[Release] openclaw/openclaw v2026.4.10: openclaw 2026.4.10

OpenClaw v2026.4.10 introduces major features including Codex provider integration for GPT models, an Active Memory plugin for context-aware conversations, experimental local MLX speech support for macOS, and expanded QA testing capabilities with Matrix and Telegram lanes. The release includes extensive security hardening across browser navigation, tool execution, and WebSocket handling, plus numerous fixes for Microsoft Teams integration, WhatsApp media delivery, gateway startup stability, and agent-based tool use.

Key Points

• •Codex provider integration: GPT models now use Codex-managed authentication, native threads, and model discovery while maintaining OpenAI provider path for openai/gpt-* models
• •Active Memory plugin: Automatically pulls relevant preferences, context, and past details into conversations with configurable message/recent/full context modes and live inspection via `/verbose`
• •Local MLX speech provider: Experimental Talk Mode support for macOS with explicit provider selection, local utterance playback, and system-voice fallback
• •QA testing expansion: New Matrix and Telegram QA lanes with live transport coverage, plus `--runner multipass` for running scenarios in disposable Linux VMs
• •Exec policy CLI: New `openclaw exec-policy` command with show/preset/set subcommands for synchronizing tool execution approvals with conflict detection
• •Gateway commands discovery: `commands.list` RPC enables remote clients to discover runtime-native, text, skill, and plugin commands with argument metadata
• •Security hardening: Comprehensive fixes across browser SSRF defenses, exec preflight reads, host environment denylisting, and WebSocket frame handling
• •Microsoft Teams improvements: Restored media downloads for personal DMs, OneDrive/SharePoint files, added SSO callbacks, and improved thread reply context injection
• •WhatsApp reliability: Fixed media delivery for documents/attachments, improved socket reconnection handling for replies and composing indicators
• •Agent execution: Strict-agentic embedded Pi execution contract for GPT-5 family runs ensuring plan-only turns continue acting until real blockers are hit

Workflow Diagram

Concepts

Artifacts (4)

openclaw exec-policy show
openclaw exec-policy preset <preset-name>
openclaw exec-policy set <config-path>

{
  "method": "commands.list",
  "description": "Discover runtime-native, text, skill, and plugin commands",
  "returns": {
    "commands": [
      {
        "name": "<command-name>",
        "surface": "<surface-type>",
        "arguments": ["<arg-metadata>"]
      }
    ]
  }
}

{
  "plugins": {
    "active_memory": {
      "enabled": true,
      "context_mode": "full|recent|message",
      "transcript_persistence": true,
      "verbose_inspection": "/verbose",
      "prompt_overrides": true,
      "thinking_overrides": true
    }
  }
}

{
  "models": {
    "providers": {
      "openai_compatible": {
        "request": {
          "allowPrivateNetwork": true
        }
      }
    }
  }
}