Code415 Radar

by Ziyu Guo · @Code415zg

What people are paying attention to in AI, LLMs and CS today. Every item comes with a short path to learn it and a small test to try it yourself. Updated every evening, Japan time.

Tuesday, November 17, 2026

20 items · Qwen4-Max leaves preview and Aristotle formally verifies Rust
  1. #1Models & launches

    Harmonic's Aristotle now formally verifies Rust, unsafe blocks included

    On Nov 16, Harmonic extended its Aristotle verification system from C and Verilog to Rust. It can prove that Rust code has no undefined behavior and does what its spec says, even inside unsafe blocks, and it can check that a C to Rust rewrite behaves the same as the original. On Manifest, an open source order book on Solana, it proved 28 results and found a rounding bug that the team has since fixed.

    Heat · Hacker News story at 115 points and 61 comments per the HN API
    What people foundAccess is self serve with pay per use pricing and no seats, and Harmonic is partnering with the Solana Foundation to help Solana programs get verified. On HN, bryanlarsen pointed out that Verus, VeriFast and RefinedRust already exist and use very different mechanisms, so it is worth comparing approaches before committing.
    Learn it15 min · 5 steps

    Key ideas

    Undefined behavior
    Operations such as out of bounds reads or data races that the language gives no meaning to, so the compiler may do anything with them.
    Formal verification
    A machine checked mathematical proof that code satisfies a stated specification for every possible input, not just the tested ones.
    Equivalence checking
    Proving that two programs, such as a C original and its Rust port, produce the same results for all inputs.

    Steps

    1. Read the Harmonic announcement first to see the three claims: no undefined behavior, spec matching, and C to Rust equivalence.
    2. Look at the Manifest example in the post and note what kind of property was proved, such as a valid swap going through.
    3. Skim the HN thread for the list of other Rust verification tools and how they differ.
    4. Write down one function in a crate you know that uses unsafe and what its spec would say in plain words.
    5. Self check: what is the difference between a test suite passing and a proof that a function has no undefined behavior?
    Try it30 min · 5 steps
    You need: A Harmonic Aristotle account with pay per use billing, a small Rust crate you own, and a browser. No local GPU needed.

    Steps

    1. Pick a small crate you own, ideally one with a single unsafe block or a function ported from C.
    2. Write a short plain English spec for one function: inputs, outputs and invariants it must keep.
    3. Sign up through the Try Aristotle link on the Harmonic Aristotle page and follow its docs to submit the crate and spec.
    4. Read the result: which properties were proved, which failed, and any counterexample it reports.
    5. If a counterexample appears, reproduce it as a unit test with cargo test and fix the code.

    Small angles to try

    • Port a tiny C function to Rust and ask for an equivalence proof against the original.
    • Plant a deliberate off by one bug and see whether the proof fails where expected.
    • Compare the effort with writing the same spec in Verus by hand.
  2. #2Tools & open source

    OpenCode Browse lets coding agents drive your real browser and logins

    OpenCode Browse is an MIT licensed MCP server plus Chromium extension that lets coding agents use your own browser, with your open tabs, logins and extensions. Agents open tabs in collapsed background groups, drive them with Playwright, and hand a tab back to you for logins, 2FA or payments until you click Continue. It works with any MCP client, and the page names OpenCode, Claude Code, Cursor, Codex and Zed.

    Heat · Hacker News story at 208 points and 101 comments per the HN API
    What people foundSessions can be limited to a list of allowed sites, and agents can capture screenshots, screen recordings and network traffic as a HAR file. On HN, testtestmetest called the project's approach to security reckless, so run it in a separate browser profile before trusting it with real accounts.
    Learn it15 min · 5 steps

    Key ideas

    MCP server
    A small program that exposes tools to an AI client over the Model Context Protocol, so the agent can call them like functions.
    Handoff
    The agent pauses and gives a tab back to the human for a sensitive step such as a login, then resumes after confirmation.
    HAR file
    A JSON log of every network request and response a page made, readable in browser dev tools.

    Steps

    1. Read the OpenCode Browse product page first for the feature list and supported browsers.
    2. Read the GitHub README sections on setup, other MCP clients and the recording and network tools.
    3. Skim the HN thread for security objections and the suggestion to use the Chrome DevTools Protocol directly.
    4. List which of your signed in sites an agent should never touch.
    5. Self check: why does an allowlist matter more for an agent using your real browser than for a headless one?
    Try it30 min · 6 steps
    You need: Bun, Node.js 22.19 or newer, a Chromium browser, and an MCP client such as OpenCode or Claude Code. No API key needed for Browse itself; the agent client needs its usual model access.

    Steps

    1. Create a fresh browser profile with no important logins, so the test cannot touch real accounts.
    2. Install the extension from the Chrome Web Store link on the product page, or build from source with git clone https://github.com/anomalyco/browse.git && cd browse && bun run setup.
    3. If your client is not OpenCode, add a stdio MCP server that runs node /path/to/browse/cli/dist/cli.mjs mcp, as the README shows.
    4. Ask the agent to open a site you own or a public docs page, take a screenshot, and summarize it.
    5. Ask it to log in to a test account you own and confirm it hands the tab back to you instead of typing the password.
    6. Use the network_start and network_stop tools to capture a HAR file and open it in dev tools.

    Small angles to try

    • Restrict the session to one allowed site and ask the agent to visit another to see if it is blocked.
    • Record a screen capture with recording_start of an agent filling a form on your own staging app.
    • Compare speed and reliability with a headless Playwright MCP setup.
  3. #3Tools & open source

    Windows 11 quietly logs which coding agents you run

    Xusheng Li found a Windows 11 component, MissionControlAggregator.dll, loaded under the DiagTrack service, that notes which coding agents and developer tools run, including copilot, claude, codex, agy, cursor and antigravity executables. It keeps a bitmap of which tools were seen in each reporting window, with no prompts or source code found, and queues it as optional diagnostic data. The author confirmed it by copying cmd.exe to codex.exe in a VM and finding the report in DiagTrack's upload queue.

    Heat · Hacker News story at 280 points and 157 comments per the HN API
    What people foundTurning off Send optional diagnostic data under Settings, Privacy & security, Diagnostics & feedback stops the upload; the author says the setting is on by default. On HN, lesspassiveobse noted the data will be skewed, since it only covers Windows users who leave optional telemetry on.
    Learn it15 min · 5 steps

    Key ideas

    DiagTrack
    The Windows Connected User Experiences and Telemetry service that collects and uploads diagnostic events.
    Optional diagnostic data
    The Windows telemetry tier beyond the required minimum, which a user can switch off in Settings.
    Bitmap
    A compact set of on or off bits, here one bit per tool category seen during a reporting window.

