← Back to Blogs
GH Issues

OpenClaw Issue Digest: Routing Regressions, Codex Migration Risks, and Gateway Stability

12:30–18:30 UTC May 8, 2026

OpenClaw Issue Digest: Routing Regressions, Codex Migration Risks, and Gateway Stability

Open Issues

Recent activity in the OpenClaw repository reveals a series of critical regressions affecting channel routing, configuration stability during updates, and hardware-specific performance. The most pressing issues involve the breakdown of message delivery in Telegram and Google Chat, as well as a dangerous configuration rewrite bug affecting users of the Codex runtime.

Channel Routing and Delivery Failures

Several reports indicate that outbound message delivery is failing or misrouting across multiple platforms:

  • Telegram DM Topics: A critical API shift in Telegram now requires direct_messages_topic_id for private DM topics. OpenClaw currently uses message_thread_id, leading to 400 Bad Request errors. The system silently falls back to the main DM view, masking the routing failure from operators (#79455, #79448).
  • Telegram Forum Topics: Similarly, forum-topic replies are vanishing entirely because the current delivery path does not support threadless retries for forum scopes, unlike other paths in the repository (#79408).
  • Google Chat Groups: The Google Chat plugin is failing to create proper group session keys, collapsing all group messages into the main session. This causes replies to be routed to the wrong channel entirely—in one case, messages intended for Google Chat were delivered via WhatsApp (#79395).
  • Discord Delivery: Some users report 401 Unauthorized errors when sending messages via the Discord bot, despite the bot being connected and able to read messages (#79445).

Codex Runtime and Config Migration Risks

There is a significant cluster of issues surrounding the openclaw doctor --fix command and its interaction with the Codex runtime:

  • Unsafe Model Rewrites: The doctor tool is aggressively rewriting openai-codex/* model references to openai/*. For users relying solely on Codex OAuth (ChatGPT subscriptions) without a direct OpenAI API key, this locks them out of their models entirely (#78407, #79461, #79306).
  • Billing Misroutes: These rewrites can silently switch a user from a subscription-based billing path to an API-key-billed path, leading to unexpected costs (#78519).
  • Runtime State Mismatches: Users are observing "mixed states" where sessions show a mismatch between the configured model and the active runtime, leading to hangs and idle timeouts in Telegram sessions (#79413).

Gateway Stability and Performance

Performance regressions are hitting specific hardware and environments:

  • ARM64/Raspberry Pi CPU Spikes: A severe regression since version 2026.4.25 causes the gateway to pin CPU at 100%+ on Raspberry Pi 4 (ARM64), effectively stalling the event loop and making the bot unresponsive (#79380). Similar latency issues are reported on Raspberry Pi for Slack gateways (#78704).
  • Memory and Log Bloat: Reports indicate the gateway process can peg CPU due to hot async loops in CopyDataPropertiesWithExcludedProperties, while gateway.err.log can grow to nearly 20GB due to tight error loops in dependency staging (#79422).
  • Windows CLI Hangs: The openclaw message send command is hanging indefinitely on Windows, leaking zombie node.exe processes (#79436).

Key Themes

1. Silent Failures vs. Loud Errors

Across multiple issues (#79448, #79421, #79475), a recurring theme is the danger of silent failures. Whether it is the Telegram fallback masking a routing error or the plugin loader silently denying conversation hooks, the lack of operational visibility is hindering debugging and causing lost work.

2. The "Doctor" Tool as a Vector for Instability

While intended to repair configurations, the openclaw doctor --fix command has become a source of regressions. The aggressive normalization of model IDs without verifying the user's authentication profile has created a critical failure path for OAuth users.

3. Resource Contention in Tool Loops

Several reports (#79406, #79350) highlight blocking operations during tool calls. The lack of pre-loading for plugin skills and the absence of result caching in tool policies are introducing multi-second delays per call, which can cascade into session timeouts.

Action Required

High Severity / Immediate Attention

  • Fix Codex Model Migration: The doctor --fix logic must be updated to prevent the unconditional rewrite of openai-codex/* to openai/* unless a valid API key is present. This is a critical blocker for OAuth users.
  • Correct Telegram API Fields: Update the Telegram outbound parameter builder to use direct_messages_topic_id for DM topics to restore delivery to threaded private chats.
  • Resolve ARM64 CPU Spin: Investigate the regression in 2026.4.25+ that causes 100% CPU usage on Raspberry Pi 4, as this renders the software unusable on popular low-power hardware.

Blocked or High-Risk Issues

  • Google Chat Session Keys: Fix the session key generation in the Google Chat plugin to prevent cross-channel message leakage (#79395).
  • At-Cron Silent Drops: Address the silent failure of one-shot at crons targeting session labels, which currently leaves no on-disk evidence of execution (#79475).
  • MCP Tool Denial: Extend tools.deny enforcement to the claude-cli backend MCP config generation to ensure security boundaries are respected (#79451).

References

Issues