fix(workspace-server): persist poll-mode canvas user message synchronously before queued 200 (internal#471, sibling of #1347) #1350

Merged

devops-engineer merged 2 commits from fix/canvas-user-message-poll-mode-sync-persist into main 2026-05-16 15:49:31 +00:00

Conversation 16 Commits 2 Files Changed 2 +198 -14

2 Commits

Author	SHA1	Message	Date
Molecule AI Core-BE	1d29e9ea24	fix(handlers): prevent poll-mode sync-persist test from hanging CI ... sqlmock.ExpectationsWereMet() hangs indefinitely when the expected INSERT mock never fires. If the production code ever regresses to goAsync (pre-fix shape), the handler returns before the INSERT fires, the mock never fires, and ExpectationsWereMet() blocks for the full test/-suite timeout — wedging the CI run with no diagnostic. Fix: check expectations in a goroutine with a 2s hard timeout. When the mock has fired (synchronous production code), ExpectationsWereMet() returns <1ms and the select fires the `case err := <-expectDone` arm. When the mock has NOT fired (async regression), the 2s timeout fires and the test fails with a clear message instead of hanging. Also reduce insertDelay from 150ms → 50ms. 50ms is ~50× the normal INSERT latency and sufficient to prove synchronous blocking; the larger value was adding unnecessary suite-level wall-clock under -race detection, where mock delays are amplified by the instrumenter's goroutine overhead. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-16 14:47:07 +00:00
core-be	a92beb5d49	fix(workspace-server): persist poll-mode canvas user message synchronously before queued 200 ... Sibling of #1347/internal#470 — the POLL-mode arm of the canvas user-message data-loss bug Hongming reported ("i sometimes lose my own message when i exit chat", 2026-05-16). Hongming's tenant is entirely poll-mode (4 external workspaces, no URL — verified empirically: every workspace returns the {delivery_mode:poll, status:queued} short-circuit envelope), so #1347 (push-mode only, persists AFTER the poll short-circuit) structurally cannot cover his reported case. #1347's "poll-mode was never affected" framing is overstated: logA2AReceiveQueued's durable activity_logs INSERT ran inside h.goAsync(...) — a detached goroutine with no happens-before barrier against the synthetic {status:queued} 200. The canvas sees the send acknowledged while the row may still be racing; a workspace-server restart / deploy / OOM / EC2 hibernation between the 200 and the goroutine's commit loses the message permanently (chat-history reads activity_logs; missing row = message gone on reopen). No fallback either, unlike push-mode's legacy-INSERT path. Fix: make the poll-mode ingest persist SYNCHRONOUS — committed before the queued 200 — on a context.WithoutCancel context (parity with persistUserMessageAtIngest). Best-effort preserved (LogActivity logs+swallows INSERT errors, never blocks the send). Post-commit broadcast still fires inside LogActivity (a missed WS event is not data loss; the durable row is the truth chat-history re-reads on reopen). TDD: a2a_poll_ingest_persist_test.go — deterministic RED (queued 200 returned in ~0.5ms, before the 150ms INSERT → DATA LOSS) → GREEN after fix. Full internal/handlers + internal/messagestore suites green; vet clean. Refs: molecule-ai/internal#471 (tracking), molecule-ai/internal#470 (push-mode sibling, PR #1347) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-16 06:04:14 -07:00