Vibe Coding
Pattern: Crash-Resumability Is Becoming the Table-Stakes Feature of Long-Horizon Coding Agents
Meta's Muse Code ships a local event log of model calls, tool use, approvals, and edits explicitly so a task survives a crash; Claude Code's recent releases repeatedly patch resume paths (`remote-control --continue`, `--resume` crash fixes, rewound-conversation restoration); and celld's whole pitch is durable-before-acknowledged persistent actors. Three independent stacks converging on the same primitive says long-horizon agent runs now fail often enough that restart-from-scratch is unacceptable. If you orchestrate multi-hour agent work, an append-only event log you can replay from is no longer an optimization — it's the difference between a recoverable run and a lost day.
↳ Follow the thread