Attest 1.1.0 发布：测试 AI 智能体的行为而非仅回复文本
Attest 1.1.0 是一个面向 AI 智能体的 Python 和 TypeScript 测试框架，可在 pytest 或 Vitest 中检查回复、工具调用序列和外部状态变化。
Some English translations are not ready yet. The site is still in English. Untranslated news and event text is shown in its original language until background processing completes.
#部署/工程
Hacker News AIAI score5858 Hacker News AIAI score2222 AI 将如何改变 2027 年的工程领导力
作者认为 2026 年让人们初步看到 AI 的强大，2027 年将是打磨直觉、学会何时以及如何最有效应用 AI 的一年。AI 智能体已在多数公司承担大部分实际编码工作，但生产力提升被有所夸大；作者更看好用 AI 做研究与理解问题，并建议先在可运行的人工流程中引入 AI，再逐步替换其中环节。
Hacker News AIAI score2222 Tuning a Linux laptop for AI-assisted development
Jordan Dimov describes using Claude Code to diagnose and fix performance problems on his Ubuntu Dell Precision 3591 laptop, including a 90%-full disk from retained Rust debug builds, 30-second SSD freezes caused by Intel VMD (fixed by switching BIOS storage mode from RAID On to AHCI/NVMe), and a GPU kept awake by a kernel flag. Postgres tuning, parallel test databases, and CPU affinity changes cut the full test run from 283 s to 198 s and commits from 0.58 ms to 0.04 ms. He also covers local models on the 6 GB GPU, a pgvector-based code search tool, and a log-digest script for agents.
Hacker News AIAI score5858 Groma.md: Stay in control of your architecture as AI agents write code
groma.md is a free, MIT-licensed tool that scans a repository into a live C4 architecture map stored as Open Knowledge Format Markdown in Git, updating as files are saved and keeping curation across later scans. It runs locally with no account or backend, uses deterministic code analysis, calls no AI service, and can be used entirely without AI; coding agents can curate the map via the same CLI. Architecture changes are reviewed in the same pull request as the code, and an optional GitHub Action comments on PRs with change counts and a before-and-after map.
Hacker News AIFeaturedAI score7878 GitSeed – Self-hosted control plane for AI coding agents
GitSeed is an Apache-2.0, self-hosted, vendor-neutral control plane that runs coding agents such as Claude Code, Codex, Gemini, and OpenCode inside isolated three-phase containers (prepare → agent → finalize) on your own infrastructure. It adds Ed25519-signed PR Governance Receipts, parallel three-model "Second Opinions," enforced per-run spend caps via LiteLLM, and a five-stage SDLC pipeline (Ideate → Plan → Execute → Review → Merge) or Fast-Track mode, with a CLI/TUI, a 65-tool MCP server, and a Helm chart.
Why it matters: Useful for teams weighing self-hosted agent orchestration against cloud SaaS bots, since it details concrete trust boundaries, audit receipts, and spend controls.
Hacker News AIAI score5858 FlashVMM – jail your AI agent in a MicroVM that boots in 18 ms
FlashVMM is a sandbox for AI agents that gives each agent task its own FlashVM, a custom microVM claimed warm in milliseconds that runs agent-written code behind a hypervisor boundary and is discarded when the task ends. It installs via curl or PowerShell on macOS, Windows and Linux (x86_64 and arm64), works with existing agents through `flash agents setup`, and offers an MCP server with 21 tools plus warm pools, snapshots, forking and a shared GPU pool.
Hacker News AIAI score5858 Show HN: A local-first AI agent for the desktop, no Electron, any LLM endpoint
byob is a native desktop AI agent that connects to any OpenAI-compatible endpoint (llama-server, Ollama, LM Studio, vLLM or hosted APIs) and works in a folder, reading and writing files, running code, and building spreadsheets, charts, documents and small apps beside the chat. It is a single program for Linux, macOS and Windows that draws its own GPU interface without Electron, with a browser UI compiled to WebAssembly, approval cards with diffs, per-turn undo, and a plugin system it can use to write and fix its own plugins. It ships as a roughly 60 MB download (13 MB slim build) and is licensed under MIT OR Apache-2.0.
Hacker News AIAI score3333 OpsJarvis: A self-hosted AI ops desk where humans approve every write
OpsJarvis is a self-hosted AI operations desk demo that triages inbox messages, drafts replies and queues tasks while requiring human approval before every write action. It includes a circuit breaker that pauses HubSpot CRM writes until latency drops below 800 ms, and runs on the user's own keys, .env and source code. The demo workspace uses sample data and sends no real messages.
Hacker News AIAI score5858 Show HN: db-branch – database branches in seconds for AI agents and CI, via ZFS
db-branch is a self-hosted, Apache-2.0 tool that creates copy-on-write branches of MySQL and PostgreSQL databases in seconds using ZFS clones, with nothing copied at creation. It offers a CLI, REST API, web console and MCP server, plus branch reset, TTL expiry, checkpoints and per-branch version upgrades, and is currently in Beta. It targets AI agent sandboxes, a database per pull request in CI, and version-upgrade checks.
Browser Use BlogAI score4242 Log In Once: Reuse Cookies and Profiles Across Browser Use Runs
This tutorial shows how to create an empty Browser Use profile, log in with it attached, stop the browser to save its state, and pass the same profile ID to a second run that reaches a protected page without credentials. It also demonstrates saving and reloading Playwright's storage_state file locally, plus handling expired cookies, keeping per-user profiles separate, and troubleshooting. The demo uses The Internet's public test account, and the author notes that the website still decides how long a login remains valid.
Hacker News AIAI score5858 Europa, the EU's open AI model, is aiming for nearly 1T parameters
Domyn, the Milan-based company leading the EU-backed EUROPA consortium, reported in a November 12, 2026 technical post that it ran training trials for a model with 951.9 billion parameters on the JUPITER supercomputer in Jülich, Germany, more than double the Commission's 400 billion minimum. Using a mixture-of-experts design with 384 experts (8 active per token, 32.8 billion active parameters), Domyn says it raised throughput from 90 to 1,293 tokens per second per GPU through software fixes contributed back to PyTorch and TorchTitan. The post describes only trials, not a completed training run, and EUROPA's weights, license, training data and release date remain unpublished.
Hacker News AIAI score4040 Sandlock: Confining AI Agent Code with Unprivileged Linux Primitives
Sandlock is a lightweight Linux process sandbox that compiles static, input-independent policy into kernel-enforced rules while a narrow supervisor handles runtime-dependent decisions and virtualized effects, enforcing filesystem, network, IPC, and syscall policies without root, cgroups, images, or mandatory namespaces. It supports dynamic network decisions, HTTP-level access control, TOCTOU-safe inspection of execve arguments, and reversible filesystem effects, adding roughly 5 ms of startup overhead and running Redis at bare-metal throughput on the authors' workstation. Its pipeline operator further supports per-stage confinement for separating data, network, and untrusted-content capabilities.
Hacker News AIAI score3434 Show HN: Vantage-V-AI – 2-bit ternary KV-cache with attention sinks
Vantage-V-AI (v8.2.0) is an open-source, bounded-memory inference runtime that integrates with Hugging Face AutoModelForCausalLM via a one-line vantage.enable() call to cap the active attention window and retain sink tokens. It uses a 2-bit packed ternary KV-cache (4 elements per uint8 byte, 7.97x storage reduction over FP16) plus attention sinks, and reports benchmarks on 2x NVIDIA Tesla T4 GPUs with Qwen2.5-3B-Instruct, including 100.0% exact token parity between lossless and FP8 streaming modes and 0.988 average retrieval coherence from 512 to 4,096 tokens. The repository is at github.com/khoadev-vn/VANTAGE-V-AI under the MIT License.
Hacker News AIAI score5858 Show HN: ShadowDiff – Shadow testing and SQL diffing for AI-migrated back ends
ShadowDiff is a network proxy and analysis engine that mirrors live HTTP traffic to both a legacy service and its AI-refactored candidate, then runs a 3-way diff with noise cancellation plus SQL side-effect comparison to isolate genuine regressions. It integrates with Envoy/Istio request mirroring, uses Google Gemini (gemini-2.0-flash) to synthesize regression tests, can open GitHub PRs with fixes, and offers a `shadowdiff gate` CLI for CI/CD quality gating. The project is MIT-licensed and includes Docker Compose and Kubernetes/Istio deployment examples.
Hacker News AIAI score1818 Lawyer upgrades eight-V100 private AI server to run 80% of DeepSeek v4.1 Flash
A lawyer running a private AI server on eight used Nvidia V100 GPUs moved about 80% of its work from 2-bit to 4-bit NVFP4 precision, adding 1TB of RAM and an RTX 3090. The server runs DeepSeek V4.1 Flash with a legal harness for his law practice, and code output dropped from 160–215 to 150–180 tokens per second, though the crashes that hit the new build are gone. The crashes, originally blamed on PCIe, were likely caused by a bad PSU.
Hacker News AIAI score3838 IBM Spyre AI Accelerator Card (75W PCIe Card with 128GB of LPDDR5)
The IBM Spyre Accelerator is a 75W PCIe card for IBM Z and LinuxONE systems with 32 AI accelerator cores built on Samsung's 5nm process, up to 128GB of LPDDR5 memory per card, and over 300 TOPS per card, with ensembles of up to eight cards reaching 1TB of memory. It complements the on-chip Telum-II AI accelerator for multi-model inferencing such as real-time fraud detection and code assistance, and supports on-premises generative AI use cases like Watsonx Code Assistant. IBM also describes the Appliance Control Center for deploying software appliances and the Spyre Support Appliance for managing, monitoring and diagnosing Spyre cards attached to LPARs.
Simon WillisonAI score2222 datasette-llm 0.1a11
datasette-llm is an intermediate Datasette plugin for configuring LLMs used by other Datasette plugins. Version 0.1a11 adds support for OpenAI-compatible endpoints, allowing Datasette to use an LLM hosted anywhere that speaks the OpenAI pseudo-standard web protocol, including local models run with LM Studio, Ollama or llama.cpp (which just hit v1.0.0). The post shows running Datasette via uvx with a default_model configuration pointing at a local LM Studio endpoint to get a Datasette Agent instance powered by a local model.
Hacker News AIAI score5858 Show HN: Smart Node, a self-hosted crew of AI agents for a small business
Smart Node is a self-hosted AI node that runs on one machine you control, bundling an AI assistant with task tracking, a document store, semantic memory, a model gateway, an operator panel and browser voice, while reasoning still uses a cloud model via Anthropic, Google or OpenRouter. Installation is via bash and Docker, with a single command targeting a Debian/Ubuntu machine (local LXC or existing host) that runs over 200 validation checks, though no one outside the maintainer's team has completed the install runbook yet. Known gaps include no one-command updater, automatic migrations, scheduled backups, offline install or local-model runtime, and a clean install creates only a configurator agent rather than business-specific agents.
Hacker News AIAI score3434 Hacked 2013 Recon Jet sports glasses into a Linux camera/voice AI assistant
A developer revived a barely-used 2013 Recon Jet (TI OMAP4430, 1 GiB RAM, 8 GB eMMC, 428x240 Kopin microdisplay) by replacing its dead battery, then used Claude Code to root it, port mainline Linux 6.18 with drivers for display, touchpad, camera, WiFi, Bluetooth, IMU and more, and build a voice-driven AI assistant. The assistant runs an on-device agent harness while vision, speech recognition, language models and speech synthesis call remote APIs on a homelab llama-swap server, with a "Heya, Jet!" wake word and a cranky-NPC persona using a cloned voice. The author notes the AI's world model lacked spatial coherence and that Claude excelled at low-level reverse engineering but failed at architectural hygiene.
Hacker News AIAI score4848 Show HN: AI Time Estimator – how long Claude Code will take, from your own logs
AI Time Estimator is a Claude Code plugin that has a subagent count the operations a task needs and converts those counts into P10/P50/P90 minutes via Monte Carlo simulation, using rates measured from the user's own Claude Code transcripts. A Stop hook scores each estimate against what actually happened, and after 10 scored estimates the ranges self-correct; everything runs locally with no network calls. It requires Python 3.9+ and installs as a plugin or via install.py, which adds an agent, an /estimate-here command, a hook and a CLAUDE.md section.
Hacker News AIAI score3434 The Missing Ring-0 Protections for AI Agents
This essay argues that coding agents such as Claude Code, Codex CLI and Gemini CLI act with the developer's authority while access checks sit in individual tools, so a write refused by the file tool can go through the shell, and it proposes a deterministic commit-boundary supervisor that stages file, git, network and credential effects before policy approval. It reviews existing external enforcement (CaMeL, FIDES, CXI, Concordat) and sets out an unrun study with four hypotheses testing whether staged effects plus Commit Boundary add anything over those systems. Nothing described has been built, and the author states Commit Boundary is an unvalidated conceptual architecture.
Hacker News AIAI score5858 Cowork – Self-hosted task queue and dashboard for Claude, Codex, and AI agents
Cowork is an open-source control plane that connects coding agents such as Claude Code, Hermes, Gemini and Ollama into one task inbox, with fallback chains that hand work to the next agent when an API hits its rate limit. It offers a dark single-pane dashboard with host telemetry, a zero-database design storing state as plain JSON and Markdown, one-command remote worker deployment, and 285+ built-in specialist personas across 19 departments. It runs on Node.js ≥ 20 and ships as a server install script, a skill installer for agent clients, or a Docker image exposing an MCP endpoint at /mcp.
Hacker News AIAI score5858 China builds 800MW AI data center campus to rival US Stargate project
According to a SemiAnalysis report, China is building an 800MW AI data center campus in Inner Mongolia that could challenge the economics of America's Stargate project. Envision built the first 120 MW of its Galaxy AI data center campus in under a year, at a cost researchers put at one-fifth of comparable US projects, after visiting the site in September. To secure the large amounts of electricity AI data centers need, Envision is building its own wind farms and battery storage alongside the computing facility.
Hacker News AIAI score4040 Show HN: SunsetLint – find retiring AI model IDs in your GitHub repos
SunsetLint is a GitHub App that scans repositories for AI model identifiers with announced shutdown dates and reports each match with its file, line, shutdown date and the provider's recommended replacement. It installs with read-only contents access, posts one rolling issue per repository, and offers a Pro pull-request Check Run that flags newly added references to retiring identifiers. It does not edit code or open pull requests, does not guarantee complete coverage, and does not support GitLab or Bitbucket.
Hacker News AIAI score4848 Engineering Fought Productivity Metrics. AI Ended the Fight
Rising, variable AI token costs (Uber exhausting its 2026 budget in four months, Salesforce's reported ~$300M/year Anthropic bill) have ended engineering's exemption from productivity measurement, but DORA and SPACE show no single metric works, and 2025–2026 data finds AI raises throughput while also raising bugs, incidents and review time. The article argues for attributing AI activity to shipped work and outcomes, using cost per merged pull request, autonomy rate and token yield rate instead of token consumption, and treating committed AI configuration files as maintained code.
Hacker News AIAI score3838 Why platform teams fail (and AI makes it worse)
A practitioner's essay argues that platform teams fail by sliding into permanent firefighting, conflating golden paths (products that earn adoption) with guardrails (mandatory policy), and shipping without cross-functional buy-in. Generative AI worsens this because developers can scaffold entire services from a prompt without ever discovering the platform team's hardened templates, so the author argues the catalog must become machine-readable for coding agents and guardrails must cover everything that ships. The piece also covers the infrastructure-versus-product skills gap and when an organisation should not yet form a platform team.
Hacker News AIAI score5858 Build Your Own AI Coding Agent Harness
This course walks through building TeensyCode, a TypeScript AI coding agent harness with a tool loop (read, grep, write, edit, bash, task, askUser), execute-level safety gates, structured system prompts, and a swappable Sandbox interface with local and in-memory backends. It also covers context management via pruneMessages and bounded tool output, subagent delegation, human-in-the-loop questions, sandbox lifecycle, and extensibility, across 11 modules plus a capstone. Prerequisites include TypeScript, an AI_GATEWAY_API_KEY, and Node.js 22+ or Bun.
