Back to Newsletter
v2026.4.21OpenClawSecurity
Stronger owner identity validation, session pruning, and Mattermost streaming.
This release strengthens command security with stricter owner identity checks, adds automatic session store pruning to keep things tidy, and brings streaming replies to Mattermost. Also includes dozens of reliability fixes across providers and channels.
New features
New features
- OpenClawBlueBubbles per-group system prompts — inject custom instructions per group chat with wildcard fallback support.
- OpenClawMattermost streaming — thinking, tool activity, and partial reply text now stream into a single draft preview post.
- OpenClawContext compaction notices — opt-in start and completion notices during context compaction so you know when it's happening.
- OpenClawTiered model pricing — cost estimates now support tiered pricing from model catalogs, including Moonshot Kimi K2.6/K2.5.
Improvements
Improvements
- OpenClawSessions now enforce a built-in entry cap and age prune by default — oversized stores are cleaned up automatically at load time.
- OpenClawCron job state is now split into a separate `jobs-state.json`, keeping `jobs.json` stable for git tracking.
- OpenClawTelegram polling watchdog threshold raised from 90s to 120s, reducing false stall detections. Configurable via `pollingStallThresholdMs`.
- OpenClawImage generation now falls back more gracefully when a provider/model candidate fails, with warning-level logging.
- OpenClawBundled plugin runtime dependency recovery improved via doctor paths for packaged installs.
Fixes
Fixes
- SecurityOwner identity validation now requires actual owner match or `operator.admin` status, preventing non-owners from reaching owner-only commands.
- OpenClawGateway exec in `security=full` + `ask=off` mode restored for stdin/heredoc forms.
- OpenClawCost usage cache now bounded with FIFO eviction to prevent unbounded memory growth.
- OpenClawSession `estimatedCostUsd` is now snapshotted to prevent cost compounding on repeated persist.
- OpenClawActive Memory degrades gracefully on recall failures during prompt building instead of crashing.
- OpenClawSlack thread aliases preserved in outbound sends, maintaining intended threading.
- OpenClawBrowser rejects invalid accessibility references immediately instead of timing out.
- Security`NODE_OPTIONS` and similar interpreter-startup env keys are blocked for MCP stdio servers.
Notes
Notes
- OpenClawMoonshot/Kimi default model is now `kimi-k2.6`. `kimi-k2.5` remains available but `thinking.keep = "all"` is now restricted to K2.6 only.
Running an older version?
Upgrade your instance from the dashboard to get everything above.
Open Dashboard