| Internet-Draft | knowledge-linkset Well-Known URI | November 2026 |
| Besleaga | Expires 19 May 2027 | [Page] |
This document defines the "knowledge-linkset" well-known URI, at which a web origin publishes one link set describing the knowledge artefacts it makes available: graph serializations, a JSON-LD context, an agent-facing text file, a chunk export, a change ledger, and related resources. Each artefact link may carry a SHA-256 digest, expressed with the syntax of HTTP digest fields, so that a client can check that a retrieved artefact is the one the publisher described. A profile URI identifies the conventions the link set follows. The mechanism defines no new media type and no new link relation type: a page points at the resource with the existing "describedby" relation. The profile URI is registered in the IANA "Profile URIs" registry. This document requests no IANA action.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 19 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Software agents that read the Web need two different things. They need
to find out who can act: which agent, tool server, or service is
reachable, under what name, with what capabilities. A survey of that
field is available in [I-D.jimenez-dawn-discovery-landscape]. They
also need to find out what is known: which described, versioned,
checkable artefacts an origin publishes about its own subject matter --
a graph, a vocabulary, an index, a text rendering, a change ledger.
The first question has many answers. The second has partial ones: the
Vocabulary of Interlinked Datasets (VoID) [VOID], at the registered
well-known URI void, and the Data Catalog Vocabulary (DCAT) [DCAT3]
describe datasets and their distributions as Resource Description
Framework data. None of them gives an origin one list, readable by any
Web Linking client, that names each of its knowledge artefacts, says
which serialization is which, and lets a client check that what it
fetched is what was published.¶
This document defines that list. A web origin serves one resource at
/.well-known/knowledge-linkset [RFC8615]. The representation is a
link set [RFC9264] with one link context, anchored at the
Internationalized Resource Identifier (IRI) [RFC3987] of the
published knowledge Bundle. Its links name the artefacts, each with a
media type and optionally with a digest target attribute in the
syntax of HTTP digest fields [RFC9530]. A profile URI [RFC7284],
carried on the media type or in a Link header field [RFC6906],
tells a client which conventions the document follows.¶
The design adds a path and a profile. Its extension relations
([RFC8288], Section 2.1.2) and extension target attributes
([RFC9264], Section 4.2.4.3) need no registry, digest reuses an
existing algorithm name and the structured-field value syntax
[RFC9651], and no media type ([RFC6838]) or URI scheme is defined.
A generic link set client reads the document without knowing this
specification. The precedent is api-catalog [RFC9727], a well-known
link set, identified by a profile URI, that lists an origin's APIs.¶
The artefacts, the Bundle that contains them, and the bytes their digests cover are defined by the AgenticSystemCore Specification [AGSC-SPEC]. This document specifies the discovery resource and the behaviour of a client that reads it. A client needs nothing else to read a discovery document; a publisher needs [AGSC-SPEC] only to generate the artefacts and the values of the bundle-fact attributes.¶
This document defines no agent identity or authentication of fetchers, no protocol, session, or transport between agents, no naming or resolution layer for agents, no expression of preferences about how content may be used, no read or write interface to an agent's memory, and no artefact format. Publishing a discovery document grants no rights or permissions, and asserts no property of the artefacts other than that the origin published them. Section 8 names the work that addresses each of these topics.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
Long lines in examples are folded as specified in [RFC8792].¶
The set of items and generated artefacts that one origin publishes under one anchor. The Bundle IRI is the origin's base URL with a trailing solidus.¶
One retrievable representation that belongs to a Bundle, such as a graph serialization, a JSON-LD context, a vocabulary file, a text rendering, a chunk export, a search index, a change ledger, or an attachment.¶
A profile in the sense of [RFC6906]; the profile URI is given in Section 6.¶
An origin that serves a conforming discovery document. Its operator is the publisher.¶
A node that another node names in its discovery document with the
peer extension relation of Section 4.2.¶
Said of a node whose content is not public.¶
A node MUST serve its discovery document at the path