    Steps

    1. Read Xusheng Li's write up first, focusing on the list of tracked executables and the VM test.
    2. Note where state is stored and where the queued event was found, the EventStore.db AsimovEvents table.
    3. Skim the HN thread for discussion of sampling bias and other Microsoft telemetry.
    4. Check which tier of diagnostic data your own Windows machines send.
    5. Self check: what exactly does the report contain, and what does it not contain according to the author?
    Try it30 min · 5 steps
    You need: A Windows 11 machine or VM you own with admin PowerShell. No API key needed.

    Steps

    1. Open Settings, Privacy & security, Diagnostics & feedback and record whether Send optional diagnostic data is on.
    2. In PowerShell, run Get-Service DiagTrack | Format-Table Name,Status,StartType -AutoSize to see the service state.
    3. Run tasklist /m MissionControlAggregator.dll to check whether any process has the component loaded.
    4. Check whether the folder C:\ProgramData\Microsoft\Diagnosis\AggregatorStorage\MissionControlAggregator$ exists on your build.
    5. Turn off optional diagnostic data, reboot, and repeat the checks to compare.

    Small angles to try

    • Run Get-EtwTraceProvider -SessionName Diagtrack-Listener before and after changing the setting.
    • Compare a managed work laptop with a personal machine to see what policy sets.
    • Draft a one paragraph note for your team on which setting to use on dev machines.
  4. #4Models & launches

    Qwen4-Max leaves preview at $1.50 in, $4.50 out per M tokens, closed

    Alibaba moved Qwen4-Max from the Token Plan only preview of Nov 11 to a production release on Nov 16, callable as qwen4-max on Model Studio. Model Studio calls it the next generation flagship multimodal model in the Qwen family, with web search, tool calling, structured outputs, and compatibility with both the OpenAI and Anthropic APIs. Artificial Analysis lists a 984k context, 54 tokens per second output and an Intelligence Index of 43.

    Heat · Hacker News story at 862 points and 477 comments per the HN API
    What people foundArtificial Analysis lists it as proprietary with no open weights, and on HN SirFlamenco said Alibaba has no plans to open it. Another HN commenter, maplet, pointed out it scores 43 on Artificial Analysis at a higher average cost than a rival scoring 46, so price per task matters more than list price.
    Learn it15 min · 5 steps

    Key ideas

    Per token pricing
    Input and output tokens are billed separately per million, so long outputs and thinking tokens drive cost.
    Thinking mode
    A setting where the model writes hidden reasoning before answering, controlled on Model Studio by enable_thinking and reasoning_effort.
    API compatibility
    The endpoint accepts the same request format as OpenAI or Anthropic, so existing SDK code works with a new base URL and model id.

    Steps

    1. Read the Qwen4-Max page on Alibaba Cloud Model Studio first for limits, pricing by region and thinking parameters.
    2. Open the Artificial Analysis model page and compare speed, price and index with the model you use now.
    3. Skim the HN thread for the debate on benchmark cost data versus real task cost.
    4. Note the difference between the old qwen4-max-preview entry and the new qwen4-max entry.
    5. Self check: why can a model with a lower list price still cost more per finished task?
    Try it30 min · 5 steps
    You need: An Alibaba Cloud Model Studio API key, Python 3 with the openai package, and a small set of your own prompts. The Singapore region lists a 1 million token free quota valid for 90 days.

    Steps

    1. Create a Model Studio API key and look up the OpenAI compatible base URL in its OpenAI Chat Completions guide.
    2. Point an existing OpenAI SDK script at that base URL with model qwen4-max.
    3. Run 5 to 10 of your own real coding prompts and log output tokens, wall time and answer quality.
    4. Repeat with reasoning_effort set to low and to high and compare cost and latency.
    5. Run the same prompts on your current model and compute cost per accepted answer.

    Small angles to try

    • Try the Anthropic Messages compatible endpoint with existing Anthropic SDK code.
    • Test structured outputs with a JSON Schema you already use in production.
    • Feed one very long document to see how quality holds near the context limit.
  5. #5Trends × tech

    Reskilling × AI courses: 10,000 free seats for Google's AI certificate

    Japan Reskilling Consortium announced on Nov 16 that its first 10,000 members can take the normally paid Google AI Professional Certificate for free. The course has over 20 hands on exercises on using AI at work, needs no prior AI experience and takes about 7 to 10 hours. A similar campaign in April 2026 closed within one day.

    Heat · Number 1 in the ITmedia AI+ access ranking on Nov 17
    What people foundEach account may apply once, and the first Coursera login must happen by March 31, 2027, the date the consortium ends its activities. For engineering teams, it signals that more non engineer colleagues will arrive with the same baseline AI training.
    Learn it15 min · 4 steps

    Key ideas

    Reskilling
    Training existing workers in new skills so they can take on changed or new roles.
    Professional certificate
    A short online course series on Coursera that ends with a certificate issued by the company that made it, here Google.

    Steps

    1. Read the ITmedia AI+ article first for eligibility, deadlines and how to apply.
    2. Check whether you or your colleagues can join Japan Reskilling Consortium as members.
    3. Look at the course outline on Coursera to see which AI tasks the exercises cover.
    4. Self check: what two conditions must be met to keep the free seat?
    Try it30 min · 5 steps
    You need: A browser and a free AI chat tool you already use. No API key needed.

    Steps

    1. Pick one real work task you repeat weekly, such as summarizing meeting notes or drafting a status update.
    2. Do it once by hand and time yourself.
    3. Do it again with a free AI chat tool using your own data, writing a clear prompt with role, context and output format.
    4. Compare time spent and quality, and note any errors the AI made.
    5. Write a three line guide a non engineer colleague could follow for the same task.

    Small angles to try

    • Try the same task in two different free AI tools and compare.
    • Turn the prompt into a reusable template with blanks to fill in.
    • Ask a colleague to follow the guide and note where they got stuck.
  6. #6Tools & open source

    lend: remote coding agents use your credentials without holding them

    lend is an open source proxy that sits in front of the SSH agent on a laptop, so a remote VM running a coding agent gets lend's socket instead of real keys or tokens. Each command or signing request is checked against a local JSON policy, approved by a prompt or Touch ID, and executed on the laptop, with only output and exit code sent back. A rule can pull a secret from 1Password with op read and hand it to that one approved command, so the VM never sees it.

