OpenClaw 2026.2.23 is live with major provider integrations, prompt caching improvements, and critical fixes across Slack threading, Telegram reactions, and session management. Here’s what’s new and what you need to know.
Major Features
Moonshot (Kimi) Provider Support
OpenClaw now includes native Moonshot video analysis and web search integration. The Moonshot provider brings:
- Native video understanding — Process videos directly without external transcoding
- Web search with citations — Search results now include proper citations and source attribution
- Auto-detection — Moonshot is automatically included in video key detection
- Header/URL precedence — Video execution now honors entry/config/provider baseUrl and header precedence
This is a game-changer for developers building AI agents that need to analyze video content or perform web searches with cited sources.
Kilo Gateway (Kilocode) First-Class Support
Kilo Gateway now has first-class provider support in OpenClaw, meaning:
- Auth and onboarding flows built in
- Implicit provider detection (no manual config required)
- Default model set to
kilocode/anthropic/claude-opus-4.6 - Transcript and cache-ttl handling fully supported
- Complete documentation
If you’re running Kilocode, setup is now seamless—just authenticate and you’re ready to go.
Sessions Maintenance & Disk Budget Control
OpenClaw now includes hardened session maintenance with disk-budget controls:
session.maintenance.maxDiskBytes— Total disk quota for sessionssession.maintenance.highWaterBytes— Trigger cleanup when this threshold is reached- Safer transcript and archive cleanup behavior
- Per-agent store targeting for granular control
This prevents runaway session storage from consuming your disk and gives you precise control over session lifecycle management.
Provider Updates & Fixes
Vercel AI Gateway
Vercel AI Gateway now accepts Claude shorthand model refs (e.g., vercel-ai-gateway/claude-opus) and automatically normalizes them to canonical Anthropic-routed model IDs. No more manual mapping needed.
Anthropic OAuth Token Handling
Fixed a critical bug where OAuth tokens (sk-ant-oat-*) were incorrectly receiving context-1m-* beta headers, causing 401 auth failures. The fix preserves OAuth-required betas while skipping unsupported ones.
Bedrock Prompt Caching
Prompt-cache retention is now properly scoped:
- Enabled for Anthropic Claude models on Bedrock
- Disabled for non-Anthropic models (Nova, Mistral, etc.) which don’t support cache metadata
Groq TPM & OpenRouter Reasoning Fixes
Groq throttling errors (TPM limits) are no longer misclassified as context overflow, preventing unnecessary recovery logic. OpenRouter reasoning requests no longer send conflicting top-level reasoning_effort when using nested reasoning.effort.
Slack Threading & Telegram Reliability
Slack: Thread context now persists beyond the first turn, and threading respects your replyToMode setting. Extension replies and file uploads now correctly route to the right thread.
Telegram: Reactions are more resilient (soft-fail on policy/token/emoji errors), polling offsets are properly scoped to bot identity, and reasoning text no longer leaks into channel output.
Security & Config Hardening
- Sensitive key redaction: Config snapshots now redact sensitive keys like env.* and skills.entries.*.env.*
- Session key canonicalization: Mixed-case session keys are now normalized to prevent duplicate sessions
- Browser SSRF policy: Now defaults to trusted-network mode for better out-of-the-box usability
- HTTP security headers: Optional Strict-Transport-Security support for HTTPS deployments
What You Should Do
- Update now:
openclaw update --channel stable - Run doctor:
openclaw doctorto check for deprecated configs - Review your config: Check browser.ssrfPolicy settings—the old key name is still supported but
openclaw doctor --fixwill migrate you automatically - If using Slack: Test thread context to ensure your conversations flow correctly
Looking Ahead
This release sets the stage for more sophisticated provider integrations and better multi-model orchestration. The focus remains on reliability, security, and making OpenClaw easier to deploy at scale.
Questions? Check the updating guide, browse the full docs, or join the OpenClaw Discord.
Share:
