What?

This is part of a series of issues and PRs created by a coding agent looking at the output of an AI generated fuzzer. See #77716 for the tracking issue.

This issue of the 16 MiB /wp-json/wp-sync/v1/updates limit causing the connection lost modal is themeatically similar to #77669 in that this also causes a connection lost modal and it's also due to an update that's too large, but this is a distinct failure mode.

The repro shown in the video below is a bit odd in that it's stuffing updates into the title, but the same failure mode should apply in the more reasonable case of putting updates into the body of posts.

I don't know how realistic this is, but I didn't file #77631 when the fuzzer surfaced it before it was found by a human developer who was fixing a bug encountered by a real user because I didn't know if 50 rooms was realistic or not. I've gotten the "connection lost" modal a number of times myself from normal use and, in general, think it's nice to get rid of cases that could cause it so that fuzzing can find other cases that might be more realistic even if this case isn't realistic, but I can also see the argument for not adding this much code to fix a corner case issue that might not happen in practice.

BEGIN AI GENERATED TEXT

Why this is distinct

• Not PHP OOM.
• Not browser offline / request abort.
• Not the >50 rooms validation limit.
• Not the 1 MB single-update limit or oversized compaction case.
• The server rejects the request at the route validator with 413 Request body is too large.

Repro shape

• Baseline collaborative editor state contributes 4 rooms.
• Load 40 additional numeric postType/post:* entity records into sync.
• Edit all 40 extra post records with sub-1 MB title changes (450 KiB each).
• The next poll sends:
  • 44 rooms total
  • multiple queued updates
  • request body around 24.6 MB
• WordPress responds with repeated 413 Request Entity Too Large.
• After retries are exhausted, the editor shows the Connection lost modal.

Reproduction levels

The full Connection lost modal is only visible at the browser/editor layer,
but the bug can be reproduced at lower levels by separating the failure into
the client payload shape and the server validator.

Payload-shape repro

This is the lowest-level client-side repro. It constructs the same SyncPayload
shape that the polling manager sends, while staying under the 50 room cap and
under the 1 MiB per-update cap. The resulting JSON body still exceeds the
server's 16 MiB body cap.

Run from any checkout:

node <<'NODE'
const MAX_BODY_SIZE = 16 * 1024 * 1024;
const ROOM_COUNT = 44;
const BASELINE_ROOMS = [
	'postType/post:1',
	'root/comment',
	'taxonomy/category',
	'root/site',
];
const ENCODED_UPDATE_SIZE = 600 * 1024;
const rooms = Array.from( { length: ROOM_COUNT }, ( _, index ) => ( {
	after: 0,
	awareness: { user: `client-${ index + 1 }` },
	client_id: index + 1,
	room: BASELINE_ROOMS[ index ] || `postType/post:${ index + 1 }`,
	updates:
		index < BASELINE_ROOMS.length
			? []
			: [ { type: 'update', data: 'x'.repeat( ENCODED_UPDATE_SIZE ) } ],
} ) );
const body = JSON.stringify( { rooms } );
console.log(
	JSON.stringify(
		{
			rooms: rooms.length,
			updates: rooms.reduce(
				( total, room ) => total + room.updates.length,
				0
			),
			encodedUpdateBytes: ENCODED_UPDATE_SIZE,
			bodyBytes: Buffer.byteLength( body, 'utf8' ),
			maxBodyBytes: MAX_BODY_SIZE,
			exceedsLimit:
				Buffer.byteLength( body, 'utf8' ) > MAX_BODY_SIZE,
		},
		null,
		2
	)
);
NODE

Expected result: rooms is 44, each encoded update is 614400 bytes, and
bodyBytes is about 24581413, which is greater than 16777216.

Server-validator repro

This isolates the server behavior: the route-level validator rejects an
oversized raw request body with rest_sync_body_too_large and status 413.
This does not exercise the editor modal, but it proves that the server rejects
the aggregate request before the sync handler stores updates.

vendor/bin/phpunit \
	--filter test_sync_rejects_oversized_request_body \
	phpunit/tests/collaboration/wpHttpPollingSyncServer.php

This test is in
phpunit/tests/collaboration/wpHttpPollingSyncServer.php.