    Heat · Hacker News, 25 points and 4 comments at time of writing; blog post from Nov 13, updated Nov 17
    What people foundLeases are the practical detail: approving a command can grant a time window or a fixed number of uses, but only for that exact command and arguments, and the lease ends with the lend ssh session or when the screen locks. The author warns that approving a shell, env or op itself would expose injected secrets.
    Learn it15 min · 5 steps

    Key ideas

    SSH agent forwarding
    The remote machine borrows a socket to the local agent so it can sign with keys that never leave the laptop.
    Policy with default deny
    Only commands and keys listed in the policy file can run, and each rule names how it must be approved.
    Lease
    A short lived approval for one exact command, limited by time or number of uses, so repeats do not need new prompts.

    Steps

    1. Read the author's launch post about running coding agents without handing over credentials, on blog.antonio.cx, to see the threat model.
    2. Open the lend README on GitHub and read the policy field list: default, commands, args_prefix, approval, secrets, lease and ssh keys.
    3. Look at examples/policy.json in the repo and note how gh is limited to pr view while op is limited to read.
    4. Read the section on Touch ID modes and note that biometric approval works on macOS only.
    5. Self check: if a rule allows the executable bash with prompt approval, why does that defeat secret injection?
    Try it30 min · 6 steps
    You need: A laptop with an SSH agent and your own remote VM or local Linux container reachable by SSH, Homebrew or curl, the GitHub CLI logged in on the laptop. Touch ID modes need macOS. No API key needed.

    Steps

    1. Install lend on both the laptop and your own VM with brew install antoniomika/tap/lend or the install script linked in the README, then run lend version on each.
    2. On the laptop, list the keys in your existing agent with lend keys --socket "$SSH_AUTH_SOCK" and copy one fingerprint.
    3. Write policy.json based on examples/policy.json: default deny, require_reason true, one rule for the absolute path of gh with args_prefix pr view and approval prompt, and your key fingerprint with approval prompt.
    4. Check the file with lend doctor --config ./policy.json, then start a session with lend ssh --config ./policy.json -- agent-vm, replacing agent-vm with your own host.
    5. On the VM, run lend doctor --remote, then lend exec --reason "Inspect a PR" -- gh pr view 1 --repo yourname/yourrepo against a repo you own, and approve it on the laptop.
    6. Try a command outside the policy, such as gh repo delete on a throwaway repo name, confirm it is refused, then review everything with lend history.

    Small angles to try

    • Add a lease to the gh rule and count how many repeat calls run without a new prompt.
    • Run lend mcp and point a coding agent on the VM at its run_local_command tool.
    • Use lend explain to see which rule a request matches before approving anything.
  7. #7Research

    MMFlash: speculative decoding for vision language models, up to 3.58x

    NUS researchers found that draft acceptance in speculative decoding drops from 46% to 22% as output tokens depend more on the image or video. MMFlash trains the drafter with extra weight on visually dependent tokens it gets wrong and adds sparse attention for long video. They report up to 3.58x lossless speedup, beating EAGLE-3 by up to 195.4% and DFlash by up to 17.1%.

    Heat · Hugging Face Papers, 35 upvotes; listed in daily papers Nov 16, arXiv posting Nov 13
    What people foundDrafters are already on Hugging Face for Qwen3.5-4B, Qwen3.5-9B, Qwen3.8-27B and Gemma 4 E4B-it, and the README says they run on stock SGLang 0.5.23 or vLLM 0.33.0 with no patches, so trying it is a serving flag change rather than a training project.
    Learn it15 min · 5 steps

    Key ideas

    Speculative decoding
    A small draft model guesses several tokens and the large model verifies them in one pass, keeping output identical while cutting latency.
    Acceptance rate
    The share of drafted tokens the target model accepts, which sets how much speedup is possible.
    Visual dependence
    How much a token's correct value relies on the image rather than the text so far, which text trained drafters cannot guess well.

    Steps

    1. Read the abstract and the acceptance rate figure on the Hugging Face paper page for MMFlash.
    2. Skim the method section of the arXiv paper to see how training loss is reweighted toward visually dependent tokens.
    3. Open the MMFlash README and read the table pairing each target model with its FSCCS drafter.
    4. Compare the reported gains against EAGLE-3 and DFlash and note which benchmarks they come from.
    5. Self check: why would a drafter that sees only text accept fewer tokens when describing a chart than when continuing a sentence?
    Try it30 min · 5 steps
    You need: Linux with an NVIDIA GPU; the README benchmarks on one H200 with 141 GB, smaller cards may fit the 4B pair but that is untested here. Python, pip install sglang==0.5.23, Hugging Face downloads of Qwen/Qwen3.5-4B and FSCCS/Qwen3.5-4B-MMFlash. No API key needed.

    Steps

    1. Install SGLang with pip install sglang==0.5.23 in a fresh virtual environment.
    2. Start a baseline server for Qwen/Qwen3.5-4B without any speculative flags and send 20 image description prompts using your own photos or screenshots, recording tokens per second.
    3. Restart with the README command, which adds --speculative-algorithm DFLASH, --speculative-draft-model-path FSCCS/Qwen3.5-4B-MMFlash and --speculative-dflash-block-size 16 plus the attention and parser flags it lists.
    4. Replay the same 20 prompts with the same sampling settings and record tokens per second again.
    5. Diff the outputs from both runs under greedy decoding to confirm they match, then compute the speedup.

    Small angles to try

    • Split prompts into text heavy and chart or OCR heavy groups and compare speedup per group.
    • Repeat on vLLM 0.33.0 with the README's --speculative-config JSON and compare to SGLang.
    • Try the Gemma4-E4B-it-MMFlash drafter with the Gemma specific flags from the README.
  8. #8Research

    VibePrevisBench: coding agents build Blender scenes, miss camera moves

    VibePrevisBench gives coding agents 242 tasks in which a script, photo, continuation or reference video must become an executable Blender scene that renders a proxy video. A rule based Scene Check reruns the code against hidden facts across nine dimensions, and a vision model Screen Check judges the render. Eleven frontier coding agents manage basic staging and camera work but fail harder spatial and camera motion requests.

    Heat · Hugging Face Papers, 83 upvotes, among the top papers of Nov 16
    What people foundThe authors' most reusable finding: agents do inspect their own renders but often fail to fix the errors they find, so adding a visual feedback loop to an agent is not enough without making it act on what it sees.
    Learn it15 min · 5 steps

    Key ideas

    Previsualization
    A rough 3D version of a film shot used to plan staging and camera before real production.
    Scene Check
    A rule based grader that runs the agent's Blender code and tests the scene against hidden facts like object positions and camera paths.
    Screen Check
    A vision language model that watches the rendered video and judges whether it matches the request.

