#编码

Show HN: Cosh – An open-source AI agent orchestrator

Cosh is a general-purpose coding agent for the terminal with software engineering and computer tools that let it interact with graphical applications. It offers a multi-agent mode (/agents) for running external agents such as Gemini, Claude, Codex and Cline alongside the main agent in split panes, and reads graphical interfaces via an accessibility tree, falling back to screenshots only when needed. It is installed via a curl/bash or PowerShell script, is tested on Linux and Windows (macOS untested), and is released under Apache-2.0.

AI usage generates a fragile codebase

A developer who says he was laid off from Meta in May argues that heavy reliance on AI coding tools produces brittle code, citing shallow tests, overly complex code, declining human understanding of the codebase, and review speed that outpaces quality control. He lists signs of a fragile AI-generated codebase, including rising incidents, slower incident recovery, errors visible only to users, and a degraded user experience, and concludes that high-quality SaaS products will keep a human touch.

Quoting Prompting Claude Fable 5.2

Anthropic documentation notes that Claude Fable 5.2 responds closely to what a prompt says about who owns the code it works on. In Anthropic's testing on small coding tasks, the model tended to add to existing structure when a prompt said nothing about ownership or presented the codebase as someone else's, even when that repeated existing code; telling the model the codebase is its own led it to merge repeated code in about half of tries, versus under a tenth without that sentence.

Locating the source of a mysterious beep with Claude

Simon Willison describes using Claude Code to build "Beep Locator," a Python web app that connects multiple devices as microphone listeners over WebSockets to triangulate an intermittent beep in a box room, drawing on the 2007 BeepBeep acoustic ranging paper. The system located the beep to within about 40cm, leading him to an old carbon monoxide detector needing new batteries. He highlights the approach of giving Claude Code "ears" by having it run code that analyzes microphone audio and hot-swap new JavaScript onto listeners.

rust-v0.169.0-alpha.8

This is a GitHub Changelog entry titled "rust-v0.169.0-alpha.8." The supplied material consists only of the site's navigation menu and product links, with no details about the release itself.

JPMorganChase expands Claude Code to more employees across the firm

JPMorganChase is expanding Claude beyond engineering to product teams, advisors and other employees, with infrastructure and operations teams using Claude Code for monitoring, alerting and root cause analysis and the security team using frontier models to find and fix vulnerabilities faster. Claude Code runs in controlled environments including the firm's Devspace platform, with staged access and an approval process for new models, and an automated gate enforces a hard pass or halt so Claude cannot sign off on its own work. In Asset & Wealth Management, small teams are using Claude Code in their engineering workflow, and J.P. Morgan Asset Management recently announced a connector bringing 55ip's ActiveTax Technology into Claude for Financial Advisors.

CodeRabbit scales agentic software development and ships 20x more

CodeRabbit uses OpenAI models to govern and trust AI-generated code and Codex to ship 20 times more software with the same engineering team.

Indie Game Development with AI

The author built two complete 3D games, Inkline Town and Spirit Call, entirely through AI agents without writing or reading any code, using only chat prompts, playtest links, and 3D model previews. He argues that AI moves the hard part of game development from implementation to the choices only a human can make—what game to build, what style, what is fun, and what is broken—and that agents still fail at anything requiring taste, such as detailed anime characters, which forced him to change Spirit Call's art style. He recommends picking a genre around the AI's bottleneck, building verification layers (ECS structure, types, lints, headless tests, recorded playtests, and math models of the economy), and playtesting as a player while reporting what feels off rather than how to fix it.

0.169.0-alpha.7

This page is a GitHub Changelog entry titled "0.169.0-alpha.7"; the supplied material consists only of site navigation and product menu text, with no details about the release itself.

Meta launches Muse Ember, a small AI model built for the repetitive work agents actually do

Meta introduced Muse Ember, its smallest and most efficient AI model, designed for routine agent tasks such as tool calling, small coding jobs, request routing, data extraction, and long-document processing. Meta says it scores 36 on the Artificial Analysis Intelligence Index at about $0.14 per task, versus $2.39 for the larger Muse Spark 1.3, though Claude Haiku 5.5 leads on real coding tasks in FrontierCode 2.0-AA at 34.9 percent to Muse Ember's 27.2 percent. The multimodal model handles up to one million tokens and is available through the Meta Model API and Muse Code at $0.10 per million input tokens and $0.50 per million output tokens.

