Dify Community Reports Resent Reasoning History Anomaly; Agent App May Mimic Tags and Fabricate Tool Calls

A compatibility report about Dify’s agent execution flow warrants attention from deployment teams. On October 28, a user reported that the newer Agent App, running on a self-hosted 1.18.0 deployment via dify-agent and paired with the Tongyi plugin and qwen3.8-max, puts reasoning from the previous turn back into a regular assistant message. The report says the model then outputs extra tags and may even cite job IDs that were never returned by a tool. These observations are currently limited to a single deployment. Community issue

The directly verifiable technical issue lies in the history-message conversion layer: when _map_model_response_to_prompt_message() encounters a ThinkingPart, it wraps the content in <think>…</think> and adds it to the text content; an actual ToolCallPart is converted separately into a structured tool call. As a result, a reasoning block that originally had distinct semantics becomes ordinary text visible to the model in the next turn. Dify code

This distinction has a clear interface rationale. Alibaba Cloud’s official documentation says historical reasoning for qwen3.8-max should be passed back through the reasoning_content field, and explicitly warns that concatenating reasoning into content may degrade model performance. preserve_thinking controls whether historical reasoning is included; enabling it also increases input-token charges. The framework and plugin therefore need to preserve field semantics together; retaining only the text visible on screen is not enough to ensure consistent multi-turn agent behavior. Official API documentation

As of this review, the issue remains open, with no confirmed official fix. A contributor also confirmed the code behavior that resends tags and produces a stray closing tag, but said they had not independently reproduced the production tool hallucination. Engineering teams should next focus on history-message round-trip tests, tag splitting in streaming responses, and whether tool calls remain correctly paired with their returned IDs. When evaluating a fix, they should measure task success rates and token costs separately, rather than treating the disappearance of tags as proof that agent reliability has been restored. Investigation and fix planning