v0.5.1

Changed

• Migrated audio capture dependency from deprecated @analyticsinmotion/micstream to decibri (v1.0.0). The package was renamed; the API is identical with no functional changes.

Security

• Bumped undici from 7.22.0 to 7.24.1 — fixes 6 CVEs including 3 high-severity WebSocket and HTTP vulnerabilities (PR #8).
• Bumped flatted from 3.3.4 to 3.4.2 — security patch for devDependency (PR #9).
• Bumped picomatch from 4.0.3 to 4.0.4 — security patch for devDependency (PR #10).

v0.5.0

Fixed

• Native engine dependencies (sherpa-onnx, micstream, sentencepiece-js) are now correctly bundled in the published .vsix. The v0.4.0 CI build omitted engine/node_modules because it was gitignored and never installed in CI. SherpaEngine failed with MODULE_NOT_FOUND on all platforms. Fixed by adding cd engine && npm install to each CI job before packaging.
• Duplicate wake phrase detections within 3 seconds are now suppressed. A buffered DETECTED message could arrive on stdout after the engine process was killed, causing a second command to fire.

Changed

• Release CI now produces platform-specific .vsix files (win32-x64, darwin-arm64, linux-x64), each containing the correct native audio binaries for that platform.

v0.4.0

Added

• Cross-platform speech engine (SherpaEngine) using sherpa-onnx keyword spotting. Supports Windows, macOS, and Linux. Runs as a child process under system Node.js so native audio addons load against the correct Node.js ABI, with no Electron conflicts.
• Cross-platform microphone capture via @analyticsinmotion/micstream (PortAudio). This resolves a fundamental blocker: VS Code's Electron runtime cannot load native audio addons, so there was previously no way to capture microphone audio on macOS or Linux. Running micstream under system Node.js in the engine child process bypasses this entirely.
• wakeWord.engine setting to select the speech engine: auto (platform default), windows (Windows System.Speech), or sherpa (sherpa-onnx, cross-platform). Defaults to auto.
• wakeWord.nodePath setting to override the Node.js executable path used by SherpaEngine. Useful when Node.js is installed via nvm, fnm, or a non-standard location not on VS Code's PATH.
• Engine indicator in the status bar showing the active engine (Windows or Sherpa) while listening. Click the indicator to open engine settings.
• Changing wakeWord.engine or wakeWord.nodePath in Settings now takes effect immediately; the engine restarts without reloading the window.
• Status bar countdown during cooldown: after a wake phrase fires, the status bar shows a live second-by-second countdown (Wake: 30s → Wake: 29s → ...) instead of a static "Wake: Active" message. Gives clear feedback on how long until listening resumes.

Changed

• Countdown uses a clock icon ($(clock)) rather than the spinning sync icon. The spinning icon reset its CSS animation every second when the status bar text was updated, causing a visible jerk. The clock icon is static and appropriate for a timed countdown.
• Documentation updated to be editor-neutral: "VS Code" replaced with "your editor" or "the editor" in settings descriptions, README subheading, and How It Works section. The extension works in Cursor, Windsurf, and other VS Code forks. The docs now reflect that. Technical content (platform requirements, command IDs, architecture) is unchanged.
• Both engines now implement a shared ISpeechEngine interface, enabling clean engine switching and shared event handling.

Fixed

• Switching the speech engine during an active cooldown no longer cancels the countdown or breaks subsequent detection. The countdown continues normally; the new engine starts when it expires.
• Model download now follows HTTP redirects. GitHub release URLs return 302 → CDN. Downloads previously failed silently on first install.

v0.3.0

Added

• Dedicated "Wake Word" output channel in the Output panel for all logging (debug, warnings, errors, detections).
• Startup diagnostics logged automatically: route count, threshold, OS, VS Code version.
• "Show Log" action on error toasts — opens the output channel directly.
• Dual logging: output channel always, debug console when running via F5 (dev mode).
• Pause on focus loss (wakeWord.pauseOnFocusLoss) — pauses listening when VS Code loses focus, resumes on regain. Off by default.
• Phrase aliases: phrase field now accepts a string or array of strings, mapping multiple trigger phrases to one command.
• Per-route cooldown: optional cooldownSeconds on each route entry, overrides the global setting.
• SVG guard in npm run lint — catches blocked SVG references in README.md before commit.

