[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7BH5EB8HZSDPS1EVW8CS0FC":3,"skill-same:01M7BH5EB8HZSDPS1EVW8CS0FC":60},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7BH5EB8HZSDPS1EVW8CS0FC","windows-mcp",null,"Driving the Windows desktop on `workstation` through the windows MCP server. Use before any `mcp__windows__` tool, when a desktop call blocks or returns \"Desktop busy\", when Snapshot or Screenshot come back empty, or when this agent has no desktop tools and the task needs them.","hadronomy\u002Fskills","hadronomy","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F17086478?v=4","01M7BH4GMWYV32S7AXN0AYYRZC","Pablo Hernández","01M7BH4GP7A5Z0MQX2JNY7BGJN",0,"MIT","https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fin-progress\u002Fwindows-mcp\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: windows-mcp\ndescription: Driving the Windows desktop on `workstation` through the windows MCP server. Use before any `mcp__windows__` tool, when a desktop call blocks or returns \"Desktop busy\", when Snapshot or Screenshot come back empty, or when this agent has no desktop tools and the task needs them.\n---\n\n# Windows desktop\n\nOne physical Windows machine, `workstation`, reached over SSH from this Mac. Its tools\narrive as `mcp__windows__*`.\n\n## Which half you have\n\nTwo ports serve different toolsets. Read your own tool list rather than assuming:\n\n- **Desktop port** — 37 tools, including `Click`, `Type`, `Snapshot`, `Screenshot`, `App`.\n- **Tool port** — 16 tools: `PowerShell`, `FileSystem`, `Registry`, `Service`, `EventLog`,\n  `Network`, `Package`, `Scrape`. No UI tools at all.\n\nWith no desktop tools, UI work is not yours to do. Say so, and let the agent on the\ndesktop port take it.\n\n## The desktop is someone's real machine\n\n`workstation` is in daily use, not a scratch VM: real apps, real windows, a real\nlogged-in session. Leave it as you found it — restore the window you switched away from,\nclose what you opened, and type only into a window you deliberately focused.\n\n## The working loop\n\n`Snapshot` returns the foreground window as a labeled accessibility tree, with element\nnames and refs you can act on directly. That is the primary sense.\n\n1. `Snapshot` to see what is there.\n2. Act: `Click`, `Type`, `Invoke`, `Shortcut`.\n3. `Snapshot` or `Assert` to confirm it landed.\n\n`Screenshot` costs roughly 700 KB a call and tells you less than the tree. Reach for it\nwhen you need actual pixels — a rendering bug, a canvas, an image — not to find a button.\n\n## The lease\n\nOne physical desktop, so desktop calls are serialized across every connected agent. The\nrelay grants a lease on your first desktop call and holds it while you keep working.\n\n**Keep multi-step UI work tight.** The lease releases after 8 seconds idle, or 3 seconds\nwhen another agent is queued. Pause longer than that mid-sequence and another agent can\ntake the desktop between your `Snapshot` and your `Click` — and your click lands\nsomewhere else. For anything past a few steps, use `Plan` then `Apply`: `Apply` runs the\nwhole sequence inside one call, holding the lease throughout, which makes it atomic.\n\n**A busy desktop is a wait, not a failure.** A blocked call waits up to 2 minutes, then\nreturns `isError: true` saying the desktop is busy. Nothing was sent. Wait a few seconds\nand retry the same tool — that is the path that respects the lease. Driving the UI by\nanother route, such as SendKeys through `PowerShell`, bypasses the lease and corrupts\nwhichever agent holds it.\n\n`Kill` and `GuardrailStatus` sit outside the lease and always answer.\n\n## When the desktop looks empty\n\n`Snapshot` returning nothing, or `DisplayInventory` reporting 1024x768, means the server\nis running outside the interactive session: it sees no desktop, and input goes nowhere\nwith no error. This is a session problem, not a transient one — retrying will not fix it.\nThe relay must run in session 1, which needs the user signed in. If they signed out, ask\nthem to sign back in.\n\n## What cannot be automated\n\nThe sign-in screen, the lock screen, and UAC elevation prompts live on the Windows secure\ndesktop, which no user-session process can see or drive. When a UAC prompt appears the run\nis blocked: tell the user and let them click it. Elevated apps cannot be driven from here\neither.\n\n## Long-running work, and showing progress\n\n`PowerShell` blocks and caps at 600 seconds, and it returns everything at the\nend. A cargo build hits that ceiling and gives no feedback until it does. Codex\ncaps its tool calls at 300 seconds, which is tighter still.