OpenClaw Update: Optimizing Telegram Caching, Hardening Sandbox Security, and Expanding Model Support
Merged PRs
- perf(telegram): append reply-chain cache records Original PR
- fix(backup): keep temp manifest outside source paths Original PR
- [Fix] Bound CLI runner supervisor output Original PR
- feat(config): support Qwen thinkingFormat values Original PR
- fix(image-gen): propagate ssrfPolicy into provider HTTP calls (#79716) Original PR
- fix(model): restore image input for google multimodal fallback models (#79750) Original PR
- fix(sandbox): include container path hint in sandbox-root escape error (#79712) Original PR
- fix: harden agent recovery failures Original PR
- fix(telegram): guard DM bindings from being parsed as topics Original PR
- fix(launchd): set ProcessType=Interactive to prevent macOS App Nap Original PR
- fix(status): show Codex usage for Codex harness Original PR
- Unify thinking and fast run controls Original PR
- fix(failover): classify assistant-prefill rejection as format error (#79688) Original PR
- fix(agents): ignore ACP-only streamTo and treat default model as unset Original PR
- feat(mistral): add mistral-medium-3.5 model Original PR
- fix(qqbot): route gateway websocket through ambient proxy agent Original PR
- fix(gateway): harden macOS update restart lifecycle Original PR
- fix(agents): classify stream_read_error as transient Original PR
- fix: install Codex for OpenAI onboarding Original PR
- fix(models): use effective auth health for status Original PR
- refactor: use PI Codex Responses transport Original PR
- fix(memory): warn on unverified qmd path conflicts Original PR
- fix(backup): retry on tar EOF race and skip known volatile files Original PR
- fix(agents/cli): bridge CLI assistant deltas into channel preview (#76869) Original PR
- fix(feishu): stop automatic mention cascades Original PR
- docs: update stale sub-agent context injection details Original PR
- fix(agents): suppress DeepSeek DSML stream artifacts Original PR
- feat(matrix): expose MessagePresentation metadata in Matrix events Original PR
- fix: annotate message-tool-only replies in Codex tool spec Original PR
Key Changes
Performance and Infrastructure
- Telegram Cache Optimization: The Telegram reply-chain cache now uses append records instead of rewriting the full cache file for every inbound message. Benchmarks show massive speedups, reaching up to 88.3x for larger record sets.
- macOS Stability: To prevent the gateway from being suspended by macOS App Nap,
ProcessType: Interactivehas been added to the launchd plist. Additionally, the macOS update restart lifecycle was hardened to prevent stale restart states from poisoning future requests. - Backup Reliability: Fixed a critical issue where
openclaw backup createwould fail due totarEOF races when files (like logs or transcripts) were modified during the backup process. The system now skips known volatile files and implements a retry mechanism.
Agent and Model Enhancements
- Expanded Model Support: Added support for
mistral-medium-3.5and enabledthinkingFormatvalues for Qwen models (qwenandqwen-chat-template), allowing better control over reasoning capabilities in local OpenAI-compatible backends. - DeepSeek Stream Filtering: Implemented a filter to suppress provider-private DSML (DeepSeek Markup Language) artifacts from assistant-visible output, ensuring that control syntax does not leak into user transcripts.
- Failover and Recovery: Improved agent recovery by classifying
assistant-prefillrejections andstream_read_erroras specific error types (format and transient, respectively), preventing infinite retry loops and enabling proper fallback to alternative models. - Sandbox Ergonomics: Sandbox escape errors now include a container path hint, helping LLM agents pivot to the correct workspace path rather than looping on rejected host paths.
Integration and Channel Fixes
- Telegram & Feishu: Fixed a bug where Telegram DM bindings were incorrectly parsed as topics. In Feishu, automatic mention cascades were disabled to prevent error fallbacks from notifying all mentioned users unnecessarily.
- CLI Streaming: Bridged CLI assistant deltas into channel previews, allowing users of CLI-backed agents (like
claude-cli) to see live text streaming in Telegram instead of waiting for the full message. - Matrix Metadata: Matrix events now expose
MessagePresentationmetadata, with a plain-text fallback for divider-only payloads.
Impact
These changes significantly improve the operational stability of OpenClaw, particularly for macOS users and those performing live backups. The Telegram performance optimization drastically reduces I/O overhead for high-volume bots. From an AI perspective, the refined error classification and sandbox diagnostics reduce agent "hallucination loops" during tool execution and failover. Finally, the expanded support for Qwen and Mistral ensures that users can leverage the latest reasoning-capable models with precise control over their thinking processes.