Repository navigation
Add outputSchema to ModelContextTool and RegisteredTool - #254
beaufortfrancois wants to merge 8 commits into
Conversation
Codify WebMCP JSON Schema subset (Draft 2020-12) per 2026-03-05 CG resolution.
Meta-validation in registerTool() (TypeError on invalid schema) and input
validation in executeTool() parallel block before tool execute steps
(DataError with {path,reason,schemaPath} for agent self-correction).
Spec-only per review; WPT split to web-platform-tests/wpt (tentative).
Coordinates with webmachinelearning#254 (outputSchema, same dictionaries).
|
Is this still planned on being added? |
a7e008f to
a988ee0
Compare
I hope so :) 🤞 @domfarolino I've started experimenting with |
a988ee0 to
ec56f58
Compare
markafoltz
left a comment
markafoltz
left a comment
There was a problem hiding this comment.
A few questions/comments about how the serialization result is passed around (might be missing some key idea)
markafoltz
left a comment
markafoltz
left a comment
There was a problem hiding this comment.
Looks good % a few optional suggested edits
| <dt><code><var ignore>tool</var>["{{ModelContextTool/outputSchema}}"]</code></dt> | ||
| <dd> | ||
| <p>A JSON Schema object describing the expected output structure for the tool [[!JSON-SCHEMA]]. | ||
| Tool output is expected to conform to the provided schema, and [=agents=] may reject outputs |
There was a problem hiding this comment.
If this can be validated at the User Agent level, should this go a step further and throw an exception or fail the executeTools promise if the output does not match the schema?
There was a problem hiding this comment.
Good question! I guess it's related to #92
|
Thanks for working on this. One thing I'd like to see said explicitly is that leaving out Related, and a question more than a request: since results are serialized as JSON, a tool that returns a string comes back as a JSON string with quotes and escaped newlines. Is that the intended shape for text results, or is a plain text result something to look at separately (maybe together with the typed results work in #41 and #282)? On sjdallst2's question about validation, I'd say a result that doesn't match should reach the host as a tool error, so the host can tell it from a normal result and decide what to do. Also, a schema is written by the page, so the description fields inside it are model-visible text from the page and should get the same treatment as tool descriptions. |
|
@domfarolino I think we're good now. |
This PR adds an optional
outputSchemamember to theModelContextToolandRegisteredTooldictionaries to enable browser agents and in-page agents to help LLMs reliably reason about the return values of tools.While
responseConstraintwas previously discussed, that name is too specific to the Built-in AI / Prompt API. TheoutputSchemaname aligns directly with the MCP Tools specificationoutputSchemaused to define expected structured outputs with JSON Schema.FIX #9
Preview | Diff