realvco Docs

v2026.4.22 — xAI Image/Voice Trio + Multi-Platform STT

Released: April 22, 2026 Theme: Multimedia provider expansion + Codex harness deepening Breaking Changes: None

Core Highlights

1. xAI Full Media Support (Image + TTS + STT)

xAI provider gets all three media generation modes at once:

  • Image gen: grok-imagine-image and grok-imagine-image-pro, with reference-image edits
  • TTS: six xAI voices, MP3/WAV/PCM/G.711 formats
  • STT: grok-stt batch transcription + xAI realtime transcription (for Voice Call streaming)

2. Multi-Provider Realtime STT Streaming

Voice Call streaming transcription expands to:

  • Deepgram
  • ElevenLabs (also adds Scribe v2 batch transcription)
  • Mistral

Joining the existing OpenAI and xAI realtime STT paths.

3. TUI Local Embedded Mode

TUI (terminal interface) adds local embedded mode:

  • Run terminal chats without a Gateway
  • Plugin approval gates still enforced

Good for quick local trials or restricted environments.

4. Onboarding Auto-Installs Plugins

Setup auto-installs missing provider and channel plugins:

  • First-run config doesn’t need manual plugin recovery
  • Smoother flow

5. Tencent Cloud Provider

Bundled Tencent Cloud provider plugin:

  • TokenHub onboarding
  • Docs
  • hy3-preview model catalog
  • Tiered Hy3 pricing metadata

6. WhatsApp + Group System Prompts

WhatsApp groups and direct chats now support per-group / per-direct systemPrompt config:

  • Config injects into inbound context GroupSystemPrompt on every turn
  • Supports "*" wildcard fallback
  • Account-scoped overrides at channels.whatsapp.accounts.<id>.{groups,direct}

7. /models add From Chat

New /models add <provider> <modelId> command:

  • Register a model from chat without restarting the gateway
  • /models becomes a simple provider browser
  • Clearer add guidance and copy-friendly examples

8. Codex Harness Deepening

  • Native Codex app-server turns route through before_prompt_build, before_compaction, after_compaction
  • Bundled-plugin Codex app-server extension seam (async tool_result middleware)
  • llm_input, llm_output, agent_end all fire
  • Codex MCP tool approval elicitations route through OpenClaw plugin approvals

Direct OpenAI Responses models now automatically use OpenAI’s native web_search tool (when web search is enabled and no managed search provider is pinned); other choices like Brave keep the managed web_search.

10. Gateway Subprocess OOM-Friendly

On Linux, gateway-managed supervisor, PTY, MCP stdio, and browser child processes wrap in a tiny /bin/sh shim that raises the child’s own oom_score_adj:

  • Under cgroup memory pressure, the kernel prefers killing transient workers over the long-lived gateway
  • Opt out with OPENCLAW_CHILD_OOM_SCORE_ADJ=0

11. WhatsApp Outbound Dedupe Fix

Outbound holds an in-memory active-delivery claim while a live send is in flight:

  • Concurrent reconnect drain no longer re-drives the same pending queue entry
  • Cron messages no longer duplicate 7-12x after the 30-minute inbound-silence watchdog fires mid-delivery

12. Codex / OAuth Profile Fixes

  • Stale openai-codex:default OAuth profile no longer shadows newer main-agent identity-scoped profile
  • ACPX OpenClaw bridge commands route through MCP-free runtime path (even when wrapped with env, with bridge flags, or resumed from persisted session state)

Upgrade Recommendations

  • Customers wanting Grok image/voice: upgrade and use immediately
  • Voice Call customers: Deepgram / ElevenLabs / Mistral streaming transcription available
  • TUI users not running gateway: local embedded mode is here
  • WhatsApp group routing users: per-group systemPrompt now configurable
  • Linux gateway operators worried about OOM: subprocess OOM-friendly behavior helps
  • Anyone affected by WhatsApp outbound duplication: upgrade fixes it

← v2026.4.21 · v2026.4.23 → · Back to list