mantra-kmp/docs/subgroups.md

Subgroups

A group makes another group, and the child can prove where it came from. This is how the four ceremonies fit together, what the parent's signature actually covers, and the one thing a subgroup cannot inherit.

Read shared-key-derivation.md and shared-key-ceremony.md first. Everything below is built on the property they state — that a room's id is the key it signs with — and on the ceremony they describe, which a subgroup runs again from scratch.

Built, phases 1-9, one commit each. The phases are kept as written because they are the reasoning, and the code reads better against the argument it came from than against a summary of itself. Where the implementation chose differently the section says so, and it did so six times worth reading:

what the plan said	what it turned out to be
SubgroupManager.openCeremony	never built. The picker calls createNip17ChatRoom and proposeRitual directly, exactly as robust-group creation does -- a wrapper over two repository calls would have been a third name for one act
MarmotGroupCreation.create called from the view model	reached through ChatRepository.createMarmotGroup. View models here talk to repositories and managers take the database; the first cut reached for a database the interface does not expose, and should not
a pure SubgroupGuardsTest in commonTest	jvmTest. Every refusal reads the database -- a key-holding session, an admin flag, a ceremony in the derived room -- so a pure version would have tested less, not differently
Phase 4 adds one tag	two. subject fixes a bug older than subgroups: getOrCreateNip17ChatRoom has always read a subject off the payload and the ritual path never wrote one, so every robust group arrived nameless on every device but its creator's
stateFrom reads the parentage with two parsers	three, and a wrapper. A tag present and unreadable looks absent through a parser, so claimsParentage reads the tag names; and "refused" has to be told from "none claimed", which a bare nullable cannot carry
Phase 8 adds three refusals to a settled function	it short-circuited the tests that were already there. canSign and isAdmin run before the picking rules, so the fixtures had to grow a parent that can actually sign and a coordinator who is actually an admin -- see the note in that phase

Two things the plan got right that are worth keeping if any of this is rewritten. The key-package check moved to the picker before a line of it was built, on review, and it is the difference between a subgroup failing in a second and failing after three ceremonies. And the founding-roster rule -- that certifies must not compare the name or the p tags against anything current -- has a test whose job is to fail the day somebody adds the comparison that looks obviously missing.

Written against the code as it stood at b50b1762.

What a subgroup is

A subgroup is an ordinary Marmot room with an ordinary threshold key, made by an ordinary ChillDKG ceremony, plus a birth certificate — the parent group's signature over the child's room id — carried on the child's key state.

The key state grows two tags for it, not one, and only the first is an artefact:

tag	what it is
birth_certificate	the parent's signed statement, whole. The claim itself
parent_group	the parent's room id — an index into the certificate, which already carries the same value in its own parent_group tag and as its author

The second is redundant by construction and is kept anyway, so a reader can answer "whose child is this" without parsing an event out of a tag value. What stops it being a second, weaker source of truth is Phase 3's rule that a state carrying one without the other, or the two disagreeing, is dropped entirely. There is no state in which the index is believed and the certificate is not.

That is the whole of the relationship. The child does not derive its key from the parent's, does not share a share with it, and cannot be signed for by it. What the certificate buys is a checkable claim, and only that claim:

The group holding key P said, with a quorum, that the room C is its child.

Everything else about a subgroup — how it signs, who administers it, what it holds — is the same as any other robust group in this app, and deliberately so.

Why the key is fresh rather than derived

SharedKeyDerivation can already walk a group's key to any path, and m/9420/1/0 would give a child room for free with no ceremony at all. It is the wrong answer for one reason, stated in that file:

Anyone who learns one derived private key recovers the threshold key and can sign as the group without any quorum at all.

A derived child is the parent wearing a different hat. Its members would be the parent's members, its quorum the parent's quorum, and the parent's admins would be able to sign as it. A subgroup exists precisely so that a different set of people — including people the parent does not trust to administer the parent — can act on their own. A fresh ceremony is what makes the child's authority the child's.

The cost is that a subgroup is not cheap: four multi-party rounds, every one of which needs every selected member to show up. That is named again under what this does not do.

The order, and why it is forced

  1. ChillDKG among the new admins       -> threshold key K, and so C = marmotGroupId(K)
  2. birth certificate, in the parent    -> parent's quorum signs C
  3. key state, among the new admins     -> child's quorum signs "C signs with K", carrying the certificate
  4. the Marmot room                     -> coordinator creates C and welcomes the admins

Nothing here is a policy choice. Step 2 needs C, which does not exist until step 1 produces K. Step 3 carries the certificate, so it needs step 2. Step 4 is gated on the key state for the same reason DkgRitualViewModel.createAdminGroup is today — a room created before its group has agreed what it signs with is a room whose founding fact is settled after the founding.

The coordinator is the parent admin who pressed the button. They coordinate the child's ceremony, propose the certificate in the parent, propose the key state in the child, and create the room. ChillDKG and FROST both treat a coordinator as untrusted, so this buys them nothing but work — the same bargain the existing ceremony makes.

Only step 1 is theirs alone. After the ceremony fixes the participant set, every remaining step is open to somebody else: the certificate can be proposed by any parent admin holding a share of the parent's key, and the key state and the room by any of the child's admins, who by then all hold shares of K and know C. So a coordinator whose phone dies after the ceremony does not strand a subgroup — the flow is resumable by anyone who was in it, which is why every step reads its state off stored rows rather than off a session object. The UI should make that reachable rather than merely true: Phase 7's rung is offered to whoever opens the screen and can act, not only to the member who started.

