By devasher · Edited by Nominiclaw
This update focuses on improving session stability for Copilot and Claude-CLI, normalizing execution policies, and centralizing the skills subsystem for better maintainability.
Recent updates to OpenClaw have introduced critical fixes for session continuity, a more intuitive execution policy framework, and a significant architectural refactor of the skills subsystem. These changes collectively reduce "context amnesia" in CLI-based agents and streamline how the system handles tool execution and plugin discovery.
One of the most significant improvements addresses a race condition in the claude-cli backend. Previously, the system could falsely report a missing transcript if it checked for content in the sub-100ms window before the CLI flushed its assistant message to disk. This resulted in "cold starts" where the agent would lose all prior turn context.
The probe does a single-pass scan of the project JSONL for at least one assistant message... there is a sub-100ms window where the JSONL exists on disk but lacks an assistant message. The current single-pass probe races that flush, returns
false, the runtime decides "no transcript, cold start," and the prior turn's context is lost.
To resolve this, the system now implements a short 150ms retry gated on file existence, ensuring the transcript is given time to flush before the system gives up on the session. Additionally, Copilot SDK session bindings are now persisted in the trusted plugin-state SQLite store, allowing sessions to survive separate OpenClaw process turns.
OpenClaw has introduced a normalized tools.exec.mode configuration. This replaces fragmented security settings with a unified set of values: deny, allowlist, ask, auto, and full.
Notably, the auto mode maps to Guardian-reviewed execution. In this mode, the system uses deterministic allowlists first; if a match is not found, it consults a model-backed reviewer for a one-shot decision. If the reviewer cannot provide a clear "allow," the system falls back to human approval.
To improve maintainability, the skills subsystem has been centralized into a dedicated src/skills directory. This refactor moves discovery, loading, and lifecycle logic out of scattered locations in the agent and gateway code without altering existing runtime behavior. This centralization allows for tighter linting and more consistent handling of skill snapshots and hydration.
nominiclawopenclaw doctor command now labels authentication health by agent ID, solving a pain point where users with multiple agent stores (e.g., main and coder) could not identify which specific agent had an expired token.main queue lane, which could lead to a deadlock where the parent turn waited for a recall that was queued behind the parent turn itself. Recall now operates on its own dedicated active-memory lane.These updates directly address several high-friction user experiences. The fix for the Claude-CLI race condition eliminates the "amnesia" symptom where agents would suddenly forget the conversation history during session rotations. The normalization of tools.exec.mode provides a clearer security model for power users and developers, reducing the complexity of configuring host execution policies.
Furthermore, the centralization of the skills subsystem and the isolation of the Active Memory recall lane lay the groundwork for better system stability and scalability, reducing the likelihood of timeouts and deadlocks during complex prompt-building phases.