\n\nDo not reach for the MCP tool for this. The same machine answers plain SSH:\n\n```bash\nssh windows-mcp 'powershell -NoProfile -EncodedCommand \u003Cbase64>'\n```\n\nOutput streams line by line as it is produced, with no timeout of its own, and\nit bypasses the desktop lease entirely because no UI is involved.\n\n**Encode the command.** Quoting does not survive the trip through this shell,\nsshd, cmd, and powershell — it silently collapses and the command runs wrong or\nnot at all. Base64 of UTF-16LE has nothing to escape:\n\n```bash\nenc() { python3 -c 'import base64,sys;print(base64.b64encode(sys.stdin.read().encode(\"utf-16-le\")).decode())'; }\nssh windows-mcp \"powershell -NoProfile -EncodedCommand $(enc \u003C script.ps1)\"\n```\n\nSet `$ProgressPreference = 'SilentlyContinue'` at the top of every script, or\nPowerShell writes CLIXML progress records onto stderr and they land in your\noutput.\n\n**`Start-Process -RedirectStandardOutput` does not work here.** Launched from a\nnon-interactive SSH session it returns a PID and writes a zero-byte file\nforever. Stream over the live connection instead.\n\n### Progress without context cost\n\nEvery line a monitor emits is a conversation message. A build that reports 40\ntimes costs 40 messages. Route progress to the status line instead, which\nrenders in the UI and never reaches the model:\n\n```bash\nscripts\u002Frun-progress.sh inari windows-mcp 'cd C:\\Users\\pablo\\Repos\\inari; cargo build'\n```\n\nRun it with the Bash tool's `run_in_background`. The pipeline overwrites one\nstate file per run in `~\u002F.claude\u002Frun-progress\u002F`, the status line reads that file\nonce a second, and the whole build costs exactly one completion notification.\n\nCargo hides its progress bar when stdout is not a terminal, and draws it with\ncarriage returns rather than newlines. `run-progress.sh` handles both: it sets\n`CARGO_TERM_PROGRESS_WHEN=always` and splits the stream on `\\r`, which turns\n`Building [====>    ] 96\u002F240: tokio` into a live bar.\n\nRequires `statusLine` in `settings.json` pointing at a script that reads\n`~\u002F.claude\u002Frun-progress`, with `refreshInterval: 1`. `scripts\u002Fstatusline-progress.sh`\nis that segment, and composes with an existing status line.\n\nThe protocol-level answer is MCP progress notifications: the client sends a\n`progressToken` and the server emits `notifications\u002Fprogress` against it. That\nneeds changes in `windows-mcp-server` and a client that sends the token, and the\ndocumented fallback when either is missing is to skip silently. The status line\nneeds neither.\n\n## Blast radius\n\n`PowerShell`, `Registry`, `FileSystem`, `Process` and `App` have full system access with\nno sandbox, on a machine that matters. Confirm with the user before deleting, overwriting,\nor changing system state. Every call lands in a tamper-evident audit log that cannot be\nswitched off, so treat your actions as reviewable.\n\n`Credentials` injects a secret into an app without revealing it and has no read mode.\nNever ask for the plaintext.\n","# Windows desktop\n\nOne physical Windows machine, `workstation`, reached over SSH from this Mac. Its tools\narrive as `mcp__windows__*`.\n\n## Which half you have\n\nTwo ports serve different toolsets. Read your own tool list rather than assuming:\n\n- **Desktop port** — 37 tools, including `Click`, `Type`, `Snapshot`, `Screenshot`, `App`.\n- **Tool port** — 16 tools: `PowerShell`, `FileSystem`, `Registry`, `Service`, `EventLog`,\n  `Network`, `Package`, `Scrape`. No UI tools at all.\n\nWith no desktop tools, UI work is not yours to do. Say so, and let the agent on the\ndesktop port take it.\n\n## The desktop is someone's real machine\n\n`workstation` is in daily use, not a scratch VM: real apps, real windows, a real\nlogged-in session. Leave it as you found it — restore the window you switched away from,\nclose what you opened, and type only into a window you deliberately focused.\n\n## The working loop\n\n`Snapshot` returns the foreground window as a labeled accessibility tree, with element\nnames and refs you can act on directly. That is the primary sense.\n\n1. `Snapshot` to see what is there.\n2. Act: `Click`, `Type`, `Invoke`, `Shortcut`.\n3. `Snapshot` or `Assert` to confirm it landed.\n\n`Screenshot` costs roughly 700 KB a call and tells you less than the tree. Reach for it\nwhen you need actual pixels — a rendering bug, a canvas, an image — not to find a button.\n\n## The lease\n\nOne physical desktop, so desktop calls are serialized across every connected agent. The\nrelay grants a lease on your first desktop call and holds it while you keep working.