Three ceremonies, two quorums, one coordinator

step	runs in	signed by	transport
ChillDKG	a NIP-17 room over the child's admins	n/a — every participant	gift wraps
birth certificate	the parent Marmot room	the parent's quorum, as the parent's key	MLS group events
key state	the child's NIP-17 ceremony room	the child's quorum, as the child's key	gift wraps

The middle row is the only new shape, and it is not new machinery: FrostSigningManager already runs on both transports and already gates each signer's approval behind the existing proposal UI. What is new is a kind for it to carry and a reason for the parent's admins to say yes.

Informed consent, and what the parent's admins are actually signing

The spec for the certificate is "the hex of the new groupId, signed with the existing groupId's key". Taken literally that is 32 opaque bytes: a parent admin would be asked to put the group's signature to a number they cannot check, produced by a ceremony most of them were not in, on behalf of people they have only the coordinator's word about.

So the content is exactly the new group id, as specified, and the tags carry what makes it checkable. The signature covers both — an event id hashes over its tags — so nothing is added to the claim by putting it there, only to what a signer can see before agreeing:

kind:    30329
pubkey:  <the parent room's id>            (set by the signing session, not the proposer)
content: <the child's room id, 32-byte hex>
tags:
  ["d",            <the child's room id>]        addressable; newest certificate per child wins
  ["parent_group", <the parent room's id>]
  ["subgroup_key", <the child's threshold key, 33-byte hex>]
  ["frost_path",   "m/9420/0/0"]
  ["name",         <the subgroup's name, as the coordinator typed it>]
  ["p",            <each of the child's admins>]

With those, a parent admin's device can check the thing that actually matters before it signs — marmotGroupId(subgroup_key, frost_path) == content — and ProposedEvent can render "Translation team, for Alice, Bob and Carol" rather than a hash. A coordinator who lies about who is in the child is then lying in a field the parent's signature covers, which is the difference between a mistake and evidence.

The name is in there for the same reason and not for the room's benefit: the room takes its name from MarmotGroupData like every other Marmot room, and this copy exists so that what the parent approved is legible in the transcript and in the proposal screen. Two consequences to hold:

The name and the p tags are the founding roster, and they are frozen. A certificate is signed once and members move afterwards — somebody is added to the child, somebody leaves, the room is renamed. None of that reaches a signature already made, and none of it should: the certificate says who the parent certified and under what name, which is a historical fact and stays true. So the UI must never render either as the current state of the subgroup. Where the room exists locally, its own row is the live answer; the certificate is what it was born as. The Phase 7 list follows that rule and the plan says so again there.

A rename therefore drifts from what was signed, on purpose. The alternative — leaving the name out so nothing can drift — buys consistency by making the parent's admins approve an unnamed hash, which is the problem this section exists to fix. Drift in a historical record is not an error; an unreadable approval is.

Phase 1 — the certificate, and checking one

New package nostr/subgroup/, no database, no coroutines — the same shape as GroupKeyStateEvent, and testable the same way.

SubgroupBirthCertificateEvent at kind 30329, the next free kind after ChronicleRequestEvent (30328). Sits in the private 303xx range with everything else that never leaves an encryption.

object SubgroupBirthCertificateEvent {
    val KIND: Kind = 30329

    fun assembleTags(
        subgroupChatRoomId: String,
        parentChatRoomId: String,
        thresholdPublicKey: HexKey,
        adminPublicKeys: List<HexKey>,
        name: String,
        path: List<Long> = SharedKeyDerivation.MARMOT_ADMIN_GROUP_PATH
    ): Array<Array<String>>

    fun parseSubgroupChatRoomId(tags: Array<Array<String>>): String?
    fun parseParentChatRoomId(tags: Array<Array<String>>): String?
    fun parseThresholdPublicKey(tags: Array<Array<String>>): HexKey?
    fun parsePath(tags: Array<Array<String>>): List<Long>?
    fun parseAdminPublicKeys(tags: Array<Array<String>>): List<HexKey>
    fun parseName(tags: Array<Array<String>>): String?

    /** Whether [event] is a certificate the room [parentChatRoomId] actually signed for [subgroupChatRoomId]. */
    fun certifies(event: Event, subgroupChatRoomId: String, parentChatRoomId: String): Boolean
}

certifies is where the whole of the trust in a parent link lives, and it is six questions, all answerable from the event:

1. the kind is 30329;
2. content == subgroupChatRoomId, and the d tag agrees with it;
3. the parent_group tag is parentChatRoomId;
4. SharedKeyDerivation.marmotGroupId(subgroup_key, frost_path) == content — the id really is that key's room, so a certificate cannot be pointed at a room the key does not derive;
5. GroupKeyStateEvent.isSignedByRoom(event, parentChatRoomId) — the author is the parent's id, the id is the hash of the fields beside it, and the signature verifies;
6. everything above inside a runCatching, because every input is off the wire.

Point 5 is GroupKeyStateEvent.isSignedByRoom used verbatim, not reimplemented. It already asks "did this room sign this", and a room id is a public key here — that is the whole economy of docs/member-chronicle.md and it applies unchanged.

The name and the p tags are deliberately not among the six. They are covered by the signature — everything in the tags is — but nothing downstream may require them to match anything, because they are the founding roster and the world moves. A certifies that compared the p tags against the room's current members would start rejecting a valid certificate the first time somebody joined the child, and the failure would look like a forgery.

Two tag classes beside it, in nostr/subgroup/tags/, matching FrostDerivationPathTag's shape: SubgroupParentTag (parent_group) and SubgroupKeyTag (subgroup_key). The d and p tags are quartz's; name is a bare two-element tag with no class of its own, since nothing parses it but the proposal screen.

