OpenClaw 2026.7.1: what shipped, and the upgrade that breaks

OpenClaw 2026.7.1 is a substantial release. It is also, for a subset of existing installs, one that will not start. Both of those things are true at once, and most coverage only mentions the first, so this covers both.

What actually shipped

Models and providers

GPT-5.6 is recognised across the catalog, capability and runtime selection paths, so the new model family is selectable rather than merely reachable. Nemotron Super's 1M context window is now used properly instead of being clamped, and explicit OpenRouter authentication headers are preserved rather than being rewritten in transit. The release also adds a bundled ClawRouter provider plugin with credential-scoped model discovery, OpenAI-compatible and native Anthropic and Gemini transports, and managed budget reporting.

The new attach command

openclaw attach launches an external harness against an existing gateway session. In practice that means you can drop into a running session with a Codex-style interactive workflow, inspect it, and resume it, rather than starting something new and losing the context. If you work across a terminal and a UI, this is the change most likely to alter how you actually use the tool day to day.

Event-driven cron

Scheduling gains an on-exit kind: instead of running an agent at a fixed time, you can wake it when a watched command exits. Session-targeted runs can also detach cleanly now. Together these make OpenClaw a plausible reaction layer for build pipelines and long-running jobs, not just a timer.

Control UI

A session-first sidebar, a compact context meter, a warm light theme, a reasoning-effort slider, a streamlined composer and a slash-command picker. The context ring now shows context-window progress, the latest run's input and output tokens, and the active model. Sessions get generated titles from their first message, with a separate lower-cost utilityModel route for that title generation so naming a session does not bill against your main model.

Messaging and native apps

Telegram can start Codex pairing with /login, steer active Codex runs and recover final replies across transient API failures. iMessage gains native poll creation, reading and voting. Signal picks up target aliases. iOS adopts the iOS 26 visual system with reworked Chat, Talk, onboarding and reconnect flows, the macOS app can install and start a local gateway on its own, and localization expands across Apple and Android surfaces.

Capability profiles

Quieter than the rest but worth noting: capability profiles lay the groundwork for per-conversation tool and access boundaries without weakening the default profile. If you run one agent across contexts with different trust levels, this is the direction that eventually matters.

The part that is not in the highlights

Node 24.15 or newer

2026.7.1 requires Node >=24.15.0 <25, or the 22.22.3 and later 22.x line, or 25.9.0 and above. On an older runtime the CLI refuses to run outright with an engine error. That is a clean, obvious failure and an easy fix, so check your Node version before you start:

node --version
openclaw --version

The migration bug, which is not clean or obvious

With a correct Node version, some installs upgrading from earlier releases hit a gateway that crashes on boot and refuses to report ready. The errors look like this:

Left plugin install index in place because shared SQLite state has
conflicting plugin install metadata for: brave, codex

Skipped Memory Core legacy memory index import ... legacy memory meta rows
conflict with canonical memory index rows

OpenClaw startup migrations did not complete cleanly; refusing to report
the gateway ready.

This is a documented cluster of known issues rather than a local misconfiguration. The relevant reports are #102749, where legacy-state migration never converges when a .migrated archive already exists, #107133, where a Memory Core embedding-cache conflict blocks the gateway permanently, and #90040, which covers the conflicting plugin install metadata with no native reconcile path.

The error message suggests a repair that does not repair it

The crash points you at the doctor's repair mode. On this particular failure it does not fix the migration, and #107220 and #107227 record exactly that. Worse, it can rewrite openclaw.json into the newer format, which then blocks rolling back, since the older binary refuses a config last written by a newer version. It has also been observed removing autostart entries.

So the practical advice is narrow and specific: back up ~/.openclaw before upgrading, and do not run an automatic repair as a first response to this particular crash. A backup is what makes the rollback a five-minute operation instead of a rebuild.

What to do

Situation	Recommendation
Fresh install, no existing state	Upgrade freely. The migration bug needs legacy state to fail on.
Existing install, want the new features	Back up the whole state directory first, then upgrade. Test that the gateway reports ready before you trust it.
Existing install, gateway already crashing	Roll back to your previous version and restore the config from your backup. Do not attempt manual state surgery on a live state directory.
Production, uptime matters	Wait for a stable release that carries the migration fixes, and upgrade on a copy first.

The honest summary

2026.7.1 is a good release with a bad upgrade path for a specific class of existing installs. If you are starting fresh, none of the second half of this article applies to you. If you are carrying state from earlier versions, the single most valuable thing you can do is take a backup you can actually restore from, because the failure mode is a gateway that will not start and the recovery is a rollback.