OpenClaw Gateway Stability and Security Hardening: May 16 Digest

The latest set of merged pull requests for OpenClaw focuses heavily on the reliability of the gateway's lifecycle management and the hardening of sensitive data handling. From resolving complex race conditions during startup to ensuring that credentials never leak into diagnostics, these changes stabilize the core orchestration layer while refining how the system interacts with external messaging channels.

Merged PRs

• [Fix] Queue startup restart signals during gateway startup #82660
• fix(agents): harden auth refresh redaction #82670
• fix(telegram): normalize announce group targets #81229
• fix(update): preserve channel config across package repair #82549
• fix(gateway): redact credential-bearing diagnostics #82677
• Fix chat session picker agent switching #81858
• fix(discord): bind delayed identify to socket generation #82225
• fix: bypass npm freshness filters during updates #82641
• fix(acp): refresh runtime handles on config changes #82237
• test(matrix): add state-after E2EE QA coverage #82655
• fix(gateway): scope session data lookups by agent [AI] #81386
• fix(telegram): cache startup bot info #82634
• Fix exec allowlist wildcard target normalization #75723
• [Fix] Improve gateway restart readiness diagnostics #82603
• fix(auth): avoid keychain creation for oauth profile secrets #82652
• fix: preserve generated media completion attachments #82636
• fix(matrix): avoid state-after sync opt-in #82631
• fix(infra): drop in-flight approval delivery after onStopped #82482
• fix(sessions): estimate local transcript usage #82317
• fix(auto-reply): log suppressed message-tool-only finals #82609
• fix(slack): route DM thread replies to main session instead of thread-scoped session #82418
• fix(codex): keep app-server turns progress-aware #82601
• fix(auto-reply): preserve session model display for heartbeat usage #82267
• fix(agents): log detail-less responses failures #82593
• OC Path: restore YAML support #81436
• Strip inbound metadata from replayed user turns #82614
• fix(tui): update model display during fallback #82296
• fix: honor Codex auth order for OpenAI PI #82605
• fix(gateway): bound traced channel startup handoff #82592
• fix(ollama): skip think for non-reasoning models #82445

Key Changes

Gateway Orchestration & Lifecycle

Several PRs address critical timing and diagnostic issues during gateway startup and restart. A significant fix was introduced to handle restart signals that arrive while the gateway is still starting up; the run loop now queues these requests and flushes them once the state is valid, preventing inconsistent recovery during update churn. Additionally, diagnostics have been improved to distinguish between HTTP bind readiness and full gateway readiness, providing clearer performance benchmarks.

Security & Credential Hardening

Data privacy and secret management received substantial updates:

• Redaction: Gateway diagnostic strings and OAuth refresh failure messages are now more aggressively redacted to prevent credential material from appearing in logs, even when logging.redactSensitive is disabled.
• Keychain Management: On macOS, the system no longer automatically creates OAuth profile master keys in the Keychain, avoiding intrusive OS modals in headless environments.
• Data Scoping: Session data lookups (artifacts, memory, usage) are now strictly scoped by agentId, preventing request-scoped APIs from performing overly broad discovery across different agents.

Channel-Specific Improvements

• Telegram: Fixed a regression where sessions_send announce delivery failed for groups because targets weren't normalized to numeric chat IDs. Bot identity is now cached to skip redundant getMe calls on restart.
• Slack: DM thread replies are now routed to the main session instead of creating invisible thread-scoped sessions, improving conversation continuity.
• Matrix: The Matrix SDK's state_after sync opt-in is now stripped from /sync requests to avoid issues where incomplete state omits encryption data, leaving outbound crypto unconfigured.
• Discord: Fixed a race condition where the IDENTIFY payload could be sent over a replacement WebSocket instead of the original socket that received the HELLO signal.

Agent & Model Runtime

• Codex: App-server turns are now progress-aware, ensuring that active turns don't time out as long as they are producing activity, while unrelated requests no longer mask stalled turns.
• Ollama: To prevent 400 errors from servers that reject thinking for non-reasoning models, the think parameter is now skipped for models marked reasoning: false.
• Prompt Optimization: Inbound metadata envelopes are now stripped from historical user turns during LLM replay, reducing prompt bloat (which could reach 77-79% of the prompt in some cases) without losing current-turn context.

Impact

These changes collectively reduce the operational overhead of managing OpenClaw gateways, particularly during updates and restarts. Users will experience more reliable channel connectivity and a more responsive Control UI, with the TUI now correctly reflecting model changes during provider fallback.

For developers and operators, the security hardening ensures that sensitive tokens and keys are less likely to be exposed in logs, and the agent-scoping of memory and session lookups provides a more robust multi-tenant architecture. The restoration of YAML support in oc-path also expands the utility of the universal addressing model, allowing agents to interact with workflow files and YAML configurations as seamlessly as they do with JSON or Markdown.