Summary

• Cedar's CRM push is diff-based against its own locally stored field values, with nothing verifying those values still match what's actually committed on the CRM side. Confirmed incident: a Forecast Category value that had never actually been pushed to Salesforce was treated as "unchanged" during a stage advance, so it was silently excluded from the outbound push, and Salesforce's own validation then rejected the stage move because the live field was blank.
• Adds resyncLinkedDealFromCrm, a timeout-bounded (9s), never-throwing wrapper around the existing pullRefreshCrmDealsForConversations, wired in as a pre-step right before the crm-updater agent starts reasoning about a linked deal , skipped when the subagent is disabled or the event is itself a CRM sync, and falling back to today's behavior (no resync) if the call fails or hangs.
• Scoped with a new fieldsOnly flag so this pre-execution gate only refreshes field values, never the owner-reassignment / active-deal-marking side effects pullRefreshCrmDealsForConversations also performs for its other callers.
• Freshly resynced fields are also merged into the LLM's own <current_fields> view (all 8 core metadata fields, not just a subset), so the agent reasons from live data on the same turn, not just the later push-diff. That re-read phase has its own, smaller timeout ceiling (3s) so it can't hang the run either.

Full incident writeup and design: docs/design/crm-updater-pre-execution-resync.md.

Test plan

• pnpm --filter @zero/server run types , clean, no errors
• pnpm deps:check , no dependency violations
• Scoped vitest across every touched/related test file , 39 tests passing:
  • src/services/crm/__tests__/crm-pull-refresh.resync.test.ts , resyncLinkedDealFromCrm's hard no-throw/no-hang contract (real DB failure + genuine hang, not mocked away), and the fieldsOnly guarantee (proves zero owner-directory fetch / reassignment calls under fieldsOnly: true, plus a regression check that omitting it preserves the original behavior for the two pre-existing callers)
  • src/services/crm/__tests__/crm-resync-field-overrides.test.ts , direct coverage of fetchResyncedFieldValues's metadata/custom-field split and null-handling, plus a genuine-hang test proving the new 3s re-read timeout actually bounds the phase (not just typechecks)
  • src/mastra/tools/event-execution/__tests__/crm-updater-enabled-flag.test.ts , end-to-end proof through the real runCrmUpdaterTool.execute call site that a resync failure never affects the run, and that the resync correctly skips when the subagent is disabled or the event is external_crm
  • src/mastra/utils/context-formatting/__tests__/extract-conversation-fields-resync-overrides.test.ts , unmocked extractConversationFields, proving all 8 core metadata fields (not just custom fields) correctly reflect a resync override
  • src/services/crm/__tests__/conversation-field-comparison.test.ts , regression check on the shared type guard used to eliminate a cast in both this file and the new one
• Pre-existing, unrelated failure noted and confirmed out of scope: apps/mail/tests/modules/chat-store/chatHistoryList.test.ts (chat-history date bucketing) , zero files under apps/mail are touched by this diff.

Verify in prod

Two structured log families to watch in Axiom (cedar-prod dataset) or CloudWatch (/aws/ecs/aws-prod-api/*-service):

• crm_pre_execution_resync_complete (info) / crm_pre_execution_resync_failed (warn) , the live-CRM fetch + field-write phase, fields: user_id, conversation_id, org_id, updated_fields (complete) or error (failed, including "timed out after 9000ms" on the 9s ceiling).
• crm_resync_refetch_timed_out / crm_resync_refetch_failed (warn) , the post-resync re-read phase that feeds the LLM's prompt, same user_id/conversation_id/org_id fields plus error.

['cedar-prod']
| where _time > ago(1h) and message in ("crm_pre_execution_r