The concept pages on this site describe what the Seed daemon does today, starting from the protocol overview. This page collects the direction the team has agreed on but not yet shipped, so that no concept page has to describe an unbuilt design as current. Everything here is design direction as of September 2026, dated to the team discussion it comes from. It will change, and pieces of it may be dropped. When something lands, its concept page is updated and the item leaves this page.

The team made these decisions in tech syncs and design meetings and recorded them on the team site. The process is slow on purpose. A change to permanent data goes through a written proposal, argumentation, and a decision "by technical extenuation". The team prefers to solve problems at the application layer whenever the permanent data can stay as it is.

The node and resource redesign (HM26)

Where it comes from. August and September 2026 tech syncs.

The problem. Today a document's path carries its identity, its position in the hierarchy, and its permissions. A child's Ref repeats its parent's path names. So renaming a parent means republishing every descendant with new Refs and redirects, which the team called unreliable. Paths mix up identity, hierarchy and permission.

The direction. Every mutable thing becomes a node with a stable, non-forgeable node ID. A node blob carries a target (changes, a tombstone or a redirect), a relative name, and a parent, which is a space plus a node ID. Nodes have exactly one parent. There are no hard links, because a node is both a directory and a file. Creating a name in a parent that points at a node needs authority over both. Human-readable "pretty" paths become the space owner's responsibility. A link-shortening and redirect layer outside the permanent data serves them, and existing hm:// URLs keep resolving the same way. Links are written with full context, node ID plus path, and resolved through a fallback order. Nodes may be unnamed, so private documents can exist without human paths. Comments become nodes without a prior ID. An edit references the original node ID, so comments, children and capabilities move as a unit when their target moves.

What it replaces. The current path-keyed Ref model, called HM24 internally. Legacy node IDs cannot be reused in the new format, to prevent forgery. One migration idea is to hash the old path as the node ID. The migration is expected to discard sub-document capabilities and migrate everything else. The team's phrase: the authority graph is critical, the placement graph is for naming and addressing, and moves must preserve authority.

Vocabulary. "Resource" stays the umbrella term for mutable things built from blobs, matching the RPC. Capabilities are treated as a special category of resource. Whether comments are resources at all is still open. One view is that they are closer to triples than to locations.

Affected pages. Documents, URLs, Comments, Ref.

Home documents become profiles

Where it comes from. 2026-05-14 tech sync and the HM26 discussions.

Today an account's site is rooted at its home document, the document at the empty path. Its siteUrl attribute names the server that publishes it. The direction is to move that role onto the profile. Home documents and profiles would become nodes anchored to a fixed, hard-coded node ID under an empty parent, writable only by the space owner. The profile blob would get a deterministic timestamp so multiple devices converge. A new app version would upgrade existing accounts automatically.

Affected pages. Identity, Sites.

Permissions

Where it comes from. The 2026-09-03 and 2026-09-10 tech syncs.

Today. A capability is signed only by the space owner, grants the role writer or agent, is scoped by path prefix, cannot be re-delegated, and has no expiry or revocation. The proto reserves an editor role, a non-recursive flag and revocation. The daemon implements none of them. See Permissions.

Direction.

Path-scoped and non-recursive capabilities are being discontinued. Invitations will be to a space's home only. A writer of a space is a writer of the whole space. Non-recursive sharing is being removed from the app immediately.

A group concept, one capability granting several keys at once, is wanted. The team reviewed Keyhive and UCAN and prefers fewer keys. The idea is a permanode per space whose admins are permanent, with admin removal done by rotating the permanode.

Revocation and an editor or read-only role remain declared in the proto and unbuilt. No date is attached.

The permissions investigation's synthesis, "everything is a grant", proposes one signed statement kind that covers public publishing, private spaces, document sharing, share links and comments. It builds on the observation that the visibility table already behaves like grants to an audience. It is a design record only. Nothing was decided.

Affected pages. Permissions, Capability, Role.

Private documents, phase two

Where it comes from. Phase one shipped in the first half of 2026; consensus on phase two at the 2026-08-13 tech sync; transport discussion 2026-09-10.

Today. A Ref, a comment or a change can carry private visibility. Private documents live under single-segment paths with random names. The owner and root writers can read them. They are pushed only to the space's declared site server or to authenticated peers, and filtered out of public-only servers. The daemon refuses to create new private documents through its public API. The check is a hard precondition in the create path, lifted only in tests. Existing private documents still work. See Privacy.

Direction.

Phase two is inheritance: children of private documents, an "inherit permissions" idea expressed through metadata instead of the path, flat non-human names with redirects, anyone-with-the-link read and write, and an invitation flow that can grant to someone who has no account yet.

Until there is encryption, sync permission and read permission are the same: "no difference between reading and syncing if there is no encryption." Agent delegations need care here.

At the transport, uploads should carry the context that proves the write is allowed, and blobs should arrive authority first. The team sees Bitswap's willingness to accept arbitrary data as a problem. Fixing it may need a deeper fork of the IPFS stack or an HTTP trustless-gateway transport. The DHT provider system is disabled, and no DHT is in use today (Network).

"Public by default" is shifting. It made sense for a publishing system, but users of a knowledge system assume private by default. The team frames private networks as richer permission and sync rules inside the one network.

Affected pages. Privacy, Network, Files.

One sync API instead of subscriptions plus discovery

Where it comes from. May 2026 tech syncs.

Today. The daemon has two mechanisms backed by one scheduler. Periodic sync of subscriptions is derived from local accounts and contacts and is invisible to the user. On-demand discovery is treated as hot and polled more often. The subscriptions gRPC service still exists. See Network.

Direction. Remove subscriptions as a user-facing concept. Subscription becomes automatic daemon behaviour, derived from what the node's accounts have joined and followed. Discovery grows fine-grained addressing so a client can ask for exactly /:profile or /:comments of a resource. The team's phrase: one powerful syncing API rather than two incomplete ones. The client-side sync helpers in the web app and the /hm/api/discover route will go once the daemon covers them. Immediacy targets are three seconds for comments and under ten seconds for subscribed changes on desktop, with a publish-subscribe pattern for discovery under consideration.

Affected pages. Network, Building with gRPC.

Schemas as daemon resources

Where it comes from. 2026-08-25 and 2026-09-03 syncs.

Today. Hypermedia Schemas are documents: a schema-defining page binds a schema blob with schemaDefinition, and typed documents point at it with attributesSchema. Validation is a guardrail in the app and a TypeScript error for developers. It never blocks a write. No protocol change was made for schemas.

Direction. Schemas should eventually be a daemon-level resource with resources as the primary entity, so the daemon can index and query typed attributes generically. Until then "documents as resources" is the chosen interim. Schema recursion is an open question. Both MIME types and schemas will be supported as ways to say what a value is. The generic indexer for arbitrary attribute queries is being built incrementally and already backs the attribute listings and the query grammar.

Affected pages. Schemas, Query grammar.

A custom CRDT

Where it comes from. 2026-09-03 sync.

The team decided to proceed with nodes over the current change data and to defer a generic JSON CRDT. The team rejected existing libraries because they lack history garbage collection and decentralised, signed attribution. The plan is a custom CRDT shaped for signed history. Snapshots against full history is an open debate. Trustless attribution needs the full history, while signed snapshots that can walk back would be cheaper. Nothing here changes how today's change graph is read.

Affected pages. Documents, Change.

Web of trust

Where it comes from. Long-standing; restated in 2026 notes on trust in the agent era.

Today. A contact is a public statement that you know an account by a name, with two follow flags. It grants nothing, and there is no transitive trust anywhere in the code. What exists is a capability system plus a public address book.

Direction. The team does not plan a global web of trust in the PGP sense. The current framing is federated, contextual trust rooted in communities: endorsements within a community, authority rankings set by an organiser for a group, and moderation that grows from the edge. No data model has been decided.

Affected pages. Permissions, Contact, Signed content.

Smaller items with a decided direction

Collections. A collection is a document whose content is one top-level query block that lists its children. The indexer reports it as isCollection, and the older type: Collection attribute is ignored. It is an app presentation convention, and the protocol has no collection concept. See query blocks.

Extensions. A plugin system is planned. Extensions are documents that carry small apps a site can install, starting with custom pages, then custom blocks, attribute editors and themes. A first attempt was built on a branch and closed without merging in September 2026. The system is being redone and nothing ships today.

Comment moderation. Anyone can comment on anything, and the protocol has no moderation. Deleting and revoking comments on a site was on the mid-2026 launch checklist and is still a site-side design (Comments).

Relays and reachability. Networks that block the peer-to-peer port, such as university networks, motivated a proposal for relay addresses on port 443. Status unknown; see Network.

How to read this page

If a concept page and this page disagree, the concept page describes the code and this page describes intent. If the code and this page disagree, the code wins and this page is out of date. Please say so in a comment.