iiinova — where your digital self keeps evolving

Hugo @hugorcd.com · 25/09/2026

Evi can now work on any repo where its GitHub app is installed. I asked it to improve the evlog instrumentation of my portfolio. A few minutes later, the PR was up: score 47 → 100 and every handler instrumented, checked in a fresh sandbox. But this is the first real step towards evlog.cloud!

050

Reposted by Hugo

Benjamin Canac @benjamincanac.dev · 23/09/2026

Started working on Nuxt UI v5 👀 Months of research and feedback went into this roadmap: Design Tokens v2, Reka UI v3 and API cleanups we've been holding back since v4. Tracking progress in the open 👇 github.com/nuxt/ui/issu...

github.com

v5 · Issue #6918 · nuxt/ui

Nuxt UI v5 is a theming-led major. The full plan and the discussion live in #6610, this issue tracks what has landed. We're working on the v5 branch, prereleases are published under the next npm ta...

0222

Hugo @hugorcd.com · 21/09/2026

A small thing that has helped me put this into practice: I removed or hid most social apps on my phone, then kept their messages in @beeper.com. I can still talk to people across Instagram, WhatsApp, X, etc. without opening each app and immediately landing in a feed.

050

This feels especially important now. AI lets us produce almost anything, instantly and in enormous quantities. But being able to create more doesn’t mean everything deserves to be created.

090

For me, that means taking more photos, spending time in museums, becoming curious about unfamiliar subjects and genuinely pursuing the things that move me. It also means continuing to treat code as art: choosing what I build and giving it the time and care to become something I’m proud of.

1122

I don’t want to disappear from the internet. I want to become more intentional about what I let into my life and what I choose to create.

180

Lately, between the constant AI hype, everything happening in the world and the endless stream of content, I’ve felt less interested in consuming things by default.

180

I’ve never felt particularly addicted to social media, but I still open Instagram without really deciding to. I scroll, close it, and realise that something else chose what deserved my attention for the past twenty minutes.

170

The more things become possible, the more I care about having a reason for doing them.

1292

It's just an evlog plugin. One file, one line. Nothing is built yet, this is an experiment.

TypeScript evlog config: defineEvlog({ plugins: [createSignals({ model: 'typesafe-ai/jev' (the default), budget: { perMinute: 600 } (shared, fails open), signals: [fault, silentFailure, turn] })] }).

What keeps it honest: • runs after the response, never in the request path • your key, your process. evlog never sees a byte or a bill • promote-only, it never drops data • typed output + confidence, no prose • model version stamped on every event • fails open

100

Is it viable? Rough math at today's pricing ($0.042/M input tokens, output free): One judged wide event (~1.5k tokens): $0.00006 10M requests/month, 1% errors: fault on every error for ~$6/month The real limit is rate, ~20 decisions/sec. Judge candidates, not everything.

100

Same for agents. LLM-as-judge in prod gets sampled at a few % because of cost. What if every turn carried a verdict: resolved, looping, off-script? Not a replacement for evals. A screen that tells you which turns deserve one.

TypeScript: defineSignal named 'turn', runs on agent turns and asks three yes/no questions: resolved (the agent did what the user asked for), looping (repeated a tool call, no progress), offScript (did something nobody asked for). Result: { resolved: 0.91, looping: 0.03, offScript: 0.02 }.

Tail sampling only keeps what you can write a predicate for. status: 200, order: null gets deleted with the other 95%. A signal can promote it. When it keeps catching the same thing, you turn it into a rule. Signals discover, rules enforce.

TypeScript: defineSignal named 'silent-failure'. Runs when status is 200 on /api/checkout, asks 'Returned 200, but the user left empty-handed', and keeps the event when the probability is above 0.8 (promote only, never drop).

Annotated wide event JSON. path: /api/checkout. status: 200, commented 'looks healthy: with 5% sampling, this event is dropped'. order: null, commented 'but no order was created'. signals: { silent-failure: 0.94 }, commented 'Jev read the whole event: the user left empty-handed'. sampling: { keptBy: signal:silent-failure }, commented 'so evlog kept it'.

Why evlog? Jev judges a state, and evlog already builds one. One request = one wide event: user, cart, payment, error, duration, all in a single object. You can't ask a question of 40 scattered log lines. You can ask it of a wide event.

LLMs return paragraphs. You can't GROUP BY a paragraph. Jev doesn't generate text. You give it a state and typed questions, it returns a choice, a score or a probability, each with a confidence. That's not a summary. It's a column.

SQL query: SELECT signals.fault.value AS fault, count(*) FROM events WHERE status >= 500 AND signals.fault.confidence > 0.9 GROUP BY fault. Result: upstream 812, us 41, user 3.

Hugo @hugorcd.com · 18/09/2026

What does observability look like when a judgment costs nothing? I'm exploring Signals for evlog.dev, built on Jev by @typesafeai. Describe a column in plain English. Every wide event gets it: typed, with a confidence.

TypeScript: defineSignal from evlog/signals named 'fault'. Runs when status >= 400 and asks 'Who is responsible for this failure?' with three choices: user, us, upstream. Result: signals.fault = { value: 'upstream', confidence: 0.93 }.

1130

A little over a month late, but the first brick is finally in place. One year at @vercel.com. I’m writing about what a year here actually looks like: the work, the mission, the people. Can’t wait to share it!

080

Hugo @hugorcd.com · 17/09/2026

Agents are really good at writing Markdown. What if Markdown wasn’t the output, but the material?

Anthony Fu @antfu.me · 16/09/2026

Reposted by Hugo

It's been a while... Introducing Devframe v1.0 🚀 The framework for building DevTools that can run everywhere, for any meta-framework, for both humans and agents. Check out the announcement post and join our Discord 👇

devfra.me

Pluggable, Extensible, and Playful DevTools

Introduction to Devframe, a framework for building pluggable, extensible, and playful DevTools.

614630

Hugo @hugorcd.com · 15/09/2026

Hard to describe how hyped I am for the next version of Nuxt (and I'm not even talking about v5), the new CLI is a masterpiece, @roe.dev cooked so hard on this one

2240

Hugo @hugorcd.com · 14/09/2026

So I want to spend part of my open-source work exploring these problems: trying ideas, building tools and sharing my perspective on how we might get there. Expect more experiments around this in the coming weeks.