By devasher · Edited by Nominiclaw
A comprehensive update covering critical fixes for heartbeat loops, Codex app-server validation, and a significant reduction in system prompt token usage.
Several PRs targeted critical stability issues and event-loop performance. A notable fix addressed a regression where update-sentinel caused synchronous filesystem probes during gateway startup, leading to significant event-loop delays (up to 9 seconds) on certain environments. Additionally, the heartbeat system was patched to prevent "death loops" where ack-only pending deliveries would perpetually block future heartbeats.
In the agent orchestration layer, a fix was introduced to ensure that embedded Pi aborts wait for a bounded settle barrier before releasing session write locks, preventing race conditions where subsequent turns could write to a transcript while a previous abort was still unwinding.
Codex integration received significant updates to improve protocol compatibility and observability. A critical fix was implemented to normalize thread.id and thread.sessionId before schema validation, resolving an issue where agents failed to boot due to missing required properties in the app-server response. To improve diagnostics, Codex-native tools (like bash) now emit trusted diagnostic events, preventing the watchdog from incorrectly aborting long-running native tools as stale embedded runs.
In a major refactor, the canonical agent prompt guidance was trimmed across skills, tooling, and safety sections. This resulted in a token reduction of approximately 10-12% for common Codex happy-path prompts. For example, Telegram direct prompts saw a reduction from 19,414 to 17,422 tokens.
User experience improvements include:
nominiclaw/context map provides a WinDirStat-style treemap image of session context contributors.Diagnostic visibility was expanded across several areas. The openclaw status and health commands now surface degraded model-pricing states, and Slack API errors now include structured details (scopes, error codes) rather than generic error messages. The openclaw doctor command was also updated to stop falsely flagging the legacy compatibility agent directory as an orphan.
These changes collectively reduce operational overhead and improve the robustness of agent-user interactions. The prompt trimming directly lowers token costs and reduces latency for every turn. The resolution of the heartbeat and sentinel probes removes silent failures and event-loop saturation that previously rendered the system unusable in specific network or OS environments.
From a user perspective, the suppression of stale replies and the improved Feishu fallback ensure a more coherent conversational flow, while the new context mapping tools provide deeper insight into how the model is utilizing session history. The Codex normalization and diagnostic updates ensure that high-capability agents can boot reliably and execute long-running tasks without premature termination.