OpenClaw Issue Digest: Model Resolution Regressions and Channel Instability

Open Issues

Recent activity in the OpenClaw repository reveals a cluster of high-impact regressions affecting model resolution, provider authentication, and channel stability. Most notably, several users are reporting critical failures in the OpenAI and OpenRouter provider paths, where internal provider prefixes are not being stripped from API requests, leading to HTTP 400 errors and silent degradation to local fallbacks.

Further instability is evident in the Gateway's event loop. Production monitoring via the Heimdall system has flagged repeated event loop delays exceeding 2 seconds, often coinciding with subagent/embedded-run load. This pressure is manifesting as WebSocket handshake timeouts and connection drops in channels like WhatsApp (408 Request Time-out). Additionally, a significant bug in the Codex runtime is causing non-OpenAI model selections to be routed into the Codex app-server, where they fail immediately because the harness cannot resolve non-OpenAI providers.

On the channel side, Telegram and Feishu users are reporting intermittent delivery failures. In Telegram, some group topics only respond to direct @mentions despite requireMention: false being configured. In Feishu, thread-based replies are being silently dropped when the API call fails, rather than falling back to a standard group message.

Key Themes

Model Resolution and Auth Failures

There is a recurring theme of "silent failures" in the model routing layer.

• Prefix Stripping: In version 2026.5.7, OpenRouter requests are sent with the openrouter/ prefix still attached, which the API rejects.
• Auth Header Gaps: A critical bug in the QuickStart Docker flow results in OpenAI requests being sent without an Authorization header, despite the API key being correctly saved to auth-profiles.json.
• Codex Runtime Mismatch: The Codex harness is incorrectly attempting to process models from providers like MiniMax and Anthropic, leading to Model provider not found errors.

Gateway Performance and Stability

Resource contention during subagent execution is a primary concern.

• Event Loop Saturation: High CPU utilization and event loop delays are causing "handshake timeouts" for new WebSocket connections.
• Connection Overhead: Every agent run currently performs a full model-resolution stage, which can take 7-8 seconds for custom OpenAI-compatible providers due to a lack of HTTP connection pooling.
• Zombie Processes: In containerized environments, the lack of a PID 1 init system (like tini) is leading to an accumulation of zombie processes from exec tool calls, eventually poisoning lock files.

Channel-Specific Regressions

• Telegram: Regression in group topic behavior where plain messages are ignored unless the bot is @mentioned.
• Feishu: Threaded replies are failing silently without fallback, and some users report the bot failing to respond to @mentions entirely due to "unavailable message" errors.
• Slack: Thread sessions are missing the parent/root message content, leaving agents without the necessary context to understand user replies.

Action Required

High Severity / Blockers

• OpenAI Auth Header Fix (#79358): Immediate attention is needed for the Docker QuickStart flow where the Authorization header is missing, rendering the installation unusable for OpenAI users.
• OpenRouter Prefix Stripping (#79309): This is causing widespread silent degradation to local fallbacks across multiple agents.
• Codex Runtime Routing (#79279): Fix the routing logic to prevent non-OpenAI models from being sent to the Codex app-server.

Blocked or High-Priority Fixes

• Subagent Process Lifecycle (#79355): Address the issue where runTurn hangs forever if a child process exits before a terminal event is emitted.
• Event Loop Optimization (#78851): Implement HTTP connection pooling for provider connections to eliminate the 7-8s model-resolution overhead per run.
• Matrix Case Sensitivity (#75670): Resolve the case-normalization bug in Matrix event IDs that is causing duplicate stuck sessions and delivery failures.

Contributor Attention Needed

• iMessage Catchup (#78649): Implement a bounded, cursor-persisted catchup for inbound messages received while the gateway was offline.
• TUI State Labels (#56856): Improve the TUI to distinguish between Idle, Generating, Working, and Waiting states to reduce operator confusion.