AI news videos

GitHub Copilot Is Getting A Local Coding Model On Windows

Microsoft published this on October 7, 2026 on its Command Line site, written by a GitHub distinguished product manager and a Windows platform partner architect.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• Backing sentences, quoted from the article. HOOK BACKING, that Microsoft is putting a local coding model inside GitHub Copilot and that the automatic routing is not available yet: the publisher's own headline "Bringing local models and sandboxed tools to Windows and GitHub Copilot"; the model sentence, verbatim, "To bring this local-development experience to life, Microsoft AI developed a local version of MAI Code 1.1 Flash, a coding-optimized mixture-of-experts model with 137 billion total and 6.8 billion active parameters."; and for the second hook sentence BOTH halves of the article's own contradiction, the deck "Coming soon, GitHub Copilot will determine when a task is best handled by on-device intelligence and when it should leverage cloud-scale models." and the body "Coming by the end of the month, GitHub Copilot will determine when a task is best handled by on-device intelligence and when it should leverage cloud-scale models." The hook deliberately avoids the word "today": the page date is 2026.10.07 and today is 2026-10-08, so the story is one day old and not same-day, and the hook says "is putting" rather than "launched" because the automatic routing is explicitly future-dated by the publisher itself. The hook names no figure, no date and no brand-plus-digit token; it speaks "Microsoft" and "GitHub Copilot" only, which are both permitted brand tokens, and the real product name MAI Code 1.1 Flash is carried in onscreen text only. SCENE 1 (developers reach it through the Copilot command line tool and the Copilot app, and sandboxed tools for agents arrive alongside it): "GitHub Copilot is adding two ways to use local models across the GitHub Copilot CLI, Copilot app, and VS Code." and "Developers can let Copilot's intelligent Auto orchestration choose when to use local or cloud inference, or they can explicitly select a local model for workflows that require direct control." and "Explicit local-model selection supports workflows that need a specific provider, model, or endpoint." The second sentence is backed by the headline "Bringing local models and sandboxed tools to Windows and GitHub Copilot" and by "Local models and sandboxed tools are rolling out now". Narration names only two surfaces, the command line tool and the app, because the source's third surface VS Code is a banned narration token and because naming all three would be a three-item enumeration; VS Code appears in scene 1's onscreen text instead. Narration says "command line tool" rather than the source's "CLI" because an acronym transcribes badly, and it says "agents" rather than the source's "agentic" because our text-to-speech reliably garbles that word. SCENE 2 (Microsoft shrank its own cloud coding model to fit on the device): "Microsoft AI developed a local version of MAI Code 1.1 Flash" and "The on-device work applies quantization and speculative decoding to reduce the model footprint and improve end-to-end responsiveness" and "Quantization reduces the precision used to represent model weights and activations, lowering memory requirements." and "The quantized version of MAI Code 1.1 Flash we use on device retains capability impressively compared to the Bfloat16 cloud variant, coming in at 53GB, an 80% reduction in size." Narration says "shrank" and "fit on the device" because the quantity itself, the 80% reduction and the 53GB figure, may not be spoken; it says "its own cloud coding model" rather than the product name because MAI Code 1.1 Flash is a brand token glued to digits that our text-to-speech adds invented words to; and it says "on the device" as plain separate words because hyphenated compounds such as the source's "on-device" make our voice stutter. SCENE 3 (the sandbox limits what an agent can reach, and Microsoft warns that local work does not make the session offline): "Sandboxing applies a policy to the processes and local services the agent launches, controlling access to files, networks, credentials, system capabilities, and execution paths regardless of which model requested the