Крымский мост

@thdxr @dillon_mulroy I have a granular set of permissions, when model spawns subagents sometimes i see notification that permission access needed (i use plugin) but i don’t see menu in the tui with accept btn, because it’s inside of the subagent session and the whole thread stuck. Will it be fixed?

@dillon_mulroy that isn't right you can come up with funny looking cases for everything subagents lower cost in net and models are very very good at calling them efficiently

this is also why i don't add asub agent tool to pi and force it explicitly to use herdr + new pi instance when i want subagent like behavior this kind of thing is only going to keep pushing ent/api token spend customers to open weight models

@opencode I have a granular set of permissions, when model spawns subagents sometimes i see notification that permission access needed (i use plugin) but i don’t see menu in the tui with accept btn, because it’s inside of the subagent session and the whole thread stuck. Will it be fixed?

grok 4.5 is now available in OpenCode Zen also if you're a SuperGrok subscriber you can use it right in OpenCode

@maellsilva00 @MAKHACHEVMMA Why chama had such a big booty today clown?

@MAKHACHEVMMA Your time will come, Islam.

To quit in the biggest fight of your career - is something not everybody can do. Real Le Layenda 😃

@jim_ej @chantastic Vercel has a collection of react skills which are pretty decent and outside of scope of native react docs. Actually i think those skills are the only one useful for real work i saw in last 6 months sine skills become somewhat popular.

@chantastic It's just that the average React code on the net is poor quality because good React is tough. Someone needs to write a good React skill (if they haven't already).

Fable 5 is still bad at React. Is React the hardest problem in computer science?

@thdxr Wasn’t it available before? I used it in 1.16 already referencing sibling repos. The issue is though that fff mcp doesn’t index other folders. I updated to 1.17 i think yesterday it it says it has no tool like fff when i turned fff mcp off.

OpenCode 1.17.3 can reference other git repos or local folders 𝚛𝚎𝚏𝚎𝚛𝚎𝚗𝚌𝚎𝚜: { "𝚎𝚏𝚏𝚎𝚌𝚝": "𝚐𝚒𝚝𝚑𝚞𝚋.𝚌𝚘𝚖/𝙴𝚏𝚏𝚎𝚌𝚝-𝚃𝚂/𝚎𝚏𝚏𝚎𝚌𝚝-𝚜𝚖𝚘𝚕" } gives it full access to the effect codebase, here's how we use it

@realhenry @rms80 How many customers pay for your software?

@rms80 Have you done this workflow? It’s a huge unlock. If you can’t afford the subs and tokens that’s fine. But for those who are able, they should. Done properly they’re pretty incredible

guy who sells tokens says solution to all your problems is to spend more tokens

@nexxeln @opencode I have a parent folder with multiple big repos. I open opencode inside of individual repo folder. I have refs in opencode local configs to ref sibling repos when it’s useful. Currently i use fff mcp but it is only indexing current repo folder. Will adding fff support fixes this?

in the next version of @opencode, file search is powered by fff - files the agent opens start ranking higher - tool calls reuse the same search layer instead of starting cold - less wasted context - fast af here’s me searching around the linux repo, a ~40m line codebase

@sqs Composer 2.5 is pretty good for well defined tasks in the scope of the plan and Cursor has the best in class harness for gpt models. 5.5 builds the plan and user assigns sub agents 5.5/2.5 per task. Works relatively well and it’s way cheaper than amp.

The Factory team seems very smart, and I'm eager to see how well their cost-optimizing new model router works. If they can pull it off, that is a huge accomplishment. For Amp: we may try to build a model router that picks the best model for a task, but we don't intend to build a cost-optimizing model router based on the current state of the models. Here's why. Every time we've looked into using cheaper models in Amp, we've benchmarked on tasks that reflect how people use agents for coding today. On these real tasks, the expensive frontier model was not only the best (obviously), but also usually the fastest and cheapest, when measuring end-to-end task completion. Why? Cheaper-per-token models are less capable, which means that on complex real-world tasks they spend more tokens and time fixing mistakes along the way. You can find plenty of cases where cheaper models are indeed faster and cheaper end-to-end. But such cases were rarer than we expected, and the differences were fairly small. If you can easily detect such cases, then there is an opportunity here. But even then, on the AI hedonic treadmill, once people get a taste of frontier intelligence, they don't want to go back to using those more primitive prompts where cheaper models suffice. (Which is a good part of human behavior! It's how we decided to stop living in caves!) If your tasks can be handled just as well by non-frontier models, I would strongly advise you to uplevel how you use agents and what you produce to stay competitive against people who are using frontier models. In a power-law world, with rapid intelligence advances, try to get to the frontier and stay there.

@0xglitchbyte @theo Those folks narrow worldviews are scoped by the universe of IT tooling. There are huge codebases in real industries which existed before Ai and are here to stay.

