Conversation roles must alternate" error with vLLM/Mistral after aborted responses

[jaydaftari]

When using LibreChat with vLLM-served models (Mistral, Mixtral, Llama, etc.), the error "Conversation roles must alternate user/assistant/user/assistant/..." occurs after aborting a response or when there are empty/partial assistant messages in the conversation history.

Steps to Reproduce

1. Connect LibreChat to a vLLM backend serving Mistral/Mixtral
2. Send a message to start a conversation
3. While the assistant is responding, abort/cancel the response
4. Send another message
5. Error occurs: "Conversation roles must alternate user/assistant/user/assistant/..."

Alternative reproduction:

• Any scenario where an assistant message is saved with empty content or unfinished: true

Expected Behavior

LibreChat should sanitize the message history before sending to the LLM by:

• Merging consecutive messages of the same role
• Filtering out empty/unfinished assistant messages
• Ensuring proper role alternation for strict chat template models

Actual Behavior

The raw message history is sent to the LLM, resulting in sequences like:

• user → assistant (empty/partial) → user (treated as user → user)
• user → assistant → assistant → user (consecutive same roles)

This violates the strict chat template requirements of Mistral/Mixtral and other models.

Technical Analysis

Where the issue originates:

Location	Issue
api/server/controllers/agents/request.js:114	Saves partial messages with unfinished: true that may be empty
@librechat/agents formatAgentMessages	Doesn't merge consecutive same-role messages
client.js buildMessages	Doesn't validate role alternation before sending

Evidence of awareness:

• Related discussion in Artifacts UI feature causes issues for Chat Completion on vLLM #4089 where maintainer noted: "I hope to add a new option where it will simply put the system message first"

This could be implemented in:
formatAgentMessages in @librechat/agents
Or as a pre-processing step in buildMessages
Environment
LibreChat version: [your version]
LLM Backend: vLLM
Model: Mistral/Mixtral (or other strict chat template model)
Deployment: Docker/Local
Related Issues
#4089 - Artifacts UI causes issues with vLLM (same root cause)
#11344 - Memory Agent fails with Bedrock (similar role ordering issue, fixed in #11353)
vllm-project/vllm#2112 - Conversation roles must alternate
Additional Context
This issue affects all users connecting LibreChat to:
vLLM with Mistral/Mixtral models
Any LLM backend enforcing strict role alternation
Potentially Bedrock, Azure, and other providers with strict message formatting
The fix would benefit the entire community using non-OpenAI backends that enforce stricter API contracts.