/.well-known/knowledge-linkset of its origin, over HTTPS. A node MUST
serve exactly one such document; it MUST NOT split the description of
one Bundle across several discovery resources, and it MUST NOT serve a
second, differently named manifest of the same artefacts. A copy of the
document MAY be distributed by other means, for example inside a
release archive, and a node MAY additionally serve the same bytes at a
path of its own choosing.¶
The representation MUST be a JSON link set, media type
application/linkset+json [RFC9264]. The profile URI of
Section 6 MUST be carried in at least one of two ways:¶
or, where the hosting platform cannot attach a parameter to
Content-Type, as a response header field ([RFC6906]):¶
A client MUST accept either form; whether it recognizes the profile MUST NOT depend on whether it can read response header fields.¶
A node MUST send an ETag and Cache-Control: no-cache ([RFC9111],
Section 5.2.2.4) on the discovery document. Where the publisher sets
the entity tag, it is the lowercase hexadecimal SHA-256 of the bytes
served, in double quotes ([RFC9110], Section 8.8.3). A client SHOULD
make conditional requests ([RFC9110], Section 13), and MUST NOT treat
an entity tag as an integrity statement: an intermediary may weaken or
replace it, and the digest attribute (Section 4.3.1) is
that statement.¶
The header fields and terms of this section are those of Cross-Origin
Resource Sharing (CORS) in the Fetch Standard [FETCH]. The discovery
document, and every artefact it names that the node serves publicly,
MUST be served with Access-Control-Allow-Origin: * and
Access-Control-Expose-Headers: Link, ETag, Content-Type, and MUST NOT
be served with Access-Control-Allow-Credentials. A node MUST NOT require any request
header field outside the CORS-safelisted set for the discovery document
or any public artefact.¶
A gated node MUST still serve the discovery document publicly and with
the wildcard above. It declares itself with the agsc-visibility
attribute (Section 4.3.3) and names where a reader obtains
credentials with the access extension relation (Section 4.2).¶
The document is a JSON object ([RFC8259]) whose sole top-level member
is linkset ([RFC9264], Section 4.2.1). A node that produces
canonical bytes serializes the document with the JSON Canonicalization
Scheme [RFC8785], followed by one line feed; a node that does not
MUST at least produce Internet JSON (I-JSON) [RFC7493].¶
The linkset array holds exactly one link context object, whose
anchor is the Bundle IRI, for example https://example.org/. A
client MUST ignore an additional link context object it does not
understand. A client MUST NOT take a document whose anchor is not on
the origin of the URL it finally retrieved the document from, after
redirects, as that origin's discovery document ([RFC8615], Section
4.3).¶
Every link object MUST carry href, and every href MUST be a URI
([RFC3986], Section 3), not a relative reference. Within one relation's array, link
objects are ordered by href in code point order.¶
A relation name MUST be either a registered short name from the "Link
Relation Types" registry or an extension relation expressed as a URI
([RFC8288], Section 2.1.2). The registered short names used are
describedby, alternate, license, service-doc, and author, and,
in related-system links, those of Section 4.5.¶
The extension relations form a closed set. Each is the URI
https://w3id.org/agentic-system-core/rel#<name>, where <name> and
the target are one of:¶
graph:a further serialization of the Bundle's graph¶
ontology:the vocabulary the graph uses¶
context:the JSON-LD context of the graph¶
now:the generated state page of the Bundle¶
skills:the index of exported procedure packages¶
ledger:the derived change ledger of the Bundle¶
peer:another node's discovery document¶
surface:an agent-facing surface this node serves¶
contribute:an endpoint that accepts contributions¶
access:a page that says how to obtain credentials¶
signature:a detached signature over the discovery document¶
boards:the index of the Bundle's project boards¶
A client MUST ignore a relation name it does not recognize. A node MUST NOT invent a further extension relation under the same base; a new member is added only by a later version of [AGSC-SPEC] (Section 6.3).¶
A link object MAY carry the target attributes type, title, and
hreflang ([RFC8288], Section 3.4.1; [RFC9264], Section 4.2.4.1),
and the extension target attribute profile, whose values are profile
URIs ([RFC6906]). A client MUST NOT depend on type, and MUST use
the media type of the response it actually receives.¶
Every extension target attribute value, profile included, is an array
of strings ([RFC9264], Section 4.2.4.3). A client MUST ignore an
extension target attribute it does not recognize.¶
A link object MAY carry the target attribute digest. Its value is an
