OpenClaw Issue Digest: Session Recovery, Codex Runtime Stability, and Channel Routing

Open Issues

Recent activity in the OpenClaw repository reveals several critical stability and routing issues, particularly affecting the Codex runtime and various channel integrations.

Critical Runtime & Session Failures

• Codex Harness Registration: A high-severity regression in version 2026.5.12 causes Codex-primary agents to fail with Requested agent harness "codex" is not registered (#83062). This is attributed to a naming mismatch where the resolver looks for codex while the runtime registers as openai-codex.
• Session Recovery Gaps: Feishu DM sessions are failing to recover after gateway restarts (#82974), likely due to a timing mismatch where the session store is not yet loaded when lock cleanup runs. Similarly, session restart recovery can fail silently on timeout, leading to permanent context loss (#73160).
• Codex Memory Leaks: A significant memory growth issue has been reported in 2026.5.18, where chrome-devtools-mcp sidecars accumulate under the gateway and are not properly reaped, driving cgroup memory growth (#84413).

Channel Routing & UX Bugs

• Message Tool Routing Drift: A critical bug exists where the message tool may route replies to the webchat UI instead of the originating channel (e.g., Signal) if the currentChannelProvider has drifted to webchat (#82911).
• WhatsApp Approval Failures: Manual /approve commands on WhatsApp fail to resolve exec approvals, even though the same IDs work on Telegram, suggesting a visibility mismatch in the approval-runtime client path (#84456).
• Symptom-based Failures: Telegram group chats are experiencing silent reply drops for models without adaptive thinking when streaming is enabled (#83723).

Feature Requests & Architectural RFCs

• Internal Service Identity: A major RFC proposes separating internal service identity from user authentication to resolve a long history of trusted-proxy regressions (#69066).
• Injected Context Registry: Another RFC suggests a declarative registry for plugin-injected XML tags to prevent "data leakage" where internal memory tags are promoted into long-term memory or displayed in the UI (#69541).
• ACP Integration: There is significant interest in first-class support for the Agent Client Protocol (ACP) to allow interchangeable coding agents (#83021, #84443).

Key Themes

1. The "Codex Transition" Friction

Many of the current high-severity issues stem from the refactoring of the Codex runtime. This includes the harness registration mismatch (#83062), the onboarding drift that routes users to the unstable PI path instead of the app-server (#82978), and the nested sandbox failures when running Codex inside Docker (#83018).

2. Delivery Context Contamination

There is a recurring theme of "context drift" where the gateway loses track of the correct delivery channel. This is seen in the message tool routing bug (#82911) and the subagent announce delivery issue where results are sent to parallel channel sessions instead of the originating parent session (#80201).

3. Discovery & Metadata Gaps

Several reports highlight "dark capabilities"—features that exist but are unreachable. Examples include the canvas skill lacking YAML frontmatter (#83036) and the memory-wiki linting rules conflicting with the intent of the sources/ directory (#69700).

Action Required

Immediate Attention (High Severity)

• Fix Codex Harness Registration: Resolve the naming conflict between codex and openai-codex to restore functionality for Codex-primary agents (#83062).
• Address Codex Sidecar Leak: Implement a reaper or shared pool for chrome-devtools-mcp processes to prevent gateway memory exhaustion (#84413).
• Repair Message Tool Routing: Ensure the message tool prefers session-key inferred delivery over the potentially stale currentChannelProvider (#82911).

Blocked or High-Risk

• WhatsApp Approval Path: Investigation is needed into why WhatsApp cannot resolve pending approval IDs that Telegram can (#84456).
• Feishu Session Recovery: The startup sequence needs to be audited to ensure the session store is loaded before lock cleanup and that recovery delays account for channel readiness (#82974).

Contributor Opportunities

• Frontmatter Fixes: Adding YAML frontmatter to the canvas skill is a low-effort, high-impact fix for discoverability (#83036).
• CLI UX Consistency: Implementing the list alias for openclaw sessions would align it with other CLI parent commands (#81139).