\n\n**Keep multi-step UI work tight.** The lease releases after 8 seconds idle, or 3 seconds\nwhen another agent is queued. Pause longer than that mid-sequence and another agent can\ntake the desktop between your `Snapshot` and your `Click` — and your click lands\nsomewhere else. For anything past a few steps, use `Plan` then `Apply`: `Apply` runs the\nwhole sequence inside one call, holding the lease throughout, which makes it atomic.\n\n**A busy desktop is a wait, not a failure.** A blocked call waits up to 2 minutes, then\nreturns `isError: true` saying the desktop is busy. Nothing was sent. Wait a few seconds\nand retry the same tool — that is the path that respects the lease. Driving the UI by\nanother route, such as SendKeys through `PowerShell`, bypasses the lease and corrupts\nwhichever agent holds it.\n\n`Kill` and `GuardrailStatus` sit outside the lease and always answer.\n\n## When the desktop looks empty\n\n`Snapshot` returning nothing, or `DisplayInventory` reporting 1024x768, means the server\nis running outside the interactive session: it sees no desktop, and input goes nowhere\nwith no error. This is a session problem, not a transient one — retrying will not fix it.\nThe relay must run in session 1, which needs the user signed in. If they signed out, ask\nthem to sign back in.\n\n## What cannot be automated\n\nThe sign-in screen, the lock screen, and UAC elevation prompts live on the Windows secure\ndesktop, which no user-session process can see or drive. When a UAC prompt appears the run\nis blocked: tell the user and let them click it. Elevated apps cannot be driven from here\neither.\n\n## Long-running work, and showing progress\n\n`PowerShell` blocks and caps at 600 seconds, and it returns everything at the\nend. A cargo build hits that ceiling and gives no feedback until it does. Codex\ncaps its tool calls at 300 seconds, which is tighter still.\n\nDo not reach for the MCP tool for this. The same machine answers plain SSH:\n\n```bash\nssh windows-mcp 'powershell -NoProfile -EncodedCommand \u003Cbase64>'\n```\n\nOutput streams line by line as it is produced, with no timeout of its own, and\nit bypasses the desktop lease entirely because no UI is involved.\n\n**Encode the command.** Quoting does not survive the trip through this shell,\nsshd, cmd, and powershell — it silently collapses and the command runs wrong or\nnot at all. Base64 of UTF-16LE has nothing to escape:\n\n```bash\nenc() { python3 -c 'import base64,sys;print(base64.b64encode(sys.stdin.read().encode(\"utf-16-le\")).decode())'; }\nssh windows-mcp \"powershell -NoProfile -EncodedCommand $(enc \u003C script.ps1)\"\n```\n\nSet `$ProgressPreference = 'SilentlyContinue'` at the top of every script, or\nPowerShell writes CLIXML progress records onto stderr and they land in your\noutput.\n\n**`Start-Process -RedirectStandardOutput` does not work here.** Launched from a\nnon-interactive SSH session it returns a PID and writes a zero-byte file\nforever. Stream over the live connection instead.\n\n### Progress without context cost\n\nEvery line a monitor emits is a conversation message. A build that reports 40\ntimes costs 40 messages. Route progress to the status line instead, which\nrenders in the UI and never reaches the model:\n\n```bash\nscripts\u002Frun-progress.sh inari windows-mcp 'cd C:\\Users\\pablo\\Repos\\inari; cargo build'\n```\n\nRun it with the Bash tool's `run_in_background`. The pipeline overwrites one\nstate file per run in `~\u002F.claude\u002Frun-progress\u002F`, the status line reads that file\nonce a second, and the whole build costs exactly one completion notification.\n\nCargo hides its progress bar when stdout is not a terminal, and draws it with\ncarriage returns rather than newlines. `run-progress.sh` handles both: it sets\n`CARGO_TERM_PROGRESS_WHEN=always` and splits the stream on `\\r`, which turns\n`Building [====>    ] 96\u002F240: tokio` into a live bar.\n\nRequires `statusLine` in `settings.json` pointing at a script that reads\n`~\u002F.claude\u002Frun-progress`, with `refreshInterval: 1`. `scripts\u002Fstatusline-progress.sh`\nis that segment, and composes with an existing status line.\n\nThe protocol-level answer is MCP progress notifications: the client sends a\n`progressToken` and the server emits `notifications\u002Fprogress` against it. That\nneeds changes in `windows-mcp-server` and a client that sends the token, and the\ndocumented fallback when either is missing is to skip silently. The status line\nneeds neither.