array holding one string, the serialization ([RFC9651], Section 4.1)
of a Dictionary ([RFC9651], Section 3.2) with a single member. The
member's key is the algorithm name sha-256 from the "Hash Algorithms
for HTTP Digest Fields" registry ([RFC9530], Section 7.2), and its
value is a Byte Sequence ([RFC9651], Section 3.3.5):¶
"digest": [ "sha-256=:JeigkAtOPS9f9BzKym+X6J2O4awt53GEua1AhIAiLfE=:" ]¶
The digest is taken over the canonical bytes of the target, as the publisher generated and serves them, before any content coding. A client compares it with the decoded representation data of the target.¶
Facts about the Bundle as a whole ride only on the anchor's
describedby link whose target is the Bundle's JSON-LD graph:¶
agsc-spec-version:one value, the version of [AGSC-SPEC] the node publishes against.¶
agsc-generated-at:one value, the instant the artefacts were generated, as
YYYY-MM-DDTHH:MM:SSZ (UTC, second precision, no offset). A node
that supports reproducible builds derives it from the build's source
date.¶
agsc-counts:one string per item type, of the form <type-plural>=<n>, in code
point order of the type name, counted over the published set.¶
agsc-bundle-hash:one value, the digest of the Bundle's canonical N-Quads
serialization, in the syntax of Section 4.3.1. It equals
the digest of the graph link to that serialization.¶
agsc-bundle-version:one value, the publisher's content version for the state the artefacts were generated from, a short name a person can read and cite. It is derived by the publisher and is opaque to a client, which MUST NOT order two values or infer a sequence from them.¶
Two attributes ride on the ledger link and nowhere else; the second
is optional:¶
agsc-ledger-head:one value, the chain hash of the last entry of the derived change ledger ([AGSC-SPEC]), in lowercase hexadecimal.¶
agsc-ledger-length:one value, the number of entries of the derived change ledger, in decimal.¶
Two attributes ride, beside the digest, on the surface link whose
agsc-surface value is chunks, and nowhere else. That link names the
chunk export's main file, which lists the other files when the export
is split; the digest covers that file.¶
agsc-chunk-count:one value, the number of chunk records in the export, across all of its files, in decimal.¶
agsc-chunk-set:one value, in the syntax of Section 4.3.1, the SHA-256 of the identifiers of all chunk records, in file order and record order, each written as 64 lowercase hexadecimal characters followed by one line feed; for an export with no record, the SHA-256 of no bytes.¶
The ledger length and the chunk attributes exist from version 1.1.0 of
[AGSC-SPEC] and are absent from a document that states an earlier
agsc-spec-version. A client MUST NOT read an attribute of this
section from a link other than the one named for it.¶
The extension target attributes below declare what a node is. They appear in the full and in the reduced form (Section 4.4).¶
agsc-surface and agsc-surface-version:on a surface link, one value each: the kind of agent-facing surface
the node serves, and, where an external specification defines the
surface, the revision of that specification the node targets.¶
agsc-access:on a surface link or the access link, one value: none,
consent, or credential.¶
agsc-visibility:on the anchor's describedby link, the single value restricted,
emitted only by a gated node.¶
agsc-contribute-mode:on a contribute link, one value naming how contributions are
accepted.¶
agsc-tombstone:on the anchor's describedby link, one instant in the form of
agsc-generated-at, present only on a node that has stopped
publishing (Section 11.5).¶
A client that meets a value it does not know in agsc-visibility,
agsc-access, agsc-surface, or agsc-contribute-mode MUST NOT fail.
It MUST treat the value as the most restrictive one: restricted,
credential, a surface the node does not serve, and a mode that
accepts no contribution, respectively.¶
A publisher without a build engine, such as a content management system
or a wiki that exports static files, publishes the same link set with
every digest attribute and every attribute of
Section 4.3.2 omitted. Such a publisher has no ledger. A
client
MUST treat a link without a digest attribute as unverified, never
as invalid.¶
A node SHOULD place, in the head element of every HTML page it
generates:¶
<link rel="describedby"
href="/.well-known/knowledge-linkset"
type="application/linkset+json">
¶
and SHOULD send, at least on the origin's root route, the equivalent response header field:¶
On the root route, the same header field may also name the discovery
document with the linkset relation ([RFC9264], Section 6).¶
describedby is defined in Appendix D of [POWDER-DR], which
registers it, and [RFC6892] registers its inverse, describes. The
target's media type and profile URI say what kind of description it is.
