6). The only areas
where we really had to work hard were documentation and a very
serious task breakdown. All of this was tested, and yes, a
review was required, but everything was within reason. The
deadline was 10 days of 24/7 work, including the review. When
attempting to submit the same task, Opus 4.7/4.6 had to be
stopped after three hours. If you have significant resources
for experimentation, you can certainly try. For us, the choice
is absolutely clear at this point.
chillfox wrote 23 hours 46 min ago:
I don't think there's any open models at the moment that can
handle the more challenging stuff.
The things that I use Opus for at work is finding bugs in about
~200k lines of microservices and libraries in a niche language.
So, we will get these bug reports that are missing context,
can't easily be reproduced on our dev server, and are usually
the result of something deep in multiple services/libraries
combining with very custom configs. I can ask Opus (max
thinking) to find what could cause the bug, and it usually
nails it in a few hours (would take me 1-2 weeks to trace it
myself). The end result will be like less than 10 lines of code
to fix it, some tests to reproduce the bug and a nice report
explaining it, so it can be checked in an hour or two.
pjerem wrote 1 day ago:
I have unlimited Claude Opus at work and itâs wonderful. Not
allozwed to use it for personal use though.
So I use Deepseek Pro on the $20 Ollama Cloud plan and itâs
really not that far behind and I never triggered the planâs
limits.
Itâs like 10-15% less powerful but costs 10 times less.
Totally worth it. I prefer Opus because my employer pays for it
but I would personally never pay 10 times more for it.
chillfox wrote 23 hours 28 min ago:
Nice,
I have got unlimited Claude Opus at work as well.
I was really having a hard time deciding between the Ollama and
OpenCode plans for personal use, I couldn't really understand
how much usage I would get with the Ollama plan, so in the end
I went with OpenCode and I have never hit the limits despite
using it most evenings and weekends for several hours.
abustamam wrote 11 hours 37 min ago:
What models do you use in open code? I too have unlimited
opus at work and I tried using my same workflow from work
using Kimi 2.6 in open code and... It's just not it, even for
relatively simple stuff.
Maybe I should try DS4p?
theanonymousone wrote 1 day ago:
Isn't caching a server-side thing? How does the agent affect it,
significantly at least?
embedding-shape wrote 1 day ago:
Say you put the current time down to the second in the system prompt,
which is the message that goes in front of the entire conversation,
then basically nothing will be cached, every agent turn needs to
ingest the entire session over and over. Contrast to not doing that,
and the backend can leverage caching all the way up to the latest
message, as nothing until then changed.
nawitus wrote 18 hours 6 min ago:
That's not necessarily true, you can have multiple cache points,
see e.g.
(HTM) [1]: https://platform.claude.com/docs/en/build-with-claude/prom...
theanonymousone wrote 1 day ago:
Yes, of course you can destroy it. But how far can you "improve",
beyond decent "common sense" behaviour.
esperent wrote 1 day ago:
Surely other agent CLIs are not dumb enough to invalidate cache on
every turn over something so obvious?
brookst wrote 1 day ago:
Probably not that exactly, but there is a tradeoff between
effectiveness of the prompt and cache hit rate. If putting the
userâs datetime in the middle of the prompt scores higher on
evals but worsens cache hits, versus at the end of the prompt
where itâs cache friendly but may not be as effective, what do
you do?
This is still art as much as science and the different harnesses
take different approaches.
chillfox wrote 1 day ago:
I don't think any the agents breaks caching on every turn, but
they might do things like current list of files, or available
tools depending upon plan/build mode... or lots of other things
that breaks caching multiple times during a session.
embedding-shape wrote 1 day ago:
Obviously not, most agents properly keep previous messages
unchanged, at least the major ones I've been digging into the
source off. Also, everything would get so much slower, that even
developers creating their own agents would notice quickly how
much slower theirs is, if they fuck this up.
hirako2000 wrote 1 day ago:
Good timing given the cost spike across other frontier models.
notjes wrote 1 day ago:
Good thing DS just made their discount permanent.