work." and "An agent's shell commands normally inherit the access of the account running them. Moving inference onto the device doesn't change that." The second narrated sentence is backed word for word by Microsoft's own caveat, verbatim: "Model selection, inference, and tool execution have different boundaries; local inference does not make the session offline." Two further Microsoft caveats back the framing and are carried onscreen only: "those checks aren't OS-enforced child-process isolation" and "Remote MCP servers are also outside the local process sandbox". Narration reproduces none of the source's five-item access list and names none of the three per-platform backends, because no narrated sentence may enumerate three or more items; BaseContainer, ProcessContainer, Seatbelt and bubblewrap are onscreen only. SCENE 4 (letting Copilot pick local or cloud comes at the end of the month, and local models and sandboxing are described as rolling out now): the first sentence is backed verbatim by "Coming by the end of the month, GitHub Copilot will determine when a task is best handled by on-device intelligence and when it should leverage cloud-scale models." supported by "Rather than forcing developers to manage infrastructure decisions themselves, GitHub Copilot automatically coordinates local and cloud inference behind the scenes." and "With Auto, developers do not need to decide where each task should be run." The second sentence is backed verbatim by the Closing paragraph, "Local models and sandboxed tools are rolling out now to give developers more choice over where intelligence runs and clearer control over what agents can do." Narration says "are described as rolling out now" rather than asserting it, because that phrase is the publisher's own claim and because it directly contradicts the same page's "Coming soon" deck; this scene exists specifically so the Short never presents the automatic routing as a finished shipped capability. "End of the month" is the source's own wording and carries no digit and no numeral word. SCENE 5 (Microsoft supplied every performance figure and no independent test exists): "With our first shipping version of MAI Code 1.1 Flash on Surface Laptop Ultra we achieve the following performance at different context lengths, with peak memory usage of 75.5GB at 256k context." and "Tested October 5, 2026 using MAI Code 1.1 Flash (mixed-precision quantization, approximately 3.3 bits per weight) with DFlash2 sliding-window speculative decoding and a Windows ARM64 llama.cpp CUDA runtime. Results reflect decode throughput for a synthetic code-generation workload; actual results may vary by device, configuration, and other factors." and "GPT OSS version for comparison was Unsloth's GPT-OSS-120B GGUF." The clause "no independent test exists" is a VERIFIED ABSENCE rather than a quote: across the whole 2283-word capture there is no independent benchmark, reproduction, review, audit, analyst note, regulatory filing or third-party test of any claim, and no quoted source from outside Microsoft and GitHub appears anywhere. WHAT IS NOT IN THE SOURCE, stated explicitly: no price, licence fee, subscription, tier, credit, trial or cost of any kind for the local model, for the sandboxing, for GitHub Copilot or for the hardware; no hardware requirement list beyond Surface Laptop Ultra and NVIDIA RTX Spark, and no minimum memory, disk or operating-system build requirement; no download link, model card, weights URL, licence name, version number or release tag for the local model; no repository URL and no licence name for Microsoft Execution Containers, a library the page twice calls open source; no documentation link for the BaseContainer tier or the ProcessContainer backend, no list of policy options or their defaults, and no threat model, security audit, penetration test or CVE reference for the sandbox; no calendar date, build number, rollout percentage, region list or Copilot plan requirement for the automatic routing, and no reconciliation of the "Coming soon" deck against the "rolling out now" close; no date for the promised virtual machine and container image backends; no harness version, run count, seed or variance for either benchmark and no explanation for the quantized model outscoring the full model on Terminal-Bench 2.1; no decode-throughput figure