Regression in 2026.4.27: WhatsApp session became unstable, leaked thinking traces, and fell back from Codex to MiniMax; rollback to 2026.4.23 fixed it

After updating OpenClaw from 2026.4.23 to 2026.4.27, my WhatsApp main session became unstable.

Environment

• OpenClaw before: 2026.4.23
• OpenClaw after update: 2026.4.27
• Channel: WhatsApp direct session
• Primary model expected: openai-codex/gpt-5.5
• Fallbacks: minimax/MiniMax-M2.7, zai/glm-4.7
• Host: Raspberry Pi / Debian arm64

Symptoms observed after 2026.4.27

• Gateway became noticeably slower / unstable
• Assistant responses became fragmented and inconsistent
• Internal reasoning/thinking traces were sent as normal WhatsApp messages
• The session appeared to fall back or get stuck on minimax/MiniMax-M2.7 despite Codex being configured as primary
• Session/model/thinking overrides seemed inconsistent after restart/update
• Setting thinkingDefault: off and forcing Codex helped partially, but the session remained unreliable
• Rolling back both installed OpenClaw paths to 2026.4.23 restored normal behavior

What fixed it

• Downgraded OpenClaw back to 2026.4.23
• Restarted gateway
• Confirmed:
  • CLI: 2026.4.23
  • Gateway runtime: 2026.4.23
  • Primary model: openai-codex/gpt-5.5
  • Thinking: off
• WhatsApp behavior returned to normal

Related issues that may overlap

• Gateway 7-second event-loop stalls on 2026.4.27 — createOpenClawTools synchronously rebuilds plugin registry + provider-auth env-var resolution per tool-creation pass #74860 — 2026.4.27 gateway event-loop stalls
• Feature: global default for reasoning visiblity #13607 — reasoning visibility/defaults
• Feature: Graceful error handling for LLM failures — never expose raw errors to users #39612 — thinking blocks/session corruption/errors leaking to messaging channels

Expected behavior

• Updating to 2026.4.27 should not change/freeze the effective model away from the configured primary
• Thinking/reasoning traces should never be delivered as normal user-visible WhatsApp messages unless explicitly enabled
• WhatsApp sessions should remain stable after update/restart
• Fallback behavior should be visible and deterministic, not silently “stick” to a fallback model

Actual behavior

After the update, the WhatsApp session became slow, unstable, leaked internal reasoning, and appeared to use MiniMax instead of the configured Codex primary.

Rollback to 2026.4.23 resolved the issue.

If useful, I can provide sanitized config/status excerpts.

Update from further testing (2026-05-06)

Further testing suggests the thinking-trace leakage only happened while MiniMax was in the active route.

After changing the WhatsApp main session back to openai-codex/gpt-5.5 and moving the subagents off MiniMax to Codex/GLM where appropriate, I have not observed further thinking leaks.

Current observation:

• Main session: openai-codex/gpt-5.5
• Subagents: moved to Codex/GLM depending on role
• MiniMax is no longer used as a primary model
• No new leaked thinking traces observed after that change

This suggests the leak may be model/provider-specific to MiniMax behavior or its OpenClaw integration path, rather than a general WhatsApp/session issue.