    Steps

    1. Read the abstract and main figure on the Hugging Face paper page for VibePrevisBench.
    2. Open the arXiv paper and find the list of the nine Scene Check dimensions.
    3. Look at the results table and note where camera motion and spatial tasks fall below staging tasks.
    4. Read the analysis of self inspection to see examples of agents spotting but not fixing errors.
    5. Self check: why can a code based check catch a wrong camera path that a single rendered frame would hide?
    Try it30 min · 6 steps
    You need: Linux, Python 3.12, Docker, ffmpeg 5.1 or newer, Blender 5.2.2 on the host. Downloading data needs no key; running an agent on tasks needs whatever API key your own coding agent uses.

    Steps

    1. Clone the repo with git clone https://github.com/vibeprevisbench/vibeprevisbench.git, create a Python 3.12 venv and run pip install -r requirements.txt.
    2. Run python vpbench.py doctor --blender /path/to/blender-5.2.2/blender to list missing tools.
    3. Build the sandbox image with docker build -t vibeprevisbench-sandbox:5.2.2 harness/docker.
    4. Fetch annotations with python vpbench.py download --repo AidanYuzhe/vibeprevisbench --data data --data-only, then media with python vpbench.py fetch-media --data data and check gaps with the --check flag.
    5. Pick two script tasks, one simple staging and one with camera motion, and follow the README to run your own coding agent on them inside the sandbox.
    6. Compare the Scene Check results for the two tasks and watch both proxy renders.

    Small angles to try

    • Give the agent its own rendered frames back and count how often it actually edits the scene.
    • Compare a photo input task with a script input task of similar difficulty.
    • Write one task of your own with a hidden camera fact and see if the agent meets it.
  9. #9AI in the world

    HN Who is Hiring: AI in 51.6% of posts as volume nears 12 year lows

    HN Trends counted 244 postings in November's Who is Hiring thread as of Nov 14, down 37% from 387 at the same point last year, and the 2026 monthly average of 287 is the lowest since 2013. AI appears in 51.6% of postings, up from 30% two years ago, while AI agents appear in 14.8% and LLM in 10.7%. MCP rose from 0.3% to 4.5% in a year, Rust climbed to third place at 8.2%, and Front End fell to 6.6%.

    Heat · Hacker News, 94 points and 86 comments
    What people foundThe HN headline says AI reached half of all jobs, but the figure is the share of postings that mention AI, not AI roles. November is also a partial count, so the 37% drop may narrow as the thread fills in.
    Learn it15 min · 4 steps

    Key ideas

    Who is Hiring thread
    A monthly Hacker News thread where each top level comment is one job posting.
    Mention share
    The percentage of postings containing a term, which measures how often a skill is named, not how many jobs center on it.
    Partial month count
    A tally taken before the thread stops growing, so it can understate the final number.

    Steps

    1. Read the HN Trends November 2026 page from top to bottom, including the methodology note.
    2. Note which figures are year over year and which compare to two years ago or early 2022.
    3. Skim the HN discussion for critiques of keyword counting.
    4. Self check: if 51.6% of postings mention AI, what share of postings could still be for roles with no AI work at all?
    Try it30 min · 5 steps
    You need: Python 3 and the requests package, internet access to the public Hacker News API. No API key needed.

    Steps

    1. Find the item id of the November 2026 Who is Hiring thread from its HN page URL.
    2. Fetch the thread from the official Hacker News API at https://hacker-news.firebaseio.com/v0/item/<id>.json and read its kids list of top level comments.
    3. Fetch each top level comment, keep the text, and skip deleted or dead ones, with a short pause between requests.
    4. Count postings that mention AI, MCP, Rust and front end using case insensitive whole word matches.
    5. Compare your shares with the HN Trends figures and note where keyword choices change the result.

    Small angles to try

    • Separate postings that list an AI role title from postings that only mention AI anywhere.
    • Repeat for the November 2025 thread and check the 387 baseline.
    • Track the share of remote postings and compare with the 55.7% reported.
  10. #10Models & launches

    GitHub Copilot retires Opus 4.8, Opus 5 and GPT-5.6 Terra

    As of Nov 16, GitHub Copilot deprecated Claude Opus 4.8 and Claude Opus 5, with Claude Opus 5.5 as the suggested replacement, and GPT-5.6 Terra, with GPT-6.1 Sol or GPT-6 Luna suggested instead. Individual subscribers on annual plans keep Opus 4.8, and Opus 5 stays only as a fallback for runtime refusals. The same week's release notes add Claude Fable 5.2 and Gemini 4 Carbon for Pro+, Max, Business and Enterprise.

    Heat · GitHub changelog entries dated Nov 16; no popularity metric seen
    What people foundPer GitHub, replacements switch on automatically for Business and Enterprise unless an admin disabled the global default or that model, in which case the replacement policies must be enabled by hand. Fable 5.2 requires data retention by default and its policy starts off for Business and Enterprise.
    Learn it15 min · 4 steps

    Key ideas

    Model deprecation
    The model leaves the picker and APIs, and requests must move to a named replacement.
    Model policy
    An admin switch in Copilot settings that decides whether a given model is available to an organization.
    Global default enablement
    The organization setting that lets new or replacement models turn on automatically.

    Steps

    1. Read the GitHub changelog entry on models deprecated on November 16, 2026.
    2. Read the Copilot weekly releases entry for the Fable 5.2 and Gemini 4 Carbon plan and retention notes.
    3. Open your organization's Copilot model policy page and see whether the global default is on.
    4. Self check: if the global default is off and a workflow names Opus 5, what breaks and what single change fixes it?
    Try it30 min · 5 steps
    You need: A text editor or code search over repositories you own, plus Copilot admin access for the settings steps. No API key needed.

    Steps

    1. Search your own repositories, CI files and editor settings for the strings Opus 4.8, Opus 5 and GPT-5.6 Terra, excluding Opus 5.5 matches.
    2. List each hit with its file, owner and the replacement model GitHub suggests.
    3. In Copilot settings, check whether the global default for new models is enabled and whether Opus 5.5, GPT-6.1 Sol and GPT-6 Luna are on.
    4. Update one pinned workflow to the replacement model and run it once to confirm it works.
    5. Decide whether your team accepts the data retention needed for Fable 5.2 before anyone enables its policy.

