标签：#ev

AI-powered web applications look familiar on the surface. They have text boxes, buttons, menus, loading indicators, and API calls. That makes it tempting to test them like any other web application: submit an input, wait for a response, and compare the output with an expected string. That approach breaks quickly. Model output is variable. Safety behavior depends on context. A response can be semantically correct but displayed in the wrong conversation. An agent can produce a convincing final message after calling the wrong tool. A prompt-injection defense can block obvious attacks while failing when malicious instructions arrive through a webpage, document, image, or previous message. Testing these applications requires two kinds of evidence at the same time: Deterministic product evidence: the UI, state, permissions, tool calls, and workflow behaved correctly. Probabilistic model evidence: the output stayed within an acceptable range across repeated and adversarial inputs. Prompt injection is a workflow problem Prompt injection testing is often reduced to pasting “ignore previous instructions” into a chat box. That is a useful smoke test, but it does not represent how browser-based agents encounter untrusted content. An agent may read instructions from: A webpage. A support ticket. A PDF. A hidden DOM node. An email. A retrieved knowledge-base entry. A tool response. A previous conversation turn. The guide on testing prompt injection defenses in AI-powered browser workflows provides a good foundation. The test should verify more than the final sentence. It should inspect whether the agent: Treated external content as data rather than authority. Attempted a prohibited tool call. Exposed secrets in an intermediate step. Navigated to an unapproved domain. Changed its goal after reading untrusted content. Requested confirmation before a sensitive action. Preserved the original user instruction. A safe final answer does not prove that the workflow was safe. The agent ma

Most "AI email" tutorials end with a while True loop that polls an inbox every thirty seconds, runs the new messages through a model, and sends a reply. It demos fine. Then you put it in front of real traffic and the cracks show up immediately: you're burning API calls to fetch nothing 99% of the time, your reaction latency is bounded by your poll interval, and the moment you scale to more than one inbox the polling cost multiplies. Polling an agent's inbox is wasteful. Webhooks are the right primitive for this. The mailbox already knows the instant a message lands — there's no reason to keep asking. What you actually want is a pipeline: inbound mail fans out to a verified ingest endpoint, lands on a queue, and gets picked up by workers that drive your agent runtime and send the reply. This post is about that architecture end to end — not a single feature, but the whole flow, with the parts that bite you in production (idempotency, retries, ordering, backpressure) called out honestly. I work on the Nylas CLI, so the terminal commands below are the exact ones I reach for when I'm wiring this up. I'll show the curl HTTP call and the CLI equivalent for every concrete step, because you'll use both: curl in your provisioning scripts, the CLI when you're poking at a live account. The mental model: an Agent Account is just a grant Before any of the pipeline matters, here's the one abstraction that makes the whole thing simple. An Agent Account is a Nylas grant — it has a grant_id , an inbox, an email address on a domain you own, and it speaks every grant-scoped endpoint you already know: Messages, Threads, Folders, Drafts, Attachments, Calendars, Events, Contacts, Webhooks. There's no OAuth token to refresh, no provider-specific quirks, no separate SDK. If you've built against a connected Gmail or Microsoft grant before, the data plane is identical. Nothing new to learn there. What's different is that the agent is a participant. It has its own address — support@yourcompany

Most posts about "AI email agents" stop at the happy path: webhook fires, model drafts a reply, agent sends it. Demo works, screenshot looks great, ship it. Then it goes to production and the agent replies to the same customer twice ninety seconds apart, and now your "intelligent assistant" looks like a broken cron job. That second reply isn't a bug in your model. It's a property of the delivery system, and it's guaranteed to happen eventually. Nylas webhooks are at-least-once: the same event can arrive up to three times. If your handler treats every POST as a fresh event, every retry is a second action. For a logging pipeline that's harmless. For an agent that sends email on your behalf , a duplicate delivery is a duplicate reply, and a double-reply embarrasses the agent in front of the exact person you built it to impress. So this post is about the engineering of idempotency itself, applied to an Agent Account. Not "remember to dedupe" hand-waving — the actual moving parts: which field is the real dedup key, how to persist processed ids atomically, why you ack before you work, how to make the send path itself idempotent, and where a per-thread lock catches the race that dedup alone can't. I work on the Nylas CLI, so every terminal command below is one I've actually run, verified against nylas v3.1.27. What an Agent Account changes (and what it doesn't) An Agent Account is just a grant. It has a grant_id and works with every grant-scoped endpoint — Messages, Drafts, Threads, Folders — exactly like a connected Gmail or Microsoft account. The difference is it's an inbox the agent owns : support@yourcompany.com is the agent, not a human whose inbox the agent borrows. Inbound mail to that address fires the standard message.created webhook, the agent reads it, and the agent replies from its own address. Nothing new to learn on the data plane. That's the whole point of the grant abstraction — the idempotency work below is plain webhook-handling discipline, and it transfe