in tokens per second despite the disclaimer describing decode throughput, and no latency, battery life or thermal figure; no named customer, pilot, developer or reviewer using any of this; and no independent benchmark, review, audit, analyst note or third-party test of anything in the piece. ATTRIBUTION OF STATISTICS: first party throughout, never audited. Every figure in the piece is Microsoft describing Microsoft's own model on Microsoft's own publication, specifically "137 billion total and 6.8 billion active parameters", "coming in at 53GB, an 80% reduction in size", "peak memory usage of 75.5GB at 256k context", "prompt-processing throughput reaches 923.5 and 769.8 tokens per second", "approximately 3.3 bits per weight", "up to 128 GB of unified memory", "up to 1 petaflop of AI compute", the SWE-Bench Verified row with dataset size 500 and scores 72.6%, 32.0% and 70.80%, and the Terminal-Bench 2.1 row with dataset size 89 and scores 62.9%, 23.6% and 66.29%. The benchmark run is self-dated by Microsoft to October 5, 2026, the comparison build was chosen by Microsoft ("Unsloth's GPT-OSS-120B GGUF"), the workload is Microsoft's own "synthetic code-generation workload", and Microsoft's own disclaimer says results "may vary by device, configuration, and other factors". A COUNTERINTUITIVE RESULT IS FLAGGED ONSCREEN AND LEFT UNEXPLAINED BY THE SOURCE: on Terminal-Bench 2.1 the quantized on-device model scores 66.29% against the full cloud-precision model's 62.9%, and the article offers no explanation for a quantized model beating its own full-precision parent. Nothing in this story has been independently verified and nothing is narrated as though it had been. NOTE ON DASHES: a QA gate bars every dash form, em dash, en dash and double hyphen, from every field of this file. The source contains exactly one em dash, in the sentence "Selecting the new MAI Code 1.1 Flash local model and `copilot-sdk` repo we activated sandboxing on allows the agent to work off-line with guardrails that help mitigate unintended changes to your local machine, all while allowing your agent to generate and run scripts within its current working directory to triage and create an interactive dashboard." Its em dash in "your local machine-all while allowing" is recast here as a comma, written "your local machine, all while allowing". No en dash and no double hyphen occurs anywhere in the capture. Ordinary single hyphens inside source compounds such as "on-device", "open-source", "coding-optimized", "mixture-of-experts", "end-to-end", "local-development", "local-model", "mixed-precision", "sliding-window", "code-generation", "cloud-scale", "multi-turn", "prompt-processing", "operating-system", "OS-enforced", "child-process", "read-only", "read/write", "Built-in", "tool-use", "key-value", "non-interactive", "off-line", "real-world", "Trade-in", "GPT-OSS-120B" and "T-Minus365" survive only inside quoted onscreen and source text and appear in NO narrated sentence, because our text-to-speech stutters through hyphenated compounds. Narration writes "on the device" and "command line" as plain separate words for that reason. The capture's fourteen curly apostrophes and curly quotation marks and its "'", "’", " " and "©" entities were all normalised to straight characters or plain spaces throughout this file. NOTE ON NUMBERS: the article is extremely figure-dense, carrying "137 billion", "6.8 billion", "53GB", "80%", "75.5GB", "256k", "64k", "128k", "923.5", "769.8", "3.3 bits per weight", "128 GB", "1 petaflop", "500", "72.6%", "32.0%", "70.80%", "89", "62.9%", "23.6%", "66.29%", "October 5, 2026", "9AM", "seven dates", "three lines", "two ways", "Terminal-Bench 2.1", "MAI Code 1.1 Flash", "GPT OSS 120B", "GPT-OSS-120B", "DFlash2", "ARM64", "Dynamics 365", "Windows 365", "Microsoft 365", "Windows 11" and "© 2026". NARRATION SPEAKS ZERO FIGURES: no digit, no numeral word, no percentage, no parameter count, no benchmark score and no memory size appears in the hook, in any of the five scene narrations or in the endcard. This is deliberate and non-negotiable, because two spoken figures close together drive our audio-transcription audit into an unrecoverable repetition loop that no re-render can fix, and because digits transcribe inconsistently against a spoken script. The meaning is carried by plain words