Tests — SubgroupBirthCertificateEventTest, commonTest, pure: a certificate that verifies; one whose content does not match its d tag; one whose id does not derive from its key; one signed by a different room; one whose signature is 64 bytes of nonsense; one whose key is not a point on the curve; and one whose name and admin set have nothing to do with the room's current ones, which must still verify — that last is there to fail loudly if anybody later adds the roster check the paragraph above forbids.

Phase 2 — schema 16 → 17

Four nullable columns and nothing else, so Room migrates itself and the entry joins the list in MantraDatabase.

table	column	filled from	trusted?
GroupKeyState	parentChatRoomId	the state's parent tag	yes — stateFrom verified the certificate
GroupKeyState	birthCertificateJson	the state's certificate tag	yes — same
ChatRoom	parentChatRoomId	the child's key state, as the room is created or adopted	yes
DkgSession	parentChatRoomId	a tag on the ceremony proposal	no — a hint for the UI, see Phase 4

The trust column is the point of the table. Two of these are written only after a signature has been checked and one is written from an unauthenticated claim on a wire message; a column that mixes the two is a column no reader can act on. The DkgSession one is never read for anything but a screen's title.

parentChatRoomId is a plain column, never a foreign key

Worth its own heading because the reflex points the wrong way. GroupKeyState, DkgSession and GroupSignedEvent each declare

ForeignKey(entity = ChatRoom::class, parentColumns = ["id"], childColumns = ["chatRoomId"],
           onDelete = ForeignKey.CASCADE)

and the obvious next move is to give the parent pointer the same treatment. It must not have it. A self-referential foreign key on ChatRoom with CASCADE means deleting a parent room deletes every subgroup row beneath it — and then, by their own cascades, each subgroup's messages, participants, key state, signing sessions and signed events. A user tidying up a group they have left would silently destroy a group they are still in.

RESTRICT is not the answer either: it would make a parent undeletable while any child row exists, which is a foreign key deciding a product question. The parent pointer is a reference to a room that may not be on this device at all — a member of a subgroup who was never in its parent has the id and nothing else — so it cannot be a foreign key in the first place. A dangling pointer is the normal, expected state, and readers resolve it with a lookup that is allowed to return null.

The whole certificate event is stored as JSON rather than the signature alone. That is deliberate and is argued in the appendix: a signature plus a rule for rebuilding the event it covers is a rule that breaks silently the first time the event's shape changes.

New DAO reads, all straightforward:

• GroupKeyStateDao.getByParentChatRoomId(parentChatRoomId) and an observe beside it.
• ChatRoomDao.observeByParentChatRoomId(parentChatRoomId).
• DkgSessionDao.getByParentChatRoomId(parentChatRoomId), for resuming a ceremony whose room has not been made yet.

Phase 3 — a key state that names its parent

GroupKeyStateEvent gains two optional tags, and GroupKeyStateManager gains the checks that make them mean something.

["parent_group",     <the parent room's id>]
["birth_certificate", <the signed 30329 event, as JSON>]

assembleTags takes an optional parent: SubgroupParentage? carrying both, and emits neither when it is null — so every existing call site and every state already signed is untouched.

The work is in stateFrom, which is where a state earns its row. After the two checks it already makes — the room rederives from the key, and the group signed it — a state carrying either new tag has to pass four more:

1. Both or neither. A parent tag with no certificate, or a certificate with no parent tag, is dropped. A half-claim is not a weaker claim; it is a claim with the checkable part removed.
2. The certificate parses as an Event.
3. SubgroupBirthCertificateEvent.certifies(certificate, state.chatRoomId, parentChatRoomId) — the parent signed this child, not some other one.
4. The certificate's subgroup_key equals the state's own threshold key. Redundant with (3) via the derivation, and kept because it costs nothing and the two facts are stated in different places.

A failure at any of them drops the whole state, rather than keeping it as a parentless one. A state that claims a parentage it cannot back is not a state with one field wrong — it is a device asserting a relationship the parent group never agreed to, and half-believing it is worse than believing none of it.

GroupKeyStateManager.propose takes the same optional parentage and passes it through. record and adopt carry the two verified values onto the row; they already funnel through stateFrom, so nothing else is needed to keep the row honest.

Tests — added to GroupKeyStateTest (commonTest, pure stateFrom cases) and SignedGroupKeyStateTest (jvmTest, real FROST signatures end to end): a subgroup state that verifies; one whose certificate names another child; one whose certificate is signed by a room that is not the named parent; one carrying a parent tag and no certificate; one carrying a certificate whose key is not the state's.

Phase 4 — where the subgroup's ceremony runs

The child's ChillDKG needs a room whose participant set is exactly the child's admins, and whose transport is gift wraps. There is already one: chatRepository.createNip17ChatRoom over the selected admins plus the coordinator. Its id is ChatRoom.deriveChatRoomId over the member set, so every device lands on the same room, and ChillDkgRitualManager needs no change at all — its broadcast p-tags come from that room's participants and acceptProposal derives n from those same p-tags.

This is the same move SelectChatRoomTypeViewModel.createNip17ChatRoom already makes for a robust group, and the reason step 1 costs almost no new code.

