Summary

Updated TTS tool description to prevent models from re-echoing MEDIA paths, which caused duplicate audio delivery.

Problem

The old description "Copy the MEDIA line exactly" caused models to re-output the MEDIA path in their follow-up message. Since the delivery pipeline already dispatches media from tool results independently, this resulted in:

1. Audio sent from tool result ✓
2. Audio sent again from follow-up message ✗ (duplicate)

Fix

Changed instruction to tell models to reply NO_REPLY after successful TTS calls, since the audio is automatically delivered from the tool result.

Before: Copy the MEDIA line exactly.
After: reply with NO_REPLY after a successful call to avoid duplicate messages.

Fixes #17991

Greptile Summary

Updated the TTS tool description from instructing models to "Copy the MEDIA line exactly" to "reply with NO_REPLY after a successful call to avoid duplicate messages." This prevents models from re-echoing the MEDIA: path in follow-up responses, which was causing audio to be delivered twice - once from the tool result and again from the assistant's message. The system already extracts and delivers media from tool results automatically via splitMediaFromOutput in src/media/parse.ts, so the follow-up message should be silent (NO_REPLY) to prevent duplication.

Confidence Score: 5/5

• This PR is safe to merge with no risk - it's a simple, focused documentation change to a tool description
• The change is minimal (single line), well-documented, and directly addresses the stated problem. The instruction to use NO_REPLY is consistent with existing patterns in the codebase (seen in message tool, memory flush, and other tools). The media delivery pipeline already handles extracting and dispatching MEDIA: paths from tool results, so models should remain silent after TTS calls.
• No files require special attention

_{Last reviewed commit: ec1f198}