Rebased onto staging. The original version of this PR also fixed alias-blind direction labelling; staging's isInboundMessage (PR #2615) landed that first and did it better , it tests the SENT label as well as the address, so it also catches a self-BCC. That half is dropped. What remains is the half staging did not fix: the label was corrected, the close was not.

What's wrong

markEmailTasksCompleteByThreadId matches on task_output.threadId = <thread> AND status = 'todo'. No direction term. And it runs at line 547, ~90 lines above where direction is computed. So any message landing on the thread closes the task , the counterparty's included.

Closed response tasks over 30 days:

Respond to Aleph Patroy's billing dispute email   ← closed by "re: 20x cost inflation at Greptile"
Respond to Ben Holt , trial credits out           ← closed by "Re: Solo founder, high review volume"
Respond to Bill Adams's TPRM questions            ← closed by "Re: ElevenLabs MVP Testing - SOW"

The counterparty chased, the thread synced, and the user's obligation to reply silently left their list. Nobody replied to anything.

The comment being replaced asserted the opposite , "If this is an incoming email thread, there likely won't be matching tasks" , which is exactly wrong for response tasks: those are keyed to the inbound thread by construction, so they're the rows most likely to match.

What changed

• Direction is resolved once, above the completion, and reused for the trigger-event facts, so the gate and the announcement can never disagree about the same message.
• resolveSyncDirection / pickNewestMessage live next to the isInboundMessage predicate they wrap, so the decision that gates completion is testable.
• pickNewestMessage excludes drafts. isInboundMessage returns false for a draft , correct on its own terms, since a draft isn't a reply from someone else , but a caller reading that false as outbound turns an unsent Cedar draft into an outbound send. Latent in staging today (it only affects wording); live the moment anything gates on the value, which is what this PR does. Without the exclusion, the agent drafting a reply would close the very task the draft was written to satisfy, before the user saw it.
• Unknown direction does not close. A stale open task is visible and recoverable; a wrongly closed one is silent.
• The no-match metric is scoped to outbound, so it keeps measuring missing thread keys rather than counting every quiet inbound reply as a stamping failure. A new log line makes the gate itself observable.

Behaviour change to be aware of

If direction can't be resolved, no tasks auto-close for that sync , previously they closed regardless. That's deliberate, and the next sync recovers it, but it's a change:

['cedar-prod']
| where message == "[handleExecuteSyncThread] Task close skipped , not an outbound send"
| summarize count() by user_id, direction, bin(_time, 1d)

A collapse in outbound for a whole account means direction resolution regressed, not that the user stopped sending.

handleExecuteFromClientSend also calls markEmailTasksCompleteByThreadId and is deliberately not gated: it only runs because the user just sent something, so it's outbound by construction.

Verified

• tsc -b on @zero/server , exit 0, clean
• vitest run over services/mail, services/user-tasks/__tests__, mastra/routeHandlers , 881 passed, 3 skipped
• New sync-direction.test.ts , 10 tests: newest-message selection, alias sends via the SENT label, draft exclusion, and every null path
• pnpm deps:check , no violations (1577 modules, 8222 dependencies)

Not in scope

T