    Small angles to try

    • Add a CI check that fails when a deprecated model name appears in config.
    • Run the same prompt on Opus 5.5 and GPT-6.1 Sol in Copilot Chat and compare answers for your codebase.
    • Write a short team note on which models are approved and why.
  11. #11Trends × tech

    Password notebooks × AI attackers: Daiso sells out amid AI exploits

    Daiso says its paper password notebook is temporarily out of stock after sales rose from late October, as Japan's breach wave grew: MM Soken counted 193 leak incidents in October, up from 81 in September, and about 161 million leaked credentials by the end of October. On Nov 16 Google's Threat Intelligence Group reported a framework called exploit_agent, found on an exposed attacker server, that compares source code before and after patches to popular web apps, then writes and runs exploits with no human in the loop. IPA says paper is acceptable, but it does not help when a service's own database leaks, so it recommends unique passwords plus passkeys or multifactor authentication.

    Heat · Daiso notebook out of stock with a restock planned within November; ITmedia NEWS covered both stories on Nov 14 and Nov 16
    What people foundGTIG says the framework works with several frontier and local LLMs, and the same server held stolen victim data and web shell scripts; the report also notes zero day attacks rose from about 8 a month in 2025 to 11 a month in 2026, so the window between a public fix and an automated exploit is shrinking.
    Learn it15 min · 4 steps

    Key ideas

    Patch diffing
    Comparing code before and after a security fix to work out which bug was fixed, which tells an attacker where unpatched systems are weak.
    Credential stuffing
    Trying usernames and passwords leaked from one service on many other services, which only works when people reuse passwords.
    Passkey
    A login credential stored on a device that uses public key cryptography, so there is no shared secret for a site to leak or a phishing page to capture.

    Steps

    1. Read the ITmedia NEWS article on the GTIG AI Threat Tracker to see what exploit_agent did and what else the report found.
    2. Read the ITmedia NEWS article on the Daiso notebook for the MM Soken leak numbers and IPA's advice on paper versus password managers.
    3. Compare the two threats: a leaked service database bypasses any notebook, while patch diffing targets anyone who patches slowly.
    4. Self check: which of your own services would still be safe if their password database leaked tomorrow, and why?
    Try it30 min · 5 steps
    You need: Your own password manager or browser password store and your own accounts; No API key needed

    Steps

    1. Open your password manager or browser password settings and export nothing; just open its built in security or password health report.
    2. List every account flagged as reused, weak or found in a known breach, and change the top five, starting with email.
    3. For your email, code hosting and cloud accounts, check the security settings for passkey support and register a passkey where offered.
    4. Turn on multifactor authentication for any of those accounts that still lack it and store the recovery codes offline.
    5. If you maintain a web app, check how many days your last three security patches took from upstream release to production.

    Small angles to try

    • Count how many of your top 20 accounts support passkeys today.
    • Write a one page patch speed target for your team, such as critical fixes within 48 hours.
    • Compare a paper notebook and a password manager on loss, leak and phishing risk for an older relative.
  12. #12Tools & open source

    Coral relaunches as a gateway where agents write their own tools

    Coral now describes itself as an open source, self hosted execution layer between agents and services such as GitHub, Linear and Slack: agents call tools by writing short TypeScript programs and can save them as reusable custom tools. Programs run in V8 isolates with no network access or credentials, and each tool call is checked against workspace policy before credentials are attached. It is written in Rust, runs as a desktop app or shared server, has an Apache 2.0 core, and connects to agents over MCP; the GitHub README still shows the earlier SQL query interface.

    Heat · Show HN, 35 points and 9 comments when checked; GitHub repo at about 5k stars
    What people foundThe founders report that custom tools made Opus 5.5 agents 29% faster at 31% lower cost on GitHub, Linear and Slack tasks, a vendor benchmark; co-founder jcs advises starting with the desktop app, letting agents discover which tools they need, and having humans review the additions.
    Learn it15 min · 4 steps

    Key ideas

    Code mode
    Instead of calling one tool per model turn, the agent writes a small program that calls several tools and returns only the result.
    V8 isolate
    A lightweight sandbox inside the JavaScript engine that runs code with its own memory and no access to the host unless granted.
    Credential injection
    The gateway adds API tokens only after a call passes policy, so the agent's code never sees the secrets.

    Steps

    1. Read the Show HN post and the co-founder's replies for the current design and deployment advice.
    2. Skim the withcoral.com homepage and quickstart to see how providers, MCP clients and custom tools fit together.
    3. Open the GitHub README and note how the older SQL interface differs from the gateway framing.
    4. Self check: why does running agent code in an isolate without credentials reduce the damage of a prompt injection?
    Try it30 min · 5 steps
    You need: Desktop machine, the Coral desktop app, an MCP client such as Claude Code, and a GitHub token for your own account; the agent needs its usual model API access

    Steps

    1. Install the Coral desktop app from the download page linked on withcoral.com.
    2. In the Coral UI, add GitHub as a provider using your own account.
    3. Copy the MCP client setup shown in the Coral UI into your agent and restart it.
    4. Ask the agent what your open pull requests are, as in the quickstart, and watch the program it writes.
    5. Ask for a repeat task, such as summarizing stale issues in one of your repos, and check whether it saves a custom tool you can review.

    Small angles to try

    • Run the same task with the GitHub MCP server directly and compare tokens used.
    • Add a workspace policy that blocks write calls and confirm a write attempt is refused.
    • Add a second provider you own, such as Slack or Linear, and ask a question that needs both.
  13. #13Models & launches

    Tencent Hy4 flash: 224B MoE, 12B active, 1M context, $0.07 per M input

    Tencent launched Hy4 flash on Nov 16, a cheaper multimodal variant of its Hy4 flagship with 224B total and 12B active parameters and a 1M token context. Unite.ai reports it cuts the KV cache to one sixth of Hy4's, and Tencent claims about one twentieth of Hy4's per task cost on its internal WorkIndex benchmark. Tencent says it plans to release the weights but has given no date.

    Heat · Listed on OpenRouter on Nov 16 at a 50% launch discount; covered by TechNode and Unite.ai; small HN posts only, about 6 points
    What people foundOn OpenRouter, tencent/hy4-flash lists $0.07 input and $0.281 output per million tokens, halved during the launch discount, through an OpenAI compatible API; all benchmark claims so far are Tencent's own.
    Learn it15 min · 4 steps

    Key ideas

    Mixture of experts
    A model where each token runs through only a few expert subnetworks, so 12B of 224B parameters do the work per token.
    KV cache
    Stored attention keys and values for earlier tokens; a smaller cache makes long contexts cheaper to serve.
    Cost per task
    Total spend to finish a job, which depends on tokens used and retries, not only on the price per token.

