OpenClaw Digest: Enhancing Agent Control and Channel Reliability

This update covers a series of critical refinements to the OpenClaw ecosystem, focusing on strengthening the "emergency brake" mechanisms for runaway agents, improving the reliability of channel integrations—particularly Telegram—and refining how model runtime policies are resolved across different providers.

From fixing silent money-burn bugs caused by unbound API keys to resolving event-loop starvation in the Codex extension, these changes ensure that the gateway remains responsive and cost-effective under heavy agentic workloads.

Merged PRs

• [codex] Unify OpenAI auth provider picker (Original PR)
• fix(auto-reply): abort active text stop runs (Original PR)
• Fix gateway auth logout aborting active runs (Original PR)
• fix(codex): yield app-server notification projection (Original PR)
• Adopt Proxyline for managed proxy routing (Original PR)
• refactor(agents): share subagent cron fallback selection (Original PR)
• fix(cron): honor subagent model fallbacks (Original PR)
• fix(telegram): retain transcript-backed truncated finals (Original PR)
• fix(cron): separate failure notification delivery (Original PR)
• fix: surface stalled Telegram ingress backlog (Original PR)
• fix(doctor): materialize group allowFrom fallback (Original PR)
• Treat ambient group chatter as room events (Original PR)
• fix(codex): use stable hooks feature flag (Original PR)
• fix(auto-reply): restrict NO_REPLY to automatic group replies (Original PR)
• fix(agents): honor wildcard model runtime policy (Original PR)
• fix(discord): harden read message results (Original PR)
• fix: scope local agent gateway dispatch (Original PR)
• fix(agents): honor OpenAI completions token aliases (Original PR)
• fix(codex): scope user MCP servers by agent (Original PR)
• fix(openai): honor streaming usage compatibility (Original PR)
• fix(gateway): raise lifecycle hook timeout defaults (Original PR)
• fix: route crabbox proof through brokered aws (Original PR)

Key Changes

Agent Control and Safety

One of the most critical updates addresses a high-severity bug where LLM calls continued even after an API key was unbound in the Gateway control plane. This led to "silent money burn," where users incurred significant costs from rogue sessions that didn't terminate upon credential removal. The fix introduces models.authLogout, which now actively aborts matching in-process runs when auth is removed.

Additionally, the /stop command on Telegram was restored to its "fast-abort" behavior. A previous regression had caused /stop to be queued as a regular message, meaning the agent would finish its entire turn before the abort command was processed, defeating the purpose of an emergency stop.

Channel Reliability and UX

Telegram integration received several stability and UX improvements:

• Ingress Backlog Visibility: The gateway now correctly marks Telegram polling as unhealthy if the inbound spool is blocked by a stale handler, even if the Bot API polling itself is succeeding.
• Ambient Chatter: Non-trigger group messages are now treated as "room events." This allows agents to maintain context of group conversations without inappropriately replying to every message unless the message tool is explicitly called.
• Streaming Fixes: A regression where partial streaming replies were finalized as truncated snapshots (ending in ...) has been resolved by retaining transcript-backed drafts.

Infrastructure and Codex Extension

• Proxy Routing: OpenClaw has adopted Proxyline for managed proxy routing, replacing the previous internal shim with a dedicated runtime to improve network stability.
• Codex Event Loop: A critical fix was implemented in the Codex extension to prevent notification handlers (such as rate-limit updates) from synchronously blocking the Node.js event loop, which previously caused the gateway to become unresponsive to health checks.
• MCP Scoping: User MCP servers in Codex are now scoped by agent, ensuring that agents only have access to the servers specifically assigned to them.

Model Runtime and Compatibility

• Wildcard Policies: The system now honors provider/* wildcard runtime policies. This ensures that dynamic or self-hosted providers (like vLLM) correctly apply runtime settings to all matching models unless a specific model override exists.
• Token Alias Normalization: To support a wider range of OpenAI-compatible providers (including DashScope Kimi), OpenClaw now canonicalizes token aliases like max_completion_tokens and max_tokens to ensure request budgets are honored across embedded Pi runs.

Impact

These changes significantly reduce the operational risk for users by ensuring that credential revocation and stop commands are instantaneous and effective. The resolution of the event-loop starvation in the Codex extension and the hardening of Telegram's ingress health monitoring mean that production deployments will experience fewer "silent wedges" and unexpected downtime.

For developers using self-hosted LLMs or specialized OpenAI-compatible endpoints, the improved wildcard runtime policies and token alias handling provide a more seamless configuration experience, removing the the need for exhaustive per-model manual entries. Finally, the transition to Proxyline provides a more robust foundation for complex network environments requiring managed proxy routing.