A page may carry other describedby links, such as one to an
[LLMSTXT] file. A client selects this one by
type="application/linkset+json" and, having fetched it, confirms the
profile (Section 3.1). Where several candidates remain,
a client SHOULD fetch the one whose href resolves to
/.well-known/knowledge-linkset on the origin.¶
The profile URI of the discovery document is
https://w3id.org/agentic-system-core/profile/agentic-knowledge.¶
A client that ignores the profile reads the document as a link set
([RFC6906], Section 3). A client that recognizes it may rely on the
single link context anchored at the Bundle IRI
(Section 4.1), the closed set of relations (Section 4.2),
the form of digest (Section 4.3.1), the links that carry
the bundle-fact, ledger, and chunk attributes
(Section 4.3.2), and extension target attribute values
that are arrays of strings (Section 4.3).¶
The profile URI, and every extension relation URI of Section 4.2, resolves through w3id.org, a persistent-identifier service, with an HTTP 303 redirect ([RFC9110], Section 15.4.4) to a published specification page. That page carries one fragment identifier for the list of relations and one per relation name, so every extension relation URI dereferences to its own documentation ([RFC8288], Section 2.1.2).¶
The profile URI carries no version number. It names the same profile
for every version of [AGSC-SPEC] that has the same major version
number. A document in the full form states its version in
agsc-spec-version (Section 4.3.2).¶
A later minor version of [AGSC-SPEC] only adds:¶
extension relations under the same base URI, each with its own fragment on the specification page;¶
extension target attributes with the agsc- prefix;¶
related-system relations, once they are registered;¶
values of the declaration attributes.¶
The client rules of Section 4.2, Section 4.3, Section 4.3.3, and Section 4.5 let a client read such a document without error and without change. Any other change is made only in a new major version of [AGSC-SPEC].¶
A client that fetches one document from an origin it was given applies
Section 7.1 and Section 7.3. A client that follows
peer links applies all of this section.¶
A client MUST refuse any URL that is not an absolute https URL,
except an http URL to a loopback address when the client has been
explicitly placed in a development mode.¶
Before connecting, a client MUST resolve the host and MUST refuse any resolved address that is not globally reachable according to the IANA IPv4 [IANA-IPV4-SPECIAL] or IPv6 [IANA-IPV6-SPECIAL] Special-Purpose Address Registry, other than a loopback address in the development mode above. A client MUST equally refuse any IPv6 address that embeds an IPv4 address -- IPv4-mapped, IPv4-compatible, NAT64 (IPv6-to-IPv4 translation), Teredo, and 6to4 addresses -- whatever IPv4 address it embeds, and any multicast address. A host that resolves to several addresses MUST be refused if any one of them is refused.¶
A client MUST give every textual form of one address the same classification: compressed or expanded, and with an embedded IPv4 address written as a dotted quad or in hexadecimal.¶
A client MUST connect to one of the addresses it classified, MUST NOT re-resolve the host between classification and connection, and MUST re-verify the remote address of the established connection against the same rules, closing the connection on a mismatch.¶
A client MUST bound the number of redirects it follows, MUST re-apply
the rules above to every hop, and MUST NOT follow a redirect to a
non-https URL other than the development loopback case. The final
URL of a redirected fetch is not substituted for the URL that was
declared.¶
A client MUST bound the size of a discovery document it reads, measured after content decoding, and treat a larger response as unreadable without buffering it; one mebibyte suffices for the documents this document describes.¶
A client SHOULD bound the time it waits for a response.¶
The target of a peer link is the peer's discovery document URL. Two
nodes are peers of each other when each names the other; nothing else
establishes the relation.¶
A client that walks from a starting node MUST:¶
bound the depth of the walk, the starting node being depth zero;¶
consider at most a fixed number of peer links per discovery
document, in document order, ignoring the rest;¶
issue at most a fixed number of requests per walk, counting every request including redirects;¶
keep a visited set keyed by the canonical well-known URL, and never fetch a member of it twice;¶
treat a peer whose fetch times out, returns a non-2xx status, exceeds the size bound, or yields a document that is not a conformant discovery document as unreachable: record it, skip it, and never retry it within the walk.¶
A client that stops because a bound was reached, or that ignored links beyond its per-document bound, MUST report the result as partial, with the set of nodes it did visit. A client that completes without touching any bound MUST report the result as complete.¶
Suggested defaults, which a deployment may lower: depth three, fifty
peer links per document, five hundred requests per walk, ten seconds
per request, three redirects per fetch.¶
Everything a client obtains from a node other than the one it is acting for is untrusted data. A client MUST carry that fact with the data: a search hit, a retrieved item, a chunk, or a quotation derived from another node MUST be labelled untrusted and MUST name the origin it came from, through every interface the client offers and every export it produces. A client MUST NOT merge another node's prose into its own published artefacts other than through a reviewed proposal that records the other node's IRI as the source.¶
A node's query surface is the set of artefacts it publishes. A node MUST NOT fetch another node while building its own artefacts, and MUST NOT expose an endpoint that fetches another node on a caller's behalf.¶
This section is informative.¶
llms.txt [LLMSTXT] is a community convention for a Markdown file at
/llms.txt that presents an origin's content to language models. Its
version 2 recommends rel="describedby" for the file and
rel="alternate" with type="text/markdown" for a page's Markdown
rendering; a node may name such a file with alternate.
