Kenn Costales

Building and

[Kenn Costales @kennyfrc]

@Akashi203 Hi! I upgraded to Builder Tier but I noticed that the workspace token limit is still 4M toks/min and did not change. The other limits increased but this one did not.

[Kenn Costales @kennyfrc]

@preshing @antirez he has something better: a free c course! https://t.co/ls66FMwOyN

[Kenn Costales @kennyfrc]

@VictorTaelin @0xSero you can try this out: https://t.co/B0QcMhwx3W

testable in https://t.co/xBVBXLPIT4.

[Kenn Costales @kennyfrc]

@mweinbach glm

[Kenn Costales @kennyfrc]

@donvito Very generous now compared to a few months ago. I never hit my limits nowadays

[Kenn Costales @kennyfrc]

@__morse @VictorTaelin less doom loops and less 503s starting 2 days ago

[Kenn Costales @kennyfrc]

@beyang @benhylak It seems to stop early more often than sonnet 4.5.

[Kenn Costales @kennyfrc]

@burkov I’m a big fan of @DeepInfra . Their Qwen 3 Turbo and Deepseek 3.1 Terminus instances are good

[Kenn Costales @kennyfrc]

@steipete agree. i went to the point where it’s just one shared module for both web & cli, and it emits events for the route handlers and command handlers to consume. testing is so much easier

[Armin Ronacher ⇌ @mitsuhiko]

If you are using claude code/codex for system administration/debugging on your local mac I encourage you to enable the pam_tid.so pam.d module. That enables touch ID for sudo which works even when an agent sudos.

[Kenn Costales @kennyfrc]

@steipete I moved from Sublime Merge to Fork as it has better worktree support and is equally fast!

[Kenn Costales @kennyfrc]

@steipete GLM mostly does fine, the only thing that bugs me with GLM is that it’s bad with editing / writing json and it can randomly reformat files while editing (changing single to double quotes, removing indentation) that’s not related to the task (which makes reviews diffs hard)

[Kenn Costales @kennyfrc]

@alexpotato @FoundersPodcast dwight howard

https://t.co/WtgtPLq1jW

[Kenn Costales @kennyfrc]

@thdxr it's my go-to when claude gets overloaded lol. it just needs a lot of help with context like https://t.co/6t5iVULzeN / todo.json stuff due to its small context window (seems to eat up tokens faster too)

[Kenn Costales @kennyfrc]

@jandotai Is there a specific prompt template that we need to use with the MCPs? It seems to be not doing any search tool calls - it makes up mock tool calls in the response as opposed to actually invoking it. Can't seem to replicate with a few test SimpleQA questions.

[Kenn Costales @kennyfrc]

@RepoPrompt @steipete cool! i'll try it out again soon. i really liked it just had stability issues (that was about 5 months ago)

[Kenn Costales @kennyfrc]

@Shpigford Definitely slow. 5s-20s TTFT based on Openrouter.

I switched to gpt-5-chat-latest and it’s so much better (<1s). The only problem there is the lack of function calling support, and you’ll need to define your own schema / tool call parser for that

[Тsфdiиg @tsoding]

Software is Infrastructure

[Kenn Costales @kennyfrc]

@steipete The only difference I see between Opus and Sonnet is that Opus is diligent in completing long todo lists (e.g. 20+ items) whereas Sonnet gets lazy and stops early. I have a command that breaks specs down to long, detailed todo lists. Otherwise, it’s almost unnoticeable