Hitting the limits of frontier reasoning
After every batch of specification edits, I have Claude run a bunch of checks on the specification itself, decision record, and the supporting conformance files. Lately, these have begun to run into [reasoning_extraction] blockages. I’ve had to downgrade the subagents to Sonnet 5, which apparently doesn’t have much reasoning to extract? The same prompts that succeed with Sonnet 5 are blocked by both Opus and Fable.