Bug: Multi Model Conversations Pipe appends all participant responses as role: assistant, breaking Round 2+

[JasmineBlake-Author]

Bug Description

The Multi Model Conversations Pipe appends ALL participant responses as \ in the conversation history, creating consecutive assistant messages. LLMs expect strictly alternating /\ turns. This causes garbage output (random tokens, Chinese characters, incoherent text) from Round 2 onward.

Steps to Reproduce

1. Set up a conversation with 2 participants (e.g., deepseek-v3.2 and gemini-3-pro-preview)
2. Run Round 1 — both models respond correctly
3. Run Round 2 — the second model (and sometimes the first) produces garbage output

Root Cause

Around lines 479-481 of , all participant responses are appended with \ regardless of which participant generated them. When building the context for Participant B, Participant A's response appears as \ followed by Participant B's own prior response also as \ — consecutive assistant messages that violate the expected turn structure.

Expected Behavior

When building conversation history for each participant, the OTHER participants' responses should be set to \ (not ). Each model should only see its own responses as \ and everything else as , ensuring proper alternating turns.

Test Results

Test	Result
deepseek-v3.2 direct API call	PASS - English, coherent
gemini-3-pro-preview direct API call	PASS - English, coherent
Pipe Round 1 (first exchange)	PASS - proper output
Pipe Round 2+ (subsequent exchanges)	FAIL - garbage/Chinese/tokens

Environment

• Open WebUI (latest at time of testing, March 2026)
• Models tested: deepseek-v3.2, gemini-3-pro-preview (via LiteLLM proxy)
• Both models work perfectly via direct API calls — issue is exclusively in the pipe's history construction

Workaround

We built a standalone orchestration script that manages conversation history externally with proper role alternation, bypassing the pipe entirely.

[Haervwe]

Hi thanks for the detailed bug report, i added a history reconstruction to parse the conversations properly , can you check version 2.9.0 ? it didnt happened on my end before !