Vercel says durable-harness works with Pi, OpenCode, Codex, and Claude Code without changing the harness itself. It records each step of a turn in a durable log, recording intent before model or tool calls and the outcome after completion. A queue backs the turn; when a function dies, a replacement reads the log and resumes from completed work. Mechanism: A durable log records call intent and result, while a queue restarts interrupted turns; tool replay labels distinguish safe retries from potentially completed side effects that require checking. The log distinguishes completed, in-progress, and not-yet-started calls; tool-level replay labels govern whether interrupted work can be retried or must be checked rather than repeated.
Why now: Vercel describes cloud coding-agent sessions lasting minutes or hours and argues that cloud compute alone does not preserve call completion state or resume interrupted turns. This release packages recovery for existing harnesses, and Vercel says the same interface can target its hosted backend, local files, or self-hosted storage such as Postgres. First implication: Teams can evaluate durability as a layer around a chosen harness, instead of treating a harness migration as the prerequisite for crash recovery. Teams can assess crash recovery as an adapter around an existing harness rather than adopting a new harness.
Evidence boundary: Claims and architecture are described in Vercel's announcement; no independent reliability evaluation or general exactly-once guarantee is provided. These are capabilities and deployment options described by Vercel; the announcement is not an independent reliability evaluation, and it does not establish exactly-once behavior for arbitrary external side effects. Tool policies and integration determine recovery behavior. First watch condition: Whether deployments show logs and replay policies avoiding duplicated or lost work after crashes, especially for external side effects. Whether real deployments report that replay policies and durable logs prevent duplicated or lost work across crashes, particularly for tools with external side effects.