Sealed-write — E2EE ingestion from a keyless writer

This is the answer to: can an external webhook post into an end-to-end-encrypted space without holding the space's keys? Yes — by publishing a space write key and letting the webhook encrypt to it.

The problem

Starfish's delegated encryption wraps content with a shared symmetric CEK that every reader also holds. To produce a valid ciphertext you need the CEK — so in that model, anyone who can write can also read. A webhook is exactly the party you do not want to grant read access: it should be able to inject a message, never to decrypt the history.

The mechanism

Sealed-write builds a write-only party on the keyring's existing anonymous-sender primitive (seal/unseal: ephemeral X25519 → HKDF-SHA256 → AES-256-GCM):

1. Mint a space write key.

   import { generateSpaceWriteKey } from "@drakkar.software/starfish-webhook"
   const { kemPubHex, kemPrivHex } = generateSpaceWriteKey()

   • kemPubHex (public) is published openly — e.g. in a plaintext space-config document any reader (and the webhook) can fetch.
   • kemPrivHex (private) is distributed to members out of band — in practice wrapped into the space keyring alongside the CEK, so every member recovers it exactly the way they already recover the CEK.

2. The webhook seals to the public key. Configure the route with seal + a sealer identity:

   routes: {
     bridge: {
       secret,
       transform,
       target: "/push/pubspaces/owner/space/streams/room",
       author: { edPubHex: BOT_PUB, edPrivHex: BOT_PRIV },
       seal:   { recipientKemPubHex: kemPubHex },
       sealer: { edPubHex: BOT_PUB, edPrivHex: BOT_PRIV },
     },
   }

   The message is sealed at the receiver, so a plain (none) collection stores only ciphertext ({ entry, ct }). The server never sees the cleartext.

3. Members open it.

   import { openSealedDocument, isSealedBlob } from "@drakkar.software/starfish-webhook"
   
   for (const el of pulledItems) {
     const msg = isSealedBlob(el.data)
       ? await openSealedDocument(el.data, kemPrivHex, { requireSealer: BOT_PUB })
       : el.data // a plaintext write from a normal client
   }

   requireSealer pins provenance: the open throws unless the blob's wrap entry was signed by the webhook's key, so a member can trust who injected the message against any outside party. (The signature commits to the wrapped content key, not the message bytes; forging a different message under a pinned sealer would require the per-message content key, which is fresh-random and known only to the sealer and key-holders. A member — who can already post — is therefore out of this guarantee's scope; it defends against non-members, not against insiders.)

What each party can do

Party	Holds	Can write?	Can read?
Webhook	kemPubHex + its own signing key	✅ seal to the space	❌ never
Member	kemPrivHex (via keyring)	✅ (and decrypt)	✅
Server	nothing	—	❌ ciphertext only

Revocation

The space write key rotates like the CEK: mint a new pair, publish the new public half, re-distribute the new private half to current members. A removed member who kept the old kemPrivHex can read webhook messages only up to the rotation — the same property the CEK already has.

Trust boundary

The E2EE guarantee is "from the receiver onward". An external sender already transmits cleartext to whatever URL it's given, so the webhook receiver is the encryption edge: it holds only the public key, seals there, and from that point the Starfish server stores and serves ciphertext that only members can open. For this to be real E2EE the sealing must happen at that edge — not inside the core sync server, which would otherwise observe plaintext. Metadata (timing, size, target room) remains visible, exactly as for any message.

Mixing sealed and plaintext

A collection can hold both: normal clients write plaintext (or CEK-encrypted, in a delegated collection), and the webhook writes sealed blobs. isSealedBlob lets a reader branch per element, so the two coexist in one log.