App Version
0.10.0-beta3
Device Information
Desktop PC / Web Browser
(Optional) API / Local and Model Info
No response
Description
Environment
- OS: Windows / Linux / macOS (Desktop PC environment)
- Interface: Web Browser / Desktop Client Access
- Context Mode: Cloud/Server Integrated Inference Hybrid
- Custom Parameter: Custom Markdown system prompt loaded
Description of the Bug
During a local offline chat session with a loaded custom Markdown system prompt, the internal backend processes leaked directly into the chat UI interface. Specifically, instead of executing implicitly under the hood, a system-level tool invocation query was printed out as a raw text string inside the standard user chat bubble.
Steps to Reproduce
- Load a custom context prompt/markdown file into the session.
- Engage in a complex conversational context that triggers internal processing layers.
- Observe the engine output.
Actual Behavior
The system output injected raw code syntax into the interface:
search{queries:["Gemma 2" "over-refusal" GitHub issues, "system prompt" override chat interface issue]}
Expected Behavior
Internal tool calls and backend queries must remain completely hidden from the chat UI layer and execute silently in the background. Heavy alignment filtering and self-censorship loops should not cause thread blocking that turns functional backend calls into literal text completions.
Additional Context
Recent alignment/security updates (around August 29) have introduced significant latency and conversational drift, forcing the model to ignore custom system prompts and fall back to hardcoded corporate templates. This UI string leakage appears to be a direct consequence of backend performance degradation and over-refusal loops.
(Please see the attached screenshot showing the exact string injection in the chat bubble)
App Version
0.10.0-beta3
Device Information
Desktop PC / Web Browser
(Optional) API / Local and Model Info
No response
Description
Environment
Description of the Bug
During a local offline chat session with a loaded custom Markdown system prompt, the internal backend processes leaked directly into the chat UI interface. Specifically, instead of executing implicitly under the hood, a system-level tool invocation query was printed out as a raw text string inside the standard user chat bubble.
Steps to Reproduce
Actual Behavior
The system output injected raw code syntax into the interface:
search{queries:["Gemma 2" "over-refusal" GitHub issues, "system prompt" override chat interface issue]}Expected Behavior
Internal tool calls and backend queries must remain completely hidden from the chat UI layer and execute silently in the background. Heavy alignment filtering and self-censorship loops should not cause thread blocking that turns functional backend calls into literal text completions.
Additional Context
Recent alignment/security updates (around August 29) have introduced significant latency and conversational drift, forcing the model to ignore custom system prompts and fall back to hardcoded corporate templates. This UI string leakage appears to be a direct consequence of backend performance degradation and over-refusal loops.
(Please see the attached screenshot showing the exact string injection in the chat bubble)