@filigran/chatbot 3.15.1 → 3.15.2

package/CHANGELOG.md

@@ -8,7 +8,28 @@ the versions follow [semantic versioning](https://semver.org/spec/v2.0.0.html).
 This file starts at 3.10.0; earlier releases are on the
 [releases page](https://github.com/FiligranHQ/filigran-ui/releases).
 
-## 3.15.1 — unreleased
+## 3.15.2 — unreleased
+
+### Fixed
+
+- The thread keeps the end of the conversation in view when what an answer
+  carries takes its height after the answer rendered (an image, a chart, a file
+  card): it used to stop short of it (#460).
+- A message sent while the agent writes its answer shows above that answer,
+  where the backend stores it, instead of below it with the rest of the answer
+  streaming in above; the same for a message of another tab relayed during a
+  relayed answer (#460).
+- A turn whose stream drops is awaited for as long as the event stream takes
+  to reopen (a backend restart), showing "Reconnecting…", instead of being
+  marked stopped 15 s after the drop; an `error` or `cancelled` of the event
+  stream naming it ends it at once, and one that ended out of sight is
+  answered or stopped as the history read back says (#460).
+- A stream its server keeps alive (comments at a keep-alive cadence) that then
+  falls silent for 45 s is taken for dropped, so a stream a proxy left
+  half-open is handed over instead of hanging, when the event stream can pick
+  its turn up (#460).
+
+## 3.15.1 — 2026-10-27
 
 ### Fixed
 

package/README.md

@@ -1058,8 +1058,18 @@ never relay a panel's own turn back to it; frames naming the `turn_id` of the
 panel's own turn are dropped too. When the panel's own stream drops mid-turn,
 or ends without the turn's last frame (`done`, `error`, `session_expired` or
 `cancelled`), it takes a new id and picks the turn up from the event stream,
-keeping the segments it closed (or reads the history back when nothing comes
-within 15 s). A read-back that waited for a turn to end joins the thread in
+keeping the segments it closed. It shows "Reconnecting…" for as long as the
+event stream takes to reopen (a backend restarting, up to 15 minutes), then
+waits 5 s for the snapshot to replay the turn; an `error` or `cancelled` of
+the event stream naming that turn ends it as a relayed turn's end would, and
+with nothing in time the turn ended out of sight, answered or stopped as the
+history read back says. When the event stream cannot open (no conversation
+to follow yet) or answers an error other than a gateway's (502, 503, 504),
+the turn is not awaited that long. A stream whose `: keepalive` comments came
+at a keep-alive cadence and that then sends nothing for 45 s (a proxy that left
+it half-open) is taken for dropped the same way, when the event stream can
+pick its turn up; a backend without one keeps a silent stream open, as before. A message sent while a turn writes its answer shows above that
+answer, where the backend stores it. A read-back that waited for a turn to end joins the thread in
 place: each stored message missing from the screen goes after the one stored
 before it, so nothing lands above messages it follows.