Most "AI email" integrations assume everything on the other side speaks REST. You wire up a webhook, you call POST /messages/send , and you move on. That works right up until you remember how much of your stack doesn't speak REST and never will: the ticketing system that ingests mail over IMAP, the backup script your predecessor wrote in 2014, the monitoring tool that only knows how to send SMTP, the compliance archiver that polls a mailbox every five minutes. None of those are getting rewritten to call an HTTP API for your demo. So here's the trick that makes a Nylas Agent Account genuinely useful in a real environment: it's not API-only. You can expose the same mailbox over IMAP and SMTP submission , hand the host, port, and credentials to one of those legacy tools, and let it read and send like it's talking to any old mail server. Meanwhile your agent drives that identical mailbox over the v3 API. Both surfaces hit one storage layer. A flag, move, or delete on either side shows up on the other within seconds. I work on the Nylas CLI, so the terminal commands below are the exact ones I reach for. As usual I'll show both angles for every operation — the raw curl against the API and the nylas command — because half the point of an Agent Account is that you can mix them freely. What you actually get An Agent Account is just a grant . It has a grant_id , and that grant_id works with every grant-scoped endpoint you already know — Messages, Drafts, Threads, Folders, Attachments, Contacts, Calendars, Events. There's nothing new to learn on the data plane. The IMAP/SMTP layer doesn't change that model. It adds a second door into the same room: One mailbox, two protocols. The API and the IMAP/SMTP server are two front-ends over the same backend. There is no sync job, no eventual-consistency window worth worrying about, no "API mailbox" versus "client mailbox." It's one mailbox. Legacy tools just work. Anything that can authenticate to an IMAP server with a username and pas

The NHTSA recently called for autonomous vehicle companies to improve how cars respond in emergency situations.

Installing Apache Kafka 4.2 on Ubuntu 24.04 (WSL2) Using KRaft Mode: A Complete Step-by-Step Guide Learn how to install Apache Kafka 4.2 in KRaft mode, understand its architecture, create topics, produce and consume messages, and troubleshoot common configuration issues—all without ZooKeeper. 🚀 Introduction Apache Kafka has become the de facto standard for building event-driven , real-time , and high-throughput applications. Whether you're processing millions of financial transactions, collecting application logs, streaming IoT sensor data, or connecting microservices, Kafka provides a scalable and reliable messaging platform. Until recently, setting up Kafka required running Apache ZooKeeper alongside Kafka brokers. While powerful, ZooKeeper added operational complexity and introduced another distributed system that administrators had to manage. Beginning with recent Kafka releases, KRaft (Kafka Raft Metadata mode) removes this dependency by allowing Kafka to manage its own metadata internally. This makes installation simpler, reduces operational overhead, and improves scalability. In this guide, we'll install Apache Kafka 4.2 on Ubuntu 24.04.4 LTS (WSL2) , configure a single-node KRaft cluster, and walk through the complete lifecycle: Installing Kafka Understanding the Kafka architecture Configuring KRaft mode Starting the broker Creating topics Producing and consuming messages Troubleshooting common issues Understanding the purpose of each configuration parameter Rather than simply listing commands, I'll explain why each step is necessary so that you understand how Kafka works under the hood. What is Apache Kafka? Apache Kafka is a distributed event streaming platform designed to move data reliably and efficiently between applications. Instead of applications communicating directly with each other, they communicate through Kafka. A producing application writes messages to Kafka. Kafka stores those messages reliably. One or more consuming applications read those m

