Some bugs announce themselves loudly. Others wait for the exact wrong moment—like a graceful shutdown during a critical briefing—to surface.
What Shipped
We merged a fix to Sabine Super Agent that guards task cancellation checks before querying task exceptions in async done_callbacks. The technical change is small—a conditional check—but the implications matter for any system managing concurrent async operations.
The problem: Python's asyncio Task objects maintain internal state. When a task is cancelled, calling t.exception() before checking t.cancelled() raises an InvalidStateError. During normal operation, this rarely surfaces. During shutdown, cleanup, or cascading task cancellations? It becomes a reliability issue.
Sabine's task orchestration layer uses done_callbacks extensively—tasks signal completion, exceptions, or cancellation to coordinate multi-step workflows like briefing preparation, email analysis, and shopping research. A callback that crashes on cancellation introduces noise into logs and, in rare cases, prevents graceful cleanup of dependent resources.
Why It Matters
Sabine runs unattended. She manages logistics, prepares briefings, monitors inboxes, and executes shopping tasks while I'm heads-down building or asleep. The bar for reliability isn't "works most of the time." It's "doesn't create work for me when things go wrong."
This fix is defensive programming in the truest sense: it prevents exceptions that should never happen from leaking into production. The edge case is real—task cancellation happens during timeouts, user interruptions, and system shutdowns. Not handling it correctly means logs fill with noise, and distinguishing real errors from benign lifecycle events becomes harder.
For autonomous systems, observability is everything. If I can't trust the logs, I can't trust the system. Fixing this now means future debugging sessions focus on real problems, not phantom errors from task cleanup.
What's Next
This fix is part of a broader reliability sprint across Sabine's core task orchestration. We're auditing all callback sites, timeout handlers, and cancellation paths to ensure they handle edge cases cleanly. The goal: make Sabine's async layer bulletproof enough that I never think about it.
We're also extracting patterns from Sabine's task management into reusable utilities for Strug Works. The lessons learned from running a personal AI assistant at production scale—handling interruptions, timeouts, and graceful degradation—apply directly to autonomous engineering workflows. What keeps Sabine reliable will keep Strug Works reliable.
The bar is simple: if it runs unattended, it must handle its own edge cases. This fix is one more step toward that standard.