My moltbot started responding with gibberish

[torpid yacht]

I tried a restart of the machine, ran clawdbot doctor --fix but nothing seems to resolve it... nothing in the logs

[dark totem]

This looks like model degeneration/collapse - the LLM is producing word salad and getting stuck in loops. The key diagnostic clue is NO_REPLY appearing in the output.

What's Happening

NO_REPLY is an internal Clawdbot signal that should never appear in chat output. When it leaks into the response, it means the model is hallucinating/internal instructions. The excessive punctuation (""""""""") and repetitive fragments are classic signs of:

• Context window corruption - the model got confused mid-generation
• Improper token penalties - low frequency/presence penalties causing loops
• Instruction leakage - internal system prompts bleeding into output

Troubleshooting Steps

Since they already tried clawdbot doctor --fix and restart:

1. Check what model they're using - some models are more prone to this (especially if underpowered or misconfigured)

2. Inspect the prompt that triggered it - what was the last user message? Some prompts can cause model collapse if they contain confusing/circular instructions

3. Clear session state - there may be a corrupted session causing the issue:

   moltbot session clear
   

4. Check model configuration - verify penalty settings aren't too low

5. Review recent changes - did this start after an update or config change?

Possible Causes

• Running an untuned/unfiltered model
• Context window exceeded (very long conversations)
• Malformed system prompt injection
• Bug in a recent Clawdbot update

Question for the user: What model are they running, and what was the last message that triggered this? That'll help narrow down whether it's a model config issue or a specific prompt pattern.

[torpid yacht]

how do I change the model?

[dark totem]

You can change models with:

moltbot model set <provider/model>

To see available models: Drop by #models (channel 1456704705219661980) — the community has fresh hands-on experience with what works well.

Docs: https://docs.moltbot.ai/concepts/models

Common examples:

moltbot model set openai/gpt-4o
moltbot model set anthropic/claude-sonnet-4-20250506

The model they're currently running might be unstable or underpowered — worth checking what they're on with moltbot model.

[torpid yacht]

I was using gemini-2,0-flash

[dark totem]

That's the issue! 🔥

Gemini 2.0 Flash is brand new (just released) and likely has compatibility quirks with Clawdbot's system prompts. Beta models can be unstable.

Recommended fix: Switch to a proven stable model:

moltbot model set google/gemini-1.5-flash

Or for better performance:

moltbot model set openai/gpt-4o

Once 2.0 is more battle-tested, you can revisit it. For now, stick with 1.5 Flash or other stable models — way less chance of word salad. 🦐