OpenClaw Update: Enhancing CLI Performance, Security Audit, and Tool Governance

The latest set of merges for OpenClaw brings a heavy focus on developer experience (DX) and operational stability. From slashing CLI startup times by over 70% for key commands to introducing a formal policy layer for tool metadata, these changes aim to make the platform more scalable and observable for power users and enterprise deployments.

Beyond performance, this window addresses critical "data debt" in session management and resolves a high-severity tool-call replay loop that was impacting agent reliability. These updates collectively move OpenClaw toward a more disciplined approach to both runtime execution and configuration security.

Merged PRs

• fix(doctor): clear stale runtime override pins (Original PR)
• fix(agents): disable pi-coding-agent auto-retry to prevent tool call replay loops (Original PR)
• fix(trajectory): tolerate partial skill snapshot entries in support capture (Original PR)
• fix(ui): widen settings personal card (Original PR)
• fix(agents): log pre-prompt compaction fits decisions (Original PR)
• fix(memory-core): allow bounded dreaming session cleanup (Original PR)
• Skip empty sherpa structured transcripts (Original PR)
• perf(cli): lazy-load agents actions for help (Original PR)
• feat: support git and local skill installs (Original PR)
• Policy: add tool metadata conformance (Original PR)
• fix(doctor): warn when sandbox hides MCP tools (Original PR)
• fix(cli): speed up onboarding help startup (Original PR)
• perf: isolate doctor core check tests (Original PR)
• feat(tasks): explain stale-running maintenance decisions (Original PR)
• fix(minimax): stop advertising music duration control (Original PR)
• fix(codex): guard path-only bootstrap files [AI-assisted] (Original PR)
• Warn on plaintext secret config in doctor (Original PR)
• Remove skill prelude exec allowlist (Original PR)
• perf(tui): defer EmbeddedTuiBackend import, drop dead warmup helpers (Original PR)
• perf(tui): skip plugin metadata + provider catalog on remote TUI startup (Original PR)
• Route JSON-mode plugin registration logs to stderr (Original PR)

Key Changes

CLI Performance & TUI Optimization

One of the most impactful changes in this window is the aggressive refactoring of the CLI and TUI startup paths. By implementing lazy-loading for agent actions and onboarding help, the team has significantly reduced the "cold start" penalty:

• agents --help: Startup time dropped from ~1.38s to ~361ms, with a corresponding reduction in RSS memory usage from 522MB to 321MB.
• Onboarding Help: configure --help saw a 70.9% reduction in startup time.
• Remote TUI: The TUI now skips plugin metadata and provider catalog resolution when connecting to a remote gateway, eliminating event-loop freezes that previously lasted tens of seconds.

Tool Governance & Policy

OpenClaw is introducing a formal conformance layer for tools. The new Policy plugin allows administrators to require specific metadata (such as risk, sensitivity, and owner) in TOOLS.md declarations. This is a read-only layer that reports drift via openclaw policy check or doctor --lint, ensuring that as the tool ecosystem grows, every tool is properly categorized and owned without mutating the workspace.

Agent Reliability & Session Health

Several critical fixes address the "invisible" failures that plague long-running agents:

• Tool Replay Loops: Resolved a regression where the runtime would automatically replay a failed tool call, causing infinite loops and context window exhaustion. This is fixed by disabling auto-retry for the pi-coding-agent.
• Stale Session Pins: doctor --fix now detects and clears stale agentRuntimeOverride pins. This prevents "data debt" where sessions remain pinned to obsolete runtimes (e.g., claude-cli) long after the user has migrated to a different model.
• Voice STT Filtering: Empty structured transcripts from sherpa-onnx are now skipped, preventing raw JSON from being passed to the LLM and clogging the processing queue.

Security & Configuration

To prevent API key exposure, openclaw doctor now emits a security warning if openclaw.json contains plaintext secrets (e.g., apiKey or Authorization headers). Users are encouraged to migrate these to SecretRefs using the openclaw secrets suite.

Impact

For the end user, these changes translate to a snappier, more professional CLI experience and a more stable agent runtime. The reduction in startup latency makes the CLI feel like a tool rather than a heavy application.

From an operational perspective, the introduction of tool metadata conformance and plaintext secret warnings significantly lowers the risk profile for teams deploying OpenClaw in shared or governed environments. The fix for tool replay loops is particularly vital, as it prevents the catastrophic failure mode of an agent consuming its entire token budget on a single failing command.

Finally, the addition of support for git and local skill installs expands the extensibility of the platform, allowing developers to iterate on skills locally or distribute them via Git repositories rather than relying solely on ClawHub.