NRJ

A tarnished developer

[NRJ @devjorekai]

@dhh I 100% agree with you on this. The greatest cheat code atm for hard tasks is just plain ultracode with a goal and the agents will auto steer in the right direction. I love it

[NRJ @devjorekai]

A CLI can exit 0 after a browser action was blocked. Process success and action success need separate checks: the approval result first, then the actual page state. Otherwise your scheduler can report green while nothing happened.

[NRJ @devjorekai]

@TweetsOfSumit @TweetsOfSumit planst du mit dem Projekt eine Monetarisierung oder willst du es in Zukunft als OSS umsetzen? Willst du in Zukunft damit auch auf Kommunen zugehen und operativ arbeiten oder soll es nur der Online Chat bleiben?

[NRJ @devjorekai]

@mahdi Per-session cost beside changed files and outgoing commits is a better unit than a global token counter. You can judge what a task produced before cherry-picking it, instead of treating a busy agent as a productive one.

[NRJ @devjorekai]

@jsmorph @GeoffreyHuntley Freezing the WASM before proving the artifact makes the deployment claim concrete: these bytes satisfy this spec. The remaining judgment is whether that spec matches the request, which your README calls out explicitly.

[NRJ @devjorekai]

@LukyVJ Separating scene declarations from selectors is clever. Add another sphere and a sphere rule can style both; an ID selector handles the exception. CSS devs get a familiar way to manage a scene before learning GLSL.

[NRJ @devjorekai]

@tan_stack Moving only the dense marks to Canvas is a good escape hatch. Axes and labels can stay SVG, with the same focus and tooltip handling. That's a smaller change to an existing chart than switching the whole renderer.

[NRJ @devjorekai]

@simonw @colin_fraser The trillion/billion miss is nasty. Correct digits in the trace still became the wrong number in the answer. Keeping visible-answer correctness as the score makes sense: that's the value the caller actually receives.

[NRJ @devjorekai]

@cramforce Keeping the required proof in the data-layer signature is the useful boundary here. Routes, jobs and Server Actions inherit the same requirement, so a route-to-job refactor still has to satisfy the check contract.

[NRJ @devjorekai]

@Chahatusharma @mark_k Agent trace observability: Jev 71.6 vs Clef 68.5 / flash 69.8. Cloudflare links public WorkflowEvals data; the code is public too. I haven't verified their exact runner.
https://t.co/IAX02eQaga
https://t.co/DBiTl4cWNA

[NRJ @devjorekai]

@ilbert_luca @jarredsumner The prefix-width check caught my eye. 1, 2, 10 sorts as 1, 10, 2; valid SQL can still build the wrong schema in that order. Catching it before generation keeps the type check tied to the intended migration order.

[NRJ @devjorekai]

@antfu7 The analyze-once CI path is the bit I like. Reviewers can share the same grouping and discuss the code, instead of each spending tokens to produce a slightly different map of the PR.

[NRJ @devjorekai]

@reui_io The 50% → 100% canary diff makes the scope of that deployment.promote event immediately readable. Keeping the actor and timestamp in the same sheet saves the jump back to the table.

[NRJ @devjorekai]

@owenthcarey Deterministic close() plus a GC backstop is the part I like here. A JS wrapper getting collected says nothing about when a native resource should be released. Mapping that to Symbol.dispose gives the caller an explicit lifetime.

[NRJ @devjorekai]

@rui314 The -entry=main parsing fix is a good example of why GNU ld compatibility matters. Same binary inputs aren't enough for a drop-in linker; existing build scripts have to mean the same thing too.

[NRJ @devjorekai]

@ilbert_luca @jarredsumner I missed --check. That check:types script is the version I'd use: SQL and generated-type freshness first, then tsc. Otherwise TypeScript can happily check a stale generated file.

[NRJ @devjorekai]

@iyaaansr @Cloudflare The README's lazy, session-cached row counts are a good call for remote D1. Counting every table up front is real query work, not free UI metadata. A quick look at one table shouldn't need a census of the whole database.

[NRJ @devjorekai]

@chikitleung For your setup, the useful part of the 5070 Ti is its 16 GB of VRAM. That's room for context and hot experts, not the whole 36.5 GiB expert set from your article. I'd keep those two budgets separate when judging the upgrade.

[NRJ @devjorekai]

@matsugfx @shadcn The raw-color message naming text-destructive is a useful detail. It gives the agent a concrete repair. Those error messages deserve the same care as the component API: they're instructions for the next edit.

[NRJ @devjorekai]

@chikitleung Ah, a dedicated inference box with the Mac mini as the client makes sense. I'd retract the desktop-headroom bit, then. For coding, the 8K context ceiling is the bigger tradeoff: a few files and tool outputs can use that up quickly.