Natural language attributes

[Starborn]

Not sure where to add this #22

Notes on the https://github.com/webmachinelearning/webmcp/tree/declarative-api-explainer

This explainer makes a strong case for the declarative path and the toolautosubmit attribute in particular is a useful design. I want to flag a knowledge representation concern that runs through several of the open questions here.
The proposal relies on natural language attributes -- tooldescription, toolparamdescription -- as the primary mechanism for agents to understand what a tool does. This works for human-readable discovery but is insufficient as a semantic contract when toolautosubmit is in play. An agent authorized to submit a form on a user's behalf without manual review needs more than a natural language description to make that trust decision safely -- it needs a formal, verifiable account of what the form does, what its preconditions are, and what its side effects will be.
This is the declarative-layer version of the gap raised in issue #96 regarding delegation context and audit. The correlation ID proposed there addresses post-hoc accountability; what's missing here is pre-invocation semantic grounding.

Two concrete suggestions for the open questions:
On input schema synthesis (currently TODO): the reduction of form elements to JSON Schema is a necessary but not sufficient semantic representation. It captures structure but not meaning -- a with options "economy" and "business" reduced to an enum says nothing about what selecting each value commits the user to. It may be worth considering whether a lightweight semantic annotation layer (perhaps via existing or JSON-LD conventions already referenced in the response section) could travel alongside the synthesized schema. On the toolactivated/toolcanceled event target question: firing at the rather than Window is the stronger choice from a semantic containment perspective -- it keeps the tool's identity and its event surface co-located, which matters for the kind of per-tool audit context discussed in #96. The W3C AI Knowledge Representation Community Group has vocabulary coordination in scope and would be glad to contribute to the input schema semantics question as it develops. Happy to discuss.