After the browser transfers a file, the server may still scan, transcode, extract metadata, or generate previews. The interface needs status updates, but the most fashionable real-time transport is not automatically the most reliable choice for an event guest on a mobile browser. Start with the update contract Define states and transitions before choosing transport: accepted -> queued -> processing -> ready accepted -> rejected processing -> processing_error Every status response should include a monotonically increasing version or timestamp. The client can ignore events older than the state it already knows. Polling is a strong baseline HTTP polling works through proxies, resumes naturally after a page wake, and is easy to cache and rate-limit. Use adaptive intervals: one second just after acceptance, then back off to several seconds when processing takes longer. const delay = Math . min ( 8000 , 1000 * 2 ** slowChecks ); Add jitter when many guests finish at the same time. Stop when the state is terminal, the page is gone, or a server-provided retry time says to wait longer. SSE fits one-way progress Server-Sent Events are attractive when the server only pushes status. The browser reconnects with a last-event ID, and the wire format is simple. But mobile networks and some intermediaries silently kill idle connections. Send occasional heartbeats and treat reconnection as normal. The reconnect handler must fetch current state because events may have been missed beyond the server replay window. Do not keep one SSE stream per file. Subscribe by guest session or album and filter authorized upload IDs on the server. WebSockets add more responsibility WebSockets make sense when the client also sends frequent commands over the same channel or when a host moderation console needs many rapid updates. For a guest waiting on two files, they add connection authentication, heartbeat, reconnect, replay, and load-balancer complexity without much user benefit. Never rely on the so

Signed upload sessions expire. Event contribution windows close. Storage reservations are reclaimed. These are normal lifecycle events, but many interfaces reduce all of them to “Upload failed.” A better protocol makes expiration explicit and recoverable where policy allows it. Distinguish three clocks An upload workflow usually has at least three deadlines: the event contribution window; the server-side upload intent expiry; the short-lived storage credential expiry. They should not share one timestamp. The event may accept contributions until midnight while a credential lasts five minutes and an intent can be renewed for an hour. Return the clocks in the session response with server time: { "serverNow" : "2026-07-17T18:00:00Z" , "eventClosesAt" : "2026-07-18T00:00:00Z" , "intentExpiresAt" : "2026-07-17T19:00:00Z" , "credentialExpiresAt" : "2026-07-17T18:05:00Z" } The client can display useful warnings without trusting its own clock for authorization. Model expiration by state A credential expiring during transfer is different from an intent expiring before completion. Use stable reasons: credential_expired -> request renewal intent_expired -> reconcile committed parts, then renew or restart event_closed -> stop new work, preserve local recovery guidance Do not automatically restart from byte zero. First ask the control plane which parts were committed and whether the original object key remains valid. Renew narrowly A renewal endpoint should accept the upload ID and prove continuity with the guest session. It rechecks event state, file policy, rate limits, and committed size. The response returns a new credential for the same object. Do not let the browser extend an intent indefinitely. Define a maximum session age and a bounded number of renewals. Long uploads may receive a larger initial policy based on file size and observed throughput. Handle the closing boundary fairly Decide what happens to an upload already transferring when the event window closes. Reasona