Changed

• Speech engine protocol: DETECTED:<phrase> now includes confidence as DETECTED:<phrase>|<confidence>.
• npm run lint now also runs scripts/check-readme.js to catch vsce-blocked SVGs.

v0.2.1

Fixed

• Consent flow now uses try-finally to ensure isStarting flag is always reset, even if the dialog throws.
• Removed duplicate isActive state tracking in favour of speechEngine.isListening.
• pause() now clears pending retry timers to prevent unexpected restarts during mic handoff.
• resume() resets the retry counter so a fresh session gets the full 3 retries.
• killProcess() now removes all process listeners before nulling the reference, preventing listener accumulation across crash/retry cycles.

v0.2.0

Added

• Configurable confidence threshold (wakeWord.confidenceThreshold) — adjustable from 0.1 to 0.9 with safe clamping.
• Automatic retry with exponential backoff when the speech engine crashes (up to 3 retries at 2s, 5s, 10s delays).
• Non-Windows activation guard: commands register as stubs with an informational message on macOS/Linux.
• Troubleshooting section in README with common problems and solutions.
• npm run lint step added to the release CI workflow.

Changed

• Speech engine switched from async RecognizeAsync() to synchronous Recognize() polling loop. The async approach crashed with exit code 2 when spawned from the VS Code extension host.
• Error reporting from the PowerShell script now uses stdout (ERROR: prefix) instead of stderr to avoid PowerShell CLIXML wrapping issues.

Fixed

• Grammar construction: each wake phrase is now added individually to Choices via a foreach loop instead of passing the array directly. The previous approach created a sequence grammar (all phrases as one utterance) instead of alternatives.
• deactivate() no longer crashes on non-Windows platforms where speechEngine is not initialized.

v0.1.1

Added

• AGENTS.md with project conventions, commands, architecture, and boundaries for AI coding agents.
• ESLint configuration (.eslintrc.json) with TypeScript support.
• CHANGELOG now follows Keep a Changelog format with semantic versioning.

Changed

• Updated extension display name to "Wake Word".
• Updated extension description to "Voice control for VS Code with customizable wake word. Fully local, zero config."
• Updated extension icon to fill the full 128x128 frame with dark background.
• Bumped @typescript-eslint/parser and @typescript-eslint/eslint-plugin from ^6.18.0 to ^8.56.1.
• Activation event changed from onStartup to onStartupFinished for deferred loading.
• Auto-start now deferred by 1 second to allow VS Code to finish initialising.
• Consent dialog no longer disables enableOnStartup when dismissed.
• Configuration change listener now only restarts the engine when wakeWord.routes changes, not all settings.

Fixed

• PowerShell injection vulnerability: switched from double-quoted to single-quoted strings for user-provided phrases in generated PowerShell script.
• executeCommand is now properly awaited with correct error handling for async failures.
• Removed premature isActive = true assignment in resumeListening() -- state is now set by the "started" event handler.
• Removed duplicate setStatusBar("handed-off") call on wake word detection.
• Added concurrency guard (isStarting flag) to prevent multiple consent dialogs.
• stop() no longer emits "stopped" when already stopped.
• resume() no longer clears paused state prematurely -- defers to "started" event.
• Added dispose() method to SpeechEngine for proper event listener cleanup on deactivation.
• Added route validation to filter out entries with empty phrase, label, or command.
• Fixed unnecessary escape character warnings in PowerShell split patterns.
• Fixed catch (err: any) to use catch (err: unknown) with proper type narrowing.
• Buffered stderr to handle PowerShell CLIXML chunked output, parsed on process exit.
• Added _killedIntentionally flag to suppress spurious error messages from intentional process kills.
• Removed dead customText parameter from setStatusBar().

v0.1.0

Changelog

0.1.0

• Initial release
• Zero-config wake word detection: no API keys, no accounts, no model downloads
• Local speech recognition via Windows built-in System.Speech.Recognition engine
• Constrained grammar matching for fast, accurate wake phrase detection
• Multi-phrase routing: map any spoken English phrase to any VS Code command
• Default routes for "Hey Copilot", "Hey Claude", and "Computer" out of the box
• Mic handoff: pauses wake word listening and releases the mic when an assistant is triggered
• Automatic resume after configurable cooldown period
• First-run consent flow with clear microphone usage disclosure
• Status bar indicator with listening state
• Auto-start on VS Code launch (after consent)
• Compatible with VS Code, Cursor, Windsurf, and other VS Code forks
• Published to VS Code Marketplace and Open VSX Registry