OpenClaw Digest: Strengthening Codex Context, Gateway Stability, and Diagnostic Tooling

This update covers a series of critical fixes and feature enhancements merged into OpenClaw between May 17 and May 18, 2026. The changes primarily target the stability of the Codex extension, the reliability of the gateway process under various supervision topologies, and the expansion of the doctor command's diagnostic capabilities.

Merged PRs

• Fix recent context loss in long Codex sessions (Original PR)
• Keep legacy Codex OAuth sidecar profiles usable (Original PR)
• Fix denied Codex app-server tool policy (Original PR)
• fix(gemini-transport): validate thought_signature base64 before forwarding to Gemini (Original PR)
• feat(browser): surface observed dialogs (Original PR)
• Fix Discord verbose tool progress delivery (Original PR)
• Fix Telegram stop debounce bypass (Original PR)
• fix(gateway): keep unmanaged restarts in-process (Original PR)
• docs: clean up building plugins guide (Original PR)
• Log Telegram outbound delivery success (Original PR)
• Doctor: add health-check contract and --lint validation (Original PR)
• ci(mantis): allow ClawSweeper telegram proof agent (Original PR)
• fix(doctor): detect stale session snapshot paths (Original PR)
• fix(agents): persist subagent registry before returning accepted (Original PR)
• fix(codex): preserve streamed command output (Original PR)

Key Changes

Codex and LLM Transport Enhancements

Several PRs addressed critical failures in how Codex handles long-term session state and provider-specific data:

• Context Preservation: A bug was fixed where long Codex sessions were losing recent context because the projection guard was truncating the most recent turns instead of the oldest. The system now ensures the newest turns survive when clipping is necessary.
• Security Policy Enforcement: The Codex app-server now strictly honors deny: ["*"] policies. Previously, some native tool surfaces (like code mode and built-in environments) remained available even when a sender was denied all tools.
• Gemini Transport Stability: To prevent HTTP 400 errors during context compaction, the Gemini transport now validates thought_signature Base64 strings. If a signature is truncated mid-token during compaction, it is dropped silently rather than causing the entire assistant turn to fail.

Gateway and Infrastructure Stability

Significant work was done to prevent crash loops and improve process management:

• In-Process Restarts: A critical bug was fixed where the gateway would detach-spawn a replacement child when restarting under a custom supervisor (e.g., a bash loop or Kubernetes probe). This orphaned the process and led to EADDRINUSE crash loops. Unmanaged restarts are now kept in-process to maintain PID stability.
• Telegram Responsiveness: The stop and /stop commands now bypass inbound debounce buffers immediately, ensuring that abort requests are processed without delay and cancel any pending buffered prompts.

Diagnostic Tooling (doctor)

The doctor command has been evolved from a procedural repair tool into a structured health-check framework:

• Linting Contract: A new detect() / repair() contract separates diagnosis from application. This enables a new openclaw doctor --lint mode for read-only workspace validation.
• Stale Snapshot Detection: The doctor can now detect stale session snapshot metadata that references old runtime roots (e.g., after an install path change), preventing "random" path breakage in long-lived sessions.

Impact

These changes collectively reduce the operational friction for both end-users and maintainers. The fix for gateway restarts eliminates a permanent failure mode for users running OpenClaw under custom supervisors, while the Codex context and Gemini transport fixes ensure that long-running AI sessions remain coherent and stable.

From a developer perspective, the new doctor --lint capability allows for better CI integration and preflight checks, ensuring that workspace health can be validated without the risk of accidental mutation. Additionally, the restoration of verbose tool progress in Discord improves the transparency of long-running operations, addressing a common user pain point where silent tool execution was perceived as a system hang.