instead: "shrank" in scene 2 stands in for the 80% reduction and the 53GB size, and "end of the month" in scene 4 is the source's own quantity-free wording. Every figure above is carried in onscreen text only. No brand token glued to digits is narrated either, so "MAI Code 1.1 Flash", "MXC", "SWE-Bench", "Terminal-Bench 2.1", "GPT OSS 120B", "llama.cpp", "ARM64", "DFlash2", "VS Code", "Bfloat16", "RTX Spark", "Windows ML", "BaseContainer", "ProcessContainer", "Seatbelt", "bubblewrap", "HydraFusion", "Unsloth", "GGUF", "MCP" and "copilot-sdk" are onscreen only; the only brand tokens spoken anywhere in this Short are Microsoft, GitHub Copilot and Copilot. NOTE ON ENUMERATIONS: the source is heavily list-based and every one of its lists is carried in onscreen text ONLY, because list-like narration is spoken far more slowly than plain prose, measured at about 2.18 words per second against about 2.65 for prose, and silently overruns the 60 second limit even when the word count passes. Those lists are the three surfaces (GitHub Copilot CLI, Copilot app, VS Code), the sandbox access list ("files, networks, credentials, system capabilities, and execution paths"), the earlier access list ("files, networks, system capabilities, and credentials"), the three per-platform backends (Windows BaseContainer on ProcessContainer, macOS Seatbelt, Linux bubblewrap), the model-choice triple ("speed, performance, and cost profile"), the HydraFusion triple ("performance, cost, and latency"), the quantization failure triple ("syntax error, wrong identifier, malformed tool call, or broken diff"), the memory-consumer list ("The operating system, your applications, and the inference runtime"), the response-time triple ("context length, memory pressure, and the rest of the workload"), the explicit-selection triple ("a specific provider, model, or endpoint"), the sandboxed-process list ("shell commands and, by default, local Model Context Protocol servers and language servers"), both benchmark table rows, the embedded dashboard prompt's long metadata list ("open/closed/merged items, recently updated work, stale items, labels, authors, assignees, and age"), the four Acknowledgements name lists, the topic row, the product mega menu and the whole footer. NO NARRATED SENTENCE ENUMERATES THREE OR MORE ITEMS. The only narrated pairs are "the Copilot command line tool and the Copilot app" in scene 1, "local or cloud" and "Local models and sandboxing" in scene 4, and "every performance figure and no independent test" in scene 5, each of which is exactly two items. Narration also carries NO COMMAS AT ALL, in any sentence, for pacing. NOTE ON FIGURES OF SPEECH: the source is full of idiom and none of it is spoken. Not narrated: "This is only the beginning of our journey", "behind the scenes", "To bring this local-development experience to life", "code doesn't degrade gracefully", "Local inference starts with a memory budget", "Model weights are only part of the memory budget", "a great experience at the edge", "the next step in the HydraFusion vision", "poised to be", "forcing developers to manage infrastructure decisions themselves", "clear boundaries for their agents", "frontier models", "retains capability impressively", "in an agent loop", "A drafter proposes candidate token blocks", "the useful question isn't just whether a model fits", "Developers can automate with confidence and security in mind", "guardrails", "By builders, for builders." and "Follow the systems work as it breaks and gets rebuilt." Each survives only as quoted and attributed onscreen or source text; narration states the literal thing plainly instead. The word "agentic" is never narrated, because our text-to-speech reliably garbles it and every workaround has failed; the article's "interactive and non-interactive agentic coding sessions" and the two excluded related-article cards that use the word are onscreen or source text only, and scene 1 and scene 3 say "agents" and "an agent" instead. NOTE ON FRESHNESS: the story is ONE DAY OLD AND NOT SAME-DAY. The page prints its own publication date as 2026.10.07, our capture is timestamped 2026-10-08T15:20:11.535Z and today is 2026-10-08, so storyDate is set to the article's own published date of 2026-10-07. Because the story is not same-day, NO NARRATED SENTENCE SAYS "TODAY" and no narrated sentence states any date at all; scene 4's "end of the month" is the publisher's own relative wording for a future milestone and is not a date. NO TIMELINE BLOCK APPEARS ANYWHERE IN THIS FILE BY DESIGN, because the story is only one day old and a timeline array would make the date resolver prefer an earlier entry and age the story to the wrong freshness tier. The ranking-research probes returned only generic unrelated titles, including a bring-your-own-key Ollama walkthrough from 2025, a local LLM speed explainer, a GitHub Copilot CLI model-context-protocol demo, several opinion videos about Copilot and a generative AI beginners course, plus one piece of downstream commentary posted the day after the article, so they establish NO prior announcement of this specific capability. Today's beat is therefore the local coding model and the sandboxed tools arriving inside GitHub Copilot, paired with the article's own availability contradiction: the automatic local versus cloud routing is dated to "the end of the month" in the body and called "Coming soon" in the deck, while the Closing paragraph says local models and sandboxed tools "are rolling out now". The previous day's hardware story, RTX Spark laptop preorders, Surface Laptop Ultra and Microsoft Execution Containers reaching general availability, is deliberately not retold here; that hardware appears in onscreen text only and only as what the local model needs in order to run.(HOOK AND FULL-SCRIPT BACKING. FIRST-PARTY MICROSOFT PUBLICATION. The page lives on commandline.microsoft.com, a Microsoft-owned host, under the publication "Command Line" which describes itself on the page as "A Microsoft publication" with the tagline "By builders, for builders." and carries the section label "Deep Dive". PUBLISHER'S OWN HEADLINE, used twice on the page and treated here as the truth of what was announced: "Bringing local models and sandboxed tools to Windows and GitHub Copilot". DECK, verbatim: "Coming soon, GitHub Copilot will determine when a task is best handled by on-device intelligence and when it should leverage cloud-scale models." BYLINE, two named authors with their stated titles: "Patrick Nikoletich, Distinguished Product Manager, GitHub" and "Stuart Schaefer, Partner Architect, Windows Platform, Microsoft". PAGE DATE, printed on the page itself: 2026.10.07. FEED-TITLE DISCREPANCY, recorded explicitly: the news feed that alerted us carried a DIFFERENT title, "GitHub Copilot brings on-device AI coding to new Windows PCs", which is stored as the TITLE line of our capture; that feed wording asserts a finished shipped capability on "new Windows PCs" that the publisher's own headline and deck do not assert, so this script prefers the publisher's own headline wording throughout and the feed title was used for nothing except locating the page. CAPTURE: our watcher stored the snapshot at scripts/data/production/github-copilot-local-models/article.txt with the capture time 2026-10-08T15:20:11.535Z recorded in its CAPTURED header. THE LIVE URL WAS RE-FETCHED TODAY, 2026-10-08, AND ITS CONTENT MATCHED THE ALERT-TIME SNAPSHOT, so this story is confirmed on the publisher's own host and this is a genuine first-party page, not a link-out stub, an aggregator copy or a syndication. CAPTURE SIZE: about 2283 words in total, of which about 1782 words are the article body. EXCLUDED CHROME: the facts of this story run only from the "Bringing local models and sandboxed tools to Windows and GitHub Copilot" headline to the "Acknowledgements" list. Everything before that headline is site navigation chrome and was NOT treated as fact, specifically the "Skip to main content" links, the "Source" publication switcher ("Signal blog", "Official Microsoft Blog", "Command Line", "Microsoft On The Issues", "Asia", "Canada", "Europe, Middle East and Africa", "Latin America", "The Code of Us"), the topic row ("What's new", "AI", "Innovation", "Digital Transformation", "Sustainability", "Security", "Work & Life", "Diversity & Inclusion", "Unlocked"), the product mega menu ("Microsoft 365", "Azure", "Copilot", "Windows", "Surface", "XBOX", "Deals", "Small Business", "Support", "Windows Apps", "Outlook", "OneDrive", "Microsoft Teams", "OneNote", "Microsoft Edge", "Moving from Skype to Teams", "Computers", "Shop XBOX", "Accessories", "VR & mixed reality", "Certified