speaklisted

Give the agent a voice using a local neural text-to-speech engine. ## Core Rule: Never Block on Audio Audio playback takes real-world seconds. Never run speech synchronously or wait for playback to complete before continuing your work. Always fire speech as a background process (WaitMsBeforeAsync: 500 or background shell job) so you can proceed to your next tool call immediately. ## When to Speak - **Task start**: One sentence confirming what you are starting. - **Milestones**: When finishing a major step during a long task. - **Completion**: Clear summary of what was accomplished and verified. - **Errors**: Immediate audible notification if a build, test, or process fails. - **Key findings**: Verbal summary when research or diagnosis yields a crucial discovery. Keep messages short (1 to 2 sentences). Avoid reading raw code blocks, file diffs, or terminal stack traces out loud. Summarize the intent in plain language. ## Reading Configuration Before speaking, check for configuration in this order: 1. `./.speak.md` (repo-local configuration) 2. `~/.config/speak/config.md` (user-global configuration) 3. Hardcoded default fallback: Pocket-TTS at `http://localhost:8000/tts`, voice `eponine`, tone `conversational`. If the active configuration has `frequency: silent` or disables notifications for the current task category, skip speaking and proceed silently. ## How to Call the Active Engine Run the platform driver script in the background to handle auto-starting, volume