代理框架
OpenClaw 2026.9.3 Rehearses Upgrades in Isolation, but Simultaneously Breaks Compatibility with Node and Plugin Interfaces
The new OpenClaw release validates the core, plugins, and migrations in an ephemeral candidate state before activation, with a separate check serving as the final gate. It also retains records of failures and rollbacks. At the same time, the upgrade drops support for Node 22/25 and several SDK interfaces; one report indicates that an agent may still lose semantic session continuity even after every structural check passes.
[OpenClaw 2026.9.3](https://github.com/openclaw/openclaw/releases/tag/v2026.9.3) turns upgrades into a two-stage process: the core, plugins, and supported data migrations are first rehearsed in an isolated candidate state, and the production Gateway is switched over only after validation succeeds. If candidate validation fails, the system can use the existing inference configuration to perform bounded repairs in an ephemeral environment, after which a separate check determines whether to activate the candidate. If it cannot be repaired, the failure and rollback outcomes are retained; cases requiring manual configuration changes are handed back to the operator. For services that preserve long-term memory, schedules, and multi-agent state, this is closer to a transactional release than directly overwriting a live environment.
This is not a low-risk patch. The minimum runtime has been raised to Node 24.16.0 or Node 26.1.0 and later, with Node 26 officially recommended. Reasons include silent TEXT truncation at embedded NUL characters in some older versions of `node:sqlite`, as well as SQLite WAL safety requirements. Node 22, Node 25, and older Node 24/26 releases are no longer supported, which also ends some installation paths for older macOS systems and Linux ARMv7. The [compatibility documentation](https://docs.openclaw.ai/install/node-compatibility) recommends verifying the SQLite library actually loaded at runtime instead of relying solely on the reported Node version.
Plugin authors must also address breaking changes to execution-policy, approval, search, and directory-result callbacks. Ownership of Workshop skills has likewise shifted from the workspace to a writable collection maintained separately for each agent. On performance, the new release attempts to preserve a warm prompt cache, reduce redundant work in cold-session updates and memory searches, and reuse worker builds across sessions, but the release notes provide no quantified results.
An isolated rehearsal still cannot prove that application-layer semantics are fully correct. One [P0 issue report](https://github.com/openclaw/openclaw/issues/142580) says that after upgrading from 2026.7.1-2, Doctor, SQLite integrity checks, and Gateway health checks all passed, yet the primary agent could not continue using its existing memories. So far, this has been observed in only one environment, and the root cause remains undetermined. Production deployments should still clone representative long-running sessions for semantic regression testing and retain both a pre-upgrade snapshot and an operational environment capable of running the previous release.