[I-D.arsentev-llm-context-discovery] defines the discovery of one
such publisher-curated context file through a well-known URI, a link
relation, and a robots.txt record.¶
[I-D.serra-mcp-discovery-uri] defines an mcp URI scheme and the
discovery of Model Context Protocol servers through a well-known path
and DNS TXT records. [MCP] defines a protocol between a client and a
tool server, and [A2A] an agent card, served at a well-known path,
that describes an agent's capabilities and interfaces; a node may
declare either as one of its surfaces (Section 4.3.3).¶
The dawn working group [DAWN-CHARTER] addresses the discovery of
agents by name, and [I-D.jimenez-dawn-discovery-landscape]
catalogues the mechanisms in that space. The agentproto working
group [AGENTPROTO-CHARTER] addresses the management of dialogs
between users, agents, and tools. The W3C AI Agent Memory
Interoperability Community Group [AIMEM-CG] addresses the
interoperability of protocols for AI agent memory.¶
[I-D.ietf-aipref-vocab] defines a vocabulary for preferences about
AI usage of content, and [I-D.ietf-aipref-attach] the means of
attaching them in HTTP; a publisher attaches preferences to the
artefacts a discovery document names by those means and by
robots.txt [RFC9309]. HTTP message signatures for automated traffic
[I-D.ietf-webbotauth-httpsig-protocol] identify the client making a
request.¶
The "Knowledge Bundle" of the Open Knowledge Format [OKF], a set of
Markdown files with front matter and an index file, is a different
object from the Bundle of this document. A VoID or DCAT description
(Section 1) is named with a further describedby link
(Section 4.5).¶
[I-D.aiendpoint-ai-discovery] requests the well-known URI suffix ai
(or ai-discovery) for a description of a service's capabilities, and
[I-D.car-ai-txt-wellknown] requests ai.txt and ai.json for
declarations of AI usage preferences and licensing. The suffix used
here names the class of resource served (Section 9.1) and
does not begin with ai or agent.¶
This document requests no IANA action. It creates no registry, and requests no media type, no URI scheme, no link relation type, and no well-known URI registration.¶
No registration of the "knowledge-linkset" suffix in the "Well-Known URIs" registry is requested at this time. A later request would use the template of [RFC8615], Section 3.1, with these values:¶
URI suffix: knowledge-linkset¶
Change controller: Andrei N. Besleaga (andrei.besleaga.nicolae@gmail.com)¶
Specification document(s): This document.¶
Status: provisional¶
Related information: Used with the "https" URI scheme. The
resource is an application/linkset+json document [RFC9264]
carrying the profile URI
https://w3id.org/agentic-system-core/profile/agentic-knowledge,
whose registration is described in Section 9.2.¶
The suffix names the class of resource served -- a link set of the knowledge artefacts an origin publishes -- rather than a subject area, in keeping with the precision [RFC8615], Section 3, expects of a registered name.¶
The following profile URI was registered in the "Profile URIs" registry ([RFC7284], Section 4) on 2026-10-05; that registry's policy is First Come First Served. The fields are given here so that this document, which specifies the profile, records them:¶
Profile URI:
https://w3id.org/agentic-system-core/profile/agentic-knowledge¶
Common Name: AgenticSystemCore knowledge link set profile¶
Description: Identifies an application/linkset+json document
[RFC9264] that describes a published knowledge Bundle: one link
context whose anchor is the Bundle IRI, links to the Bundle's
artefacts -- graph serializations, a JSON-LD context, a vocabulary, a
chunk export, an agent-facing text file, a change ledger -- each
optionally carrying a digest target attribute in the syntax of
[RFC9530], and the bundle-fact and declaration target attributes
this document defines. Served at /.well-known/knowledge-linkset;
linked from pages with rel="describedby" and
type="application/linkset+json".¶
Reference: This document; and [AGSC-SPEC], which defines the artefacts and their canonical bytes.¶
Notes: The profile does not change the semantics of
application/linkset+json for a client that ignores it ([RFC6906],
Section 3); it adds integrity, bundle-fact, and declaration target
attributes that a client MAY use. The same profile URI serves every
version of the AgenticSystemCore specification with the same major
version number; a document in the full form states the version it