Two tags are added to the ceremony proposal, on DkgRitualEvents.PROPOSAL only, and both are cosmetic in the strict sense that nothing acts on either.

parent_group is parsed by acceptProposal onto DkgSession.parentChatRoomId. It authenticates nothing — anyone can claim any parent — and is read for exactly two things: the ritual screen saying "a subgroup of Ekklesia" instead of "a shared key ceremony", and a member finding their way back into a flow they closed the app halfway through. The load-bearing claim is the certificate, three steps later, and the code should say so where the column is declared.

subject is the fix for a gap this flow inherits and makes worse. NostrDao.getOrCreateNip17ChatRoom already builds the receiving side's room with subject = decryptedGiftWrapPayload.parseSubject(), and ChillDkgRitualManager.broadcast writes no subject tag — so today a member selected for a ceremony watches an unnamed chat room appear on their device with a key ceremony already running in it. That is survivable when the user just agreed to make a group with those people; it is not when the room is a means to an end they were not consulted about. The proposal carries the subgroup's name, the receiving room gets it for free through a reader that already exists, and the change is one tag on one broadcast.

It is worth fixing for the existing robust-group flow in the same commit, since SelectChatRoomTypeViewModel passes a subject to createNip17ChatRoom that only ever reaches the creator's own device.

The collision this buys, and it is real

A NIP-17 room's id is a pure function of its members, so one admin set gets one ceremony room, forever. Two consequences, and both need a guard rather than a comment:

• Selecting an admin set that already holds a completed ceremony returns that ceremony (proposeRitual hands back anything not FAILED), which would make the "new" subgroup the old group under a new name — same key, same room id.
• Selecting every member of the parent lands in the parent's own ceremony room, which is the room over all its members.

The second is not refused, and the first is refused more narrowly than it first shipped. A subgroup over everybody is legitimate -- a subgroup is a logical division rather than a smaller membership -- and sharing a room was never the problem; sharing a ceremony was. DkgSession.parentChatRoomId tells two ceremonies in one room apart, and DkgSessionDao.getLatestSessionFor is scoped by it, so what Phase 8 refuses is only the same parent asking twice over the same admins. See that phase.

Why not the parent's Marmot room

The obvious alternative is to hold all three steps in the parent room: no sibling room, no member-derived id, and the collision above disappears. It is a better design and it is not available yet. What follows is why, because "we did not" without "and here is the price" is how a decision gets re-litigated every six months.

It is one decision, not three. The certificate already runs in the parent room and is the only one of the three free to run anywhere. The key state is not:

// GroupKeyStateManager.propose
check(chatRoomId == signingRoomId || key.chatRoomId == signingRoomId)

// FrostSigningManager.signingPath, the one case that cannot be self-checked
if (localChatRoom.chatRoom.mlsGroupState == null && key.chatRoomId == chatRoomId) {
    return SharedKeyDerivation.MARMOT_ADMIN_GROUP_PATH
}

The child's room does not exist, so chatRoomId == signingRoomId cannot hold and the key state has to be signed where its ceremony ran. Both guards say so independently. So the ceremony and the key state move together or not at all, and the question is only ever "which transport does the child's ChillDKG ride".

What moving it would cost, mechanically. Three things, all enumerable:

1. ChillDkgRitualManager.broadcast writes only a GiftWrapPayload; it needs the two-transport shape FrostSigningManager.broadcast already has, plus a Marmot inbound arm in NostrDao beside the one dispatching FROST signing at kind 30320-30325. The p-tags stay on both transports, unlike FROST's — for a ceremony the p-tag set is the participant set every device derives n from, and inside MLS encryption naming the participants leaks nothing.
2. DkgSession.chatRoomId stops identifying a ceremony. Three queries key on it and all three break: proposeRitual's one-ritual-at-a-time guard would return the parent's own completed session and refuse to open a subgroup ceremony at all; observeLatestSessionForChatRoom would show the child's ceremony on the parent's shared-key screen; and completedKey's third fallback would resolve the child's key for the parent room, for any parent member holding no key state row — which is every member welcomed after the parent's own ceremony. Fixable with a subjectChatRoomId column and three scoped queries, but that third one is a load-bearing fallback and the failure is silent.
3. signingPath's special case is gated on mlsGroupState == null and would have to admit a Marmot room. Still device-local — both inputs are read from this device's own database, so a proposer still chooses nothing — but it widens the one function whose contract is that the path can never come off a proposal.

Why not, even after all that. docs/mls-skipped-keys.md names ChillDkgRitualManager.proposeRitual in its table of reliable triggers — "proposal, then the host key" — and today that costs nothing, because the ritual runs on gift wraps and the bug is in MLS. Moving it onto group events puts a ceremony on the one transport where a message arriving a moment late is dropped for good.

A DKG is the worst possible payload for that. It cannot finish until every participant takes part, so one lost round-1 message stalls it permanently for everybody, and the ritual's own resume machinery cannot help: relays redeliver gift wraps and replayStoredMessages re-feeds the backlog, while a group event whose inner event never materialised leaves a stored MarmotGroupEvent row and nothing to replay. FROST signing survives the same transport only because it needs t of n and routes around the signer whose nonce was lost.

MlsGroupCache covers the burst-in-one-sync case and nothing else: a restart loses the skipped keys, and any other writer in the room — a chat message, a member being added — invalidates the cache mid-ceremony.

So: revisit when the quartz fix lands. Once saveState/restore carry the skipped keys, the parent room is strictly the better home for all three steps, the collision above stops being a refusal and becomes a non-issue, and DkgSession.parentChatRoomId stops being needed at all. Until then the child's ceremony rides the transport that tolerates reordering, and pays for it with one sibling room and one refusal.

