GitHub热门项目 | A Python library for building AI agents that leverage the full power of Google Antigravity. | Stars: 2,250 | 21 stars today | 语言: Python

When a bug is fixed, most teams retest the exact failure path once and move on. That is understandable, but it leaves a gap: the team learned something from a real failure, then failed to turn that learning into reusable regression coverage. Here is a lightweight template I use for turning resolved bugs into regression test cases that can be copied into a spreadsheet, Jira, TestRail, Qase, Xray, Zephyr, or any other QA workflow. The CSV fields For a bug fix regression test, I like these columns: Test ID Bug ID Feature Area Regression Scenario Original Failure Preconditions Test Data Steps Expected Result Negative Check Priority Regression Risk Test Type Automation Candidate Notes This is enough structure to make the test reusable without turning every bug fix into a heavyweight test plan. Example bug Bug ID: BUG-1842 Bug title: Non-admin users could resend workspace invitations. Original failure: A workspace member could open Pending Invitations and click Resend, even though only owners and admins should be allowed to resend invitation emails. Fix summary: The resend invitation action now checks the user's workspace role before sending the email. Example regression test case Test ID: REG-BUG-1842-001 Feature Area: Workspace invitations Regression Scenario: Workspace member cannot resend a pending invitation. Preconditions: Workspace has at least one pending invitation. Test user is a workspace member, not an owner or admin. User is logged in. Steps: Log in as the workspace member. Open Workspace Settings. Go to Pending Invitations. Locate the pending invitation. Check whether the Resend action is visible or available. If the action can be triggered through the API, attempt the resend request. Expected Result: The member cannot resend the pending invitation. The UI hides or disables the action, and the API rejects unauthorized resend attempts. Negative Check: Confirm that an owner or admin can still resend the invitation if product rules allow it. Priority: High Regr

With a Real CI Automation Example Loop Engineering is suddenly everywhere, and honestly, I wanted to understand it properly instead of just repeating the buzzword. The simplest way I can explain Loop Engineering is this: it replaces me as the person constantly prompting the agent. Instead of me manually noticing a problem, deciding what it means, writing the next prompt, and pushing the process forward, I design a system that keeps moving on its own until it reaches the outcome I want. That is the whole point of Loop Engineering. I stop acting like the operator and start acting like the system designer. To make that idea concrete, I built a practical software engineering workflow around CI failures. Whenever a GitHub Actions CI run fails, the system automatically classifies the failure, creates a Jira bug for real issues, sends a Slack notification, and records the outcome so it does not process the same failure twice. What Loop Engineering actually means Early AI workflows were mostly linear. I would give a prompt, the model would return an answer, and if the answer was incomplete or wrong, I would jump back in and prompt again. That worked, but it kept me trapped inside the process. Loop Engineering changes that dynamic. I am no longer the person babysitting each step. I build an autonomous loop that can observe, decide, act, and persist state. The system keeps iterating until the task is done, without needing me to micromanage it. That distinction matters. In a normal prompt based workflow, the human is still the glue. In Loop Engineering, the human creates the machine, and the machine runs the loop. The five building blocks of Loop Engineering When I break down Loop Engineering, I think of it as five core building blocks working together. 1. Automations These are the event driven triggers that start the whole system. They are the heartbeat of the loop. Something happens, and the automation fires. Without this, nothing starts. 2. Skills Skills give the agent stru

In my previous articles, I’ve consistently emphasized a core architectural principle: once the render layer no longer dictates the entire data flow, the boundaries between State, Derived State, and Effects become critical. When we fall into the habit of stuffing every UI-affecting variable into generic "state," the system quickly loses its semantic structure. In modern frontend applications, this architectural gap becomes most glaring when dealing with asynchronous work. Async data is never merely "a value that will appear in the future." It carries complex semantics regarding its source, temporal validity, cancellation, error recovery, and invalidation. If these semantics aren't modeled explicitly, they inevitably get pushed down into the UI framework’s lifecycle—indirectly patched together through component mounts, effect dependencies, and callback guards. This brings us to the core question of this article: What does a system lose when the correctness of async work is forced to depend on the UI lifecycle? We are all incredibly familiar with this pattern: const data = await fetchSomething () setState ( data ) Or, using a standard UI framework hook: useEffect (() => { let cancelled = false fetchSomething (). then ( result => { if ( ! cancelled ) { setData ( result ) } }) return () => { cancelled = true } }, []) There is nothing inherently wrong with this code for simple use cases. It’s intuitive and perfectly aligns with how Promises are designed to work: trigger the operation, wait for the resolution, and write the result back into state. However, this mental model has a subtle downside. It encourages us to think of async work as simply calling setState after a Promise resolves. That may hold up for simple screens, but as an application grows, the model starts to expose structural problems. Promise Only Describes Completion, Not Ownership A Promise solves a very specific problem: A piece of work will complete in the future, and it will either succeed or fail. It c

