feat: add tool support for antropic streaming #14758
Conversation
|
Tested with 'Coder' and 'claude-3-5-sonnet-latest' on your branch with the following prompt: '@coder Implement the interface MCPServerManager in the same file (packages/ai-mcp/src/common/mcp-server-manager.ts). Don't ask me anything, jut do it.' got: |
|
hmm, the tool parameters were not extracted. I will debug. |
|
one tool had no schema defined as it doesn't need parameters, openai accepts undefined in that case, but anthropic doesn't. |
eneufeld
force-pushed
the
pr/anthropic-tools
branch
2 times, most recently
from
b1da705 to
0ebd8be
Compare
eneufeld
force-pushed
the
pr/anthropic-tools
branch
from
0ebd8be to
e4ddde2
Compare
| @@ -75,7 +75,11 @@ export class ToolCallPartRenderer implements ChatResponsePartRenderer<ToolCallCh | |||
| let responseContent = response.result; | |||
| try { | |||
| if (response.result) { | |||
| responseContent = JSON.stringify(JSON.parse(response.result), undefined, 2); | |||
| let resultObject = response.result; | |||
There was a problem hiding this comment.
Should we do this conversion in the model implementation? Is this a corner case?
There was a problem hiding this comment.
As I found out, MCP tools behave differently to our tools.
Our tools return the result usually as a string. But MCP returns it in the format that Anthropics API expect, which is an object with a content field which can be a string or an array of objects with content.
So the rendering completely exploded for me.
I can move the logic to the backend and make sure that the result written to the tool result object is a string.
What it does
This adds tool support to anthropic models when using streaming.
How to test
Follow-ups
Breaking changes
Attribution
Review checklist
Reminder for reviewers