When the tts tool runs, it returns a tool result containing [[audio_as_voice]]\nMEDIA:/path/to/audio.mp3. The delivery pipeline dispatches this correctly as media.

The problem is that the model then generates a follow-up assistant message that re-echoes the same MEDIA: path and adds text on top of it. The delivery pipeline processes both: the tool result's media dispatch AND the follow-up message. This results in the audio being sent twice + unwanted text alongside it.

Example from a session log

1. Model calls tts tool with text
2. Tool result: [[audio_as_voice]]\nMEDIA:/tmp/tts-.../voice-123.mp3 — dispatched as media ✓
3. Model follow-up: [[tts:rate=+5%,pitch=+2%]]\nGoodnight, my love...\n\nMEDIA:/tmp/tts-.../voice-123.mp3 — dispatched again as text + duplicate media

Root cause

Two things contribute:

1. The TTS tool description says "Copy the MEDIA line exactly" — this tells the model to re-output the path in its reply, but the delivery pipeline already sends the tool result's media independently.

2. filterMessagingToolDuplicates only deduplicates text content, not MEDIA: URLs. So even though the same file path appears in both the tool result and the follow-up, it's sent twice.

Suggested fixes

• Remove "Copy the MEDIA line exactly" from the TTS tool description, or change it to instruct the model to reply with NO_REPLY after a successful TTS call
• Extend filterMessagingToolDuplicates to also track and deduplicate MEDIA: paths across tool results and follow-up messages

Workaround

I added explicit instructions in TOOLS.md telling the model not to re-output MEDIA lines from TTS results. This helps but relies on model compliance — a runtime dedup would be more reliable.

Environment

• v2026.2.14 (c1feda1)
• kimi-coding/k2p5 model (but likely affects any model)
• Discord channel, Edge TTS