Willison’s web-edit: a measured hotspot and an agent-to-agent repair handoff

October 12–13, 2026 follow-up to the credential-editor release. In his profiling report, Willison had Codex and GPT-6 Astra produce a Python 3.15 sampling-profile demo for the unmodified app; all-thread collection mattered because worker threads represented 95.1% of recorded stacks. Preserving mixed line endings appeared in 64.8% of 15,296 sampled stacks during a busy edit workload. His 0.1a1 release reports feeding that profile to Claude Opus 5.5 for performance changes. This is an unusually concrete handoff of measurement artifact from one model session to a second, rather than an unprompted rewrite. Neither short release gives before/after user-visible latency, profiler workload representativeness, reviewer touch time or real-credential access results; do not conflate CPU-stack proportions with latency savings or improved security. It is a subsequent engineering iteration on a personal tool, not established maintained value. The older 0.1a0 code audit still refers specifically to the October 12 snapshot.