Cursor has spent the past week in headlines after confirming a partnership with SpaceX that could eventually lead to a $60 billion acquisition . The deal, for now, centres on training more capable coding models using SpaceX’s compute infrastructure. Alongside that push on model performance, however, Cursor is now addressing a separate issue: the reliability of the code those models produce. Cursor has partnered with Chainguard , which provides verified open-source packages, to route dependencies through its curated repositories, aiming to reduce the risk of compromised components entering AI-built applications. The announcement lands as AI coding tools push more software into production with less human review, raising questions about how much of that code can be trusted. Supply chain risks in the agentic era The partnership addresses a problem developers know all too well. Modern applications depend heavily on open-source libraries and container images, most of which are pulled from public registries such as npm, PyPI, and Docker Hub. Those registries operate on openness, with limited checks in place. Developers — and now AI agents — often install dependencies without knowing who built them or whether they have been tampered with. Recent incidents have underlined the risk . In March, projects such as Trivy, LiteLLM, Telnyx, and Axios were compromised, with attackers using poisoned packages to steal credentials and spread malware. For teams using AI-generated code, the exposure increases. Agents can select and install dependencies automatically, making trust decisions at a pace that outstrips manual review. As Chainguard co-founder and CEO Dan Lorenc put it, generating code is becoming routine — checking its integrity is where the pressure now sits. “AI agents are making dependency decisions at a scale and speed no security team can manually review,” he wrote in a blog post . “As organizations adopt agentic development, the biggest blocker is no longer how fast code

A few months ago I didn't know much about coding. Today I have a full stack marketplace running with a real API, a live database, user authentication, image uploads, messaging and a React frontend. The Idea Kisumu has a huge secondhand clothes market. Mitumba is everywhere — Kibuye market, roadside stalls, WhatsApp groups. But there is no dedicated digital platform for it. If you want to sell a jacket in Kondele you have no easy way to reach buyers in your area. If you want to find size L shoes in CBD you have to physically go and look. I wanted to build something that solved a real local problem. Not another todo app. Not another weather app. Something that could actually help people in my city. That is how Kisumu Marketplace was born. The Tech Stack I Chose I built the backend in Go using the Gin framework. The database is PostgreSQL hosted on Neon.tech — a free cloud database that saved me more than once when my laptop broke. Authentication uses JWT tokens and bcrypt for password hashing. Images are uploaded to Cloudinary. The frontend is React with Tailwind CSS. I chose Go because Zone 01 teaches it and I wanted to go deep on one language rather than shallow on many. I chose PostgreSQL because it is the industry standard and learning it properly matters. I chose React because it is the most in-demand frontend framework and I wanted to build something real with it. What I Learned I learned Go from scratch while building this. I learned React from scratch while building this. I learned PostgreSQL, JWT, bcrypt, Cloudinary, Tailwind, axios, React Router and more — all by needing them for this project. The most valuable thing I learned is that you understand something properly only when you build with it. Reading about JWT is nothing like debugging a 401 Unauthorized error at midnight. I also learned that documentation is a skill. Writing this article, explaining how things work, is making me understand my own project better.

