OpenClaw 2026.4.20 hardens cron delivery, session pruning, and subagent failure reporting
TL;DR
OpenClaw's 2026.4.20 release landed on April 20 with a focus on stability under load: session prune-on-load prevents gateway OOM from cron backlog, cron delivery handles mode: none correctly, subagent failures now carry role and timing context, and PowerShell tool allow-list parsing finally works correctly on Windows.
What changed
What shipped
OpenClaw 2026.4.20 is the latest weekly release, landing April 20, 2026. The changes are unglamorous but operationally significant for anyone running OpenClaw under real load.
Session and gateway stability
Session prune-on-load. The gateway now enforces the built-in entry cap and age-prunes by default, and prunes oversized stores at load time. This prevents accumulated cron or executor session backlogs from crashing the gateway before the write path gets a chance to run. If you have been running OpenClaw with cron-heavy workloads and hitting intermittent gateway failures, this release should fix it.
Cron improvements
OpenClaw's cron subsystem received an unusually large set of fixes this release:
- Runtime state split. Jobs-state.json is now separate from jobs.json, so git-tracked job definitions stay stable while runtime execution state moves independently.
- delivery.mode: "none" handled correctly. Previously, no-delivery cron jobs could persist false delivery failures. Now they do not.
- Isolated cron chat delivery. Channel: "last" targets resolve correctly from the gateway. Delivery previews show up in cron list/show. Duplicate fallback sends after direct message-tool delivery are eliminated.
- Telegram dedupe per execution. Isolated direct-delivery dedupe is now keyed to each cron execution rather than reusing the session id, so recurring Telegram announce runs no longer silently skip later sends.
- Ambiguity rejection at validation time. Multi-channel or invalid target-id provider settings now fail at cron add or update time, not after persisting.
- Heartbeat target preservation. Cron main-session heartbeat.target: "last" is preserved through deferred wake queuing, gateway wake forwarding, and same-target wake coalescing.
Subagent improvements
Failure payloads now carry context. Subagent failure payloads include the requested role and runtime timing. Parent agents can correlate failed or timed-out child work instead of inheriting an opaque failure.
Windows-specific fix
PowerShell --tools allow-list parsing. Cron add and cron edit now parse PowerShell-style --tools allow-lists the same way as comma-separated input. Windows users who have seen "exec read write" persisted as one combined tool entry: this release fixes that.
Model and provider updates
Kimi model updates, tiered pricing, better OpenAI handling, improved plugin handling. The 2026.4.20 release notes emphasise broadened support across model providers, though the headline fixes are in the cron and session subsystems.
GitHub status
OpenClaw passed 355K GitHub stars in roughly five months, making it one of the fastest-growing open-source AI agent projects of 2026. The weekly release cadence continues with community-submitted PRs making up a large share of each release.
Upgrade path
No breaking changes called out in this release. Pull the latest version through the standard update mechanism; the session prune-on-load applies automatically on next boot.
Who this matters for
- Vibe Builder: Cron delivery is significantly more reliable in this release. If you run OpenClaw with scheduled Telegram or WhatsApp check-ins, update and expect fewer silent failures on recurring runs.
- Basic User: Session prune-on-load means your OpenClaw instance is less likely to crash under accumulated cron backlog. Just update; no config changes needed.
- Developer: Subagent failure payloads now carry requested role and runtime timing. Debugging parent-agent workflows with multiple failed children just got easier. Windows PowerShell tool allow-list parsing is finally correct.
What to watch next
OpenClaw's 2026.4.20 is a stability release, and those are the ones that matter most in production. Anyone running cron-heavy OpenClaw workloads has probably hit the gateway OOM failure mode at some point; the session prune-on-load fix removes that entire class of bug.
The cron hardening is the real story. Six distinct fixes to cron delivery, dedupe, heartbeat preservation, and validation in one release. This is what a maturing open-source project looks like: the early releases add features, the middle releases fix the integration edges, and the late releases harden production. OpenClaw is clearly in the middle-to-late phase on cron.
For Windows users, the PowerShell tool allow-list parsing fix is a small but real quality-of-life win. Anyone who has tried to configure cron jobs with specific tool allow-lists on Windows knows the pain of watching your three-tool permission persist as one combined string. Fixed now.
The subagent failure payloads carrying role and timing is the fix most people will not notice until they need it. When a parent agent has ten subagent failures across a complex workflow, the difference between "subagent X failed at 14:03 after 47s as role Y" and "subagent X failed" is the difference between debugging in minutes versus hours. Small feature, significant operational impact.
The 355K GitHub stars in five months puts OpenClaw in a rare category. Worth watching how the team balances feature velocity against the production concerns that arise when a tool becomes the de facto choice for self-hosted personal AI agents.
by Harsh Desai