Hacker News AIAI score5858 Show HN: Gamma PDF – AI-native PDF reader, self-hosted
Gamma PDF is an open-source, self-hosted PDF annotator with outliner notes, live Markdown and LaTeX, and an AI that answers questions about papers with cited answers using your own Anthropic or OpenAI key or a paid ChatGPT plan. It runs on Windows, macOS, Linux, iPad, Android and Docker, stores data as SQLite and files in one folder with no account or telemetry, and supports shared workspaces, Zotero/Obsidian/Logseq/Notion imports, and read-only MCP access for Codex and Claude Code.
Hacker News AIAI score3434 Transforming Ripgrep's Documentation with AI Automation and MkDocs
The author describes a third-generation documentation automation workflow that expanded ripgrep's two documentation files into 55 pages, using automatic page splitting, per-page visual enhancement, and Mermaid diagram validation via a new tool called mermaid-sonar. The workflow runs on MkDocs Material and adds 87 Mermaid diagrams, 358 admonitions, and 180+ code annotations, with a strict build passing. The post also compares the workflow to earlier mdBook-based versions and outlines steps for adapting it to other projects.
Hacker News AIFeaturedAI score7878 Anthropic upgrades Claude Code Projects to run parallel AI workers
Anthropic has updated Claude Code Projects so that users describe a goal and Claude scopes the request, delegates work to parallel threads, coordinates them, reviews outputs and assembles the result, with progress steerable from a phone. Each thread is a Claude Code cloud session on its own branch and repo copy, and threads now share a project memory and a library of added files and artifacts. The beta is available to select Claude Pro and Max subscribers using cloud sessions who have no existing web or desktop projects, with broader access over the coming week and Team and Enterprise plans after that.
Why it matters: It details a concrete shift in how Claude Code Projects delegate and coordinate multi-thread agentic work, plus the staged rollout limits.
Hacker News AIAI score3636 Sick of Herding AI Cats?
Clowder is a native macOS status board that shows each running Claude session as a coloured tile — red when blocked and needing you, green when done, blue while working — so you can see at a glance which session needs attention. It reads five local sources (Claude Code, Code, Cowork, Design and chats) with no account or telemetry, and clicking a tile jumps to the exact session. It is free for four tiles, or £24 a year for unlimited tiles, and requires macOS 14 or later on Apple silicon.
Hacker News AIAI score3838 ChipAgents – The agentic foundation for semiconductor design (AI for IC design)
ChipAgents presents itself as an agentic foundation for semiconductor design, offering full-flow autonomy for digital and analog chip design and a path to tapeout backed by Meet IC-star. The page cites 200+ semiconductor enterprise deployments, 50% design cycle reduction, a 20,000+ global user community and 40% annual LLM spending savings, and includes a founder story video with Founder & CEO William Wang. It also states enterprise-grade security, including SOC 2 Type II, with sensitive design IP kept in place.
ElevenLabs BlogAI score2222 Open source TTS vs paid APIs: How to choose wisely
This article compares open source and paid text-to-speech APIs, noting that both can produce high-quality speech and that the choice depends on team capacity and workload scale. It explains that self-hosting offers control and data sovereignty at the cost of infrastructure and engineering overhead, while managed APIs provide turnkey, scalable access with per-character pricing. It also distinguishes open weight from open source models, covers licensing considerations, and offers criteria for evaluating voice quality, latency, reliability, and total cost of ownership.
Hacker News AIAI score4242 Doc-foundry – AI writes Security detection rules, then proves they work
Doc-foundry is an adversarial multi-agent detection-engineering workflow (LangGraph, Pydantic, FastMCP, Claude) that turns a threat such as a CVE, CISA KEV entry, incident writeup or ATT&CK technique into a validated first-version Sigma detection rule, a threat report with round-by-round convergence history, and a proof bundle containing a positive corpus, benign baseline, adversarial mutation set and a computed coverage label (real-coverage/brittle/broken) with FP/FN numbers and a reproducibility hash. Four LLM agents (Threat Analyst, Blue Detection Engineer, Telemetry, Red-Team Adversary) propose, while a deterministic referee handles FP/FN evaluation, adversarial evasion, the coverage verdict, red-team verification and the deployment gate, so the same input yields the same verdict and hash. A second pipeline, the Review Chain, runs draft → Red Team → revision → SOC Lead → final rule and produces a reviewed but explicitly unvalidated rule with the full trail of findings that shaped it.
