Silent 30-second timeout halts Cursor agent workflows under load in pstack-claude

An on-call engineer found that a refactoring job run by the pstack-claude Cursor agent harness stopped silently after 30 seconds, with no error logged, while queued requests piled up in Redis. The Cursor timeout setting of 120 seconds had no effect because the programmatic HTTP client inherited httpx's 30-second default. Under concurrent load, sequential sub-agent calls totaled about 32 seconds and triggered the cutoff, and the error handling swallowed the exception. Adding an explicit 120-second timeout and surfacing errors through logging eliminated silent failures and cut the queue depth and P95 latency sharply. The author recommends setting explicit timeouts sized to the workflow's SLA rather than relying on library defaults.