Bug Description

When messages.tts.auto is set to "inbound", the model's tts tool call generates audio successfully, but the MEDIA: directive in the reply is not processed as an audio attachment. This creates a conflict where:

1. Auto-TTS is skipped because it detects the MEDIA: directive in the reply
2. The MEDIA: directive itself is not processed into an audio attachment

Result: No audio is sent to the user. The MEDIA: path appears as plain text.

Environment

• OpenClaw version: 2026.2.3-1
• Channel: Telegram (streamMode: "partial")
• TTS config: auto: "inbound", provider: "edge"
• Default model: minimax-cn/MiniMax-M2.1 (OpenAI Completions API)
• Test model: anthropic/claude-opus-4-5 (Anthropic Messages API) — switched via /model
• OS: macOS (Apple Silicon)

Steps to Reproduce

1. Configure messages.tts.auto: "inbound" with Edge TTS provider
2. Set default model to a non-Anthropic provider (e.g., MiniMax via OpenAI Completions API)
3. Switch active model to anthropic/claude-opus-4-5 via /model opus
4. Send a voice message from Telegram
5. Model calls tts tool → gets MEDIA:/path/to/voice.mp3
6. Model includes MEDIA: line in reply

Expected: Audio sent as voice note attachment
Actual: MEDIA: path appears as plain text, no audio played

Systematic Test Results

All tests performed under anthropic/claude-opus-4-5 (switched from default MiniMax):

Root Cause Analysis

The TTS docs state:

"When enabled, OpenClaw: skips TTS if the reply already contains media or a MEDIA: directive."

This creates a deadlock when the model uses the tts tool:

User sends voice → auto: "inbound" triggers
  → Model calls tts tool → reply contains MEDIA: line
  → Auto-TTS detects MEDIA: → skips auto conversion
  → MEDIA: directive itself is NOT processed as attachment
  → No audio delivered

The design conflict: the system prompt instructs the model to call the tts tool when the user sends voice, but doing so blocks auto-TTS while the MEDIA: directive fails to deliver audio on its own.

Additional Observations

• In previous sessions, MiniMax (OpenAI Completions API) occasionally succeeded with MEDIA: lines
• Gemini Flash (Google API) also succeeded in one instance
• Anthropic API (Opus) failed consistently across all MEDIA: delivery methods
• The message tool with filePath + asVoice: true works 100% as a workaround

Suggested Fix

Consider one or more of:

1. Process MEDIA: directives reliably regardless of model API type
2. Auto-send audio when the tts tool is called successfully, instead of relying on the model to place the MEDIA: line correctly
3. Clarify the intended flow — if auto: "inbound" handles TTS at the system level, the system prompt should not instruct the model to also call the tts tool (as this blocks auto-TTS)

Workaround

Use the message tool to send the TTS-generated audio directly:

1. Call tts tool → get MEDIA: path
2. Use message tool: action=send, filePath=<path>, asVoice=true

This bypasses both the MEDIA: directive and auto-TTS mechanisms entirely.