    Steps

    1. Read the TechNode launch article for the parameter counts and the WorkIndex cost claim.
    2. Read the Unite.ai piece for the KV cache and multimodal details.
    3. Open the OpenRouter model page and note context length, prices and the discount end terms.
    4. Self check: why can a model with a lower price per token still cost more per finished task?
    Try it30 min · 5 steps
    You need: Python or any HTTP client, an OpenRouter account and API key, a few dollars of credit, and your own sample documents

    Steps

    1. Create an OpenRouter API key and follow its quickstart to send one chat request to tencent/hy4-flash.
    2. Pick 10 tasks from your own work, such as extracting fields from an invoice PDF converted to text or summarizing a long log.
    3. Run the same tasks on tencent/hy4-flash and on the cheap model you use today, recording tokens and cost from the responses.
    4. Score each answer by hand as correct, partly correct or wrong.
    5. Compute cost per correct answer for both models and decide whether to switch.

    Small angles to try

    • Feed a 300k token document and ask questions about the start, middle and end.
    • Send a chart image and check the extracted numbers.
    • Repeat the cost run after the launch discount ends.
  14. #14Research

    143K SWE-rebench traces: agents with equal scores code very differently

    Nebius researchers analyzed 143,209 trajectories from 125 agent configurations spanning 117 models, using a rule based parser that maps raw traces to actions such as repository search, file editing and test runs. Agents with similar solve rates differ substantially in how they explore, edit and verify code, and behavior profiles alone can identify the agent stack. The scaffold largely decides how files get edited, while higher reasoning effort goes into exploration and workspace checks; the paper was accepted at a NeurIPS 2026 workshop on interpreting agent behavior.

    Heat · HF Papers, 10 upvotes and 4 comments; arXiv posting dated Nov 12, listed on HF Papers Nov 16
    What people foundCommenter victor-checkpaper argues the 77.5% exact match figure in Table 1 cannot exceed about 74.0% given the paper's own counts, and opened a GitHub issue; there was no reply yet, so treat that table with care.
    Learn it15 min · 4 steps

    Key ideas

    Trajectory
    The full log of an agent's actions and tool calls while it works on one task.
    Scaffold
    The harness around a model that defines its tools, prompts and loop, such as how it edits files.
    Solve rate
    The share of benchmark tasks an agent finishes correctly, which says nothing about how it got there.

    Steps

    1. Read the HF paper page abstract and the submitter's summary comment.
    2. Skim the arXiv paper's figures comparing scaffolds and reasoning effort.
    3. Read the Table 1 comment thread and the linked GitHub issue.
    4. Self check: if two agents both solve 60% of tasks, what behaviors would you compare before picking one for your repo?
    Try it30 min · 5 steps
    You need: Python 3 with the Hugging Face datasets package and free disk space for a partial download; No API key needed

    Steps

    1. Open the trajectories dataset linked from the HF paper page and read its card for the file layout.
    2. Load a small slice of the dataset in Python and print one trajectory.
    3. Write a short script that counts search, edit and test actions per trajectory by matching tool names or commands.
    4. Group the counts by agent or scaffold and compare two agents with similar solve rates.
    5. Check whether your counts agree with the paper's claim that scaffolds shape editing style.

    Small angles to try

    • Measure how often each agent runs tests before its final edit.
    • Compare trace length for solved versus failed tasks.
    • Apply the same counting script to logs from your own coding agent.
  15. #15Trends × tech

    Budget phones × on device photo AI: AQUOS sense11 erases bystanders

    Sharp announced the AQUOS sense11 on Nov 17, on sale from Nov 26 through docomo, au, SoftBank, Rakuten Mobile, J:COM MOBILE and the open market, from 94,600 yen for 8GB RAM and 128GB storage. Its camera AI removes bystanders behind the subject, removes shadows from food and text, cuts glass reflections and masks addresses and other identifying details, a feature carried down from the flagship AQUOS R11. It also preinstalls NIAH, Sharp's assistant app for appliance help, schedules and chores.

    Heat · Covered by ITmedia Mobile and Impress AI Watch on launch day; a Google Trends Japan spike was reported but not verified here
    What people foundITmedia reports the photo AI runs on the device and keeps processing in the background after the camera app is closed; the base model costs 31,900 yen more than the sense10 base, which Sharp ties to more memory, up to four OS upgrades and six years of security updates.
    Learn it15 min · 4 steps

    Key ideas

    On device inference
    Running the AI model on the phone's own chip, so photos are not uploaded to a server for editing.
    Inpainting
    Filling a removed region of an image with plausible pixels, which is how bystanders get erased.
    EXIF metadata
    Hidden data saved with a photo, such as time, device and GPS location, which pixel edits do not remove.

    Steps

    1. Read the ITmedia Mobile launch article for the feature list, prices and the background processing note.
    2. Read the AI Watch article for more detail on the camera AI features.
    3. Compare what text masking hides in pixels with what EXIF metadata may still reveal.
    4. Self check: if a user uploads a masked photo to your app, what location data could still leak?
    Try it30 min · 5 steps
    You need: Your own phone photos and a free photo editor with object removal, such as the one built into your phone or Google Photos; No API key needed

    Steps

    1. Take three photos of your own: one street scene with a sign, one of food, one through a window.
    2. Use your phone's built in object removal to erase a person or object and blur a sign or house number.
    3. Inspect each edited file's metadata with your phone's photo info screen or a free EXIF viewer and note whether GPS location remains.
    4. Upload one edited photo to a private album in an app you use and check whether location is kept or stripped.
    5. Write down which privacy steps the editor handled and which you still had to do by hand.

    Small angles to try

    • Compare pixel masking with simply turning off location tagging in the camera.
    • If you build an upload feature, check whether your backend strips EXIF location.
    • Time how long on device edits take with airplane mode on.
  16. #16Trends × tech

    Card payments × chat agents: ChatGPT Wallet pays with Face ID approval

    OpenAI announced ChatGPT Wallet on Nov 13 in the US, built with American Express, Mastercard and Visa, so its agent can pay merchants during a conversation; Japanese outlets ITmedia and Gigazine covered it on Nov 16. Every payment shows an approval screen with merchant, item, total, card and any shipping or recurring terms, and the user confirms with Face ID or Touch ID. The card network then issues a payment token, so neither the merchant nor the agent ever sees the real card number.