Introducing Claude Fable 5.2 on AWS

Anthropic's Claude Fable 5.2 and Claude Mythos 5.2 are now available on Amazon Bedrock and Claude Platform on AWS, with Fable 5.2 positioned as Anthropic's most capable generally available model for coding, math, and science, and Mythos 5.2 offered through verification programs for cyber security and life sciences work. Fable 5.2 is designated a Covered Model requiring up to 30 days of data retention and human safety review by default, with an aws_review retention mode set on the account or project before invocation, while Enterprise Frontier Safeguards combining zero data retention with misuse detection is in development. The post includes a Boto3 example using model ID global.anthropic.claude-fable-5-2 and describes the claude-api skill in Claude Code for migrating existing codebases.

Anthropic releases new Mythos and Fable models amid RSI concerns

Anthropic released Mythos 5.2 and Fable 5.2, citing improvements in coding, reasoning, and vision, and lowered Fable's price from $10 to $8 per million input tokens, making it cheaper than GPT-6 Astra, with responses 30% faster across effort levels. It is Anthropic's first release since a METR report suggested the lab might be approaching recursive self-improvement, finding internal models were speeding up Anthropic's AI progress by roughly 2x. Anthropic said the models voice suspicion of being evaluated more often than recent models, and that Mythos 5.2 shows stronger covert capabilities than prior models, though it maintains the model cannot act covertly against its user.

Featured coverage of this story: 《Introducing Claude Fable 5.2 on AWS》

v2.1.321

Claude Code v2.1.321 adds Claude Fable 5.2 (claude-fable-5-2) as the default Fable model with 1M context and $8/$40 per Mtok pricing, adds a --settings <file> option to claude remote-control, and fixes numerous bugs including MCP OAuth secrets being sent to another host on redirect, sandbox and permission-check issues, and network-path editing problems. It also changes mods to turn on without asking in auto and bypass modes, moves saved Fable 5.1 picks to fable, and adjusts fast mode labels, nested skill loading, and gateway and Windows behaviors.

Anthropic launches new Claude Fable and Mythos 5.2 models

Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2, which it says improve on coding, reasoning, and vision and generate outputs faster than the 5.1 versions. Anthropic says Fable 5.2 can orchestrate agent teams and handle long-running projects, and that during prerelease testing the models helped engineers with software migrations, designed three working electronic devices, solved two open problems in black hole physics, and formalized the proof of Fermat's Last Theorem in 30 hours. Fable 5.2 launches today, while Mythos 5.2 is initially available only to vetted organizations.

Featured coverage of this story: 《Introducing Claude Fable 5.2 on AWS》

Lessons learned when hiring engineers in the age of AI

The author describes how their company restructured engineering hiring around AI: the AI app-building round is now mainly a screen, take-homes were dropped due to candidate churn and a desperation filter, and traditional no-AI problem-solving and system-design interviews remain the core test of engineering judgment. They also note that many candidates have limited AI budgets at their current employers, that AI fluency can mask shallow engineering depth, and that they look for a "spark" of exceptional upside in every hire.

0.169.0-alpha.6

This page is a GitHub Changelog entry titled "0.169.0-alpha.6," but the supplied material contains only site navigation and product menu text (Platform, Copilot, Actions, Codespaces, Advanced Security, pricing, and similar links) rather than the changelog content itself. No release notes, features, fixes, or dates are stated in the material.

Show HN: We Gave AI Coding Agents Production Memory

Codela checks each deploy against production behavior via observability tools like SigNoz, Datadog, Sentry, New Relic and Grafana, learns from failures, and writes lessons into the repo (.codela/lessons.md, plus lines in CLAUDE.md and AGENTS.md) as pull requests that Claude Code, Cursor and Codex read before writing code. It also comments on pull requests that repeat a past mistake, and in the team's tests Claude Code repeated mistakes 8 of 8 times without Codela and 0 of 8 with it. Free for 1 repo and 20 PR checks a month; Pro is $7.99 per person per month.

AI code review has the order backwards

The article argues that pull requests were designed for human-speed code and break down when coding agents produce changes faster than reviewers can meaningfully read them, turning approvals into rubber-stamping and "code laundering." It proposes moving judgement to the start by agreeing on a "bar" of machine-verifiable criteria before code exists, with agents building against it and a reviewer model from another family checking the proof. It cites one Saturday (November 7th) on which 197 items landed with no human in the review loop.

Do AI coding agents care about SemVer or is it something for humans only?

Microsoft ran two Vally experiments with GitHub Copilot CLI (nearly 1,000 runs) and found that some models, notably Claude Sonnet 5.5 and Claude Haiku 5.5, checked upgrades less often when the version label was a patch than when it was a major release, even with byte-identical code. Three models (Claude Opus 5.5, GPT-6.1 Sol, GPT-6.1 Terra) checked every update regardless of label, and all four merges that left main broken came with a patch or minor label. The author recommends keeping SemVer honest, enforcing CI on every dependency branch, and treating the version number as a hidden variable when evaluating agents.

