openclaw - 💡(How to fix) Fix Codex runtime sessions should recover when persisted app-server thread id is missing

Environment

• OpenClaw 2026.5.22 (a374c3a)
• Main agent using openai-codex / Codex runtime
• Trigger path: Telegram direct-message dispatch

Impact

A direct-message session became permanently pinned to a missing Codex app-server thread id. Incoming messages reached OpenClaw, but dispatch failed before normal agent execution and users only saw the generic failure response:

Something went wrong while processing your request. Please try again.

Outbound Telegram transport was otherwise healthy.

Evidence

• Inbound Telegram DM was received and routed to a session key shaped like agent:main:telegram:direct:<redacted-chat-id>.
• Dispatch failed before normal agent execution:

Preflight compaction required but failed: thread not found: <redacted-app-server-thread-id>

• The OpenClaw session store row still existed and the transcript file still existed at agents/main/sessions/<redacted-session-id>.jsonl.
• The Codex app-server sidecar still existed at agents/main/sessions/<redacted-session-id>.jsonl.codex-app-server.json and still pointed at the missing app-server thread id.
• Moving only that stale sidecar aside caused OpenClaw to create a fresh app-server thread id, after which the same session key succeeded again with openclaw agent --session-key agent:main:telegram:direct:<redacted-chat-id> ....

Expected behavior

If a persisted Codex app-server thread id is missing during preflight compaction, OpenClaw should retire or refresh the app-server sidecar and continue with a new thread while preserving the OpenClaw transcript. A stale sidecar should not leave channel dispatch permanently broken behind a generic user-facing error.

Suggested fix

• Detect thread not found during preflight compaction.
• Mark the sidecar stale and create a replacement Codex app-server thread for the same OpenClaw session.
• Preserve the transcript and record the recovery in logs.
• Consider extending openclaw sessions cleanup to detect stale Codex app-server sidecars, not just missing transcript files.

Local recovery used

Backed up the stale sidecar under agents/main/sessions/stale-thread-backups/ and let the next run create a new sidecar. That restored dispatch for the existing direct-message session key without deleting the transcript.