Subduction Policy

SubductionPolicy: authoritative reference

This skill covers the policy (authorization) surface. For subduction sync reliability — the 60 s round-timeout amplifier, the O(N²) bulk-sync cliff, why connection replacement must close/evict immediately (drain is harmful, 200× slower), the in-flight gate + heal re-drive, the edge-replicator Effect layer-ordering bug (silent serviceOption→none), and the workerd dead-context alarm pump — see references/sync-reliability.md.

SubductionPolicy is the JS-imported policy object the Subduction Wasm bridge consults for every authorization decision. The TypeScript shape (verbatim from subduction_wasm/src/policy.rs's TS_POLICY block):

export interface Policy {
  authorizeConnect(peerId: PeerId): Promise<void>;
  authorizeFetch(peerId: PeerId, sedimentreeId: SedimentreeId): Promise<void>;
  authorizePut(requestor: PeerId, author: PeerId, sedimentreeId: SedimentreeId): Promise<void>;
  filterAuthorizedFetch(peerId: PeerId, ids: SedimentreeId[]): Promise<SedimentreeId[]>;
}

Convention: resolving allows, throwing (or returning a rejected promise) denies. All peerId arguments are subduction-level Ed25519 identities derived from the signer (MemorySigner.peerId() / WebCryptoSigner.peerId()), NOT automerge-repo's PeerId string.

This skill is the source of truth. It supersedes any older guidance that called authorizeFetch "server-side fetch-RPC only" or filterAuthorizedFetch "completely dead". Those were both wrong; see below for what they actually do.

Where each hook is consulted

Verified against inkandswitch/subduction@main (subduction_core/src/{subduction.rs,subduction/peers.rs,storage/powerbox.rs}).

authorizeConnect(peerId)

• One call site: Subduction::add_connection (subduction.rs ~line 840).
• Fires once per addConnection / acceptTransport / connectTransport call, i.e. once per handshake.
• Denial = the connection is rejected and never added. No subsequent hook fires for that peer until they reconnect.
• Cycling the underlying transport (e.g. reconnectAdapters in DXOS tests) drives a fresh handshake → fresh authorizeConnect.

Use it when you want a per-peer kill-switch ("never talk to this peer at all"). Don't use it when you want per-document gating — authorizeConnect is all-or-nothing for that peer.

authorizeFetch(peerId, sedimentreeId)

• One call site: StoragePowerbox::get_fetcher (storage/powerbox.rs), invoked from Subduction::send_requested_data (subduction.rs ~line 2603), which is the single function used for sending sedimentree data to a peer.
• send_requested_data is called from THREE places, all inside the bidirectional batch-sync protocol (subduction.rs lines 1084, 1990, 2265). Every batch-sync round between two peers exchanges fingerprints, BOTH sides discover what the other is missing, BOTH sides call send_requested_data for the requested half.
• Therefore authorizeFetch fires on the holder for every outbound data send, including:
  • explicit fetch RPCs from the peer,
  • the connect-time fullSyncWithAllPeers,
  • any syncWithAllPeers (which is what addBatch triggers internally — see below for the "new commit" path which is different).
• Denial → that peer does not receive that sedimentree on that sync round. The peer sees DataRequestRejected; their entry transitions to a failed state and will only retry via heal-retry exponential backoff or a fresh connection.

Use it when you want per-peer, per-sedimentree control over what data leaves this node. This is the only Policy hook that gates outbound serving of sedimentree contents to a connected peer.

authorizePut(requestor, author, sedimentreeId)

• One call site: StoragePowerbox::get_putter (storage/powerbox.rs), invoked from inbound sync paths in subduction.rs (lines 1914, 1953, 2201, 2240) and subduction/ingest.rs (lines 93, 137). All inbound-data ingest paths.
• requestor is the immediate sender (the peer whose connection delivered the bytes).
• author is the original signer of the commit, recovered from VerifiedAuthor (PeerId::from(*author.verifying_key())). Empirically verified in the 3-peer chain test (E1: requestor === immediate sender, author === original signer).
• Denial → that specific commit/fragment is dropped. Logged in subduction_core as WARN ... policy rejected ... policy denied.
• NOT called for the holder's OWN writes: the local add_commit / add_fragment / add_batch paths use local_putter ("the node trusts itself — no policy check is performed"). So a permissive holder cannot use authorizePut to gate its own writes; that's by design.

Use it when you want to refuse inbound data:

• by immediate sender (per-peer): match on requestor.
• by original author (per-source, even through a relay): match on author.
• by document (per-sedimentree): match on sedimentreeId.