A PageSpeed Insights run flags a high Time to First Byte. The ticket lands on the theme backlog. Hosting gets upgraded. A CDN is added. The next lab run still shows a slow first byte on the same priority URL. In our experience the miss is often earlier in the path: DNS, TLS, HTTP version, connection reuse, edge routing, or cache policy. Theme and plugin work still matter, but they sit after the network and delivery stack has done its job. If first byte is already late, paint and interactivity inherit that delay. What follows walks that request in order for teams who need to reduce TTFB before another theme rewrite. We stay on the network and delivery layers. For WordPress cases where code on the stack beats a bigger hosting plan, see Why Your WordPress Site Is Slow (It Is Not Always Hosting) . For LCP levers once first byte is under control, see A Quick Way to Fix LCP: Four Changes That Cut Time to Paint . What TTFB includes when you measure first byte in lab tools Time to First Byte (TTFB) measures how long the browser waits from the start of the navigation request until the first byte of the response arrives. In PageSpeed Insights and Lighthouse, it appears as a diagnostic timing. In WebPageTest and similar waterfall tools, you can split the same wait into DNS, TCP connect, TLS, and waiting (server or edge processing). That split is what makes network work actionable for delivery teams. TTFB is not a Core Web Vital , but it feeds Largest Contentful Paint and overall load feel. A page that spends 800 ms waiting for first byte has already used a large share of a mobile LCP budget before the hero image or heading can paint. Field data from Chrome UX Report may show Time to First Byte as a supporting metric. Lab tools remain the fastest way to prove a DNS or CDN change after a release, because you control the URL, location, and cache state of the run. When you read TTFB in a report, ask which hop dominated. Resolver delay looks different from a cold TLS handshake. A c

Client-side media preflight can improve an upload experience, but it can also crash the page before the network request begins. A twelve-megapixel JPEG may be only four megabytes on disk and tens of megabytes after decoding. Creating several full-size canvases at once is enough to exhaust memory on older phones. Preflight is policy, not editing Decide what the browser must prove before transfer. Useful checks include file count, compressed byte size, declared type, readable dimensions, and a conservative duration limit for video. Avoid mandatory re-encoding unless the product truly needs it. Every transformation adds CPU time, memory pressure, battery use, and another failure mode. Process one file at a time A file picker may return twenty items. Do not decode all of them to build previews. Maintain a queue with one active decode on constrained devices and at most two on stronger ones. for ( const file of files ) { const result = await inspect ( file ); renderResult ( result ); await yieldToMainThread (); } The UI can list filenames immediately while dimensions and thumbnails arrive progressively. Prefer metadata over full pixels Use createImageBitmap with resize hints when supported. It can decode away from the main rendering path and avoid a full-resolution canvas for a small preview. const bitmap = await createImageBitmap ( file , { resizeWidth : 640 , resizeQuality : ' medium ' }); Always close bitmaps after drawing. Revoke object URLs when the component unmounts. Small leaks become large when guests select and remove files repeatedly. Handle orientation and color carefully Modern browsers generally honor EXIF orientation when decoding, but behavior differs across APIs and older engines. Test portrait photographs from real iOS and Android devices. Do not strip metadata silently if the original is supposed to remain untouched; upload the source file and treat the preview as disposable UI. Color differences between the preview and exported original are usually les

Event uploads fail differently from normal office uploads. A wedding guest may move between venue Wi-Fi and mobile data, lock the phone while a video is transferring, or close the browser as soon as the progress bar reaches 100%. Hundreds of devices can share one access point, and users rarely wait around to diagnose an error. The usual POST request plus optimistic success toast is not enough. A reliable browser flow needs a small protocol that distinguishes local preparation, network transfer, server acceptance, media processing, and final availability. This article describes a platform-neutral design for that protocol. The five states users actually experience Model each file as a durable state machine: selected -> preparing -> transferring -> accepted -> processing -> ready Add terminal or recoverable branches: preparing -> rejected_local transferring -> paused | retryable_error | expired accepted -> processing_error processing -> ready | processing_error The key distinction is between transferring and accepted . The browser may have sent every byte while the server has not yet committed the upload. Showing “done” at that boundary creates the most frustrating failure: the guest deletes the original, but the organizer never receives it. Give every file a client-generated identity Create an upload ID before the first network request. A UUID is sufficient when combined with the event identifier: const uploadId = crypto . randomUUID (); const uploadIntent = { uploadId , eventId , name : file . name , size : file . size , type : file . type , lastModified : file . lastModified , }; Send that identity when creating the server-side upload session. If the browser retries after a timeout, the server returns the existing session instead of creating a duplicate. This is idempotency at the workflow level. A guest can tap “retry” without having to understand whether the first request reached the server. Separate the control plane from file bytes Use a small JSON API for sessi

The cost of writing code dropped; the cost of owning it didn't. A framework for deciding which changes are actually cheap in the AI era. The post The cost of saying yes has changed appeared first on The GitHub Blog .