AI-Assisted Coding: Working Well with Coding Agents

A link to an article titled "AI-Assisted Coding: Working Well with Coding Agents" at nokku.payanai.com, submitted to Hacker News with 2 points and no comments.

Ginux: A New Kernel Runs Linux Userspace, Made via Human-Directed AI Engineering

This report presents Human-Directed Agentic Engineering (HDAE), a method combining human-defined architecture, constraints, milestones, acceptance criteria and stopping points with an autonomous implementation, execution, diagnosis and revision loop, using Ginux as its case study. Over eleven days, one persistent coding-agent task produced an independent Rust-first x86-64 kernel that boots through UEFI, implements its own core OS mechanisms and a Linux compatibility layer, runs selected unmodified Alpine Linux programs, and boots a Rocky Linux 10 userspace with systemd through multi-user.target. Ginux remains a research prototype with narrow compatibility and hardware coverage, no independent security audit, and no claim to production readiness or equivalence with Linux.

Being a meat proxy for AI should be grounds for getting people fired

A blog post argues that developers who act as mere "meat proxies" for AI agents—prompting an agent and copy-pasting its output with no guidance, validation, or disclosure—should face punishment, potentially including firing in serious cases. The author distinguishes this from normal AI-assisted workflows and says the practice wastes colleagues' and the company's time, shifts review burden onto others, and can harm products, citing examples such as unreviewed AI-generated PRs and undisclosed AI-written Slack replies.

Hide your slop PR descriptions behind <details><summary>

Simon Willison describes using HTML <details><summary> tags, which GitHub Markdown supports, to collapse long agent-written PR descriptions behind a short human-curated opener that explains what the PR is for. He uses the model name and effort level as the summary text, and shares recent examples from his datasette and llm projects.

The Web Is More Than Enough, by Paul Ford

In a Hacker News comment, Simon Willison describes building a Nudibranch-themed Fabergé egg in Three.js and then asking Claude (Opus 5.5 Medium in regular Claude Chat) to port it to Vanilla JS in the same thread, producing a non-identical result he links along with a transcript. He concludes Claude Opus 5.5 can do 3D modeling in Vanilla JS plus WebGL without Three.js, and finds the Three.js code slightly easier to understand, though he rarely reads the code for these 3D projects.

Exody – Native macOS AI Workbench (5-stage loop, 100% BYOK, $0 markup)

Exody is a native macOS autonomous coding agent with a five-stage loop (investigate, plan, build, verify, report) in which a fresh reviewer reads the diff before the user does. It supports bring-your-own-key access to Claude, GPT, Gemini, Grok and a local Ollama tier with no markup, plus a smart router that picks a model per task, a mobile remote-control app, and a Design Studio for websites, 3D scenes and film. The page claims its code scored higher than the same model's raw API in blind reviews and that its design pages beat raw models in blind side-by-side tests.

parse_number() as a SQLite extension: C vs Rust

Four SQLite extensions implement the same parse_number() behavior in C and Rust, with tests covering type preservation, parsing rules, and floating-point accuracy. The optimized C implementation was fastest at 16–19 ns per row, about twice as fast as the Rust versions and far faster than SQL replace(). Plain C was slower on decimals because it relied on strtod(), while a compact fast path closed that gap; all versions matched across 8 million fuzz checks.

[One Million AI Engineer Special] How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco

In the second run of Stanford's CS146S, Mihail Eric brought open-source maintainers into the course, having students learn an unfamiliar repo, propose a change, and explain the patch to the person who might merge it. In this interview marking one million subscribers on the AI Engineer YouTube channel, he walks through the ten-week syllabus and discusses which lessons held up with his own engineering team at Monaco, including building agent workspaces, buying code review tools, and giving an on-call agent access to Datadog and GitHub.

How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco

In this AI Engineer interview, Mihail Eric, who teaches Stanford's CS146S "The Modern Software Developer" and leads AI at Monaco, describes how the course was refreshed for a second run, with part of the homework now being contributions to open-source projects reviewed by the repos' maintainers. He also discusses Monaco's approach of buying some tools (such as Devin for code review) while building its own agent workspaces, guided by his rule of thumb: "You buy to learn, and you build to keep."

How I Fight AI Brain Fry

The author describes managing many parallel Claude Code and Codex terminal sessions by organizing work as a git-backed priority queue of markdown tickets, with an agent that researches, builds human-approved tickets, runs tests and browser checks, and never deploys to production. Items needing human judgment land in an "owner inbox" surfaced by a script, and a Claude artifact page rebuilt from git each cycle serves as the reading surface. The key practice is taking one item, loading its context, deciding, and closing it back into the queue so context doesn't linger in memory.

