v0.72.0

0.72.0 - 2026-10-06

Highlights

• Start cached Cloudflare containers in about 1–2 seconds, then checkpoint and fork them. Native container execution speeds up startup, while checkpoints capture reusable workspaces and forks start new leases from them. PR 2625, PR 2713. Thanks @altaywtf.
• Use class-sized AWS disks without a 400 GB image minimum. The promoted eu-west-1 Linux developer image now has a 32 GB root snapshot, so tiny leases request 40 GB; a live lease reported a 38 GB root filesystem. PR 2668, PR 2704, PR 2714.
• Get clearer AWS acquisition and Windows readiness results. Reclaimed instances fail promptly, recorded Windows leases use Windows readiness, and --os windows-server:2022|2025 selects the server version explicitly. PR 2721, PR 2722, PR 2720, Issue 2716, Issue 2715, Issue 2717. Thanks @saftall.
• Keep run output through transient coordinator append failures. Atomic, idempotent event appends and bounded retries preserve event order and recover lost acknowledgements without duplicating output. PR 2723, Issue 1561.
• Find the right on-demand computer and access workflow more easily. The README, crabbox.sh home page, provider catalog, and desktop/VNC/browser guides now explain the current agent workflow and supported providers together. PR 2697, PR 2699, PR 2698. Thanks @shakkernerd.

Upgrade notes

• Stop kept Cloudflare leases before redeploying the runner: the one-time cutover deletes the old classes' Durable Object state. Then run crabbox cleanup to retire old claims and delete the six legacy container applications as described in the Cloudflare upgrade guide. --type basic now maps to the higher-cost standard-1 with a warning; --type lite is rejected because the bundled image cannot start on it. PR 2625. Thanks @altaywtf.
• Explicit target or Windows-mode flags that contradict a recorded lease now fail before guest access. Remove the conflicting flags or select the lease's recorded platform; release-only resolution remains available for cleanup. PR 2722, Issue 2715. Thanks @saftall.
• Repository contributors must install the shared autoreview skill from openclaw/agent-skills; the repository no longer carries its own executable helper. Existing copy-mode consumers should reinstall to receive updates. PR 2663. Thanks @vincentkoc.

Added

• AWS Windows leases accept --os windows-server:2022 or windows-server:2025 (also os / CRABBOX_OS) in direct and brokered mode; the default remains Server 2022 and promoted Windows AMIs still win without a Windows selector. PR 2720. Issue 2717. Thanks @saftall.
• Checkpoint and fork Cloudflare containers. crabbox checkpoint create captures a running Cloudflare lease as a cloudflare-container-snapshot in about 8 seconds, and checkpoint fork starts new leases from it in seconds, retrying while a fresh snapshot propagates. Delegated-run providers can now implement native checkpoints through DelegatedCheckpointBackend. Cloudflare has no snapshot delete API; snapshots expire after 30 days and checkpoint delete --local-only removes the record. Cloudflare lease commands keep using the runner URL and workdir the lease was created with, and a create or fork whose runner response is lost keeps its claim so status and stop can still reach the container. Stopping a lease while it starts keeps it stopped; unresolved creates retain cleanup custody regardless of elapsed time or a temporary stopped status during snapshot retries, and retries reuse containers already started by concurrent requests. PR 2713, PR 2626. Thanks @altaywtf.
• Add optional vCPU and memory minimums for new AWS Linux leases while preserving existing class fallback behavior. PR 2710. Issue 2709. Thanks @shakkernerd.

Changes

• Use one shared autoreview installation from openclaw/agent-skills; repository entrypoints receive upstream fixes without copied helpers or tests. PR 2663. Thanks @vincentkoc.
• AWS Linux leases now get their class root-disk size (40 GB for tiny instead of 400 GB): the promoted eu-west-1 developer image was rebaked from stock Ubuntu with a 32 GB root. PR 2668, PR 2704, PR 2714.
• Cached Cloudflare containers start in about 1–2 seconds. The cloudflare runner moves to the Containers durable_object scheduling policy: one Durable Object class starts each lease's image and instance type through ctx.container, commands and uploads use native exec(), and the in-container Go runner, six per-type classes, and max_instances: 4 cap are gone. --type basic now uses standard-1 with a warning, --type lite is rejected because the bundled image cannot start on it, --cloudflare-image selects a named runner image, canceled commands stop their process group, and a lease whose container stops ends instead of continuing in an empty workspace. Upgrading is a one-time cutover: stop kept leases first, because redeploying deletes the old classes' Durable Object state. PR 2625. Thanks @altaywtf.
• Rewrite the README around on-demand computers for agents, with every supported provider listed. PR 2697.
• Redesign the crabbox.sh home page around on-demand computers for agents, with a provider wall generated from the catalog, a new crab-in-a-box icon, and a tokenized visual system shared by the docs and Features pages. PR 2699. Thanks @shakkernerd.
• Refresh the docs against the current CLI and provider catalog, including a single guide to desktop, VNC, browser and portal access. PR 2698.

Fixes

• Accept native Tenki SSH certificates on the first connection and during refresh by comparing the decoded session extension directly, while retaining key, signature, and expiry checks. PR 2727.
• DigitalOcean: check size availability before creating resources and retain definite fixed-lease rejections as retryable, locally releasable attempts. PR 2725.
• Scaleway: accept the same SSH public key after API comment normalization, allowing first fixed-lease acquisition and safe cleanup of stranded key-only attempts. PR 2724.
• Boat (ASCII Box): release absent sandboxes without waiting for blocked background data purges, while preserving the purge operation ID in fixed-lease terminal receipts. PR 2726.
• Probe resolved SSH leases with their recorded platform, preserve Linux bootstrap gates, and reject contradictory explicit target flags. Issue 2715, PR 2722, PR 2718. Thanks @saftall.
• Fail AWS acquire promptly when EC2 terminates an instance or closes its Spot request during network, SSH, Windows, or coordinator SSM readiness, preserving the reclaim reason and existing cleanup ownership. PR 2721, Issue 2716. Thanks @saftall.
• Retry transient run-event append failures with stable IDs, atomic sequence updates, and bounded publication deadlines while preserving ordered output and capped queues. PR 2723, Issue 1561.
• Blacksmith Testbox: run --id tbx_... and stop read only the exact claim file instead of scanning every local claim. PR 2719.
• Apply explicit lifecycle target flags before provider validation so Linux defaults cannot block macOS stop/release, pause/resume, list, or cleanup. PR 2712, Issue 2706. Thanks @coygeek.
• Stop coordinator alarm rearm loops from stale prewarm retries, clamped overdue work, and controller-owned wakes, with diagnostics and bounded backoff that survives concurrent heartbeats without postponing future deadlines. PR 2711.
• Transfer only verified seeded-worktree deltas on POSIX SSH targets, prune exclusions from fresh verified seeds, include reverted and remote edits on reused leases, and witness rsync receivers directly so failed transfers do not strand a detached workspace guard. PR 2707.
• Reject AWS developer-image mints outside the requested region and clean up failed unpromoted candidate checkpoints and snapshots while preserving promotion rollback. PR 2704.
• Keep lease reads, heartbeats, and releases responsive during maintenance history discovery by releasing the lifecycle mutex and Cloudflare storage input gate while scanning candidates; retain fenced rereads before mutations. PR 2705.
• Keep the homepage workload router's Copy button fully visible for short command snippets. PR 2703. Issue 2701. Thanks @shakkernerd.
• Advertise Scaleway's existing fixed-lease-id support in the provider catalog and enforce capability parity across registered backends. PR 2702.