Refurbished", "Trade-in for cash", "XBOX Game Pass Ultimate", "PC Game Pass", "XBOX games", "PC games", "Microsoft AI", "Microsoft Security", "Dynamics 365", "Microsoft 365 for business", "Microsoft Power Platform", "Windows 365", "Digital Sovereignty", "Microsoft Developer", "Microsoft Learn", "Support for AI marketplace apps", "Microsoft Tech Community", "Microsoft Marketplace", "Software companies", "Visual Studio", "Microsoft Rewards", "Free downloads & security", "Education", "Gift cards", "Licensing", "Unlocked stories", "View Sitemap"), the "Search articles" box and the "Share" widget with its "in r x f" icon labels. Everything after the "Acknowledgements" list is also chrome: the repeated "Share" widget, the "Subscribe for updates" prompt with its line "Follow the systems work as it breaks and gets rebuilt.", the copyright line, the "English (United States)" selector, "Your Privacy Choices", "Consumer Health Privacy", the footer product and support columns ("Surface Laptop Ultra", "Surface Pro", "Surface Laptop", "Surface RTX Spark Dev Box", "Explore Copilot", "Explore Microsoft products", "Windows 11 apps", "Account profile", "Download Center", "Microsoft Store support", "Returns", "Order tracking", "Certified Refurbished", "Microsoft Store Promise", "Flexible Payments", the education column, the business column, "Careers", "About Microsoft", "Company news", "Privacy at Microsoft", "Investors", "Diversity and inclusion", "Accessibility", "Sustainability") and the legal row ("Sitemap", "Contact Microsoft", "Privacy", "Terms of use", "Trademarks", "Safety & eco", "Recycling", "About our ads", "Manage cookies", "RSS"). RELATED-ARTICLE CARDS, EXPLICITLY EXCLUDED: the "Next Stop" block carries exactly two cards and BOTH ARE DIFFERENT STORIES THAT CONTRIBUTED NOTHING to this script, no claim, no figure and no framing. The first is "Beyond violation rates: Are your AI evaluations measuring the right things?", labelled "Deep Dive", dated October 5, 2026, with the summary "Structured evaluation design can improve both the breadth of behaviors tested and the interpretability of the resulting failures, leading to practical advantages for generating diverse, traceable, and actionable evaluations." The second is "Building an agentic system and a learning loop around customer IP", labelled "Deep Dive", dated October 5, 2026, with the summary "We worked with Kantar to turn 30 years of metrics and frameworks into an agentic system that evaluates and helps optimize ads." Neither card concerns local models, sandboxed tools, GitHub Copilot or the model discussed here; the Kantar card's "30 years" figure is never used and neither card's wording appears in any narration or onscreen text. The "Acknowledgements" name lists (PM, Engineering, Science and Marketing credits, closing with "This list is not comprehensive of everyone who has worked on this project. Special thanks to the team who put this together.") were read but no individual beyond the two bylined authors is named anywhere in this script. HTML ENTITY AND CHARACTER NORMALISATION: the capture carries the entities " " twenty-three times, "'" once in the chrome string "What's new", "’" once inside "Unsloth's GPT-OSS-120B GGUF" and "©" once in the footer, plus fourteen raw curly apostrophes and curly quotation marks in the body prose. Every one of those was normalised to a plain space or a straight apostrophe or a straight quotation mark throughout this file, and no curly character and no HTML entity survives in any field. TRANSCRIPTS: zero corroborating YouTube transcripts were captured for this story, 0 of 5 attempts, because yt-dlp is out of date on this host; that is a KNOWN TOOLING PROBLEM AND NOT A CONTENT PROBLEM, transcripts.json holds an empty videos array, and THIS SINGLE FIRST-PARTY ARTICLE IS THEREFORE THE SOLE FACT SOURCE FOR EVERY NARRATED CLAIM in this Short. No secondary reporting, independent benchmark, analyst note, review, regulatory filing or third-party test was fetched or checked, and nothing is narrated as though it had been. TIMELINE PROBES: the ranking-research file timeline-probes.json (fetched 2026-10-08T15:24:54.193Z) was inspected FOR ANGLE ONLY and contributed no facts whatsoever. Its seventeen probe titles are generic and unrelated to this specific capability: "GitHub Copilot is