(HTM) [1]: https://x.com/deepseek_ai/status/2057854261699195173
skeledrew wrote 1 day ago:
Not a fan of that page. The animated typing and resulting continuous
resize of the example keeps moving the content beneath it down and up.
Such bad UX.
m4rkuskk wrote 1 day ago:
Claude design AI slob.
embedding-shape wrote 1 day ago:
Agents or no agents, people still need to test their websites on
different resolutions or at least window width, but seems this is
becoming a lost art.
mirekrusin wrote 1 day ago:
Yeah, doesnât look designed for people who want to read it beyond
animated typing animation.
canadiantim wrote 1 day ago:
So what's best low cost coding agent these days? Kimi 2.6? Qwen's
latest closed model? Composer 2.5? DeepSeek?
throw10920 wrote 1 day ago:
Cursor with Composer 2.5 seems to be competitive with frontier models
(Opus and GPT-5.5) for a significant price discount. Benchmarks are
gamed, as always, but $0.55/task vs $11.02 a task definitely
indicates that there's some cost advantage.
(HTM) [1]: https://cursor.com/evals
abalashov wrote 1 day ago:
Although I have little interest in agentic coding, when I do use it,
I have found Kimi K2.6 to give Opus-quality output, and have switched
entirely to it for pretty much everything.
throw10920 wrote 1 day ago:
I've used Opus extensively and tried K2.6 on a few projects, and
the gap is huge. K2.6 is nowhere near the performance of Opus.
That's fine because it's also far cheaper, but public benchmarks
line up with my own personal experience that they aren't comparable
in terms of intelligence.
(that is, different places on the Pareto efficiency graph)
abalashov wrote 1 day ago:
No two uses are alike, I suppose. For me, whatever difference is
a wash. However, I probably tend to shy away from throwing
high-complexity/long-horizon tasks at the model.
stavros wrote 1 day ago:
For me, it's by far Deepseek. It's many times cheaper than
competitors, and about as good as Sonnet 4.6.
fouric wrote 1 day ago:
I'd generally agree about Deepseek being as good as Sonnet - but I
have extreme trouble with prompt compliance with V4 Pro in a way
that I've never had with Sonnet. I'll tell it "find the bug, but
don't fix it" or "please use this tool I just developed" and it'll
ignore me a high fraction of the time.
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?
stavros wrote 1 day ago:
I have Opus make a fairly detailed plan, then Deepseek
implements, and GPT reviews. With that setup, I have zero issues,
probably because what you mention is handled (the plan keeps it
on track and the reviewer catches any issues).
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.
e2e4 wrote 1 day ago:
Which CLIs are you using for each of the steps?
stavros wrote 1 day ago:
OpenCode for everything:
(HTM) [1]: https://www.stavros.io/posts/how-i-write-software-wi...
e2e4 wrote 23 hours 43 min ago:
thank you; will read your post
passive wrote 1 day ago:
I've gone through ~600m tokens in Xiaomi Mimo though Claude, and it's
been the most effective use of an agent I've had yet. It's very
capable, but generally not ambitious, picking simple but effective
solutions to most problems I give it.
Going to write something longer about the experience when I get to a
billion tokens.
Alifatisk wrote 1 day ago:
I do have my eyes on the coding plan, which is quite generous.
(HTM) [1]: https://mimo.mi.com
gandreani wrote 1 day ago:
Are you using Mimo 2.5 pro?
passive wrote 1 day ago:
Yes. I tried a couple of weeks with non-Pro, and it was pretty
good, but I had too many spare tokens, so I switched back to Pro.
:)
skeledrew wrote 1 day ago:
Seems to be DeepSeek.
