Pre-navigation discovery for page-exposed tools: an RFC 9264 link set at a well-known URI

WebMCP tells an agent what a page can do once the agent is on the page.
There is a complementary question one step earlier: how does an agent
learn, *before* loading anything, that a site exposes tools at all, which
revision they target, and whether the bytes it is about to read are the
bytes the publisher meant? I have been specifying a static answer to that
question and would like the group's view on whether it belongs anywhere
near WebMCP.

The mechanism is one file, /.well-known/knowledge-linkset, served as
application/linkset+json with the RFC 9264 §5 profile parameter (or the
equivalent RFC 6906 Link: …; rel="profile" header, either being acceptable
to a reader). Its sole top-level member is the RFC 9264 linkset array,
holding one link context anchored at the site's Bundle IRI with
relation-keyed link arrays. At conformance Level 2 and above every artefact
link carries a digest target attribute — an RFC 9530 sha-256 value in RFC
9651 Byte Sequence syntax over the canonical bytes — so a browser agent can
verify what it fetched before parsing it. The description link uses the
already-registered describedby relation; no new relation name is invented.
Each agent surface the node actually serves is declared as its own link
with a surface name (webmcp, mcp, chunks, llms-txt, …) and, for external
surfaces, the revision of that external specification it targets. All
public artefacts are served Access-Control-Allow-Origin: * with no
credentials, so a page-side agent can read the whole thing with a simple
request.

Read against the WebMCP explainer, this sits outside its Non-Goals, which
say WebMCP "is designed to complement, not replace, existing
backend-focused protocols like MCP". Discovery there starts at
document.modelContext; this link set stops at the door.

*Question:* is pre-navigation, site-level discovery in scope for the group,
or deliberately left to the host — and if the latter, would an informative
pointer in Prior Art be useful?

Specification: https://agenticsystemcore.com/specs/ · Internet-Draft
draft-besleaga-agentic-knowledge-wellknown-00:
https://datatracker.ietf.org/doc/draft-besleaga-agentic-knowledge-wellknown/ ·
preprint DOI: https://doi.org/10.5281/zenodo.23052710.

This is an independent release candidate, 1.0.0-rc.6 (not a standard of the
IETF, the W3C or any other body; its registrations are requested or to be
requested, and none is granted). Planned for the next version, with no
date: a semantic layer over the typed links (a stated meaning for each, and
checks over it), a graph view, and a richer composition output.

Thank you,
Andrei Besleaga.