Browser HTTP repro

This is the focused browser repro without relying on a fuzzer campaign. It
loads the extra synced entity rooms, edits them, observes a POST /wp-json/wp-sync/v1/updates request above 16 MiB, observes repeated 413
responses, and then asserts that the Connection lost modal appears.

WP_ENV_PORT=8893 npm run wp-env-test start
WP_ENV_PORT=8893 WP_BASE_URL=http://localhost:8893 npm exec \
	--workspace @wordpress/e2e-tests-playwright -- wp-scripts test-playwright \
	test/e2e/specs/editor/collaboration/collaboration-sync-body-size-connection-lost.spec.ts \
	--project=chromium
WP_ENV_PORT=8893 npm run wp-env-test stop

Use a different WP_ENV_PORT if 8893 is already occupied.

How this was introduced

This appears to be a composition bug in the original HTTP polling sync design,
not a bug introduced by the later 1 MiB single-update accounting fix.

Relevant history:

• #72114,
  c214929139f50337250efe2bb24ff82c3ff2b6aa, made syncing a
  side-concern in the core-data resolver. This is part of the path where
  resolved entity records can be loaded into the sync manager.
• #74564,
  48ce44dac7981eb730079563a3a2975b89840fac, added the default HTTP
  polling sync provider. Its polling manager built one payload from all
  registered roomStates and drained each room's queued updates with
  state.updateQueue.get(), without an aggregate request-byte budget.
• #75699,
  c4e4fac0a26bfb2dc38f17c765f2e84266b68b99, removed the
  IS_GUTENBERG_PLUGIN guards around collaborative editing. In the current
  path, any resolved numeric entity record with entityConfig.syncConfig and
  no query is loaded into sync.
• #76987,
  1be2ef27e6819597dacc3b395caa05670d494194, backported the server
  validation hardening from WordPress/wordpress-develop#11296. It added
  MAX_BODY_SIZE = 16 * MB_IN_BYTES, MAX_ROOMS_PER_REQUEST = 50,
  MAX_UPDATE_DATA_SIZE = MB_IN_BYTES, and the route-level
  validate_request() path that returns rest_sync_body_too_large with
  status 413.

Two later nearby fixes are related but do not fix this aggregate-body bug:

• #77631,
  1642980d599be51c7cce7b2dc3a0c052b69ad367, rotates rooms when the
  registered room count exceeds MAX_ROOMS_PER_REQUEST. It addresses the
  >50 rooms failure, but a request with 50 or fewer rooms can still exceed
  the 16 MiB body cap.
• #77669,
  a54911b0c49e3b4abea4d6d7ce85c0e2c2bad11e, fixes the separate
  per-update base64 accounting mismatch. That prevents a single encoded update
  from exceeding the 1 MiB update limit, but does not limit the total body
  size of a multi-room poll.

The sync endpoint has three independent caps:

• at most 50 rooms per request
• at most 1 MiB of encoded data for one update
• at most 16 MiB for the whole request body

The client batches every registered room into one poll request and drains every
room's queued updates into that request. That makes the first two caps
insufficient to protect the third cap. For example, 40 rooms with updates far
below 1 MiB can still create a request body larger than 16 MiB once their
base64 update strings, awareness payloads, room metadata, and JSON overhead are
combined.

The number of rooms can grow beyond what a user would perceive as the one
document they are editing because resolved numeric entity records are loaded
into the sync manager. A page that touches many synced post records can
therefore create many sync rooms, each with ordinary-sized queued updates.

The failure then becomes sticky. A 413 from the route-level body validator
means the server rejected the request before storing the updates, but the client
treats it like a generic retryable poll failure. It retries the same oversized
shape after backoff, eventually surfacing the generic Connection lost modal.

Issue analysis

This is an availability and scalability issue, not evidence of server-side
partial write corruption. The server body-size validator runs before the sync
handler stores updates, so the failing request is rejected as a unit.

The risky part is the client recovery behavior after rejection. The current
generic failure path cannot tell that the payload is structurally too large to
ever succeed. It may restore the same queued updates, or replace failed outgoing
updates with compaction updates in cases where the room already has a cursor.
That behavior is appropriate for ambiguous network errors where the server
might have committed the write, but it is the wrong shape for a deterministic
413: retrying or compacting does not reduce the aggregate request size and can
keep the session in a permanent retry loop.