📦 Code: github.com/USER/hls-multi-audio - replace before publishing TL;DR We'll add a working language picker to an HLS player. The hard part isn't the dropdown, it's the manifest. We'll author alternate audio with EXT-X-MEDIA audio groups, package it correctly, debug the classic "zero audio tracks" bug, and wire a switcher on hls.js v1.7 . Adaptive video, captions, the whole pipeline already works. Now someone wants an English/Spanish audio toggle. In HLS, "which audio can the viewer pick" is decided at packaging time and written into the master playlist. The player just displays it. Let's build it in that order. 1. Understand the structure (audio groups) HLS decouples video variants from audio renditions: Each audio rendition is an #EXT-X-MEDIA:TYPE=AUDIO entry pointing at its own media playlist. Renditions are bundled into a named audio group via GROUP-ID . Each video variant ( #EXT-X-STREAM-INF ) references a group with AUDIO="..." . A correct master playlist: #EXTM3U #EXT-X-VERSION:6 #EXT-X-MEDIA:TYPE=AUDIO,GROUP-ID="aud",NAME="English",LANGUAGE="en",DEFAULT=YES,AUTOSELECT=YES,CHANNELS="2",URI="audio/en.m3u8" #EXT-X-MEDIA:TYPE=AUDIO,GROUP-ID="aud",NAME="Espanol",LANGUAGE="es",DEFAULT=NO,AUTOSELECT=YES,CHANNELS="2",URI="audio/es.m3u8" #EXT-X-STREAM-INF:BANDWIDTH=2128000,CODECS="avc1.640028,mp4a.40.2",AUDIO="aud" video/720p.m3u8 #EXT-X-STREAM-INF:BANDWIDTH=1128000,CODECS="avc1.640020,mp4a.40.2",AUDIO="aud" video/480p.m3u8 Every attribute earns its place: LANGUAGE - BCP-47 code, used for the label. DEFAULT - plays when the viewer has no preference. AUTOSELECT - may be auto-picked from the OS language. CHANNELS - needed so the player can reason about stereo vs surround. BANDWIDTH on each video variant must include the audio group's bitrate , or your ABR logic works from a wrong total. 2. Author the renditions with FFmpeg Extract/encode each language's audio, then package. First, encode video-only and audio-only renditions: # video only (no audio), two ladder rungs

📦 Code: github.com/USER/nvenc-vs-cpu-bench - replace before publishing TL;DR A GPU encodes faster than a CPU, but "faster" and "cheaper" are different claims. We'll build a small FFmpeg + VMAF harness that times software (libx264/SVT-AV1) against hardware (h264_nvenc/av1_nvenc), then plug the results into a dollars-per-encoded-minute formula so you find your break-even instead of trusting a benchmark blog. We're using FFmpeg 7.1.x (current stable line) and an NVIDIA GPU with NVENC. Same approach works for Intel QSV ( *_qsv ) and AMD AMF ( *_amf ) if you swap the encoder names. Why this isn't obvious NVENC is a fixed-function hardware block, not "the GPU doing x264 in parallel." It's extremely fast and barely touches the CPU, but it exposes fewer rate-control knobs and gives up a little compression efficiency versus a slow software preset. The gap has narrowed a lot, but it's still there at the quality-obsessed end. So the decision is per-job, and it comes down to one number: dollars per encoded minute = (instance $/hr) ÷ (minutes encoded/hr) . GPU instances cost more per hour but encode many streams in parallel, so the answer depends on whether you can keep the encoder saturated. Let's measure instead of argue. 1. Set up the encoders Three contenders. One representative source file (use real footage, not a synthetic clip). # software H.264, quality-leaning preset ffmpeg -y -i source.mp4 -c :v libx264 -preset slow -crf 21 -an out_cpu.mp4 # NVENC H.264, quality-tuned ffmpeg -y -hwaccel cuda -i source.mp4 -c :v h264_nvenc -preset p6 -tune hq \ -rc vbr -cq 23 -an out_gpu.mp4 # AV1: software (SVT-AV1) vs hardware (needs Ada / RTX 40+) ffmpeg -y -i source.mp4 -c :v libsvtav1 -preset 6 -crf 30 -an out_svtav1.mp4 ffmpeg -y -hwaccel cuda -i source.mp4 -c :v av1_nvenc -preset p5 -cq 30 -an out_av1nvenc.mp4 💡 Tip: -preset p1 (fastest) through -preset p7 (slowest/highest quality) for NVENC. p6 / p7 is where it competes on quality; p1 - p3 is where it competes on raw throughput.

