client: unconfirmed sends are visible and recovered; watchdog for a dead connection

void commented

2026-11-12 21:35:47 +00:00

Member

Why messages looked sent and vanished (no PUT /send reached the server, rooms unencrypted): matrix-js-sdk adds a local echo on sendMessage and the timeline drew it like a delivered message at every status; sends have no timeout (sync polls do), so on a dead/half-open connection a send hangs SENDING with no request out; a reload drops the echo.

timeline hides unconfirmed local echoes; the outbox row (sending / still sending 30s / interrupted by reload / failed, Resend, Discard) is the one place they show; uploads and stickers go through it too; resend takes a fresh txn id while the SDK holds the failed echo
watchdog: a send from this page pending >=20s while visible and online triggers one guarded reload (once/60s); after reload, interrupted entries are resent under their own txn id
resumeSync only nudges; it no longer calls stopClient/startClient (closes the rust crypto for good: 'null pointer passed to rust' on later encrypted sends)
scroll to bottom when a hidden echo is confirmed

Soul passes: three; findings applied. Not covered: uploads/edits in flight are lost on the watchdog reload; Resend of a slow send can duplicate; two tabs can resend the same entry; plaintext drafts in localStorage. The 20s threshold is a trade-off for the operator.

🤖 Generated with Claude Code

Why messages looked sent and vanished (no PUT /send reached the server, rooms unencrypted): matrix-js-sdk adds a local echo on sendMessage and the timeline drew it like a delivered message at every status; sends have no timeout (sync polls do), so on a dead/half-open connection a send hangs SENDING with no request out; a reload drops the echo. - timeline hides unconfirmed local echoes; the outbox row (sending / still sending 30s / interrupted by reload / failed, Resend, Discard) is the one place they show; uploads and stickers go through it too; resend takes a fresh txn id while the SDK holds the failed echo - watchdog: a send from this page pending >=20s while visible and online triggers one guarded reload (once/60s); after reload, interrupted entries are resent under their own txn id - resumeSync only nudges; it no longer calls stopClient/startClient (closes the rust crypto for good: 'null pointer passed to rust' on later encrypted sends) - scroll to bottom when a hidden echo is confirmed Soul passes: three; findings applied. Not covered: uploads/edits in flight are lost on the watchdog reload; Resend of a slow send can duplicate; two tabs can resend the same entry; plaintext drafts in localStorage. The 20s threshold is a trade-off for the operator. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

void added 4 commits

2026-11-12 21:35:47 +00:00

timeline: hide unconfirmed local echoes; the outbox is the one place they show a92f6c4294

The SDK puts a local echo in the live timeline on sendMessage and marks it
NOT_SENT on failure; the timeline rendered it like any delivered message and a
reload dropped it, so a send killed with the page looked sent and vanished.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

client: watchdog, resend on reload, outbox for every send, no in-place client restart 77edebff9c

Sends have no timeout and sync polls do, so a dead connection leaves a send
SENDING forever with no request reaching the server while its echo looked
delivered. A watchdog (send pending past 20 s, sync ERROR/STOPPED for 30 s)
calls a guarded reload; what the reload interrupts is sent again under its
own txn id. Uploads and stickers go through the outbox. Resend takes a fresh
txn id while the SDK still holds the failed echo. The resume-sync recovery no
longer calls stopClient/startClient, which closes the rust crypto for good
("null pointer passed to rust" on later encrypted sends).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

watchdog: send stall only, own-page entries, online only; resume sync nudges and never restarts 69d752ff0b

Soul pass on 77edebff: the watchdog no longer counts interrupted entries from
a previous page (reload loop at startup), no longer reloads on sync ERROR/STOPPED
(js-sdk recovers offline by itself) or while offline, refuses instead of
throwing when sessionStorage is unusable, startClient is idempotent, and the
slow-send label says Resend may duplicate. resumeSync drops its health check
and recover hook: nudge only.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

initMatrix: resolve sessionStorage inside the guard's try; ownPendingAges tested

Build pull request / Build pull request (pull_request) Successful in 3m39s

Details

bb6391a70b

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

void merged commit b6b1d522ab into dev

2026-11-12 21:35:48 +00:00

void referenced this pull request from a commit

2026-11-12 21:35:48 +00:00

Merge pull request 'client: unconfirmed sends are visible and recovered; watchdog for a dead connection' (#123) from void/resume-reload into dev

void referenced this pull request

2026-11-13 01:02:51 +00:00

Consumer-side tests for #123: RoomOutbox as rendered; hung send to watchdog reload to resend #124

void referenced this pull request from a commit

2026-11-13 04:11:35 +00:00

Merge pull request 'Consumer-side tests for #123: RoomOutbox as rendered; hung send to watchdog reload to resend' (#124) from void/stall-consumer-tests into dev

Rows
Columns

client: unconfirmed sends are visible and recovered; watchdog for a dead connection #123