(HTM) [1]: https://news.ycombinator.com/item?id=48237663
ac29 wrote 1 day ago:
Kimi 2.6 is great. Qwen3.7-max benchmarks similarly but I havent used
it yet
lostmsu wrote 1 day ago:
Just use codex with 5.5 on low reasoning levels
bwfan123 wrote 1 day ago:
In my experience, it is claude-code paired with deepseek-v4. For
penny-pinchers like me, I can have long coding sessions with it with
no anxiety about the cost. Also, prompting it to what you want and
verifying the outputs is more important than the quality of the
model. So, I am better off with a cheaper model and taking the
responsibility for prompting it and verifying the results.
raybb wrote 20 hours 52 min ago:
How to do connect deepseek to Claude code?
qaz_plm wrote 11 hours 56 min ago:
(HTM) [1]: https://api-docs.deepseek.com/quick_start/agent_integrat...
esperent wrote 1 day ago:
It's obviously much cheaper paying by the token but how does it
compare to a codex subscription on cost?
epolanski wrote 1 day ago:
Can you quantify the actual costs in a week and the use you make?
wongarsu wrote 1 day ago:
Not GP, but for my use I'd estimate $0.10-0.30 per hour of use
per agent with DeepSeek v4 Pro
embedding-shape wrote 1 day ago:
I'm not sure you need a "DeepSeek native coding agent" to take
advantage of DeepSeeks cache, yesterday as the Codex quota usage issue
still wasn't solved for me, I wrote a tiny little bridge so I could use
DeepSeek V4 Pro via Codex, and seems most of everything I did was
basically cached as far as I can tell: [1] (2026-05-23 Input (Cache
hit): 39,123,200 tokens, Input (Cache miss) 1,692,286), and the bridge
is doing not special, just massage the DeepSeek API shape into what
Codex expects, nothing particular about caching at all.
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.
(HTM) [1]: https://i.imgur.com/7eKn6wN.png
tontinton wrote 1 day ago:
Yep exactly my thoughts, went and looked at the code for the deepseek
provider in my coding agent. and basically all of what the author
wrote there is implemented... [1] for the curios
(HTM) [1]: http://github.com/tontinton/maki
kiproping wrote 1 day ago:
This would be a better page to link to [1] 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 â cache hit rate in practice:
<20%.
So they optimize on this plus other techniques to improve cache hits,
making it cheaper.
(HTM) [1]: https://github.com/esengine/DeepSeek-Reasonix/blob/main/docs...
embedding-shape wrote 13 hours 29 min ago:
> Most agent loops reorder, rewrite, or inject fresh timestamps
each turn
I haven't seen that, it'd be crazy slow if they did this. What
"agent loops" are they talking about here specifically? The
vagueness makes it sound potentially made up.
sparkleMing wrote 15 hours 34 min ago:
The last time I heard about something like this, it was Claude Code
intentionally injecting random strings to break caching when you're
not using a Claude model. Aside from that kind of intentional
sabotage, I don't think any coding agent would just ignore prefix
caching.
ikurei wrote 6 hours 44 min ago:
I haven't heard about this, could you please share more info,
some reference on that Claude Code intentional bug?
davesque wrote 2 hours 18 min ago:
I'm not sure what the mechanism is, but I've definitely had
Claude refuse to work on sessions that were touched by other
models. Some kind of integrity check failure. Resetting the
session back to the point before I used the other model fixed
the problem.
vidarh wrote 16 hours 16 min ago:
I've never seen an agent loop "reorder, rewrite, or inject fresh
timestamps" each turn other than mostly towards the end of the
messages. Messing with a large part of the context every turn would
be a fairly crazy thing to do.
nawitus wrote 11 hours 16 min ago:
Yeah. Those claims are just some random AI slop from claude..
vidarh wrote 10 hours 12 min ago:
It's a really lazy one too - there are so many open source
harnesses, including e.g. Codex and Kimi-CLI, and of course the
leaked Claude Code source, so it's trivial to verify if someone
even just bothered to ask an agent to check actual source code
examples.
krackers wrote 22 hours 1 min ago:
>Most agent loops reorder, rewrite, or inject fresh timestamps each
turn
That's really surprising, since it'd defeat the whole point of KV
caching. I mean I buy it considering how sloppily coded the
harnesses seem to be, but this like obvious low hanging fruit.
I've also often wondered why LLMs aren't trained with a format of
having a dedicated contextual system-instruction role at the _end_,
which you could use to put context like current time or other misc
stuff.
jeremyjh wrote 9 hours 9 min ago:
Its not surprising, that doc is full of AI slop.