Phase 5 — the manager, and the four steps in order

New managers/SubgroupManager.kt, holding the orchestration that does not belong to any one of the three existing managers. It calls them; it does not reimplement them.

object SubgroupManager {
    /** The NIP-17 room a ceremony for these admins would run in. Pure. */
    fun ceremonyRoomIdFor(adminPublicKeys: Set<HexKey>, coordinator: HexKey): String

    /** Step 1: stand up the ceremony room and open the ritual in it. */
    suspend fun openCeremony(...): DkgSession

    /** Step 2: ask the parent's quorum to certify the child. Runs in the parent room. */
    suspend fun proposeBirthCertificate(
        database: MantraDatabase,
        parentRoom: LocalChatRoom,
        userPublicKey: HexKey,
        key: DkgSession,
        adminPublicKeys: List<HexKey>,
        name: String,
        path: List<Long> = SharedKeyDerivation.MARMOT_ADMIN_GROUP_PATH
    ): FrostSigningSession

    /** The newest certificate for this child that actually certifies it, or null. */
    suspend fun certificateFor(
        database: MantraDatabase,
        subgroupChatRoomId: String,
        parentChatRoomId: String
    ): GroupSignedEvent?

    /** Step 3: ask the child's quorum to sign its key state, carrying the certificate. */
    suspend fun proposeSubgroupKeyState(...): FrostSigningSession

    /** Everything the parent has certified, whether or not the room exists here. */
    suspend fun subgroupsOf(database: MantraDatabase, parentChatRoomId: String): List<Subgroup>

    /** Files a certificate that arrived, or drops it and says why. */
    suspend fun record(database: MantraDatabase, chatRoomId: String, innerEvent: Event): GroupSignedEvent?
}

proposeBirthCertificate is one FrostSigningManager.proposeSigning call in the parent room with key = null — the parent's key is resolved from the parent room the way every other signature in that room is, and a caller naming a ceremony here would be a caller choosing what the group signs as. A session of one, never batched, for the same reason GroupKeyStateManager.propose gives: a batch is only as available as its worst item, and this one is a precondition of everything after it.

proposeSubgroupKeyState is GroupKeyStateManager.propose with the parentage filled in, and it must refuse to run without a certificate that passes certifies. Proposing a state the coordinator's own device would drop is not a mistake worth letting the group discover by signing it.

subgroupsOf reads the parent's GroupSignedEvent rows of kind 30329, filters them through certifies, and joins each to the local ChatRoom if there is one. No table of subgroups: the certificates are the record, they are the group's own signed statement, and every device in the parent room already has them — FrostSigningManager.complete files a GroupSignedEvent on every device that followed the session, not only on the signers'.

Two certificates for one child is a normal outcome, not a conflict. Two parent admins can press the button on the same admin set: the second lands in the same NIP-17 ceremony room and gets the first's ceremony back, but both may go on to propose a certificate, and both sessions can complete. GroupSignedEvent is keyed on the event id, so the rows coexist. certificateFor therefore takes the newest of those that pass certifies, the way GroupKeyStateManager.signedAmong already picks the newest state — and because both certificates say the same true thing about the same child, which one wins does not matter. The d tag makes them replacements of each other rather than an accumulation, which is what it is for.

Two dispatch arms, both easy to forget

ChatMessage.applyInnerEvent needs a 30329 arm. Without one the certificate falls to else -> and every parent member gets a raw-JSON chat bubble; that arm is the failure mode docs/member-chronicle.md names for an old build meeting a new kind, and here it is avoidable. The arm delegates to SubgroupManager.record, which verifies against the room it arrived in — the event reaches this path both from FrostSigningManager.applySignedEvent, where it is already checked, and from an arriving inner event, where it is not, and the arm cannot tell which.

Unlike the key-state arm it does write a chat line. A key state is standing state whose session already wrote the transcript; a subgroup being born is something that happened, and it happened on behalf of parent members who are not in the child and will otherwise never be told.

ProposedEvent.summaryOf needs a 30329 arm too, or the parent's admins approve "Event of kind 30329" with the JSON underneath. It should read as the members and the path — the id is the thing being decided and is shown by the screen anyway.

Not chroniclable

30329 does not join ChronicleEvent.CHRONICLABLE_KINDS, for the reason that list gives for excluding GroupKeyStateEvent: verification admits an event on the strength of the group's signature, so any chroniclable kind is replayable by any member forever. A member added to the parent after a subgroup was made therefore sees an empty subgroup list — named under what this does not do, with the argument for changing it later.

Phase 6 — creating the room, once, for both flows

DkgRitualViewModel.createAdminGroup is 120 lines of Marmot room creation living in a ViewModel: resolve every member's key package or refuse, build MarmotGroupData with the admin list baked into epoch 0, MlsGroup.create, getOrCreateChatRoom, adopt the key state, add the members. The subgroup needs all of it and differs in four values.

So it moves, unchanged in behaviour, to managers/MarmotGroupCreation.kt:

suspend fun create(
    database: MantraDatabase,
    chatRepository: ChatRepository,
    groupId: String,
    name: String,
    purpose: String,
    adminPublicKeys: Set<HexKey>,
    userPublicKey: HexKey,
    keyPair: KeyPair,
    path: List<Long> = SharedKeyDerivation.MARMOT_ADMIN_GROUP_PATH,
    parentChatRoomId: String? = null
): Outcome   // Created(room, notAdded) | BlockedOn(missingKeyPackages) | Existing(room) | Failed(reason)

Both call sites then read as four values and a result. The rules that are already right stay right for the subgroup for free, and they are the ones worth not re-deriving:

• Every key package before anything exists. The id is derived, so there is exactly one room per key at this path; a half-created one occupies that address permanently and there is no second id to retry with. Better to create nothing and name who is missing.
• adminPubkeys baked into epoch 0, so a member welcomed later gets a populated group rather than chasing a bootstrap commit that predates them.
• adopt before the members are added, because filing the key state is local and certain and adding members is neither.

parentChatRoomId is the one genuinely new parameter, written onto the ChatRoom row. The subgroup's description carries the path through SharedKeyDerivation.describe, exactly as the admin room's does — MIP-01 has no field for either, and the path is what rebuilds the TweakCache.

