VS Code Extension Guide
Create, develop, and publish VS Code extensions.
For extensions with a management UI, default to a dedicated Activity Bar icon
and sidebar unless another entry point is explicitly chosen. A Marketplace
icon alone does not create that navigation; use the TreeView reference below.
For user-facing extensions, provide a clearly labeled bug-report/feature-request entry in the primary sidebar or management UI and a Command Palette fallback; a README or Marketplace link alone is insufficient. Preview minimal, non-sensitive metadata before opening a fixed feedback destination; leave submission to the user and never auto-attach prompts, raw logs, secrets or private paths.
When to Use
- VS Code extension, extension development, vscode plugin
- Creating a new VS Code extension from scratch
- Adding commands, keybindings, or settings to an extension
- Publishing to VS Code Marketplace
Quick Start
# Scaffold new extension (recommended)
npm install -g yo generator-code
yo code
# Or minimal manual setup
mkdir my-extension && cd my-extension
npm init -y && npm install -D typescript @types/vscode
Project Structure
my-extension/
├── package.json # Extension manifest (CRITICAL)
├── src/extension.ts # Entry point
├── out/ # Compiled JS (gitignore)
├── artifacts/vsix/ # Keep local VSIX archives out of the repo root
├── images/icon.png # 128x128 PNG for Marketplace
└── .vscodeignore # Exclude files from VSIX
Building & Packaging
npm run compile # Build once
npm run watch # Watch mode (F5 to launch debug)
mkdir -p artifacts/vsix
npx @vscode/vsce package --out artifacts/vsix/my-extension-1.0.0.vsix
Keep local .vsix archives under artifacts/vsix/ instead of the repository root, and prune old local builds on a schedule so release artifacts do not pile up.
Done Criteria
- The packaged VSIX installs and activates in an isolated profile; changed user-facing integrations complete their real workflow there
- Primary sidebar entry and commands work; referenced icons are present in the VSIX
- Package size < 5MB (use
.vscodeignore) - README links and feedback UI reach the intended support destination; verify the UI-to-form path with synthetic data without submitting
- Local VSIX artifacts stored outside the repo root and pruned regularly
Quick Troubleshooting
| Symptom | Fix |
|---|---|
| Extension not loading | Add activationEvents to package.json |
| Command not found | Match command ID in package.json/code |
| Shortcut not working | Remove when clause, check conflicts |
Reference Map
| Topic | Reference |
|---|---|
| AI Customization | references/ai-customization.md |
| Code Review Prompts | references/code-review-prompts.md |
| Code Samples | references/ai-customization.md and references/webview.md |
| TreeView | references/treeview.md |
| Webview | references/webview.md |
| Testing | references/testing.md |
| Publishing | references/publishing.md |
| Troubleshooting | references/troubleshooting.md |
| Notifications | references/notification-normalization.md |
Best Practices
Extension Host 境界
- Extension Host 上で動く scanner / provider / TreeView は、同じことができるなら Node 固有の
path/Buffer/ 生fsより VS Code API を優先する。Problems と実ビルドの環境差を避けやすい。 - 自分の拡張に同梱したリソースは、ユーザーのホーム配下や VS Code のインストール先を推測せず、
context.extensionUriとvscode.Uri.joinPathなど extension context から解決する。 - 他の installed extension に同梱されたリソースを読む必要がある場合も、
resources/agents|skills|prompts|instructions|hooks|mcpの既知 root と、manifest のchatAgents/chatPromptFiles宣言を優先して見る。built-in resource とは別の read-only resource として扱い、削除や再インストール導線を混ぜない。 - Runtime の診断ログは
console.logに散らさず、Output Channel ベースの logger に集約する。ユーザーがログを開ける導線も command / notification / README のどこかに用意する。 - 大きい入力の parse や集計を Extension Host スレッドで同期実行しない。上限付きの bytes を
node:worker_threadsへ渡し、cancel / opt-out ではawait worker.terminate()と listener 解除を済ませ、世代 ID が一致する結果だけで cache / UI を更新する。 - transfer するのは専有した非
SharedArrayBufferだけにし、transfer 後は送信側の view を使わない(同じ backing store の view はすべて detach される)。Bufferは pool を共有し得るので、offset 付き view や所有が曖昧な buffer は必要範囲だけを別 buffer へ copy して渡す。 - worker や別プロセスから受け取った結果をそのまま cache / 表示しない。許可キー検査は
Reflect.ownKeys、必須キーの存在確認はObject.hasOwnで行う(Object.keysは非列挙の own property を見逃し、inや素の property 参照は継承値を通す)。 - 例外の
Error.nameやFileSystemError.codeをそのままログ / UI へ出さない。既知 code の allowlist へ正規化し、未知は単一の汎用 code に落とす。path や token が名前に混ざるケースを防げる。
Manifest / Docs / Localization
package.jsonの commands、views、configuration、menus を変えたら、コード上の command ID / setting key と同時に確認する。- Marketplace 表示や設定説明をローカライズしている拡張では、
package.nls.jsonと対象言語のpackage.nls.*.jsonを同じ変更で更新する。 - In
markdownDescription, use native setting references such as`#editor.wordWrap#`, not[label](#editor.wordWrap#): the latter can leave a trailing hash in the Settings@id:filter. Guard each locale's syntax and verify the resolved target. - 設定の並び順や説明を変えたら README の設定表、manifest consistency test、release notes の必要有無までまとめて見る。
Language Model Tools
- Extension-provided LM Tools need strong
modelDescriptionintent phrases. PreferUse when the user asks to ...plus natural-language verbs for create/update/delete/toggle paths, and add a manifest/doc guard so future wording changes do not silently weaken agent-mode tool selection. - For configuration from VS Code Chat, prefer native LM Tools when no external client is required; do not call them an external MCP server. Share validated domain operations with the GUI, expose only necessary actions, and use explicit enable/disable values rather than retry-sensitive toggles.
- Keep
prepareInvocationside-effect free and request confirmation for mutations or sensitive disclosure. Revalidate input, cancellation, trust and the queried revision inside the locked write after confirmation. Host approval UI/policies still apply; custom confirmation text is not a bypass. Never accept secret values through tool arguments or silently widen execution permissions. - Page query results and omit prompt, environment and raw-output payloads by default. Make sensitive detail retrieval explicit with disclosure confirmation; describe which metadata reaches the model and treat stored strings as data, not instructions.
- Return typed outcomes from shared operations; a GUI handler that catches errors and returns nothing cannot prove a tool succeeded. Keep committed IDs/success when only presentation fails, return a sanitized warning and tell the caller to query rather than repeat the mutation. Configuration saved is not work executed.
- Before publishing an extension with LM Tools, inspect the packaged VSIX rather than trusting source files: confirm
extension/package.jsonhas the intended version, expectedcontributes.languageModelToolscount, and nosrc/, test payloads, or sourcemaps unless intentionally shipped.
Generated Sections
START/ENDmarker で囲む generated section は単一の SSOT として扱う。- 重複した marker pair を見つけたら、両方を残して追記せず、内容を統合して marker pair を1つに戻す。
命名の一貫性
公開前にパッケージ名・設定キー・コマンド名を統一:
| 項目 | 例 |
|---|---|
| パッケージ名 | copilot-scheduler |
| 設定キー | copilotScheduler.enabled |
| コマンドID | copilotScheduler.createTask |
| ビューID | copilotSchedulerTasks |
通知の一元管理
通知は共通helperへ集約し、設定値をruntimeで既知enumへ正規化する。重複抑止、ログとの分離、action、securityの設計は Notification Normalization を参照する。