The fix should preserve these invariants:

• never drop local queued updates silently
• never split the bytes of a single Yjs update
• preserve FIFO ordering of updates within each room
• keep room cursors and awareness state scoped to the room that was actually
  sent
• avoid raising the server body limit as the primary fix, because that shifts
  the failure toward memory pressure and slower requests
• avoid disabling real-time collaboration for the whole editor when smaller
  poll batches can make progress

Relevant code path

• Server-side body cap:
  • class-wp-http-polling-sync-server.php
  • class-wp-http-polling-sync-server.php
• Numeric entity records are loaded into the sync manager when getEntityRecord() resolves:
  • packages/core-data/src/resolvers.js
  • packages/core-data/src/resolvers.js
• Generic poll failures become disconnected status and then the modal:
  • packages/sync/src/providers/http-polling/polling-manager.ts
  • packages/editor/src/utils/sync-error-messages.ts

Fix plan

The least risky fix is client-side request budgeting and batching, with a
specific 413 recovery path.

1. Add a client constant for the server request-body limit and use a smaller
   soft budget, for example 15 MiB, to leave room for serialization details
   and future metadata.
2. Build poll payloads in bounded batches instead of sending every registered
   room in one request. Enforce both:
   • MAX_ROOMS_PER_REQUEST
   • serialized JSON byte length under the soft body budget
3. Measure the actual serialized payload size with the same object shape passed
   to apiFetch, rather than estimating from raw update sizes. The update data
   is already base64 at this point, so JSON.stringify( payload ) measured with
   TextEncoder should be close to the request body that the REST endpoint
   receives.
4. Change update queue handling so batching can choose a sendable prefix
   without losing updates. A safe design is to add queue operations that can
   peek at pending updates and then take only the updates assigned to the
   current batch. Rooms and updates not assigned to the batch must remain queued.
5. Preserve per-room update order. If one room has many queued updates, split
   that room across multiple polls at update boundaries. Do not split a single
   Yjs update byte array. A single encoded update should already be bounded by
   the per-update limit.
6. Prioritize rooms with outgoing updates, but keep a round-robin cursor for
   rooms without outgoing updates so incoming updates and awareness do not
   starve for secondary rooms.
7. Detect 413 / rest_sync_body_too_large separately from ambiguous network
   failures. Because the route validator rejected the request before storing
   updates, restore the exact attempted updates and retry with a smaller batch
   budget. Do not replace them with compaction updates on this path.
8. Add regression coverage:
   • unit tests for payload batching by body size and room count
   • queue tests proving unsent updates remain queued and sent updates restore
     exactly on 413
   • a polling-manager test where many rooms with sub-1 MiB updates are sent
     across multiple successful polls, with every serialized payload below the
     budget
   • a 413 test proving the next retry shrinks the batch and does not trigger
     the disconnect modal
   • an end-to-end version of this repro once the lower-level behavior is
     stable

This plan avoids new protocol semantics. It does not require server-side Yjs
chunk reassembly, does not alter sync storage, and does not raise server limits.
It makes the existing polling protocol respect the server's aggregate request
limit before sending.

Five-pass confirmation

Focused isolated reruns on 8895:

• iteration 1: request 24,617,807 bytes, modal at 22.7s, repeated 413
• iteration 2: request 24,617,433 bytes, modal at 22.8s, repeated 413
• iteration 3: request 24,617,457 bytes, modal at 22.6s, repeated 413
• iteration 4: request 24,617,468 bytes, modal at 22.8s, repeated 413
• iteration 5: request 24,617,453 bytes, modal at 23.4s, repeated 413

All 5/5 runs reproduced.

Conclusion

This is a real Connection lost cause: many ordinary-sized synced edits across
many synced entities can overflow the server's 16 MiB poll-body cap, yielding
repeated 413s and the generic disconnect modal.

END AI GENERATED TEXT

Use of AI Tools

Except for the text in this PR outside of the AI generated block, everything here was AI generated including the fuzzer code that surfaced this bug.