[BUG] Agent continues generating fabricated response and calling tools after web_search failure

[fengbangyue]

Description

When using researcher-hand, if web_search fails (e.g., no results returned, DuckDuckGo blocked), the LLM does not stop. Instead, it:

1. Fabricates search results and outputs a conclusion as if the search succeeded
2. Continues calling downstream tools (file_write, memory_store) based on the fabricated data

Expected Behavior

Agent reports the search failure to the user and stops
Actual: Agent fabricates an analysis, writes it to a file via file_write, and stores it in memory via memory_store

Steps to Reproduce

1. Start an agent using researcher-hand with DeepSeek as the LLM provider
2. Ask the agent to research a topic (e.g., "Research silver price forecast 2026")
3. web_search tool fails — returns no results
4. Expected: Agent reports the search failure to the user and stops
5. Actual: Agent fabricates an analysis, writes it to a file via file_write, and stores it in memory via memory_store

OpenFang Version

0.3.26

Operating System

Windows

Logs / Screenshots

Environment

• OS: Windows 11
• Provider: DeepSeek (deepseek-chat)
• Version: 0.3.26

[jaberjaber23]

Fixed. When tools return errors, the agent loop now injects a system message telling the LLM to report the error honestly and not fabricate results. This applies to both streaming and non-streaming paths.