follows in its agsc-spec-version attribute. Change controller:
Andrei N. Besleaga, Independent, andrei.besleaga.nicolae@gmail.com.¶
[Note to the RFC Editor: please remove this section before publication.]¶
This section records the status of known implementations of the protocol defined by this specification at the time of posting of this Internet-Draft, and is based on a proposal described in [RFC7942]. The description of implementations in this section is intended to assist the IETF in its decision processes in progressing drafts to RFCs. Please note that the listing of any individual implementation here does not imply endorsement by the IETF. Furthermore, no effort has been spent to verify the information presented here that was supplied by IETF contributors. This is not intended as, and must not be construed to be, a catalog of available implementations or their features. Readers are advised to note that other implementations may exist.¶
According to [RFC7942], "this will allow reviewers and working groups to assign due consideration to documents that have the benefit of running code, which may serve as evidence of valuable experimentation and feedback that have made the implemented protocols more mature. It is up to the individual working groups to use this information as they see fit".¶
The implementations below are the author's (Independent), who is the contact for each; the information is as of 15 November 2026. All implement this revision and version 1.1.0 of [AGSC-SPEC].¶
https://agenticsystemcore.com/, a static site that serves a
discovery document in the full form, its artefacts, and the
specification page that the profile URI and the extension relation
URIs resolve to. A second node, https://patterns.agenticsystemcore.com/,
is built by the same engine; each node names the other as a peer.
Working implementation; the engine is the npm package
agentic-system-core. Coverage: Section 3 to
Section 6. Licensing: the site's own terms.¶
validate-wellknown, a command-line tool in the npm and the PyPI
package named agentic-system-core. It checks a URL or a local file against
Section 3 and Section 4: the reduced form where it
is claimed, otherwise with the digests, the bundle facts, and the
chunk count and set recomputed. With --peer it performs the mutual
test of Section 7.2, also on local files without network
access. Working implementation; Apache-2.0.¶
JSON test vectors for Section 4, including Section 4.4, and for Section 7.2, each naming the rule it proves, against which any implementation can be run. Fixtures, released with the engine; Apache-2.0.¶
A digest attribute binds an artefact to the discovery document. It
lets a client detect that the artefact it retrieved is not the one the
publisher described: a truncated file, a cache that served an older
copy, a mirror that altered the bytes, a transfer that corrupted them.
It proves nothing else: neither the mechanisms of this document nor a
digest prove safety or the absence of a novel attack carried in the
content of an artefact, and an implementation MUST NOT claim more.¶
A digest attribute does not bind the discovery document to an author.
An attacker who controls the origin controls the discovery document and
the artefacts together, and can make them agree. A client obtains the
identity of the publisher only from the TLS certificate and the DNS
name of the origin. A publisher who needs authorship verified
independently of the origin signs the responses, for example with
[RFC9421], or names a detached signature with the signature
extension relation (Section 4.2). This document defines no signature
format: a client MUST ignore that link if it does not understand the
target, and the link affects no digest and no walk.¶
A client that keeps the agsc-ledger-head value it saw on an earlier
retrieval, and that later reads a head that does not extend the earlier
one, has observed a rollback or a replacement of the ledger, and SHOULD
treat it as such rather than as an ordinary update.¶
The only digest algorithm is sha-256; another algorithm needs a new
major version (Section 6.3). The considerations of [RFC9530],
Section 6, apply.¶
A discovery document is a list of URLs supplied by a third party. Without the rules of Section 7.1, any origin a client reads can direct it at any address its network can reach, use it to probe an internal address space, or hold it open. The re-verification of the established connection guards against DNS rebinding. Measuring the size after content decoding guards against compressed responses that expand without bound.¶