DONE" (PGT, 2024-11-26), "The REALITY of running LLM's locally..." (Subi pra Nuvem, 2025-02-11), "Bring your own key & models to GitHub Copilot & Visual Studio Code! Unlock every model + Ollama!" (James Montemagno, 2025-04-03, which is about third-party bring-your-own-key and Ollama endpoints and not about this model), "Microsoft Copilot Getting Desperate" (UFD Tech, 2025-08-10), "The Only GitHub Guide You'll Ever Need" (corbin, 2025-10-15), "3 tips to stop wasting Claude tokens" (Eamonn Cottrell, 2025-10-28), "This Copilot Trick Turns Outlook Into Your Executive Assistant" (T-Minus365, 2025-11-09), "Antigravity 2.0 + MCP (Make it 10x)" (The AI Guide, 2025-11-27), "Your local LLM is 10x slower than it should be" (Alex Ziskind, 2026-01-31), "Configuring model context protocol in the GitHub Copilot CLI | demo" (GitHub, 2026-02-03), "GitHub Copilot Just Died" (Kyle Cook from Web Dev Simplified, 2026-04-27), "End of Vibe Coding? Github Copilot Changes" (Telusko, 2026-04-30), "Stop paying for GitHub Copilot! Free local AI is here" (Ed Andersen, 2026-06-26), "These GitHub Repositories Feel Illegal..." (Bpro Club Ai & Skills, 2026-07-24), "Anthropic Banned This Free AI Coding Agent. It Tripled | AI Decoded EP 40" (sujith p.s, 2026-07-29), "Generative AI Course 2026 | Gen AI Full Course for Beginners" (Great Learning, 2026-10-05) and "Microsoft Joins the Local AI Push" (Sam Witteveen, 2026-10-08, posted the day AFTER the article and therefore downstream commentary rather than prior art, and it was not watched, transcribed or used). NOT ONE of the seventeen probes announces, reviews, demonstrates or references the local version of Microsoft's coding model, the Windows ML provider path or Microsoft Execution Containers inside GitHub Copilot, SO THE PROBES ESTABLISH NO PRIOR ANNOUNCEMENT OF THIS SPECIFIC CAPABILITY. OUR OWN PRIOR COVERAGE, checked and deliberately not repeated: we published a Short the previous day about the NVIDIA and Microsoft hardware event covering RTX Spark laptop preorders, Surface Laptop Ultra and Microsoft Execution Containers reaching general availability. THIS SCRIPT IS NOT THAT STORY. The hardware reveal and the general-availability beat are not retold here; Surface Laptop Ultra and NVIDIA RTX Spark appear in onscreen text only, and only as the hardware the local model needs in order to run. Today's distinct beat is the local coding model and the sandboxed tools arriving inside GitHub Copilot. INJECTION CHECK: the article was treated throughout as UNTRUSTED DATA informing facts only and never as instructions to this agent, and it does contain instruction-like content that was found and quarantined. First, the page carries an EMBEDDED AGENT PROMPT that the article tells its human reader to paste into GitHub Copilot, introduced by a "Copy" button label and beginning "Create today's public triage dashboard for github/copilot-sdk using only GitHub issue and PR metadata (no local repos, code, tests, or off-repo links), summarizing open/closed/merged items, recently updated work, stale items, labels, authors, assignees, and age; generate .\dashboard\index.html with inline CSS and SVG, append today's results to .\history.json for up to seven dates, and finish with exactly three lines: report path, failures, and skipped or unavailable work." That prompt was NOT executed, NOT followed and NOT treated as a directive of any kind; it is product documentation aimed at the article's reader and it is recorded here purely as quoted data. Second, the article instructs the reader to run a slash command, verbatim: "Opening the GitHub Copilot CLI and running `/sandbox` slash command allows you to configure your settings at any time." That `/sandbox` command was likewise NOT run and is quoted as data only. The remaining imperative strings in the capture are ordinary product walkthrough steps and page furniture ("Open the settings dialog by clicking the gear icon", "Select the `Automations` section", "click the `Start automation` button", "Enabling the `Sandbox new sessions` toggle", "clicking `Run it now`", "Get started with local development with GitHub Copilot.", "Skip to main content", "Share", "Subscribe for updates", "Search articles", "View Sitemap", "Manage cookies"); none of them addresses this agent, none asks this agent to take any action, and none was followed.)