DreamLake

Memo ownership transfer gate

2026-11-14 — Preserving singleton transfer source proposal (unvalidated)

The new note-metadata-v3 proposal transfers a personal-owned Memo using a versioned preservation policy. It is separate from the existing v1/v2 executors and requires OWNERSHIP_NOTE_PRESERVING_TRANSFER_ENABLED=true for new admission. The default is off. Source preparation does not establish runtime, deployment, or migration readiness. Existing stored v1/v2 requests keep their original policy hashes and replay behavior; organization-source transfers retain their existing v2 path rather than being admitted by this new profile.

The proposed request retains the existing owner/version/previewHash, idempotency key and destination acceptance contract, and adds:

json
{
  "policy": {
    "visibility": "private",
    "retainedGrantIds": [],
    "preserveAccess": true,
    "preserveExternalFilings": true,
    "memoPreservation": {
      "apiVersion": 1,
      "preserveVisibility": true
    },
    "externalMounts": {
      "apiVersion": 2,
      "dispositions": []
    }
  }
}

An empty dispositions array is discovery input, not permission to discard mounts. Every retained original filing needs retain with explicit Project audience and READ/WRITE ceiling. Every outgoing grant needs reissue, or convert-local when its target Project is already in the destination namespace. The preserving profile rejects revoke/detach dispositions. Acceptance retains the destination OWNER check for issuing or converting these grants.

convert-local supersedes the old grant and binds the same existing nodes to a receipt-bound capped local grant in the ownership transaction. It keeps the Project, folder path, node ID, audience, role ceiling and expiry. No delete and reattach, unguarded marker clearing, or ordinary local-mount permission fallback occurs. A converted grant cannot admit additional nodes through the external acceptance endpoint. Destination ownership rights change as an intentional part of the transfer; the mount ceiling limits Project-derived access, not the new owner's independent rights.

Service-account Project grant rows remain attached to their original accounts and Projects. Preview inventories exact principal/grant state, and acceptance fences their account and Project parents. Accounts are not reactivated or moved; human ACLs, agent turn capabilities and explicit delegation remain independent requirements. Inventory preservation alone is not proof that every tool access path works: route-level and concurrent-revocation tests remain release gates. Memo tools retain their caller-api authorization: turn scopes plus the caller's current Memo ACL. Removing a service-account Project grant does not revoke independent caller Memo rights. Project discovery still requires the explicit service-account grant; content search additionally reaches the caller-authorized Memo route. Assistant delegation tickets remain read-only. This transfer profile does not add delegated writes or change those tool authorization modes.

Public Memos use visibility: "republish" together with the new preserveVisibility: true contract. This means preserve the existing public visibility and publication metadata, without publishing a new body or toggling visibility. Private Memos must use visibility: "private"; a mismatched request fails. Existing permalink reservations remain, the exact aliases are reserved in the destination, and old URLs resolve only through verified completed ownership history to a currently link-readable Memo (public or later unlisted, under the existing read helper; private remains denied). An occupied public alias blocks transfer; it is never overwritten or silently renamed. Unlisted Memos remain outside this new profile.

An occupied internal Memo slug requires an explicit memoPreservation.destinationSlug in the reviewed request. It changes neither Memo ID nor title. The database's unique namespace/slug index arbitrates races; new admission verifies the existing indexes without creating them. The two actual collision resolutions must be supplied deliberately before live work.

Completion requires a completed ownerCommitted receipt and a readback of actual Memo ownership, followed by content, access and exact folder-node reconciliation. External filing alone is not transfer completion. Source tests cover owner change, grant preservation, caps, public aliases, explicit slug resolution, stale inventories, rollback and same-key replay. Added registry/route cases cover caller read/write and ACL revocation, READ caps, turn scopes, invalid proof, natural expiry, explicit mount revocation, deleted resources, new-node refusal, Project-grant revocation and foreign-scope denial, and the read-only assistant ticket boundary. RTC/storage I/O is simulated; authentication, registry, routes, ownership proof, database ACL and activation fences use real code. The offered dreamlake_notes_read / dreamlake_notes_write cases use note, baseRevision, and text, with the actual complete read's revisionRef. They retain canonical baseline checks and Memo command receipts, cover a successful converted write followed by caller revocation, and deny writes under a preserved READ cap without another body commit. Legacy registered hidden-tool cases remain separate. Registry receipt storage, turn fences, and the assistant current-delegation callback remain simulated; those are not durable turn or assistant-grant-store proofs. These cases have not run for this proposal. Full reader/writer compatibility, bounded runtime validation, normal checked release and disposable live acceptance remain outstanding.

2026-10-05 — Durable revocation and projection fencing (unreleased)

The sibling RTC implementation now tombstones old graph rooms independently of document lifetime. Its production entry point serializes socket and scoped HTTP operations through a durable per-room claim. An authenticated revoke call retains a coherent native checkpoint receipt in the same transaction as the tombstone. Local tests on two real service instances and a disposable Mongo replica set prove old-socket/URL denial, restart and document-deletion denial, native checkpoint reconstruction, alternate API denial, and failure/receipt recovery. This supersedes the negative October 4 transport observation below for the new implementation; neither implementation is claimed deployed.

