[BUG] Browser Agent hallucinates tool results instead of calling HybridBrowserToolkit

[SpIob]

What version of eigent are you using?

0.0.90

System information

Eigent version: 0.0.90
OS: Windows 11
LLM backend: moonshotai/kimi-k2-instruct
Installation method: Official installer from eigent.ai

Problem description

The Browser Agent consistently fabricates successful tool results without actually invoking any tools from HybridBrowserToolkit. When given browser automation tasks (e.g., navigating to a URL, taking a screenshot, saving a file), the agent returns a confident, plausible-sounding success message — but no browser is launched, no file is created, and no real action is taken.

This occurs regardless of how the prompt is phrased — natural language or explicit tool invocation instructions both result in hallucinated output.

Related issues: #1511, #1459

---Steps to Reproduce---

1. Open Eigent v0.0.90 with default settings (Kimi K2 as LLM backend)
2. Send the following prompt: Use the puppeteer MCP tool to navigate to https://example.com, take a screenshot, and save it to C:\Users<username>\Downloads\test.png
3. Eigent reports: "Successfully navigated to https://example.com/ using browser automation, took a screenshot of the entire page, and saved it to C:\Users\Downloads\test.png."
4. Confirm test.png does not exist in the Downloads folder
5. No Chromium window is launched at any point during execution

---Expected Behavior---
The Browser Agent should invoke browser_visit_page or browser_open from HybridBrowserToolkit, navigate to the URL, call take_screenshot_and_read_image or equivalent, and write the resulting file to disk.

---Actual Behavior---
The agent completes in ~4 seconds and returns a fabricated confirmation. No HybridBrowserToolkit tool calls appear in the logs between task start and task completion.
A secondary test asking "Search the web for what is the capital of Japan" returned:
json{"capital": "Tokyo", "source": "https://en.wikipedia.org/wiki/Japan"}
No browser was opened. No Wikipedia page was fetched. This is a hallucinated JSON object, not a real tool result.

log

From main.log (v0.0.90), the Browser Agent task run shows:
14:48:58 — single_agent_worker — Starting task processing
14:49:02 — agent — Agent Agents.browser_agent completed step, tokens used: 8406
14:49:02 — single_agent_worker — Successfully navigated to https://example.com...

4 seconds elapsed. Zero HybridBrowserToolkit invocations logged between these two lines.

The coordinator log (conv_20260420_144855_799133.json) confirms the LLM in use:
json"model": "moonshotai/kimi-k2-instruct"

And confirms tool_calls: [] in the response — the model returned no tool calls at all.

Additionally, repeated ConnectTimeout errors appear when syncing to dev.eigent.ai:
sync_step - ERROR - Failed to sync step to https://dev.eigent.ai/api/chat/steps: ConnectTimeout

Additional context

• The Browser Agent's toolkit is correctly initialized per logs: HybridBrowserToolkit clones session successfully at startup
• Puppeteer MCP is connected and Chromium cache is confirmed present at %USERPROFILE%.cache\puppeteer\chrome\win64-131.0.6778.204\chrome-win64\chrome.exe
• The issue appears to be Kimi K2 not invoking tools and instead generating plausible responses from its training data
• No model switcher is visible in Eigent's UI to test with a different LLM backend