Because no node fetches another node (Section 7.4), naming a node as a peer cannot turn it into a request forwarder. The risk is on the client side, where the size, time, and request bounds of Section 7.1 and Section 7.2 apply. A publisher SHOULD keep a discovery document small: it describes a Bundle's artefacts, not its items.¶
A client that passes retrieved prose to a language model, a shell, a query engine, or any other interpreter MUST treat it as data: never executed, never interpolated into a path, a URL, or a command string, never resolved as an identifier, and never read as an instruction to the client. Prose retrieved from another node carries the untrusted label of Section 7.3 to every place it is used.¶
A discovery document is public. A publisher MUST NOT name in it any resource that is not intended to be public, and MUST NOT let the set of links reveal the existence of resources it does not serve publicly. A gated node still serves the discovery document publicly (Section 3.3), which discloses that the node exists.¶
A gated node MUST omit the agsc-counts, agsc-bundle-hash,
agsc-bundle-version, agsc-chunk-count and agsc-chunk-set
attributes and the ledger link, and MUST omit every digest whose
target it does not serve without authentication: a digest over a gated
artefact lets anyone confirm a guess of its content, and the counts and
the content version disclose activity. It still carries
agsc-spec-version and agsc-generated-at, which say only that the
node exists and is current.¶
The path /.well-known/ on an origin MUST be under the publisher's
control ([RFC8615], Section 4.1). On static hosting, the file is
produced by the publisher's build and upload path and by nothing else.¶
A node that stops publishing MUST keep serving a valid discovery
document whose anchor's describedby link carries the
agsc-tombstone attribute of Section 4.3.3, MAY carry an
alternate link to a successor node's discovery document, and MAY omit
every other artefact link.¶
A client MUST NOT treat a successor named that way as the same node. It MUST establish the peer relation with the successor afresh before following it, and everything it derives from the successor carries the successor's own Bundle IRI as its origin (Section 7.3). A walk (Section 7.2) does not descend into a tombstoned node.¶
Serving a discovery document collects nothing, sets no cookie, carries no identifier of a reader, and requires no request header field beyond those of a simple cross-origin request (Section 3.3). A retrieval appears in the publisher's server logs like any other request for a static file.¶
The document may name an author link, and the artefacts it points at
may contain personal data chosen by the publisher. This document
requires no personal data anywhere in the discovery document. A
publisher SHOULD use a role rather than a personal contact where one
will do.¶
A client SHOULD NOT send identifying header fields beyond those [RFC9110] makes ordinary when fetching a discovery document, and a client walking peers SHOULD NOT disclose to one node which other nodes it has visited.¶
A complete discovery document for https://example.org/ in the full
form is shown below, pretty-printed. Its canonical bytes
(Section 4.1) order the members by name, have no whitespace
between tokens, and end with one line feed. The type value
application/jsonl, used for the ledger and the chunk export, is not a
registered media type; a client does not depend on type
(Section 4.3).¶
NOTE: '\' line wrapping per RFC 8792
{
"linkset": [
{
"anchor": "https://example.org/",
"alternate": [
{
"href": "https://example.org/llms-full.txt",
"type": "text/plain",
"digest": [
"sha-256=:Nfm1AxNikrZAb82mhnWgjpi3i8rGJa9K64wnI0ZiJY0=:"
]
},
{
"href": "https://example.org/llms.txt",
"type": "text/plain",
"digest": [
"sha-256=:QjxPZtSHNKLvYtiXXCOwnjJn+fYmfqkRYmash8n7+O0=:"
]
}
],
"author": [
{
"href": "https://example.org/about/",
"type": "text/html"
}
],
"describedby": [
{
"href": "https://example.org/graph.jsonld",
"type": "application/ld+json",
"agsc-bundle-hash": [
"sha-256=:/WQyaCbIwaeGFFMAFw/eMonm8tK+F7amphaKNT4BpB0=:"
],
"agsc-bundle-version": ["v1.4.0"],
"agsc-counts": [
"clusters=6", "concepts=142", "episodes=17", "gates=4",
"lessons=39", "procedures=28"
],
"agsc-generated-at": ["2026-10-09T08:15:00Z"],
