Provide tool error feedback agent #431
Conversation
…be punished when it commits mistakes
|
I just don´t know a good way to fix the breaking integration tests. It seems to be an issue with the tests themselves, not the application code |
|
Cool - thanks - this was needed. |
Yes, of course. If you could create a new tag it would be helpful, as my environment is set to use uvx to install the latest version by default |
|
I'm releasing now (just completing the e2e suite before pushing the button). |
|
0.3.15 is on pypi now including this PR. This I think is relevant too: modelcontextprotocol/modelcontextprotocol#1303 likely to be included in spec. |
| for content in tool_result.content | ||
| ] | ||
| if tool_result_contents: | ||
| if result.content is None: |
There was a problem hiding this comment.
By Looking at this, I imagine this is line is unnecessary, right?
Great, will test this |
2 participants
Today, when a tool produces an error, tool agent silently drops it and does not return the result to the AppAgent and, therefore, the user cannot send custom messages to punish the AI when the tool was misused.
This PR aims to return the fail to serve as a feedback to the AppAgent and can be passed as a custom message to the LLM or take any other decision based on the error details.