Fetching an event's content by its ID gives the latest edit content rather than the original

[edwargix]

Description

Fetching an event's content by its ID with the GET /_matrix/client/r0/rooms/{roomId}/event/{eventId} endpoint yields the content of the latest edit event rather than the content consistent with the ID.

Steps to reproduce

# send an m.text event with content "test" and then
$ EVENTID='$0_lK-IgCcbad6aPp-zdCqBaQuMg85g4-vubakBknDe8'  # ID of the sent event
$ curl -s "https://$HOST/_matrix/client/r0/rooms/$ROOMID/event/$EVENTID?access_token=$TOKEN" | jq .content
{
  "msgtype": "m.text",
  "body": "test"
}
# edit the content of the event to "test edited content" (done via Element here) and fetch again:
$ curl -s "https://$HOST/_matrix/client/r0/rooms/$ROOMID/event/$EVENTID?access_token=$TOKEN" | jq .content
{
  "msgtype": "m.text",
  "body": "test edited content"
}
$ curl -s "https://$HOST/_matrix/client/r0/rooms/$ROOMID/event/$EVENTID?access_token=$TOKEN" | jq .unsigned
{
  "age": 483762,
  "m.relations": {
    "m.replace": {
      "event_id": "$xNM2Um7jlk6jEMvGzlRPc7YW6x3nVhAOOs36WJq5kfk",
      "origin_server_ts": 1625420884120,
      "sender": "@euclid:hnitbjorg.xyz"
    }
  }
}
$ NEWEVENTID='$xNM2Um7jlk6jEMvGzlRPc7YW6x3nVhAOOs36WJq5kfk'
$ curl -s "https://$HOST/_matrix/client/r0/rooms/$ROOMID/event/$NEWEVENTID?access_token=$TOKEN" | jq .content 
{
  "m.new_content": {
    "msgtype": "m.text",
    "body": "test edited content"
  },
  "m.relates_to": {
    "rel_type": "m.replace",
    "event_id": "$0_lK-IgCcbad6aPp-zdCqBaQuMg85g4-vubakBknDe8"
  },
  "msgtype": "m.text",
  "body": " * test edited content"
}

I asked whether there's a way to get an event's original content in the #matrix-spec room without doing a full /sync and many concluded [ref] that the behavior outlined above is a bug (or at least ambiguously specified).

Version information

• Homeserver: non-matrix.org server

If not matrix.org:

• Version: 1.37.1

• Install method: official Docker image: https://hub.docker.com/r/matrixdotorg/synapse

• Platform: docker-compose

[richvdh]

Useful context from the linked discussion:

• @turt2live was surprised by this behaviour and thought it should return the original

• MSC2675 had:

  FIXME: we need to spell out that this API [/relations] should return the original message when paginating over m.replace relations for a given message. Synapse currently looks to include this as an original_event field alongside chunk on all relations, which feels very redundant when we only need it for edits. Either we specialcase it for edits, or we just have the client go call /event to grab the contents of the original?

  ... which strongly implies this behaviour is incorrect.

• @tulir linked to the place it is happening: https://github.com/matrix-org/synapse/blob/v1.36.0/synapse/events/utils.py#L401-L426

See also https://github.com/matrix-org/matrix-spec-proposals/pull/2675/files/8eec329d7d5c525386e2c06cbc9a2ba4e99519d3#r451146138 which discussed this further.

I think we'd welcome PRs to fix this (presumably by passing a apply_edits parameter into serialize_event).