fix(mcp): Make tool schemas more Gemini-compatible. #8637
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #8637 +/- ##
==========================================
+ Coverage 51.04% 51.07% +0.03%
==========================================
Files 431 431
Lines 30986 31001 +15
Branches 6367 6368 +1
==========================================
+ Hits 15818 15835 +17
+ Misses 13752 13751 -1
+ Partials 1416 1415 -1 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
| } else { | ||
| const rtext = await response.text(); | ||
| try { | ||
| console.dir(JSON.parse(rtext), { depth: null }); |
There was a problem hiding this comment.
How do we know if the response conforms to what the tool expect?
Do we need to do some sort of zod validation?
There was a problem hiding this comment.
The point of the smoke test is to make sure the Gemini API doesn't throw errors based on any of the tool input schemas provided - it's not testing actual tool calling. This is more of a sanity check than a complete solution, but is better than nothing.
There was a problem hiding this comment.
Interesting, what kind of error it may throw?
Does the Gemini API do any validation to make sure the response match the spec?
We can probably add that validation here later.
There was a problem hiding this comment.
If the tool schemas don't conform to the limited schema supported by Gemini, the request will 400 even if it's only one tool out of 35 that are broken. That's bad, so this smoke test ensures that all of our tools are vaild as far as the API is concerned.
github-project-automation
bot
moved this to Approved [PR]
in [Cloud] Extensions + Functions
github-project-automation
bot
moved this from Approved [PR]
to Done
in [Cloud] Extensions + Functions
|
Thank you for this fix! Looking forward to using :) |
Had a report from a fellow Googler that the MCP server was throwing errors when used with ADK. While I had done testing with Genkit, we do some manual massaging of JSON schemas in Genkit to work around Gemini limitations.
This brings some of those workarounds directly to the MCP server. I also added a smoke test that can be run from
scripts/mcp-teststo load all tool definitions into a simple Gemini model call so we can easily verify this going forward (automated would be great but I don't have time to figure that out today).