HenriNext

Having 4 separate limits that all are opaque and can suddenly interrupt work is not ok.

We don't know what the limits are, what conditions change the limits dynamically, and we cannot monitor our usage towards the limits.

1. 5 hour limit

2. Overall weekly limit

3. Opus weekly limit

4. Monthly limit on number of 5 hour sessions

Claude Code can already show diffs in JetBrains IDEs and VSCode ('/ide' command connects the CLI/TUI to plugin/extension running in the IDE-side).

It can also access the IDEs' real-time errors and warnings, not just compile output ('ideDiagnostics' tool), see your active editor selection, cursor position, etc.

- Forking VSCode is very easy; you can do it in 1 hour.

- Anthropic doesn't use the inputs for training.

- Cursor doesn't have $900M ARR. That was the raise. Their ARR is ~$500m [1].

- Claude Code already support the niceties, including "add selection to chat", accessing IDE's realtime warnings and errors (built-in tool 'ideDiagnostics'), and using IDE's native diff viewer for reviewing the edits.

[1] https://techcrunch.com/2025/06/05/cursors-anysphere-nabs-9-9...

Thanks, one more clarification please. The heading of point #3 seems to mention Google Workspace: "3. Login with Google (for Workspace or Licensed Code Assist users)". But the text content only talks about Code Assist: "For users of Standard or Enterprise edition of Gemini Code Assist" ... Could you clarify whether point #3 applies with login via Google Workspace Business accounts?

Claude Code was not designed from the ground up to be only an autonomous agent, but it can certainly act as one.

- It has non-interactive CLI functionality (with -p "prompt" option) in addition to the default interactive TUI, making it easy to integrate to workflows.

- It has turn-key GitHub integration (https://github.com/anthropics/claude-code-action).

- It has internal task tracking system that uses ReadTodo/WriteTodo tools to write JSON task lists to `$HOME/.claude/tasks/`, and enabling it to stay on track better than most other tools.

- It has excellent and customisable context compaction.

- And it has flexible permission system that can be used to turn all permissions questions to auto-accept when running in sandboxed environments.

Together those features enable it to be just as autonomous as any GitHub AI bot action hype thing (even though that might not have been its original or primary use).

It has fine grained permissions configuration file. And every permission question has three answer options: "yes", "yes and don't ask again", "no". And it has option '--dangerously-skip-permissions'. Out of all 20+ AI code tools I've tried/used, Claude Code has the best permission options.

AI Studio uses your API account behind the scenes, and it is subject to normal API limits. When you signup for AI Studio, it creates a Google Cloud free tier project with "gen-lang-client-" prefix behind the scenes. You can link a billing account at the bottom of the "get an api key page".

Also note that AI studio via default free tier API access doesn't seem to fall within "commercial use" in Google's terms of service, which would mean that your prompts can be reviewed by humans and used for training. All info AFAIK.

Interesting idea. But LLMs are trained on vast amount of "code as text" and tiny fraction of "code as AST"; wouldn't that significantly hurt the result quality?

Same experience. Especially the "step" comments about the performed changes are super annoying. Here is my prompt-rule to prevent them:

"5. You must never output any comments about the progress or type of changes of your refactoring or generation. Example: you must NOT add comments like: 'Added dependency' or 'Changed to new style' or worst of all 'Keeping existing implementation'."

I know half a dozen startups that let you pass in some set of keys and "build a multi-agent system"

Could you give names of those startups?

And yeah, Hugging Face is very much venture funded -- they had a modest 4.5b valuation in the last round.. (I just didn't know they had some competing product).

Thanks. Do you have a source/reference for the venture funding? (Perplexity searched 101 sites and concluded that it didn't find any public information about funding)

Curious.. which 'bring your own keys' -style competitors have venture funding?

"The need to direct boot shouldn't really be a concern", says the same man who said that the lack of expansion slots on the trashcan Mac Pro 2013 shouldn't really be a concern because you have a lot of Thunderbolt ports.

I despise GIMPs botched, unituitive and cumbersome UI so much

Btw there is a new fork of GIMP, called "Glimpse" [1], that tries to improve the UI. Their first release was just a re-branded GIMP, but maybe they'll get something done someday.

https://github.com/glimpse-editor/Glimpse

Funny that GTK was originally created for GIMP (GTK was acronym of "Gimp ToolKit"), and now they are 10 years behind the latest version of their own toolkit.