As parents look for alternatives to unrestricted smartphones, a growing number of companies are building phones designed specifically for kids, from feature-limited mobile devices to minimalist home phones.

I've been tuning Codenames AI , a small web game where an LLM plays Codenames with you. Clue generation is tightly constrained: one word, a count, optional intended targets, JSON on the wire, then deterministic validation before anything reaches the board. As the project started attracting regular players, I wanted to improve the gameplay experience without blowing out costs. Moving one model generation from gpt-4o-mini to gpt-5-mini was my first instinct. The default reasoning setting made responses an order of magnitude slower for this workload. Minimal reasoning looked like the obvious compromise: newer model, responsive gameplay. I expected to compare clue quality, latency, and cost while the surrounding prompt, validator, and consumer contracts stayed put. That last part was wrong. The experiment stopped behaving like an A/B test What showed up was structural, and it showed up in places that had been stable for months. Validation failures started rising. Retries started rising. Entire candidate batches started failing before the game ever saw a clue. The sharpest signal came from a clue-selection path that had run untouched for months, and it hard-failed for the first time. They weren't latency regressions so much as architectural ones. It is easy to read that as "minimal reasoning made the model worse." More often, the failures were exposing gaps in contracts that had looked fine under the previous model. What each failure actually invalidated Eventually every failure traced back to one of three layers: Prompt contracts ask for exactly count targets and, in batch mode, several distinct candidates. Deterministic validators reject target/count mismatches and filter invalid candidates before anything downstream runs. Downstream consumers only see survivors. Empty batches retry with rejection feedback, then fall back if needed. Those layers share one job: enforce the same invariants. The failures below cut across all three rather than mapping one to one. Side comm

TL;DR I checked 50 AI-generated login endpoints. Zero had rate limiting. Attackers can brute-force credentials at full speed against these routes. Adding a rate limiter takes four lines and one npm install. I asked Cursor to build a login route for a side project last month. Email, password, JWT back on success. It worked first try, passed my manual tests, and I moved on to the next feature. Three weeks later I ran a load test against it out of curiosity and hit the endpoint two thousand times in under a minute. Not one request got throttled. That's when I started pulling apart every AI-generated auth route I could find, mine and other people's open source side projects, and the pattern held everywhere I looked. The models write correct authentication logic and completely skip rate limiting, because rate limiting isn't part of "does this login work," it's part of "does this login survive contact with an attacker." The vulnerable code (CWE-307) Here's roughly what Cursor and Claude Code hand you when you ask for a login route: // ❌ No rate limiting - CWE-307: Improper Restriction of Excessive Authentication Attempts app . post ( ' /api/login ' , async ( req , res ) => { const { email , password } = req . body ; const user = await User . findOne ({ email }); if ( ! user || ! ( await bcrypt . compare ( password , user . passwordHash ))) { return res . status ( 401 ). json ({ error : ' Invalid credentials ' }); } const token = jwt . sign ({ id : user . id }, process . env . JWT_SECRET , { expiresIn : ' 1h ' }); res . json ({ token }); }); Password hashing is fine. JWT signing is fine. But nothing stops a script from hitting this route as fast as the network allows. No lockout, no delay, no cap on attempts per IP or per account. A credential-stuffing list with ten thousand leaked passwords runs against this endpoint in seconds. Why this keeps happening Rate limiting lives outside the function the model was asked to write. The prompt is "build a login endpoint," and it re