3uler wrote 1 day ago:
Opencode has really bad cache stability issues that they seem
uninterested in fixing at the moment.
verdverm wrote 6 hours 52 min ago:
There are some that are specific to certain models like qwen/gemma
I switched to vLLM and those went away. Need to look at my opencode
config and adjust some others based on things I see here
magicalhippo wrote 18 hours 18 min ago:
What I noticed when using OpenCode with llama.cpp, was that the
default host RAM prompt cache size in llama.cpp was way too small
for say 128k Qwen3.6 27B.
The default is just 8GB and a full 128k context for the dense model
can take most of that. So then comes an agent and causes eviction
and subsequent cache miss.
Bumped the cache size (--cram IIRC) up to 48GB and had much better
results.
estebarb wrote 1 day ago:
I'm not sure that is really the case, or relevant in practice. I
have been using OpenCode with DeepSeek lately (regular coding). For
instance, today I got 120 million input tokens hitting cache, vs
just 2.59million missing cache.
ctxc wrote 20 hours 40 min ago:
Reads like a LOT of tokens to me. What does your usage /workflow
look like? I'm v curious because although I do use Claude code,
my token counts aren't nearly as much
I want to know if I'm missing something cool!
mordae wrote 15 hours 36 min ago:
Not OP, but I routinely load 150k tokens into context. A full
sub-package to work on, select other files in the monorepo,
e.g. front-end visualization and back-end data loader. Then
work some 150k tokens, then start again.
At the end, cache hit rate is like 99.5% if Novita is not
having issues.
For official DeepSeek API, 99.9% or something.
Custom harness that never compacts or otherwise doctors the
history.
ctxc wrote 11 hours 41 min ago:
Those numbers make sense to me...120 million input tokens is
like 120 sessions of hitting the full context limit, which
seems like a lot to me though
metalspot wrote 1 day ago:
I am getting 98.6% cache hit ratio on deepseek-v4-flash with
opencode
upcoming-sesame wrote 1 day ago:
out of curiosity, how do you measure cache hit rate in opencode ?
malikNF wrote 1 day ago:
opencode stats
hackernows_test wrote 1 day ago:
The first
lugu wrote 1 day ago:
So the calculation is:
Total input token = input + cache read + cache write
Cache hit rate = cache read / total input token.
That is 71% in my very limited use of opencode.
bobkb wrote 1 day ago:
Thatâs impressive!
On the sheer performance itâs comparable to Opus ?
stavros wrote 1 day ago:
Here are my stats (from DeepSeek directly, with a script I
wrote). The prices are what equivalent Sonnet usage would have
cost, the actual amount I paid was $10. On performance,
DeepSeek V4 Pro is comparable to Sonnet for me.
./cost.py amount-2026-5.csv 0.3 3.75 15
input_cache_hit_tokens: 472,971,520 tokens -> $141.8915
input_cache_miss_tokens: 13,299,013 tokens -> $49.8713
output_tokens: 3,334,962 tokens -> $50.0244
cache hit rate: 97.27% (472,971,520/486,270,533)
cache miss rate: 2.73% (13,299,013/486,270,533)
total: $241.7872
All of this usage was with an OpenCode subagent exclusively.
dathery wrote 1 day ago:
The OpenCode devs talk about this on Twitter a lot, e.g. [1] > 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:
(HTM) [1]: https://xcancel.com/thdxr/status/2048268697790300343
(HTM) [2]: https://opencode.ai/docs/config/#compaction
awoimbee wrote 16 hours 16 min ago:
You didn't quote the interesting part:
> our implementation is it only prunes calls from > 3 user
messages ago, if context is > 40K, and only if there's at least
20K tokens to be removed
Seems reasonable to me and explains why I can have long sessions
(way longer than with zed agents) while still hitting cache.