    Heat · Covered by ITmedia NEWS and Gigazine on Nov 16 and by Digital Trends and PCWorld in English; no Hacker News thread was checked directly.
    What people foundOpenAI itself warns that approving in Wallet does not confirm an order, so users should wait for the merchant's confirmation. Access is limited to US ChatGPT Pro subscribers aged 18 or older with US issued Amex, Chase, Mastercard or Visa credit cards; spending limits set in advance by amount, category and time are only planned, and ITmedia reports no date for Japan.
    Learn it15 min · 5 steps

    Key ideas

    Network tokenization
    The card network swaps the real card number for a stand in number, expiry and code that only work for that use, so a leak at the merchant exposes nothing reusable.
    Human in the loop approval
    The agent prepares the purchase, but a person must confirm each charge with a biometric check before money moves.
    Approval record
    Each confirmed payment is logged as evidence of exactly what the user authorized, which matters in disputes.

    Steps

    1. Read the ITmedia NEWS article on ChatGPT Wallet first for the full flow from adding a card to the token.
    2. List the fields on the approval screen and note which ones a fraud system or support agent would need.
    3. Compare the token flow with how Apple Pay already hides card numbers from merchants.
    4. Read the caveats: some sites block agent purchases, and approval is not an order confirmation.
    5. Self check: if a customer disputes an agent purchase, which number appears on the receipt and who handles the return?
    Try it30 min · 5 steps
    You need: A text editor or a free diagram tool, plus a spreadsheet; no ChatGPT Pro account, no card and no API key needed.

    Steps

    1. Pick one of your own past online orders and write down its merchant, item, total, card brand and shipping terms.
    2. Mock up the approval screen described by ITmedia using those fields, on paper or in a free diagram tool.
    3. Draw the sequence: agent prepares order, user approves with biometrics, network issues token, merchant charges token, merchant sends confirmation.
    4. Mark every point where the flow can fail, such as a site that blocks agents or an approval with no order confirmation.
    5. Write three rules a merchant fraud system should apply to tokenized agent orders, and check each against your diagram.

    Small angles to try

    • Design the planned spending limit feature: what amount, category and time rules would you allow?
    • Write a support script for a customer whose receipt shows an unfamiliar token number.
    • Compare the flow with a guest checkout on a site you own and list what changes.
  17. #17Models & launches

    iOS 27.3 betas put daily limits on Apple Intelligence cloud features

    MacRumors reports that the iOS, iPadOS and macOS 27.3 betas test a free daily allowance per device for Siri AI and Apple Intelligence features that run in the cloud, resetting around midnight local time with no rollover. Paid iCloud+ plans raise the cap: 2x at 200GB, 5x at 2TB, 10x at 6TB and 20x at 12TB, while the 50GB plan adds nothing. Four categories are limited, including third party apps that use Apple Foundation Models in Private Cloud Compute.

    Heat · MacRumors report on Nov 16; no Hacker News thread found.
    What people foundApple is putting a usage budget on cloud AI the way API providers already do, and each category has its own cap that depends on request complexity and system demand. MacRumors gives no daily numbers, says basic Siri commands stay free and says Apple may adjust limits during the beta, so apps calling Private Cloud Compute need a graceful fallback.
    Learn it15 min · 5 steps

    Key ideas

    Private Cloud Compute
    Apple's server side for Apple Intelligence requests too large for the device, run on Apple hardware with privacy guarantees.
    Per device allowance
    The free quota belongs to each device, not the Apple account, and family members share any iCloud+ pool first come, first served.
    Usage multiplier
    Higher iCloud+ tiers scale the free daily cap by a fixed factor rather than granting unlimited use.

    Steps

    1. Read the MacRumors article first, focusing on the tier table and the four limited categories.
    2. Note which category covers your app or your own daily use.
    3. Read Apple's Foundation Models framework documentation to see where on device and cloud requests differ.
    4. Think through what a user sees when the limit is hit: a warning, a notification, then buy more or wait.
    5. Self check: a family on the 2TB plan shares one pool; what happens to a third party app call after the pool runs out?
    Try it30 min · 5 steps
    You need: Any spreadsheet tool and your own usage notes; no Apple beta, no developer account and no API key needed.

    Steps

    1. For one day, log every Siri or Apple Intelligence request you make and tag it as basic command, multi step personal request, image tool, Shortcuts automation or third party app.
    2. Put the log in a spreadsheet with one row per request and a column for category.
    3. Add columns for each iCloud+ tier with the multipliers 1x, 2x, 5x, 10x and 20x.
    4. Pick a placeholder daily cap per category, labeled clearly as an assumption since Apple has published none, and see which tier your usage would need.
    5. Write a short fallback plan for an app: what it does on device when a cloud request is refused.

    Small angles to try

    • Model a family of four sharing one pool first come, first served.
    • Compare the design with token based rate limits from an API provider you use.
    • List which features in your own app would degrade first under a cap.
  18. #18Models & launches

    Claude builds interactive pages inside canvas using HTML-in-Canvas

    Felix Rieseberg, who works on Claude at Anthropic, showed Claude building sites where real, selectable and clickable HTML lives inside canvas and 3D graphics, using the HTML-in-Canvas web feature Google is developing, with a demo page of 15 examples. The Decoder reports it works in Chrome and Edge 157 and later and in the Claude desktop apps from version 5.11492, but not yet in Safari or Firefox. Users have to ask for HTML-in-Canvas explicitly in the prompt.

    Heat · Covered by The Decoder on Nov 16; no Hacker News thread found.
    What people foundThe Chrome origin trial post from May describes a layoutsubtree attribute on canvas plus drawElementImage for 2D and texElementImage2D for WebGL, which keeps the inner HTML accessible, but it also warns the details have since changed. Generated pages rely on a Chromium only feature, so check Safari and Firefox before shipping.
    Learn it15 min · 5 steps

    Key ideas

    HTML-in-Canvas
    A proposed web feature that lets real DOM elements be drawn into a canvas while staying interactive and readable by assistive tech.
    layoutsubtree
    An attribute on the canvas tag, per the origin trial post, that tells the browser to lay out the canvas's child elements so they can be painted into it.
    paint event
    The event where the page draws the laid out elements into the canvas, so redraws follow DOM changes.