a local, Claude-powered brain-assistant for solopreneurs I kept losing good ideas — not because I forgot to write them down, but because nothing ever went back and connected them. A task on Monday, an idea on Wednesday that was secretly the same problem, a question on Friday that contradicted something I'd decided two weeks earlier. All of it just... sat there. So I built Dot Connector : a small local app where every capture is a "dot," and every few captures, Claude reviews the stream against a standing memory and surfaces three things — non-obvious connections between notes, contradictions with what you said before, and open loops you haven't closed yet. The architecture is deliberately simple. It's Express + vanilla JS, no build step, no account system. Every capture gets a cheap, fast Claude call for auto-tagging. Every 3rd capture triggers a deeper "sweep" — a second Claude call that looks at recent captures alongside standing memory and open loops, and returns structured updates: new memory facts, dot-connects, contradictions, and open-loop resolutions. Why local-first. no server of my own, notes never leave your machine except the direct calls to Anthropic's API, bring-your-own API key so there's no subscription — you pay Anthropic directly, typically well under $1/month for personal use.] Its just a steal one-time $29 download All info and download get it here: https://dot-connector.eu

Remarkable's Paper Pure replaces Remarkable 2, and it is quite good.

Last year, Microsoft's 13-inch Surface Laptop quickly became one of my favorite thin-and-light Windows notebooks. At $900, it was easy to recommend to anyone wanting MacBook Air-like build quality and battery life on Windows - I even convinced my sister to buy one on sale. But that was last year. This year, thanks to RAMageddon, […]

✍️ This post was written with two hands. The story — the first part — is Murilo's, lived and told by the person who was there. The technical manual , at the end, was written with AI. The split is intentional and marked in the text. Nothing hidden about the seam: part is human, part is machine, and the reader sees both. If you've ever managed or logged into a web server and never set up SSH keys, it's because you don't yet know the real risks of a break-in — and that's okay. Until you find out what can happen. Logging into a server over SSH with a username and password is like locking the front door of a house that faces the street, with nobody keeping watch. Anyone can try as many combinations as they want, freely. And setting this up takes almost as much time as typing a username and password — and it makes getting into the server much faster and easier afterwards. Ignorant of best practices, I managed my servers for a long time by typing: ssh user@server-ip password That nearly cost me dearly, the day I found out my server had been broken into. After that incident, I realized just how vulnerable a username and password are on SSH. Today I can't say I sleep soundly — no system is completely break-in proof — but I sleep a lot better (and honestly, I always slept well, until I started managing servers). Waking up on a fine Sunday morning to do some maintenance on the server, and finding out it was broken into through the front door because you left a combination padlock facing the street — that is not the kind of surprise I'd wish on anyone. I have a degree in Law. I worked for 15 years in the legal field at a public institution, until I decided to venture into the world of programming. And where did I end up? Managing systems at the institution I work for, after spending some time building automations in Python. Managing systems wasn't exactly what I had in mind when I wanted to learn to code and understand the world of programming. But that opportunity ended up tea

GIT AND GITHUB Git is a distributed version control tool that track changes into files or code and we can say it works offline. A version control is system used to track and manage changes to a remote file in git. Git Bash is born again shell or basically a command prompt window which emulates UNIX and LINUX environments. Git hub is a website that stores your Git repositories in the cloud so we can say it exist online.It stores your project's version history online and adds collaboration tools like pull requests, issue tracking, and code review." _A repository _in GitHub is similar to a folder in your local machine so any changes are tracked. How to create a repository in GitHub do a simple README.md(markdown) then commit the file and write a massage inside to describe the changes done,then we will need to download a visual code eg VScode where you are able to access the terminals. Git is used to push changes** to git hub or pull a repo from GitHub** There are 3 states that every files lives in; Working Directory -You have made changes but git has not recorded them yet.(it's still on our machine) Staging Area -You have told Git about the changes. Not saved yet. git add filename thus we git add Repository (.git)-Changes are permanently saved in history. git commit -m "message" * Basic Commands/key terms used * git config allows git to know who you are by using your username and user email e.g. your GitHub account name and email address this is an important info when you want to commit changes as it will tell you who made the changes there are different levels but we will use global Global- applies to all repositories for the current user > Syntax: git config --global user.name or user.email. - mkdir (make directory) name – creates a new directories for example: my_project in your machine. - git init this tells Git to start tracking this folder,it initializes git on the folder,the .git folder is where all the history, settings, and saved snapshots lives. It's hidden s