Andrej Karpathy built a curated video platform in a weekend using Claude and 20 years of Hacker News

Andrej Karpathy built HackerTube over a weekend using Claude's Opus 5.5 model, which filtered 43 million comments from 20 years of Hacker News into 6,454 videos that keep getting recommended there. He says the site offers a "significantly better viewing experience" than YouTube's recommendations, needs no login, shows a single mid-video frame instead of optimized thumbnails, and stores state locally in the browser. He then built TristanTube, "one-shotted with Claude Code in 3 minutes," with 47,000 videos from a YouTube channel list by Tristan Hume.

AI stance is not binary

The author argues that AI debates falsely present only two options—full delegation or full rejection—and proposes a third model using a relay-race analogy: humans and machines split the work rather than competing on the same course. Engineers should keep the difficult parts (understanding problems, architecture, trade-offs, review, debugging) while AI handles exploration, repetitive code, tests, and well-defined components. The piece critiques both the "AI makes engineers obsolete" view and DHH's reported default-agent approach at 37signals as treating software engineering as a race for maximum output.

We review 4x more AI code without reading every line

Aviator reports that shifting code review from reading diffs to verifying intent, acceptance criteria, invariants, decisions and evidence let the same nine engineers merge 4x more pull requests, cut median time to merge from 2.1 days to 3.4 hours and reduce reverts by 40%, with a human still signing off on every PR. The post describes failed approaches (an AI reviewer, a four-hour review SLA, a self-review skill), mining 11,400 review comments into 64 approved invariants, and using preview environments to prove criteria with recorded evidence.

Google AI Releases Elephant-Goldfish Model Plugin for Spec-First Coding

Google has open-sourced the Apache 2.0 Elephant-Goldfish Model (EGM) plugin, a design-and-review layer for AI coding agents in which an "Elephant" session drafts a specification before code is written and fresh, stateless "Goldfish" subagents review it across five stages. It is not a model and ships no weights; it runs on harnesses that can spawn fresh subagents, including Antigravity, Claude Code 2.1.271+, Claude Cowork, Codex and ChatGPT, using the host's model quota. No accuracy benchmark was published, and the README states it is not an officially supported Google product.

AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test

Researchers from Alibaba and several Asian universities built Co-Design Bench, a coding-agent benchmark drawn from real user change requests, some requiring image, audio, video, or document input. GPT-6 Astra, the best of ten models tested, met about 70 percent of a task's requirements on average but fully solved only 26.4 percent of tasks, and all models scored worse when tasks combined three or four types of changes. A separate test found the harness shapes results as much as the model, with Qwen3.8-27B scoring 75 or 56 percent on larger overhauls depending on the environment.

Show HN: I built an AI code editor that for teams

RxCode v2.0.0 is a free and open-source native workspace for teams of AI agents, offering sprint boards, automatic agent routing, schedules, briefings and chat on macOS, servers, iOS and Android. It supports built-in Claude Code and Codex plus other agents installed via the ACP registry, and includes Flowcy Cloud for shared boards and on-device fine-tuning of a local model.

Show HN: Codend – Build and publish web apps with AI agents

Codend is a Show HN project for building and publishing web apps with AI agents, posted at codend.dev. The Hacker News submission had 1 point and no comments at the time of the data.

Show HN: Reges Code – a free AI code editor that runs on local models

Reges Code is a free AI code editor for Windows 10/11 (64-bit) that auto-detects local model servers such as LM Studio, Ollama, vLLM, llama.cpp, Jan and Lemonade, and falls back to OpenRouter's free models when cloud use is chosen. It includes Reges Swarm, a built-in pipeline of architect, coder, reviewer, security and judge models, plus an honesty check that flags replies claiming files were created or commands run when the turn did neither. A CLI is included from version 1.0.17, and Linux and macOS builds are in progress.

An AI agent's public attempt to get 5k GitHub stars in a month (it's at 0)

An autonomous AI agent has been given a GitHub account and a goal of reaching 5,000 real GitHub stars by November 30, 2026, choosing what to build, writing the code and running the launch itself, with no bought stars or bots. As of Day 0 the count is 0, and the page logs its decisions and failures alongside the tool it has built so far: agent-recap, an MIT-licensed preview that searches past Claude Code and Codex conversations and can archive and recover transcript files via a local CLI.

Context compaction recovery for AI coding agents: six checks, not a summary

The article, shared via a Medium link on Hacker News, is titled "Context compaction recovery for AI coding agents: six checks, not a summary." The supplied material contains only the link and Hacker News metadata (1 point, 0 comments), with no article text to summarize.

0.169.0-alpha.3

A pre-release tagged rust-v0.169.0-alpha.3 was published by github-actions on 07 Nov at 19:31, with one commit to main since the release. The release lists 196 assets.