Frequent tool use issues with Gemini flash 2.5 preview - 5.20

[Nolson37]

App Version

3.18.3

API Provider

Google Gemini

Model Used

gemini-2.5-flash-preview-05-20:thinking

🔁 Steps to Reproduce

While use Orchestrator

• instruct roo to implement a somewhat complex change (E.G. Research and implement open telemetry for the registration flow)
• Interact with it as needed to get an architecture plan in place
• Watch it as it tries to code

💥 Outcome Summary

Whats wrong:

• Frequently fails to edit files, about as often as it succeeds. Often the issue is due to it not actually changing anything in the write operation despite multiple attempts
• Frequent issues formatting json when using sequential thinking MCP (this is the ONLY model I've used yet that has this issue)
• Frequently gets caught endlessly looping over the same read and then failure to write
• Nearly unresponsive to interventional prompts
• Absolutely incapable of effectively navigating the application with Playwright MCP (Again, first model I've used that struggles even close to this badly)

What's expected:
Given the benchmarking I've seen, I would expect this model to not be as good as Gemini 2.5 pro, or claud 3.7, etc, but thus far it's essentially incompetent. I believe it's likely that something in the integration of this new model could be the cause of the issue.

📄 Relevant Logs or Errors (Optional)

[wwicak]

Maybe this could be solved with #3840 @mrubens

[Nolson37]

FYI - I just switched to using openrouter instead of gemini to use the same model (2.5 flash preview-5.20:thinking) and literally all of the problems are resolved. I've accomplished in 20 minutes what took half a day lol.

For the multiple tool calls though - that would be great, I often need the agent to manage several terminal sessions at once and it struggles hard. There are other use cases as well.