Empty AI replies orphan the conversation context

[milesian01]

What happened?

Note: this is ChatGPT's analysis of the attached conversation export (Available_Tools Overview.json).

What happened?
When using the DesktopCommanderMCP server, AI sometimes returns a completely empty reply (its text field is "", no error) on the very first user message or later turns. LibreChat treats that blank reply as a valid assistant turn and advances the conversation pointer so when you re-ask, all prior history is dropped.

What did you expect to happen?
Empty replies should be retried or ignored and not advance the conversation, preserving context.

Version Information

LibreChat: v0.7.8 (ghcr.io/danny-avila/librechat-dev latest 0312c0377b84 2 weeks ago 1.02GB)

Ollama v.0.8.0

Steps to Reproduce

Where to look in Available_Tools Overview.json
Available_Tools Overview.json

At messages[1]: AI reply has "text": "" immediately after the user’s “what tools are available?”
At messages[2]: the user repeats the question
At messages[3]: AI finally replies with the tool list

Later, at the turn where the user says “let’s try the apk add approach,” the AI reply is empty; after the user repeats that request, the conversation has lost all the preceding context about which command to run

What browsers are you seeing the problem on?

Chrome

Relevant log output

From attached [Available_Tools Overview.json](https://github.com/user-attachments/files/20499079/Available_Tools.Overview.json)

messages[1]: { "sender":"User", "text":"what tools are available?" }
messages[2]: { "sender":"AI", "text":"" }
messages[3]: { "sender":"User", "text":"what tools are available?" }
messages[4]: { "sender":"AI", "text":"The available tools include …" }

messages[NN]: { "sender":"User", "text":"let’s try the apk add approach" }
messages[NN+1]: { "sender":"AI", "text":"" }
messages[NN+2]: { "sender":"User", "text":"let’s try the apk add approach" }
messages[NN+3]: { "sender":"AI", "text":"<responds—but no longer knows which package was being discussed>" }

Screenshots

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Local models have more limited context than not. It’s likely you hit 2 different context limits, which happens easily with tool outputs.

We may also need to define the proper context limit for this model and I will check that. In any case, you can adjust the max context in parameters in the right side panel to test this.

[milesian01]

Thanks for the suggestion, but I don’t think it’s a context window issue:

• I’m using qwen3:30b-a3b-q8_0 with num_ctx=40000 in Ollama and nvidia-smi shows ~38 GB used. LibreChat keeps that same model process alive for the entire chat session which means it's clearly using the full 40k window.
• This bug even shows up on the very first user message (before any tool calls or built‐up history).
• The exported conversation JSON I shared (which includes all think tokens) is under 30k tokens by OpenAI’s tokenizer, so it's an indication that the chat didn't exceed 40k.
• I’ve observed the same orphaned‐context behavior at many points in different chats.

What would you suggest as the next steps for diagnosing and fixing this?

[jhmonroe]

Seconding your issue but also Danny's reply!

1. Similar problem with empty replies sporadically. And also had context set very high (12k) and was only trying to have it read a small html file inside a folder with many as a test.

I got the MCP filesystem working and it's able to see the directories and tell me their names—so I know the MCP server is working correctly. But replies are often empty and when I ask it to read the files inside the directory I get empty replies after it said it would read the first file inside as I asked it to.

Using LMStudio + LibreChat inside Docker (which took some finangling to map the directories)

Did you try a different model? When I switched to Gemma 3 27B, turned on KV cache Quantization (to possible alleviate the caching issue) I didn't have the problem anymore!

2. To Danny's point, I did notice that even a few simple questions very quickly added up to 3k in tokens since by the last question LibreChat was sending the entire history of the conversation. LMStudio reported in logs such as:
   [CacheWrapper][WARN] Tried to trim '65' tokens from the prompt cache, but could not: Cache is not trimmable. Clearing the cache instead.