Combinations work, since you control the predicate.

filterAuthorizedFetch(peerId, ids)

• One call site: get_authorized_subscriber_conns in subduction/peers.rs.
• Invoked when the holder broadcasts a NEW commit (add_commit calls get_authorized_subscriber_conns to decide who receives the SyncMessage::LooseCommit frame; subduction.rs ~lines 1207, 1311).
• Critical caveat: only called for peers that have explicitly subscribed to that sedimentree (i.e. issued a batch-sync request with subscribe: true). If a peer is connected but has not subscribed, the broadcast falls back to "send to all connections" with NO filter consultation. From subduction.rs:

  let subscriber_conns = self.get_authorized_subscriber_conns(id, &self_id).await;
  if subscriber_conns.is_empty() {
      // No subscribers for sedimentree, broadcasting to all connections
      self.all_connections().await
  } else {
      subscriber_conns
  }
  

• The wasm shim intersects the policy's response with the input list — a buggy policy that returns extra IDs cannot expand the authorized set.
• Empirically dead in many test setups because peers don't subscribe before the holder writes; the broadcast falls into the "no subscribers" branch.

Use it when you want to gate which subscribers receive proactive broadcasts of NEW local commits. Only meaningful if peers actually subscribe (syncWithAllPeers(id, subscribe=true) or syncWithPeer(..., subscribe=true)). Don't rely on it for batch-sync flows (fullSyncWithAllPeers, syncWithAllPeers(..., false)); those go through authorizeFetch instead.

Summary table

The "can the client gate replication purely client-side?" answer

YES. Two hooks, two directions:

• Inbound gate (client refuses incoming data): authorizePut. Denial reliably drops the bytes regardless of whether they came in via proactive push or explicit fetch.
• Outbound gate (client refuses to send its own data out): authorizeFetch. Denial reliably blocks send_requested_data from shipping the bytes during any batch-sync round.

authorizeConnect is the nuclear all-or-nothing option for refusing a peer entirely (cuts the channel in both directions).

The server side does NOT need to cooperate. A passive/permissive server is fine — the client can stop both inbound writes and outbound serving entirely on its own.

Per-peer gating requires knowing the peer's Ed25519 id

peerId / requestor / author arguments are subduction-level Ed25519 identities, not automerge-repo string PeerIds. To gate per-peer you need the peer's signer public key out-of-band, e.g. by constructing peers with explicit MemorySigner instances and reading signer.peerId().toString(). In DXOS production code where the edge DO mints its own signer at startup, the client does not currently know the DO's subduction peer-id — that mapping has to be plumbed through (e.g. captured at handshake on EdgeSubductionReplicatorConnection).

Recovery after denial → allow

Subduction has a single documented fast-recovery escape hatch: repo.shareConfigChanged(). It "resets entries whose lastSyncResult === 'all-failed' to null, clears the heal-retry backoff, and triggers an immediate re-sync" (from SubductionSource.shareConfigChanged() in @automerge/automerge-repo/dist/subduction/source.js).

Empirically verified recovery behavior:

• authorizeFetch deny → allow on the holder: kick via client.shareConfigChanged() on the fetcher recovers immediately. (The fetcher is where the failed-RPC entry lives.) See 'shareConfigChanged() retries after subductionPolicy denial flips to allow' in packages/core/echo/echo-host/src/automerge/automerge-repo-subduction.test.ts.
• authorizePut deny → allow on the receiver: shareConfigChanged() on either side does NOT recover. Cycling the transport via reconnectAdapters does NOT recover (per the SKILL doc's known fork gap, lastSyncResult === 'all-failed' is not retried on connection-generation bumps). The only reliable recovery is a fresh commit on the holder which enqueues a new outbound batch that sidesteps the stuck entry. See F1: authorizePut deny → allow needs a fresh holder commit to recover.

Production implication: when flipping a client-side authorizePut gate from deny → allow (e.g. authorizing a new space), drive a no-op commit on any doc you want to pull through.

Mutating policy behavior at runtime

The subductionPolicy reference passed to Repo is captured at construction; there is no setter. Use closure-captured state inside hook bodies. The wasm bridge invokes the JS hook on every authorization decision, so changes to closure state are observed immediately for new requests.

What changed from prior guidance

If you've read older internal docs claiming any of the following, they were wrong:

Test suite

The empirical findings above are pinned by packages/core/echo/echo-host/src/automerge/automerge-repo-subduction.test.ts under describe('subductionPolicy: characterizing client-only gates') (Blocks A–H). Re-run the suite if you suspect bridge behavior has changed:

cd packages/core/echo/echo-host && pnpm vitest run src/automerge/automerge-repo-subduction.test.ts