\n\n## Blast radius\n\n`PowerShell`, `Registry`, `FileSystem`, `Process` and `App` have full system access with\nno sandbox, on a machine that matters. Confirm with the user before deleting, overwriting,\nor changing system state. Every call lands in a tamper-evident audit log that cannot be\nswitched off, so treat your actions as reviewable.\n\n`Credentials` injects a secret into an app without revealing it and has no read mode.\nNever ask for the plaintext.\n",{"name":5,"description":7},"main","8b5b348cb089a495146376d39cdff24a55d43baf","skills\u002Fin-progress\u002Fwindows-mcp","https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills","https:\u002F\u002Fgithub.com\u002Fhadronomy","hadronomy.com","Lover of all things code","2026-10-08 07:09:27",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add hadronomy\u002Fskills --skill windows-mcp",{"tool":38,"command":39},"bunx","bunx skills add hadronomy\u002Fskills --skill windows-mcp",{"tool":41,"command":42},"pnpm","pnpm dlx skills add hadronomy\u002Fskills --skill windows-mcp","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fin-progress\u002Fwindows-mcp\u002FSKILL.md",true,[46,52,56],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"scripts\u002Fcargo-progress.pl","SCRIPT","脚本",1977,"https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fin-progress\u002Fwindows-mcp\u002Fscripts\u002Fcargo-progress.pl",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"scripts\u002Frun-progress.sh",1878,"https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fin-progress\u002Fwindows-mcp\u002Fscripts\u002Frun-progress.sh",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"scripts\u002Fstatusline-progress.sh",774,"https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fin-progress\u002Fwindows-mcp\u002Fscripts\u002Fstatusline-progress.sh",[61,68,74,81,88,95],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M7BH58ZA4ESM6WM3MS571BS0","typst-author","Author idiomatic, compilable Typst documents. Use when creating or editing Typst (.typ) files, when a task involves Typst syntax, packages, templates, or bibliographies, or when a Typst build fails and the error needs a fix.","https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fin-progress\u002Ftypst-author\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M7BH56EMBDMTT82379T51HYZ","setup-hadronomy","https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fengineering\u002Fsetup-hadronomy\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M7BH50THYQKDFY428RP08JZX","rust-craft","Craft production Rust the way the top toolchain teams do. Use when writing, reviewing, or restructuring Rust — workspace layout, Cargo manifests, lints, and profiles; type, trait, and error API design; performance, allocation, and profiling work; async and concurrency; CLI construction; test, snapshot, and benchmark suites; crate selection; release, semver, MSRV, and feature flags.","https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fengineering\u002Frust-craft\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M7BH4VAWNMK94M799CXKDDK6","pr-craft","Write pull requests a reviewer can act on. Use when opening, rewriting, or reviewing the body of a pull request — title and description, the problem and approach, call-stack diffs, before-and-after screenshots, verification steps, risks, a review guide, and the tools that produce each part.","https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fengineering\u002Fpr-craft\u002FSKILL.md",[],[],{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":19},"01M7BH4NNP3PPSXHWJTVBMMJNW","effect-craft","Craft Effect TypeScript APIs with the discipline of the Effect and opencode codebases. Use when modeling data with Schema, building services with Context and Layer, naming operations with Effect.fn, exposing modules through barrels, designing hooks or HTTP surfaces, or writing JSDoc with runnable examples — new code, refactors, and API reviews.","https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fengineering\u002Feffect-craft\u002FSKILL.md",[],[],{"skillNo":96,"name":97,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":101,"publishedAt":19},"01M7BH4JB1KYAW2H8519FDG98A","docs-craft","Documentation discipline for code comments, API docs, and docs pages. Use when writing or reviewing doc comments, JSDoc or rustdoc, README sections, or framework docs pages — or when comments and docs read as filler and need cutting.","https:\u002F\u002Fgithub.com\u002Fhadronomy\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fengineering\u002Fdocs-craft\u002FSKILL.md",[],[]]