"agsc-spec-version": ["1.1.0"],
"digest": [
"sha-256=:JeigkAtOPS9f9BzKym+X6J2O4awt53GEua1AhIAiLfE=:"
]
}
],
"license": [
{
"href": "https://example.org/legal/",
"type": "text/html"
}
],
"service-doc": [
{
"href": "https://example.org/specs/",
"type": "text/html"
}
],
"https://w3id.org/agentic-system-core/rel#context": [
{
"href": "https://example.org/ns/context.jsonld",
"type": "application/ld+json",
"digest": [
"sha-256=:Dv8ogyqhEZ0jxdWQj0FkVfpXbuTnyfvWEFNwtbRTSEA=:"
]
}
],
"https://w3id.org/agentic-system-core/rel#graph": [
{
"href": "https://example.org/graph.nq",
"type": "application/n-quads",
"digest": [
"sha-256=:/WQyaCbIwaeGFFMAFw/eMonm8tK+F7amphaKNT4BpB0=:"
]
},
{
"href": "https://example.org/graph.ttl",
"type": "text/turtle",
"digest": [
"sha-256=:SylBDgXPkysmlRZkx+BvHSxVTbm3ZWGzFXskHf+kH80=:"
]
}
],
"https://w3id.org/agentic-system-core/rel#ledger": [
{
"href": "https://example.org/ledger.jsonl",
"type": "application/jsonl",
"agsc-ledger-head": [
"6acd55d78bf24c2b7866c0e46a3bb88a4b7259cea9ca0bfa6263\
81068b0d56ed"
],
"agsc-ledger-length": ["48"],
"digest": [
"sha-256=:mIn5HVeWxmn0Ecmh5Qwkc3doL+0fdO+Z7HieEDozhz0=:"
]
}
],
"https://w3id.org/agentic-system-core/rel#now": [
{
"href": "https://example.org/now.md",
"type": "text/markdown",
"digest": [
"sha-256=:cJvRMVgkkW5Usdn/z0q1iNxyBErb59kLWRv83qvpx8A=:"
]
}
],
"https://w3id.org/agentic-system-core/rel#ontology": [
{
"href": "https://example.org/ns/agsc.ttl",
"type": "text/turtle",
"digest": [
"sha-256=:kjVLSaNilITtebImxe9J+ukr7UOdmTG+NAtppRjsrH8=:"
]
}
],
"https://w3id.org/agentic-system-core/rel#peer": [
{
"href": "https://b.example/.well-known/knowledge-linkset",
"type": "application/linkset+json"
}
],
"https://w3id.org/agentic-system-core/rel#skills": [
{
"href": "https://example.org/skills/index.json",
"type": "application/json",
"digest": [
"sha-256=:NnyPQz+rxvP/Zl9ja4+0xRSCpb7gdArcAVXlAG5IoU0=:"
]
}
],
"https://w3id.org/agentic-system-core/rel#surface": [
{
"href": "https://example.org/chunks.jsonl",
"type": "application/jsonl",
"agsc-access": ["none"],
"agsc-chunk-count": ["517"],
"agsc-chunk-set": [
"sha-256=:7EWtxSlfY5u7jm7lsHVF/l/5X1a73Ueu3kVJhavUdsU=:"
],
"agsc-surface": ["chunks"],
"digest": [
"sha-256=:vAlDuXGRaRPyTeeRfLAFWa7acFyOL3Y7o+vbywRofDk=:"
]
},
{
"href": "https://example.org/llms.txt",
"type": "text/plain",
"agsc-access": ["none"],
"agsc-surface": ["llms-txt"]
}
]
}
]
}
The same node without a build engine serves the reduced form:¶
{
"linkset": [
{
"anchor": "https://example.org/",
"alternate": [
{
"href": "https://example.org/llms.txt",
"type": "text/plain"
}
],
"describedby": [
{
"href": "https://example.org/graph.jsonld",
"type": "application/ld+json"
}
],
"license": [
{
"href": "https://example.org/legal/"
}
]
}
]
}
[Note to the RFC Editor: please remove this appendix before publication.]¶
Since -00: the header is that of an ordinary individual Internet-Draft; [AGSC-SPEC] cites the released version 1.0.0; the profile URI is registered (Section 9.2); no well-known URI registration is requested (Section 9, Section 6.2); [I-D.arsentev-llm-context-discovery] is cited at its current revision. Nothing on the wire changed.¶
Since -01:¶
A document valid under -01 is valid under -02. The path, the media type, the profile URI and the extension relations are unchanged.¶
[AGSC-SPEC] is version 1.1.0: the ledger link may carry
agsc-ledger-length; the chunks surface link carries digest,
agsc-chunk-count and agsc-chunk-set, which a gated node omits;
the access link carries agsc-access.¶
Client rules match the specification: the anchor is on the retrieval origin; an IPv6 address that embeds an IPv4 address is refused whole; development mode admits loopback; every textual form of an address is classified alike; the size bound applies after content decoding; unknown declaration values have named defaults; recognizing the profile does not depend on header access; an entity tag is not an integrity statement.¶
Stated exactly: the publisher-set entity tag, the line feed after
canonical bytes, the ledger head, the chunk set of a split export, and
the related-system links. Section 5 names the
linkset relation.¶
References: [RFC7493] and [FETCH] added; [RFC9111] and the
IANA address registries normative; dawn and agentproto cited by
their charters.¶
Condensed: requirements stated twice are stated once, four subsections are merged into their neighbours, and IANA Considerations precede Security Considerations.¶