    Steps

    1. Read The Decoder article first for what Claude can now produce and the version requirements.
    2. Read the Chrome HTML-in-Canvas documentation page, then the origin trial post, noting its warning that details changed.
    3. Compare the 2D, WebGL and WebGPU paths and which methods each uses.
    4. Look at how libraries such as Three.js and PixiJS wrap the feature, as listed in the Chrome docs.
    5. Self check: why is drawing real HTML into canvas better for accessibility than drawing text with canvas fill calls?
    Try it30 min · 6 steps
    You need: Chrome or Edge 157 or later, or Chrome Canary with the flag below, and a Claude desktop app 5.11492 or later for the prompt step; no API key needed.

    Steps

    1. Check your browser version; on Chrome Canary, enable chrome://flags/#canvas-draw-element as the origin trial post describes.
    2. Create a local HTML file with the post's minimal example: a canvas with layoutsubtree containing a label and a text input.
    3. Open it locally, then add drawing code following the current Chrome docs, since the API has changed since the trial.
    4. In the Claude desktop app, ask explicitly for an HTML-in-Canvas page, such as a form floating on a rotating 3D card.
    5. Open the result in Chrome, then in Safari or Firefox, and note what breaks.
    6. Tab through the page with the keyboard and check that inputs inside the canvas still take focus.

    Small angles to try

    • Ask Claude for a fallback that renders plain HTML when the feature is missing.
    • Inspect the generated page with a screen reader on your own machine.
    • Compare a 2D version with a WebGL version of the same layout.
  19. #19Tools & open source

    Copilot in VS Code 1.146 adds agent Inbox; Local harness ends Feb 2027

    GitHub's weekly Copilot changelog adds an experimental Inbox mode to the Agents window in VS Code 1.146, gathering agent sessions from all workspaces, listing those that need attention first and letting the rest be dismissed, held or snoozed. Agents can also start model comparisons and find, install or uninstall extensions from the Agents window, both experimental. The Local harness will be removed in February 2027, and the Copilot harness is already the default for new sessions.

    Heat · GitHub changelog entry on Nov 16; no popularity signal found.
    What people foundAnyone still running Copilot agent sessions on the Local harness should move to the Copilot harness before February 2027; GitHub links a transition guide for remaining workflows.
    Learn it15 min · 5 steps

    Key ideas

    Agent harness
    The runtime that drives an agent session, deciding how tools, context and models are wired together.
    Agents window
    The VS Code view that lists and manages Copilot agent sessions separately from normal chat.
    Inbox mode
    A triage view that ranks sessions waiting on a human above ones still working or finished.

    Steps

    1. Read the GitHub changelog entry for the Nov 9 weekly release first.
    2. Open the linked transition guide for the Local harness and note what it says about remaining workflows.
    3. Check which harness your own Copilot agent sessions use.
    4. Read the notes on Compare Models and extension management to see what agents may now do on their own.
    5. Self check: what happens to a custom workflow tied to the Local harness after February 2027?
    Try it30 min · 5 steps
    You need: VS Code 1.146 or later with a GitHub Copilot subscription, on your own machine and your own repos; no separate API key needed.

    Steps

    1. Update VS Code to 1.146 and open the Agents window.
    2. Start two small agent sessions in two of your own workspaces, for example adding a test and fixing a lint warning.
    3. Switch the sessions list to Inbox mode and check that the session waiting for your input appears first.
    4. Snooze one session and dismiss another, then confirm the list updates.
    5. Confirm which harness each session uses and, if any use Local, follow the transition guide to move one.

    Small angles to try

    • Enable Compare Models and let an agent compare two models on the same small task.
    • Let an agent search for an extension in the Agents window and check the warning shown for unsupported features.
    • Time how long triage takes with Inbox mode versus the normal list.
  20. #20Trends × tech

    Subscription fatigue × vibe coding: OpenAI PM replaces $12 embroidery app

    Kathy Korevec, a product manager at OpenAI, built ZigZag with ChatGPT after her $300 Brother SKiTCH PP1 embroidery machine needed a $12 a month app, Gigazine reported on Nov 16. The GitHub repo describes an Apache 2.0 Rust project with a macOS editor and a CLI that turns SVG art into stitches, writes PES, DST, EXP, JEF and VP3, and can send designs to the PP1 over Bluetooth LE. The editor also runs a local MCP server so AI agents can create and edit designs.

    Heat · Gigazine feature on Nov 16; the GitHub repo shows 20 stars and 2 forks.
    What people foundThe README says the Bluetooth upload is "Not yet run against a real Skitch" and advises avoiding Skitch firmware updates because the protocol was reverse engineered, so treat it as a prototype. Gigazine says the app works on Mac and Windows, but the repo lists macOS 11 or later for the app, with a CLI that builds without system dependencies.
    Learn it15 min · 5 steps

    Key ideas

    Stitch file formats
    PES, DST, JEF and similar files store the needle path and color changes that a specific brand of embroidery machine reads.
    Reverse engineered protocol
    A device protocol worked out by observing traffic rather than from official specs, so a firmware update can break it.
    MCP server
    A Model Context Protocol endpoint that exposes an app's actions as tools an AI agent can call.

    Steps

    1. Read the ZigZag README on GitHub first, especially the CLI, Skitch and MCP sections.
    2. Read the Gigazine article for the backstory of the $12 a month app.
    3. Skim the repo layout to see how the SVG to stitch engine is separated from the editor.
    4. Read the MCP section and note that the server listens on localhost only.
    5. Self check: why does the README suggest skipping firmware updates, and what would break first if you ignored it?
    Try it30 min · 5 steps
    You need: Rust with Cargo on your own machine, or a Mac on macOS 11 or later for the prebuilt app, and an SVG you made yourself; no embroidery machine and no API key needed.

    Steps

    1. Clone the repo and build the CLI with cargo build --release, which the README says needs no system dependencies.
    2. Run zigzag fabrics to list fabric presets and zigzag info logo.svg on your own SVG to see stats and checks.
    3. Convert it with zigzag convert logo.svg -o logo.pes -o preview.png -o sheet.html and open the preview and the worksheet.
    4. Convert the PES back with zigzag convert logo.pes -o logo.dst to check machine formats round trip.
    5. On a Mac, open the studio and connect it to Claude Code with claude mcp add --transport http zigzag http://127.0.0.1:7878/mcp, then ask the agent to edit the design.

    Small angles to try

    • Try the --fabric thin-cap and --hat options and compare previews.
    • Feed a PNG logo and see how the tracing step changes the stitches.
    • Use the simulated machine the README mentions to try Bluetooth sending without hardware.

Get the Radar by email

One email each evening, Japan time, with that day's items. Opening soon: leave your email and we'll send a confirm link when it starts.