AWS Machine Learning BlogAI score2424 Production MLOps at GS Caltex with Amazon SageMaker Canvas, part 1
GS Caltex connects models built by process engineers in Amazon SageMaker Canvas to production workflows in Amazon SageMaker Pipelines, with ML engineers handling deployment and scheduled inference. The company automated the handoff so ML engineers now deploy models within a few hours instead of about a week, and it runs batch inference every 30 minutes, which it reports has meaningfully lowered inference costs. The post details the reference architecture, including reusing Data Wrangler preprocessing flows, registering and approving models in Model Registry, and scheduling runs with Amazon EventBridge.
Hacker News AIAI score6464 FAA tees up $875M AI tool to help manage air traffic congestion
The FAA plans to launch its AI-based SMART system to advise air traffic controllers in the congested airspace over Washington, DC, as soon as September 21, as a first step toward a nationwide rollout across the 29 million square miles of US national airspace. SMART uses AI models to predict air traffic flows and identify potential conflicts based on airline schedules, weather, airport capacity, and airspace conditions, aiming to cut fuel burn, improve on-time performance, and speed recovery from congestion. The tool is part of an $875 million, 12-year contract awarded to Boston-based Air Space Intelligence in June, which also covers a Flow Management Data and Services system to replace the current system at the FAA's Air Traffic Control System Command Center in Virginia.
Ars Technica · AIAI score6464 FAA expands AI advice to manage air traffic congestion across entire US airspace
The FAA has expanded its $875 million SMART AI system, which advises air traffic controllers, from Washington, DC airspace to all 22 centers nationwide after it helped reroute 180 flights and saved an average of 10 minutes and nearly 80 gallons of fuel per flight since September 21. The cloud-based tool combines weather, flight path, traffic flow and controller staffing data to recommend schedules, routes and arrival order, though an FAA team in DC validates its recommendations before they reach controllers. Airline executives support the rollout, but the union representing air traffic controllers says it was not involved in designing or testing the system.
AWS Machine Learning BlogAI score3636 OpenFold3 on Amazon SageMaker JumpStart: Predict biomolecular structures
Amazon SageMaker JumpStart now lets you deploy OpenFold3-preview, an open source all-atom biomolecular structure prediction model from the OpenFold Consortium and Columbia University's AlQuraishi Lab, to predict 3D structures of proteins, DNA, RNA, small-molecule ligands and their complexes. The post walks through deploying it on an asynchronous inference endpoint, submitting prediction jobs, interpreting mmCIF outputs with pLDDT, PAE and PDE confidence metrics, running parallel predictions on multi-GPU instances, supplying your own MSAs, and scaling the endpoint to zero when idle.
AWS Machine Learning BlogAI score3434 Architecture patterns for fast session startup on Amazon Bedrock AgentCore runtime
Amazon Bedrock AgentCore runtime's platformVersion "V2" snapshots a prepared container once at create or update time and restores it for each new session, making session starts faster and more consistent. The post explains how V2 starts sessions, how to measure session start time with CloudWatch metrics like OverheadLatency, and four patterns: moving initialization before app.run(), using build hooks for files the process can't hold, and pre-warming sessions on user engagement for sub-second starts.
AWS Machine Learning BlogAI score3838 Route coding agent requests to the right model on Amazon Bedrock with NVIDIA NeMo Switchyard
This AWS sample deploys LiteLLM with the open source NVIDIA NeMo Switchyard routing plugin on Amazon ECS Fargate via AWS CDK, so coding agents can send requests to a model group that Switchyard routes between a capable and an efficient model on Amazon Bedrock. It covers Stage routing (which reads recent tool activity and results to pick between two candidates) and Random routing (weighted traffic splits for A/B tests), plus a benchmark runner that measures cost per solved task against single-model baselines. The post notes the sample uses switchyard-server 0.3.0 and LiteLLM v1.102.0, is for non-production use, and that its configuration and APIs may change between releases.