Bundling is not only about performance though.

Fetching all resources in a single HTTP request (or all data in a single GraphQL query) avoids a lot of issues and workaround effort caused by non-deterministic completion order of multiple small requests.

I've seen at least in two YC videos a very clear mention that the recommendations don't have much weight, basically saying "don't worry about it".

Secondly, is YC even that homogenous? It may be skewed towards enterprise software, but they do invest in a far wider range of startups than any other VC. And the ethnical and gender diversity is not that bad either, all things considered.

Looks really nice, well done.

The video was bit too fast for my slow post-christmas brain (plus fast pace doesn't induce feeling of calm and Zen). Otherwise perfect :)

I'm not exactly against paying subscription fees.

What royally pisses me off with Adobe is that I now have pay them monthly to maintain access to my photo catalogue (edits/keywords/etc), but they don't use my moneys to improve their products.

I don't even remember when was the last useful new feature to either Photoshop or Lightroom, but I doubt it was within this decade.

Looks great. I might ditch Lightroom for this and donate the subscription moneys to you instead, as a matter of principle.

But everybody is and forever will be comparing Darktable to Lightroom, so how about:

- Add to website a clear comparison vs Lightroom.

- Add to website how to import Lighroom catalogue.

Not for me. Check your browser?

I considered this overnight. I was naive.

If the goal is to help as many developers as possible, then being profitable and/or well funded makes achieving that goal easier, through hiring more employees and developing better product faster. Which in retrospective sounds obvious.

I'll still keep perpetual license as option, but stop being arrogantly non-commercial.

Absolutely. The long term plan is to do separate edition "BugJail for CI/QA/SIT/UAT/PreProd/etc", so that when things-are-not-fine on server, you don't need to repro but just load the capture in BugJail and see what happened [edit: that in addition to your "post CI build step" use case; both uses seem compelling].

But the current beta version is not robust enough and doesn't handle long captures well enough. I want to make sure that BugJail doesn't waste people's time, hence the license allows only capturing from local workstation.

The critical problem is not the instrumentation, but the way how BugJail works: it doesn't dumb the raw events to disk, instead it reconstructs relational model of the program execution. This requires resolving raw low level events against schema of types/methods/fields/etc. With JS the schema is dynamic, so now the already hard 'reverse-engineering' has moving target. Not impossible, just hard.

And thanks for the pointer to GraalJS/Truffle, I'll have look and perhaps that provides great way for doing the instrumentation side.

My "Fair bit of compression" was massively simplified: there isn't one true canonical "compression", but dozen different places and ways where and how to compress.

BugJail does compression around storage layer, but nothing before the network hop from capturing agent to processing server.

And fully agreed of course; when BugJail graduates to CI/QA/SIT/UAT/PreProd use, it will need compression before the network hop. But that is some time away. Currently the agent and server run in same box, and compression between them has negative net effect (compression uses more CPU than what shifting around less data saves).

BugJail writes its data files in open source format, Apache ORC.

You can load the data files directly to Spark, Hive or anything else that reads ORC, and then use Python to query, visualize or get statistics from the captured program execution.

Give it a try and let me know how it goes?

Thanks, appreciated. I get this fixed.

BugJail doesn't do currently line/statement level debugging, but it already captures all the information required for full "time travel" debugging. So, coming someday.

Also, BugJail is designed so that in normal use (webapps etc) it writes fast enough to database and never runs out of memory. Though, hardcode programs like databases and compilers exceed that, and currently don't buffer to disk, but crash it hard..

If you configure BugJail to capture only your own code/packages (--primary option) the disk usage is not that bad either. If you don't, you'll burn in the magnitude of ~10MB per second.

Give it a try, and let me know how it goes?

Thanks, appreciated!

There's fair bit of compression already, but actually I haven't done that much industrial spying into time-series databases, good idea, thanks.

Also, I've tested dual machine debugging with BugJail, as part of spike towards CI/QA/SIT/UAT/Pre-Prod use. It works okey-ish but the (crappy 1GB) network became bottleneck and it was slower than single box.