This will never happen, nor should ever happen @theo is wrong here Your codebase is a contract you make with your users. Given a sufficient number of users, dependence on your codebase occurs in unexpected ways you did not intend or account for. Anything you change has unintended side effects elsewhere. That's "Hyrum's Law" in a nutshell. If you cannot account for these, you cannot specify for them. That doesn't even take into account ai problems like: - sycophantic behavior - hallucinations - non-determinism - context loss despite long windows - cannot prove generated code is correct - cannot generally prove a regenerated codebase behaves the same as the old one Each one of these alone is a difficult problem to solve, if it can be solved at all. The codebases you generate would be akin to allowing a lootbox determining your SLAs. Relevant XKCD:

@mrcldev @salauddin_n @AmpCode Its not 150 per months, it’s like $0.20 per hour, $5 max. You need more than $5 to complete even smaller task nowadays using best models. $5 is absolutely useless and actually pathetic they do even offer it lol

@salauddin_n @AmpCode You are complaining about getting free US$150 every month?

Been using @AmpCode for long time It's great tool and its harness also good but recently they reduced the fre quota 5 dollar which is bad. May they should introduce something basic plan like around 5 dollars where users can get come credits and use it instead pay as you go model

@_ChrisCovington @f_demaku @davis7 He is right though. 5.5 low reasoning is for most stupid shit only i would rather use kimi for. Medium reasoning code is way cleaner than what 5.4 produces but misses edge cases.

@f_demaku @davis7 your really going to quote benchmarks as your logic? 😂

This is very late, but I'm finally done with my 5.5 vid - use low reasoning - the name sucks - it's fast - best code I've ever seen a model write came from this model - openai's new pre-training is amazing - price looks worse than it is - over sensitive to every little thing in it's context window - feels wildly different compared to 5.4 - turn reasoning off. try it. turn the reasoning off. do it.

@MrClaustralian @thdxr This is a bold statement coming from a bogan who built nothing

@thdxr the guy who builds opencode celebrating claude rate limits lol. opencode's growth was built on piggybacking off claude's discounted auth and the second anthropic cut that off the popularity cratered. "positive impact on software" = "good for my product"

this is probably going to have a positive impact on software

@matteocollina It would consume the context and still be not followed from time to time. Seems like llm internal tendencies will always prevail.

Your coding agent keeps making the same mistakes? I built a fix. Introducing pi-self-learning: a pi extension that gives your coding agent actual memory. It extracts what went wrong after each task, scores learning by frequency/recency, and automatically injects them into context. 🧵 Here's how it works 👇

@mohammedraqeebb @ThePrimeagen You’re 🤡

@ThePrimeagen says the professional armchair commentator who has never built anything at a scale what replit has. go build something dawg.

Replit CEO so often sounds like a drug dealer

@TukiFromKL I hope your customers (if you have any) are fine with “youre absolutely right” responses lmao

🚨 Are you paying attention right now? Claude just put a price tag on your senior developer. $15–25 per code review. AI agents that start Instantly The second you open a PR. > Your tech lead makes $200K/year That's ~$400 per code review assuming 2 per day > Claude charges $15 That's a 96% pay cut > And the AI doesn't call in sick on Monday The people protecting their $200K salaries by gatekeeping code reviews are about to have a very bad quarter.

@mstockton @dillon_mulroy @dexhorthy Designing data intensive apps book has a whole chapter about that

@dillon_mulroy @dexhorthy This is actually kinda accurate - how many people on this site have any idea what map reduce is though? Simpler times man - simpler times

subagents are sparkling map reduce

@Matrix0983 @YLatynina Поэтому ты бот сбежал с украины и дрищешь изза бугра через впн бгг

@YLatynina Поэтому, к сожалению, будет до последнего русского за утопию «исконные земли». Так ненавидеть русских, россиян, как имперофилы и прочие атавизмы не умеет никто.

Позиция Европы уже год: пусть Россия прекратит огонь, а мы потом оставим ее под санкциями и ничего не договоримся. Этого не будет. Мерц прекрасно знает, что это абсолютно невыполнимое условие. С самого начала переговоров с США позиция России была проста. Договариваемся - сначала, огонь прекращаем потом. Позиция же Европы была ровно противоположна. Давайте прекратим огонь - то есть заставим Россию отказаться от стратегического преимущества - и будет договариваться до морковкина заговенья, не снимая ни санций, ничего, новый морковкин минск.

@rohanpaul_ai Ohh no ai bros pedaling bloomberg bs once again 🤡

A massive shift in the fintech landscape is happening with AI. Perplexity Finance + Perplexity Computer has the real chance to replace the $30,000 Bloomberg paywall. Using both she was able to build a terminal with real-time data to analyze Nvidia. 🤯

@thorstenball @AmpCode Seems like cursor extension already broken in deep mode. It goes into loop and not responding, not possible to stop it. After cursor restart it shows the thread content it generated. It definitely uses 5.3 based on output. 🤦🏼‍♂️

And here is the model that killed it: GPT-5.3-Codex. ampcode.com/news/gpt-5.3-c…

We believe the coding agent is dead. Soon, Amp will look very different. ampcode.com/news/the-codin…

@sqs So you’re saying that agents can now work without babysitting and, at the same time, save human resources by eliminating the need to maintain ide extensions? Hmm.

The standalone local coding agent, especially when confined to an editor sidebar that you babysit, is dead. It’s time to build what’s next. And we think we’ll be able to do that best and bring you along with us by calling out the truth. To help bring you along, we’re killing the Amp editor extension on March 5. Use the Amp CLI instead. It’s a better path to the future because it means less babysitting in your editor.