The description carries the parent too, as a hint

describe gains a second optional line, beside the path:

Admins of Ekklesia's translation work.

Shared key path: m/9420/0/0
Parent group: 4f2b…

It is there for one member and one problem: somebody welcomed into the subgroup after it was founded. The verified parent link lives on GroupKeyState, filed by adopt from the signed key state — which only ceremony participants hold — and certificates are not chroniclable, so a later joiner has a room, no key state, and no way to learn the room is anybody's child. MarmotGroupData is the one thing that does reach them: it rides in the Welcome and is baked into the epoch-0 group context, so every member gets it however late they arrive.

It is a hint and must be rendered as one. Group data is written by the room's creator and agreed by MLS, not by the parent — so this line says "this room claims Ekklesia as its parent", which is a strictly weaker statement than the certificate makes. The rule for the UI is the one Phase 2's table states: where a verified GroupKeyState.parentChatRoomId exists, it wins and the hint is never consulted; where it does not, the hint may be shown as unconfirmed and must not be written into the verified column. The upgrade path is the certificate itself — see what this does not do.

Phase 7 — the UI

Four pieces, three of them small. Every one of them is subject to ./gradlew :composeApp:m3Audit: strings in strings.xml in sentence case, spacing from MaterialTheme.spacing, colour from a role, minimumInteractiveComponentSize on anything clickable that is not an IconButton, and a decided contentDescription.

A subgroups section on ChatRoomDetailScreen, between members and the danger zone, listing SubgroupManager.subgroupsOf with one of three supporting lines per row — certified, not yet created / created, you are not a member / the room's own subject if this device is in it — and an "add subgroup" TextButton above them in the shape the existing "invite new member" one has. The row's tap target navigates into the child's room if this device has it and to the certificate's NostrEventDetailRoute if it does not.

The row's title comes from the room where there is one, and only otherwise from the certificate. That ordering is the frozen-roster rule from the consent section applied at the only place it is visible: the certificate's name and p tags are what the parent approved at founding, and a subgroup that has since been renamed or has added members would otherwise be listed under a name nobody uses and a membership nobody has. Where this device holds the child's room, the room is the live answer. Where it does not, the certificate is all there is, and the row says so by carrying its supporting line rather than pretending to be current.

Offered only where it can work: this device is an admin of the parent (Participant.adminAt != null), the parent can sign (FrostSigningManager.canSign), and the parent is a Marmot room. Hidden, not disabled, in the same way the shared-key entry is hidden on an MLS room.

A parent row on the child's detail screen, read from the verified GroupKeyState.parentChatRoomId, sitting directly under the signing-key card — where a member looks to find out what the room is.

SelectSubgroupAdminsScreen and its route and ViewModel, modelled on SelectChatRoomMembersScreen. The pool is the parent's own participants, admins and non-admins alike, with the parent's admins marked rather than filtered — the point of a subgroup is that it can be run by people the parent does not let run the parent.

A name field at the top, because nothing else in the flow can supply one. createAdminGroup synthesises "${parent.subject} (#admins)" and gets away with it — a group has exactly one admin room and the name states a relationship rather than a choice — but a group can have many subgroups and "Ekklesia (#subgroup)" names none of them. The name is required, travels on the route to Phase 6's MarmotGroupData, and is copied into the birth certificate so the parent's admins approve something legible.

Key packages, checked here rather than discovered at step 4

The single most likely way this flow fails, and the cheapest to prevent.

A key package is one-time-use: NostrDao marks the bundle consumed = true as the device processes its own Welcome, and MarmotKeyPackageBundleDao only ever hands back one with consumed = false AND rotated = false. Every group a member joins burns one. So a member who is in the parent room and has not published a fresh key package since has none left, and Phase 6 will refuse to create the subgroup for exactly the right reason — after a ChillDKG, a parent quorum and a child quorum have all completed, each of which required every selected admin to show up and approve. Three multi-party ceremonies, then a dead end, and no second room id to retry against.

So availability is a property of the picker, not of the create step:

• Prefetch every candidate's key package as the screen opens, the way SelectChatRoomMembersViewModel.scheduleKeyPackageEventSynchronization already does for group creation — by the time somebody has finished ticking names and set a threshold, the relay round trip is usually done.
• Mark a member with none, and refuse to select them, with the reason and the remedy in the same line: they need to publish a new key package. Anything vaguer makes it the coordinator's problem to diagnose.
• Re-check at confirm, because the screen may have been open a while.

Phase 6's refusal stays exactly as it is. It is now a backstop rather than the first line of defence — a member's package can be consumed by some other group between the picker and the create step — and a backstop is worth having precisely because the address it protects is permanent.

Three admins in total, the coordinator included, so the picker asks for at least two others. Three is ChatRoomType.MINIMUM_ROBUST_GROUP_SIZE, and it is the same floor for the same reason it states: below three a quorum is not a check, because two admins means every decision needs both of them and one means the coordinator is deciding alone. The coordinator counts because they are a ChillDKG participant by construction — they hold a share whether or not anybody ticked their name — so the screen shows their own row selected and locked rather than leaving them out of the count the confirm button is testing.