Opencode is just missing per-provider TTL.
arthurcolle wrote 14 hours 56 min ago:
I found that keeping current context utilization at 18% of
total context length was best for minimizing spend, across all
models with 400k context length or more
soerxpso wrote 1 day ago:
My understanding of caching with most models/providers is that a
prefix substring of the context has to be reused for a cache hit,
but not necessarily the whole entire context window. So if you
prune tool calls from the history, you're going to get one cache
miss on the newly-pruned history, and then you're going to be
getting cache hits on every subsequent turn, with a lower number
of input tokens. If you prune subsequent tool calls after that,
you would still get a cache hit for the already-pruned portion of
the context, just not the full context.
__natty__ wrote 1 day ago:
So it makes sense to first send stable prompt, reasoning and
files content, tool calls summary and actual tool calls at the
very end?
leemoore wrote 1 day ago:
The way you do this (and the way opencode does it) is you do
most of your pruning in more recent history. Last I looked at
opencode, they start pruning tool call results after 2 full
agentic turns. So you probably dont get quite as good hits on
cache for the most recent 1-5% of your turns, but after that
everything else caches fine and those tool calls that likely
aren't relavent to your session anymore are gone.
hirako2000 wrote 1 day ago:
They are. Empirical evidence on my side. Because attention is
sparse across the context. It's not truly treating a million
token the way it treats a fraction of that count. For
performance.
huqedato wrote 1 day ago:
I can't confirm this. Having utilized Opencode for a large project
over the past 10 months, with multiple models and agents, we've
never run into such 'cache stability issues'."
embedding-shape wrote 1 day ago:
That'd be really easy to spot and also fix, most likely. Any open
issue you could point us to, must surely been reported already?
3uler wrote 13 hours 6 min ago:
(HTM) [1]: https://github.com/anomalyco/opencode/pull/14743
krzyk wrote 1 day ago:
Opencode (and other coding agents) have hundreds of open issues
reported. It is quite discouraging when they are not being
closed/fixed.
verdverm wrote 6 hours 48 min ago:
These projects have also been the recipients of PR spam, lots
of duplicates and unconfirmed in there for less technical
people and clawd operators
nolok wrote 1 day ago:
> That'd be really easy to spot and also fix, most likely
Ah, reminds me of good old "There are only 2 hard problems in
computer science: cache invalidation, naming things, and off-by-1
errors."
criemen wrote 1 day ago:
> 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, ...).
franknord23 wrote 1 day ago:
That sounds like the experience of writing Containerfiles;
since steps are cached you want to pull the thing you are
iterating on as far down as possible.
verdverm wrote 6 hours 49 min ago:
It's even closer to prefix matching on super long strings
by chunk
gopher_space wrote 1 day ago:
All of this work has been done before in different
contexts. Memory management with bigger blocks and weaker
definitions that change whenever some grad student gets a
bright idea.
vidarh wrote 16 hours 10 min ago:
100%. Since you mention memory management: Generational
GC is pretty much the same idea: Keep the stuff that's
least likely to change an important property (liveness)
together.
Conceptually the underlying general idea is to sort
things based on stability if you can avoid recomputing
properties of the stable part.
himata4113 wrote 1 day ago:
this appears to be native to the terminal, as in, there's no special
application that runs or wraps an agent inside a tui. So basically
instead of commands you type plain english?
embedding-shape wrote 1 day ago:
> this appears to be native to the terminal, as in, there's no
special application that runs or wraps an agent inside a tui
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.
bwfan123 wrote 1 day ago:
> I wrote a tiny little bridge so I could use DeepSeek V4 Pro via
Codex
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.
spacedcowboy wrote 12 hours 55 min ago:
I don't think DeepSeek v4 Flash is as good as Claude for relatively
complex tasks. I ran with DeepSeek for a week, giving it the same
sort of tasks that Claude normally does, and then ran Claude and
asked it to continue. It found a whole bunch of things that had
been "overlooked" by DeepSeek, and spent some time fixing them
before wanting to move on.
