Summary

This issue tracks the follow-up activation work deferred from #1496.

PR #1496 extracted safeSendMessage, SendMessageContext, and UnknownErrorTracker into executor-shared.ts (deduplication only). As of that PR, UnknownErrorTracker remains dormant/never instantiated on both sides — executor.ts and dag-executor.ts — meaning the consecutive-UNKNOWN abort guard is available as infrastructure but not active in production.

Work required

• Wire UnknownErrorTracker into the safeSendMessage call sites in packages/workflows/src/executor.ts
• Wire UnknownErrorTracker into the safeSendMessage call sites in packages/workflows/src/dag-executor.ts

Open design questions to resolve before implementing

1. Tracker lifetime: should the tracker be scoped per-conversation or per-DAG-run?
2. Ownership location: where should the tracker be instantiated and held — at the top-level executor, per-node, or passed in via WorkflowDeps?
3. Behavior on abort: what should happen when the threshold is hit in DAG execution — fail the current node, fail the entire DAG run, or surface to the user for a retry decision?

References

• PR: refactor(workflows): consolidate duplicated safeSendMessage into executor-shared #1496
• Comment thread: refactor(workflows): consolidate duplicated safeSendMessage into executor-shared #1496 (comment)
• Requested by: @ericsoriano