The threshold is the coordinator's to set, on this screen, and only here. A stepper over ChatRoomType.quorumRange(adminCount) — 2..3 for the smallest subgroup — starting at ChatRoomType.defaultQuorum(adminCount), which is 2 of 3. It has to be settled before anything is published, and that is a protocol fact rather than a layout preference: ChillDKG hashes the threshold and the host keys into the session identity, so t is fixed the moment the proposal goes out and a group that disagrees about it gets no key at all rather than a weak one. There is no later screen where it could be changed, and the plan should not pretend otherwise.

It is also the one value the coordinator picks for other people. Consent is not lost by that: the threshold rides on the proposal in DkgThresholdTag, every invitee's acceptProposal re-checks it against quorumRange and drops a proposal outside it, and the HOST_KEY gate is where each of them agrees to the t-of-n they can now see. Joining is the agreement; the coordinator only writes down what is being joined.

Confirm stays disabled below two selections, and below a threshold outside the range, with the reason stated rather than left to be guessed.

DkgRitualRoute gains an optional parentChatRoomId, and the ritual screen grows one rung. The alternative — a parallel subgroup screen — duplicates a ladder, a threshold picker, three approval gates and a key-state rung in order to insert one step, and the two copies would drift within a release. The rung is rendered only when a parent is present, and it is where the coordinator proposes the certificate and watches the parent's quorum answer.

The approvals themselves need no new UI. The certificate is a FrostSigningEvents.PROPOSAL in the parent room and the key state one in the ceremony room; ProposalListScreen and FrostSigningScreen already show and approve both, on both transports.

Phase 8 — the refusals

A subgroup is expensive and its room id is permanent, so every one of these is checked before the button is offered and again in the manager. The UI check is what keeps a user out of a dead end; the manager check is what keeps a bug out of a room nobody can replace.

refused	because
fewer than three admins in total — the coordinator plus two	below three a quorum is not a check — the same floor ChatRoomType.MINIMUM_ROBUST_GROUP_SIZE states
a threshold outside ChatRoomType.quorumRange(adminCount)	t = 1 is a threshold key any one member signs with, and ChillDKG will happily generate one. acceptProposal already refuses such a proposal, so an unchecked one produces a ceremony every invitee silently drops
an empty or blank name	Phase 6 has nothing to put in MarmotGroupData.name, and the parent's admins would be certifying an unnamed hash
a selected member with no unconsumed key package	they cannot be welcomed into the room at step 4, and finding that out at step 4 wastes three ceremonies. Checked at the picker and again at confirm
the selection is not a proper subset of the parent's members	withdrawn. A subgroup is a logical division, not a smaller membership, so "everybody" is a normal answer. What this was standing in for is the row below, checked precisely rather than by set size
the ceremony room already holds a non-FAILED DkgSession for this same parent	proposeRitual would return that ceremony, and the "new" subgroup would be the old one — same key, same id. Scoped by DkgSession.parentChatRoomId, so a room may hold the group's own ceremony and a subgroup's
this device holds no share of the parent's key	it cannot open the certificate session; proposeSigningBatch throws, and throwing at the button is not a UI
this device is not an admin of the parent	a non-admin proposing the parent's signature is a proposal the parent's admins have to decline by hand
the parent has no signed key state and no resolvable key	there is nothing for the certificate to be signed with
a certificate already stands for this child	one certificate per child; a second is a d-tag replacement of the first, not a second subgroup

The first is the one a user will actually meet, and it says what to do: "pick at least two more people".

Why "a subgroup cannot be the whole group" was withdrawn

It shipped, and it was wrong twice.

It refused something legitimate. A subgroup is a group deciding that some of its work belongs to a differently-keyed room, not a group carving out a smaller membership — so every member being in it is an ordinary case, and no rule here had any business deciding otherwise.

And it was a proxy rather than a check. The thing it stood in for is real: a ceremony room is derived from its admins, so a subgroup over everybody lands in the room the group's own ceremony was held in, and proposeRitual handing back that ceremony would make the child the parent. But set size does not detect that. A parent whose membership has changed since its own ceremony derives a different room, so the sizes can match with no collision and differ with one.

Sharing the room was never the problem; sharing a ceremony was. DkgSession.parentChatRoomId already told them apart, so the fix was to scope the lookup by it — DkgSessionDao.getLatestSessionFor(room, parent) — rather than to forbid the selection. A room may now hold two ceremonies, and everything below that lookup already keyed off the session id, so nothing else had to learn about the second one.

One collision survives, and it is degenerate: the same parent, over the same admins, twice. Those two have nothing left to distinguish them, which is another way of saying they are one subgroup asked for twice, and that is what the message now says.

Phase 9 — the tests that prove it

The pure ones carry the weight, because every check that matters is pure:

• SubgroupBirthCertificateEventTest (commonTest) — Phase 1's seven cases, including the one that must pass: a certificate whose name and admin set no longer match the room's, which fails the day somebody adds the roster check Phase 1 forbids.
• GroupKeyStateTest (commonTest) — Phase 3's five stateFrom cases.
• SubgroupGuardsTest (commonTest) — Phase 8's table, one case each, against the pure predicates rather than the ViewModel. The key-package row is the one worth writing first: it is the guard that exists to stop three ceremonies being spent on a room that cannot be created.

Two that need a database, in jvmTest, where secp256k1 loads and Room runs:

• SignedGroupKeyStateTest extension — a real ChillDKG key, a real parent key, a real certificate signed by the parent, a real child key state carrying it, and the assertion that stateFrom accepts it and rejects each single-field mutation of it.
• SubgroupDaoJvmTest — schema 17 round-trips the four columns, and subgroupsOf returns a certified child whose room does not exist locally.