DeepSeek is good, Claude is better, at least IMHO. Deepseek is a
lot cheaper though :)
NamlchakKhandro wrote 1 day ago:
Claude code and open code are streaming piles of shit
bayesianbot wrote 1 day ago:
LiteLLM can serve OpenAI API endpoint IIRC and proxy that to other
providers like DeepSeek, should work with Codex
Den_VR wrote 1 day ago:
Iâm feeling more a novice every day, but how isnât this just
handing over your code to team deepseek for whatever they might
want
spacedcowboy wrote 12 hours 53 min ago:
Somehow I don't think DeepSeek will be that interested in a 6502
compiler [1]...
1:
(HTM) [1]: https://atari-xt.com/
dudisubekti wrote 20 hours 59 min ago:
Yeah, but it's miles better than giving Anthropic and OpenAI your
data. At least Deepseek is releasing open-weight models and a lot
of open-source libraries.
If you're concerned about espionage then the only solution is
host the models yourself, which again, only open-weight models
like Deepseek enable you to do this.
jijji wrote 1 day ago:
there's laws on the books in China that says that every company
operating in China must aid and abet the Chinese government in
espionage against the rest of the world. given those facts, I
find it deeply troubling to be using anything coming out of
China, especially a program that runs in the context of a Linux
terminal on a machine that might have something important on it.
I'd argue it's a back door waiting to happen, if not sooner than
obviously later.
zaphirplane wrote 12 hours 7 min ago:
Like which country allows companies to not follow a legal
directive. How weird
subscribed wrote 13 hours 38 min ago:
I forbidden from working on the company code with DS, but if I
have a private something that looks pretty much like one of the
thousands repositories put there, it doesn't matter that much.
cicko wrote 17 hours 56 min ago:
Yes, and the Russians are still coming!
Danox wrote 1 day ago:
The four biggest (obvious) backdoor countries in the world in
no particular order the United States, Israel, Russia, China.
Honorable mentions, North Korea, Ukraineâ¦
tim-projects wrote 1 day ago:
Is it not better to have a country far away spying on you than
your own country?
azinman2 wrote 1 day ago:
Not if itâs industrial espionage.
goobatrooba wrote 1 day ago:
As a European I have to admit I am these days more worried
about the US than China. See yesterday's article about the US
government forcing Microsoft to give them lists of Dutch
government officials. Utter madness. At least the Chinese
mainly care about the money and power levers, the US about
strange worlds of revenge and manipulation, trying to change or
influence your government. E.g. which of the two countries has
put crippling personal sanctions on staff of the international
criminal court?
Honestly I'd love to love the US again, but basically after
Obama things have just gone down and down and no soul will
trust the US again in the next generation or two.
gizajob wrote 16 hours 54 min ago:
That particular rot actually turned cancerous with Bush and
Cheney, not Obama, IMO.
c1sc0 wrote 20 hours 4 min ago:
Besides the language barrier itâs actually also just
simpler to do business with the Chinese. There are issues
like censorship but they are known & can be routed around.
Itâs best to just ignore the US and move your business
elsewhere.
monch1962 wrote 22 hours 0 min ago:
As an Australian, I completely agree with every point in your
response
jijji wrote 22 hours 42 min ago:
The situation you reference is related to a specific
investigation by US congress requesting documents about
potentially illegal censorship actions by EU officials from a
specific company (microsoft). The difference is that the laws
in china are broadly defined to include giving all
intellectual property of anyone back to the government with
no oversight, for the purposes of espionage.
The former relates to a specific investigation about
potential criminal activity, the latter relates to broad
illegal activity committed by the government itself unrelated
to any specific case.
The US has no laws on the books forcing companies to wantonly
give intellectual property and other espionage level material
back to the government. If they did, no one would use cloud
providers.
To avoid this, you can run your own hosted machine in a
colocation facility, because in the US, people do have
reduced rights when their data is controlled by a third party
versus being controlled by themselves. Its the same as if
the data was in your house, they would need a search warrant
to obtain it, but when its at a Azure or AWS datacenter not
controlled by you, your privacy rights are reduced by doing
this.
MASNeo wrote 15 hours 8 min ago:
> no one would use cloud providers
I think many are trying to move away from US providers
actually. FISA section 702 and the current administrations
liberties taken towards international law are not helping.