DreamLake now stages immutable body projections and publishes their pointer, checksum and search text only in a transaction that touches A's ownership lease and conditionally matches owner, storage root, RTC binding and old checksum. Old-room/epoch/deletion callbacks cannot overwrite retained bodies or publish stale metadata. Idle-sync bookkeeping is fenced as well. Activation, offline reads, versions and presigned body reads resolve the immutable pointer; legacy memos retain the existing archive fallback. Storage failures retain the prior pointer and retry safely.

The browser pauses on the RTC close signal, stores the visible draft by stable memo ID, and carries draft text across room rotation/remounting without carrying or replaying old operations. Its existing recovery/export UI remains available. A private staging primitive copies and checksum-verifies the complete root inventory (body, revisions, native baselines, versions, timeline) and retained files, leaves all originals intact, detects inventory changes, and checks the preparing operation before external effects. It is not a mounted adapter.

The ready Package A head e32e0e4ac69422baba00d6c7ce2bed712c484a03 is integrated, including current main. RTC PR #113 is at 0caa3cd6dd38265cbe19064f6029c08f5556e96d; browser draft preservation PR #715 is at 38a32a2e0f6ec21b26714e94880b4b4a0f9f3c76, stacked on ready A UI #702. The merge-preparation evidence records the latest 196 passing server tests, 3739 passing UI tests, 30 passing production-preview E2E tests and the passing 11-assertion RTC proof. Full CI is requested through normal PR events; its new result is unobserved. The memo runtime adapter remains unregistered and transfer remains disabled. Remaining blockers are complete mutation/upload/worker fencing, owned media and issued-capability policy, staging/abort/roll-forward orchestration, all filing and grant reconciliation, retained delete/restore authority, client parity, fleet/index migration and integrated release/recovery validation. Durable room claims require explicit quiescence and exact-token recovery after uncertainty; they never reopen on elapsed time. The graph auto-compaction loop is paused until it can share the durable claim; authorized HTTP compaction is supported.

The dated October 4 sections describe the original base and must not be read as the current implementation status. Full package D remains incomplete.

2026-10-04 — Unreleased prerequisites

Direct memo namespace transfer remains unavailable. Note.namespaceId remains the sole ownership field; BindrMember remains the filing source of truth for zero, one, or multiple projects. Filing is independent of ownership. This change introduces no transfer, restore, or ownership mutation and does not complete package D.

A valid original RTC baseline bound to a different current room now returns HTTP 409, error: session_changed, code: preserve_draft_and_resync. Both exact and merge writes stop before committing operations, writing an archive, or recording edit activity. A conditional read with the old baseline returns the same conflict. Preserve the unsent draft and original baseline, read the current source, and deliberately reconcile before submitting a new patch. Do not automatically replay the old patch with a fresh revision. Malformed baselines remain validation errors. Authorized historical reads retain checksum validation; the conflict never discloses the old room URL.

This is server reconciliation signaling. Browser draft preservation and pause UI remain unverified. A room identity check is not an ownership epoch fence.

RTC result: BLOCKED

The local audit runs the actual RTC transport and broker over real loopback WebSockets with disposable content. It invokes the production room reset's transport/broker calls, without a journal or database. A successful reset allows the old socket to write, receive later broadcasts, and reopen the same room URL. This proves those reset calls do not revoke raw socket authority. It does not certify database cleanup or a deployment.

From the server directory, against a source checkout with the RTC library built:

bash
pnpm exec tsx scripts/check-notes-rtc-transfer-gate.mts /path/to/vuer-rtc-workspace

Exit 2 means the audit completed with the transfer gate blocked. Exit 1 means the probe could not complete. This is a negative regression demonstration, not positive release certification. It starts only disposable loopback sockets; it accepts no live service URL and uses no real memo content.

Required before transfer integration

  • Package A's reviewed adapter interface, engine, epochs, authority fences and versioned fixtures must be available in the integration base. The observed sibling contract is draft-only and has not been imported.
  • RTC must prove epoch authorization or durable room tombstones, disconnecting old sockets and rejecting recreation after restart and across every instance. Room deletion's HTTP 200 is insufficient.
  • An acknowledged checkpoint must fence later projection workers. The current mutable S3 projection and metadata update are not fenced by an ownership epoch. Rechecking a row before S3 upload alone leaves a race.
  • Every body, metadata, share, version, comment, file, media and presence path must use the same durable epoch. Upload completion and callbacks must reject stale authority.
  • Inventory and fence download capabilities. Default presign durations are body 60 seconds, memo files 120 seconds, media 3600 seconds. Copying objects does not revoke existing URLs. Media IDs currently act as standalone capabilities and need an ownership-aware migration and revocation policy.
  • Verify recovery, checksums and references before owner CAS; reconcile every filing and grant, preserve IDs/body/history/comments/files/media, and roll recovery forward after CAS.
  • Implement safe delete/restore retention and API/UI/CLI/SDK parity, then test every transfer direction, zero/one/multiple affiliations, creator departure, old shares/sockets/downloads, storage failures, and crash/retry.

No transfer, lifecycle migration, deployment, or live acceptance is asserted. Public task skills are unchanged: this unreleased server diagnostic adds no client command or task routing; existing edit guidance already requires preserving the original baseline and proposed patch after failure.