One end-to-end, and it is the one that would have caught the collision in Phase 4: two ceremonies proposed over the same admin set land in the same room, and the second is refused rather than silently returning the first's key.

Rollout

Nothing here changes an existing flow's behaviour. The two shared pieces are the risk, and both are additive:

• GroupKeyStateEvent.assembleTags emits the new tags only when a parentage is passed, so every state signed before this ships verifies exactly as it did.
• GroupKeyStateManager.stateFrom's new checks fire only on a state carrying one of the new tags. A state with neither takes the path it takes today.

An older build meeting a subgroup's key state ignores two unknown tags and files the state as an ordinary one — it loses the parent link and keeps a correct answer to what the room signs with, which is the right way round. An older build meeting a certificate has no 30329 arm and writes a raw-JSON chat bubble in the parent room, once per subgroup. That is cosmetic, it is the same thing docs/member-chronicle.md reports for chronicle pages, and it is the reason to ship Phase 5's dispatch arm before anybody creates a subgroup rather than after.

Phases 1-3 can ship dark: they add a kind nothing produces and checks nothing triggers. Phases 4-6 are inert without Phase 7's entry point. So the flag, if one is wanted, is the button.

What this does not do

A subgroup cannot sign for its parent, and the parent cannot sign for it. Two independent keys, two independent quorums. The certificate is a claim about lineage, not a delegation of authority, and nothing in this design lets one group act as the other. If delegation is ever wanted it is a different mechanism and it should not be built on top of this one silently.

A member added to the parent later sees no subgroups, and a member added to the subgroup later cannot verify its parent. One cause, two faces, and both are the chronicle allowlist. Certificates are not chroniclable, for the reason Phase 5 gives, so a member welcomed into the parent after a subgroup was made holds no certificate and gets an empty list. A member welcomed into the child holds no key state either — adopt files one only from the signed event, which only ceremony participants have — so their verified parent link is null, and the description hint from Phase 6 is all they get: a claim by the room's creator rather than a statement by the parent.

Chronicling the certificate fixes both at once, and the argument for admitting it is better than it is for a key state: a certificate says "P certified C", a fixed historical fact, whereas a key state decides what a live room signs with, so a replayed certificate changes nothing a fresh one would not. It still deserves its own change — its own idempotency test, its own apply-order slot, and a decision about whether the child's chronicle may carry an event its own key did not sign, which no chroniclable kind does today. That last one is the real work, and it is why this is a follow-up rather than a line in Phase 5.

One parent, one admin set, one subgroup. Narrower than it first shipped: a room may hold two ceremonies, told apart by DkgSession.parentChatRoomId, so a group's own ceremony and a subgroup over the same people coexist. What is still refused is the same parent asking twice over the same admins, which have nothing left to distinguish them.

Widening that further needs the NIP-17 ceremony room to be distinguishable by something other than its members, and ChatRoom.deriveChatRoomId is a pure aggregate of member keys that the inbound path recomputes. Changing it is a change to how every NIP-17 room in the app is addressed.

It is also the limitation that disappears on its own: move the ceremony into the parent's Marmot room, as Phase 4 defers doing, and there is no member-derived id left to collide.

Every selected admin has to show up, twice. A ChillDKG cannot finish without all n, and the key state then needs a quorum of them. A member who never approves stalls the subgroup at step 1 indefinitely, and the only signal is the ritual screen naming who it is waiting on. This is inherited from docs/shared-key-ceremony.md and is the single biggest practical cost of a subgroup.

A subgroup of a subgroup is untested. Nothing forbids it — the child is an ordinary robust group with a key of its own, so it can certify a child in turn — and the chain is checkable link by link. But no phase here exercises depth 2, and the parent row on the detail screen shows one level rather than a path.

Removing a subgroup is not a thing. There is no revocation, and a certificate is a signature that exists forever once made. Deleting the child's room locally leaves the parent's list showing a certified child nobody has.

Appendix — what was considered and rejected

A derived child key at m/9420/1/0, with no ceremony. Free, instant, and wrong: docs/shared-key-derivation.md states that anyone learning one derived private key recovers the threshold key, and more to the point the child would be administered by the parent's members with the parent's quorum. That is a channel, not a subgroup.

The certificate as a bare 64-byte signature plus a rebuild rule. Smaller on the wire, and it makes GroupKeyStateManager.stateFrom responsible for reconstructing byte-for-byte the event the parent signed — kind, author, tags, and created_at — from fields carried alongside. The first change to the certificate's tags then invalidates every state already signed, silently, because a rebuilt event that differs by one byte hashes to an id whose signature does not verify and is indistinguishable from a forgery. Storing the whole event costs a few hundred bytes inside an encryption and removes the class.

A Subgroup table. The certificates already are the record: signed, held by every parent member, and verifiable without the table. A table would be a second copy that can disagree with them, and the only thing it would add is flow state for the coordinator — which DkgSession.parentChatRoomId covers with one column.

Running the child's ChillDKG and key state inside the parent's Marmot room. The better design, deferred rather than rejected, and argued in full in Phase 4. Short version: the two move together, the mechanical cost is three enumerable changes, and the reason to wait is docs/mls-skipped-keys.md — a ceremony needs every participant, and group events lose one of any two that arrive back to back.

A dedicated subgroup derivation path. m/9420/0/0 is reused, because the path is walked from the child's own key and two different keys at the same path derive two different rooms. A separate constant would say something true about lineage in a place nothing reads for lineage.

A parallel subgroup ritual screen. Rejected in Phase 7: it duplicates four pieces of a five-piece screen to insert one, and the copies drift.