The trust problem is real.
Not sure Iâd trust China with anything onshore. But
offshore, it does seem they play by the rules, because it
pragmatically serves the stability of the people. China has
not started wars in the past 50 years or so. By that logic
one may assume theyâd not abuse the arguably broad powers
over Chinese firms abroad to risk one now.
In a world where rules are increasingly less important how
states use power matters more to me than how they claim to
be monitored.
watwut wrote 15 hours 47 min ago:
> If they did, no one would use cloud providers.
EU has literal directive about location of data which has
to be located in the EU and not in the USA, because the
data are in danger otherwise.
miroljub wrote 15 hours 7 min ago:
Yep, and then they let US companies handle that data. One
more proof EU regime is run by ... no, I won't tell,
don't wanna get arrested.
cicko wrote 17 hours 53 min ago:
> The US has no laws on the books
Correct. They come up on Twitter daily. Pardon, this other
truth bullshit.
dominotw wrote 1 day ago:
so govt forcing a private coroporation being a big deal that
a its on the worldwide news is more scary to you than an
implicit mandate that china forces on its companies?
OtomotO wrote 1 day ago:
Exactly this.
I don't care about the US more than about Russia or China
these days.
They are definitely not our allies anymore.
dominotw wrote 1 day ago:
you dont need to be allies to do business. walmart is not
my ally.
ajuc wrote 11 hours 1 min ago:
Not enough trust to do business.
schubidubiduba wrote 12 hours 57 min ago:
The difference is that Walmart is a stable, reliable
trade partner that honors contracts and is not trying to
use propaganda to make you a fascist
OtomotO wrote 17 hours 40 min ago:
True, but then I expect them to betray me at any junction
and I'll gladly do the same.
_3u10 wrote 1 day ago:
FISA section 702 / Five eyes / Room 641A.
oldmanhorton wrote 1 day ago:
Youâre not a novice, there are a lot of us who know exactly
what we are doing and see this as a huge downside. We are just
being told to go faster, faster, faster lest we miss out onâ¦
something?
embedding-shape wrote 1 day ago:
Not everyone is working with state secrets or user personal data
(or even more closely guarded, company secrets) on a daily basis,
most of what I hack on is either FOSS already, or will be, not
much to keep secret here.
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.
jack_pp wrote 1 day ago:
either this or you work on software that even if copied won't
get you far since the business relies on network effects or
pure networking.
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
Demiurge wrote 1 day ago:
I actually agree with you, for the most part. The code I work
with actually does contain some valuable algorithms, but Im
pretty sure the effort of integrating them into a larger
system is pointless without the data. Itâs almost like
stealing half-life 2 source code without any assets.
Still, âGetting the source code of facebook or instagram
doesn't mean you could compete with them.â I think to
giants like that, having access to their source code could
open up some very interesting loop holes for manipulating the
ranking algorithms, or even security vulnerabilities.
jack_pp wrote 1 day ago:
True, haven't thought of that. However very few actual
projects / companies are in a situation where the chinese
GOVT would be interested to spend resources to hack your
platform. For the ones that are afraid of that there's
always self hosting of course
Den_VR wrote 11 hours 51 min ago:
I used to work with HVAC companies, and I noticed that
many of their customers mistakenly believed they were
purchasing air conditioners. They didnât consider these
devices, which they connected to the internet, as
computers. Despite being systems that required user
names, passwords, updates, monitoring, and other
maintenance, the prevailing attitude among these
customers was, âThis is an appliance, and why would
anyone care about my air conditioner?â
All this to say, not even subject matter experts
necessarily appreciate the risk involved in their work
embedding-shape wrote 1 day ago:
Sure, keep in mind it's a steaming pile of hacked together hacks,
probably won't work in every case, doesn't support every feature
that should be supported (like parallel tool calling, both Codex +
DeepSeek API support it), and it might make your computer catch on
fire: [1] 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.
(HTM) [1]: https://gist.github.com/embedding-shapes/eab3e63e5a95d3d78...
(DIR) <- back to front page