Qiao-yq commented on PR #1558: URL: https://github.com/apache/dubbo-admin/pull/1558#issuecomment-5728150479
> Overall, this is a solid improvement. The phase-specific prompts are clearly separated, and the tests cover rendering and tool binding well. I noticed two small points worth considering. > > The instruction to treat user-provided content as untrusted could be clearer. The runtime passes the current equest as a user message, so direct requests such as “use only logs” should still be followed. It would help to distinguish direct user requests from instructions embedded in pasted content or tool results. > > The new prompt removes the Dubbo metric and label examples, while the Prometheus tool schema does not provide metric names. For a failure-rate question, the model could query a nonexistent metric and receive an empty result even when data exists. This is a potential regression, not a reproduced failure. It may be worth retaining a small set of canonical metrics or testing a few diagnostic cases with the configured model. Thanks, both points are valid. I clarified that direct user constraints should be followed, while instructions embedded in pasted content or tool results remain untrusted. I also restored the dashboard-backed Dubbo metrics and labels, with complete regression coverage. Fixed in 839d11f. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
