LM Studio fails to handle context overflow gracefully after large tool response

[lucianommartins]

Which version of LM Studio?
LM Studio 0.4.11-1

Which operating system?
Linux (but also reproducible on Windows and MacOS)

What is the bug?
LM Studio fails silently and breaks model interactions when a tool response (e.g., a full Wikipedia page) exceeds the configured Context Length of the loaded model.
Instead of warning the user that the context window has been exceeded (e.g., showing 12526/4096 tokens in the input), the app continues processing and causes the model to output either complete garbage text (e.g., Bee<|"|> extraar zerod toara/?) or generic fallbacks like "I am a large language model, trained by Google." without answering the question.

Logs
Here is an example of the garbage output received after a massive tool response was injected:

Bee<|"|> extraar zerod toara/?

To Reproduce
Steps to reproduce the behavior:

1. Load a model with a default context length set to a low value (e.g., 4096 tokens).
2. Enable a tool or MCP server (like the Wikipedia tool) that can return a large volume of data.
3. Ask a question that triggers the tool to fetch a large amount of content (e.g., "Explain Capoeira" which returns ~12k tokens).
4. Observe that the input tokens indicator in the chat box shows that the limit is exceeded (e.g., 12526/4096).
5. See the model break and output garbage or an irrelevant generic response without any error warning from LM Studio.

[lake-effect]

Some of these problems occur separately also. Running in Strict memory protection mode on 4.16.2, I routinely get into situations like this:

The model continues to run and emit reasonable data for a while until tool calls fail with a similar failed output to the OP's near the close of a tool call tag. I suspect that the context is allowed to increase until one of the memory guardrails is hit, regardless of what the user picks for maximum context size.

Did context compaction regress in some previous version? According to #1620, maybe yes.

[ptrumpet]

I have found the a similar related behaviour when calling the LM studio directly via a HTTP POST with streaming enabled in the request.
The trigger is to supply a context too large for the model.
If the API is called with streaming on, I only get a single chat.start message and then the connection is closed with no error reported in the stream or HTTP error code.
With streaming disabled, I get the correct error reported from the HTTP POST reply.
If I make a request that is invalid in some other way (e.g. set thinking on when the model does not support it), I still get a proper error message reported but only as a full HTTP error 400 as a plain JSON error message with no streaming protocol entered into.