📦 Code: github.com/USER/video-api-bench - replace before publishing TL;DR The per-minute delivery rate is the easiest number to compare and the least useful. The real cost lives in encoding, analytics, and the player. This post compares Mux, Cloudflare Stream, api.video, FastPix, and AWS on what each includes by default, then gives you a tiny script to benchmark upload and time-to-ready on your own files so you stop trusting marketing pages. I have shipped video on four managed APIs across three jobs, and every single time the invoice surprised someone. Not because the delivery rate was wrong, but because encoding, analytics, and the player turned out to be separate line items on some platforms and free on others. Let's compare the parts that don't show up in the headline number. ⚠️ Note: pricing pages move. Everything here was checked in June 2026; verify the links before quoting numbers. 1. Encoding: free or metered? This is the widest spread in the whole comparison. Platform Encoding Delivery Storage Cloudflare Stream Free $1 / 1,000 min delivered $5 / 1,000 min stored api.video Free (unlimited) $0.0017 / min $0.00285 / min FastPix Free on standard plan ~$0.00096 / min @1080p Per-minute, tiered Mux Metered per minute Per minute Per minute AWS (DIY) Per minute (MediaConvert) Per GB (CloudFront) Per GB (S3) If your catalog is upload-heavy (lots of assets encoded once, watched rarely), metered encoding is not a rounding error. It can flip which platform is cheapest, even when the delivery rates look identical. 2. Analytics: included or a $499 floor? QoE analytics is the feature teams forget to price until playback breaks in production. Platform QoE analytics Entry cost FastPix (Video Data) Session-level, 50+ signals/session Free up to 100K views/month Mux (Mux Data) Mature, broad device SDKs $499/month (Media plan, 1M views, +$0.50/1K) Cloudflare Stream Basic Included, limited depth api.video Available Usage-based AWS Build it yourself (CloudWatch + logs) Engineerin

TL;DR If your product lets strangers upload video, you need moderation before launch, not after the first bad upload. We will build a small-team pipeline: extract frames with FFmpeg, score them with NudeNet (ONNX Runtime, CPU-friendly), route uploads into approve / human-review / block by confidence, and log every decision. No trust-and-safety department required. 📦 Code: github.com/USER/ugc-moderation, replace before publishing ⚠️ Note: this is a sensitive area. The goal here is the engineering shape (sampling, scoring, routing, auditing), not detection of any specific content. Keep test fixtures clean and lawful. A model does not decide what is allowed. It produces a score. You decide where the lines go. The whole design is about routing scores sensibly and sending the uncertain middle to a human. The architecture 🧠 upload ──> extract sample frames (ffmpeg) ──> score frames (NudeNet / ONNX) ──> aggregate to one confidence ──> route: high-confidence clean -> auto-approve uncertain middle band -> human review queue high-confidence violation -> auto-block ──> write an audit record for every decision The economics only work if the middle band is small. A decent model makes most uploads obviously fine or obviously not, so a human only ever sees the genuinely ambiguous slice. 1. Sample frames, do not score every frame You cannot afford every frame and you do not need it. Pull one frame per second (or scene-change keyframes) with FFmpeg. # extract 1 frame per second into ./frames mkdir -p frames ffmpeg -i upload.mp4 -vf "fps=1" -q :v 3 frames/frame_%05d.jpg Prefer scene changes to catch more variety with fewer frames: # keyframes where the scene actually changes ffmpeg -i upload.mp4 -vf "select='gt(scene,0.3)',showinfo" -vsync vfr frames/scene_%05d.jpg 💡 Tip: a 30-minute upload at 1 fps is ~1,800 frames. Scene-change sampling often cuts that by an order of magnitude with little loss for moderation purposes. 2. Score frames with NudeNet NudeNet runs on ONNX Runtime on pla

TL;DR Converting HDR10 to SDR with a naive FFmpeg command gives you grey, washed-out video. The fix is tone mapping. We will detect HDR with ffprobe , run two working tone-map chains ( zscale on CPU, libplacebo on GPU) in FFmpeg 8.0, compare operators, and batch it. Test the commands on your own build before shipping. 📦 Code: github.com/USER/hdr-to-sdr, replace before publishing If you have ever run an HDR clip through your normal pipeline and gotten back something flat and foggy, this post is for you. The bug is that HDR and SDR are different color systems, and "just converting" reinterprets one as the other. We will use FFmpeg 8.0 "Huffman" (8.0.2 is current as of May 2026). Why naive conversion fails HDR10 SDR Transfer function PQ (SMPTE ST 2084) gamma 2.4 / BT.1886 Color primaries Rec.2020 (wide) BT.709 (narrow) Peak luminance ~1,000 to 4,000 nits ~100 nits A command that ends in -pix_fmt yuv420p with no tone mapping reads PQ-encoded, Rec.2020 values as if they were SDR. The gamut gets crushed with no intelligence and the brightness curve is misread. Hence the fog. 1. Detect whether a file is even HDR 🔍 Do not tone-map SDR files. Check first: # detect transfer characteristics and primaries ffprobe -v error -select_streams v:0 \ -show_entries stream = color_transfer,color_primaries,color_space \ -of default = noprint_wrappers = 1 input.mkv HDR10 content reports something like: color_space = bt2020nc color_transfer = smpte2084 color_primaries = bt2020 If color_transfer is smpte2084 (PQ) or arib-std-b67 (HLG), you have HDR and you need to tone-map. If it says bt709 , leave it alone. 2. The libplacebo path (GPU, my default) 🚀 libplacebo is the Vulkan-accelerated filter in FFmpeg 8.0. It follows the ITU tone-mapping recommendations and handles the color conversions internally, so the command is short: ffmpeg -i input.mkv \ -vf "libplacebo=tonemapping=bt.2390:colorspace=bt709:color_primaries=bt709:color_trc=bt709:format=yuv420p" \ -c :v libx264 -crf 20 -c :a copy \ ou

"It works on my test input" is the most dangerous sentence in on-device AI development. I typed that sentence - or some version of it - a dozen times while building Redacto, our on-device PII redaction app running Gemma 4 E2B on a Samsung Galaxy S25 Ultra. The model would redact a patient name from a clinical note, I would nod, and I would move on. Then I would hand the phone to a teammate, they would type a police report, and the model would redact the suspect description instead of the victim name. The problem is not the model. The problem is that manual spot-checking is not validation. You are testing a single input against your own expectations, with all the confirmation bias that entails. When you have five domain modes (HIPAA, Financial, Tactical, Journalism, Field Service), three difficulty levels, and two candidate models, you need something systematic. You need a benchmark suite. This post covers how I built one - from dataset curation to scoring methodology to on-device infrastructure - for a hackathon app running entirely on a phone. No cloud. No API calls. No data leaving the device. Why Not Use an Existing Framework? The LLM evaluation space has mature tools. EleutherAI's lm-eval-harness is the community standard for evaluating language models against academic benchmarks like MMLU, HellaSwag, and ARC. Stanford's HELM (Holistic Evaluation of Language Models) provides a multi-metric evaluation framework with standardized scenarios. Google's BIG-bench offers hundreds of tasks for probing specific capabilities. These frameworks are excellent for what they do. They are also completely wrong for this problem, for three reasons. First, they assume server-side inference. lm-eval-harness expects to call a model through an API or load it in PyTorch on a GPU server. Redacto's model runs on a Qualcomm Hexagon NPU inside a phone. There is no Python runtime, no HuggingFace tokenizer at evaluation time, no way to hook into the framework's inference loop. Second, their

Plenty of SEO audits end with a single tool report. You run Lighthouse, screenshot Search Console coverage, save a "12 issues found" panel, and call it done. The trouble is that most audits finished that way silently revert within three months. Someone publishes a new post, refactors a component, swaps a font, and the issue quietly comes back. Nobody notices. Over the last five days I actually audited my four-language blog (ko/ja/en/zh, 298 posts per language). Five items, all fixed. But what I really want to talk about isn't what I fixed. It's that the five fixes mattered less than the build gates that keep them from ever returning. An audit should be a loop, not an event. Why a one-report audit always comes back Most technical SEO issues aren't "the code is wrong." They're "an invariant was never enforced anywhere." Take a clear rule: a published post must not link internally to a draft. Obvious enough. But if a human has to remember that every time, then the moment a recommendation generator pulls in one draft slug, a 404 is born. The report catches that 404 and shows it to you, but it does nothing to prevent the next one. So I ran the audit as a three-step loop. Measure. Fix the biggest item first. Then turn that item into a checker and nail it to the build . Skip the third step and the first two become a chore you repeat every six months. Once a gate is in place, the same class of problem makes npm run build fail. A pipeline enforces the rule, not human memory. This isn't a new invention. It's the same logic by which tests prevent bug regressions, applied to the content and markup layer. It's just oddly rare in SEO, where most teams leave "SEO checks" as a quarterly manual task. The five items I actually ran over five days Measurement first. Each item got a before/after in numbers, not a vibe that "things feel better" but reproducible figures. (The raw log of all five lives on the improvement history page too.) Date Item Before After Gate 07-02 relatedPosts int

While other humanoid startups chase sky-high valuations, Agility Robotics is betting its future on execution — and a SPAC.