Problem

When an LLM request fails (e.g., Anthropic API rejects due to corrupted thinking blocks in session history), the raw error message is sent directly to the user via the messaging channel (WhatsApp, Telegram, etc.) as an auto-reply.

Example: After an OpenClaw update from 2026.2.17 → 2026.3.7, a session with thinking blocks became corrupted. Every subsequent message to the KIMS agent returned:

LLM request rejected: messages.5.content.12: thinking or redacted_thinking blocks in the latest assistant message cannot be modified.

This was sent verbatim to the user on WhatsApp, who had no idea what it meant.

Expected Behavior

1. Never send raw LLM/API errors as replies to users. The error should be caught internally.
2. Graceful fallback: On LLM failure, the agent should either:
   • Retry once silently
   • Send a user-friendly message ("I'm having a technical issue, please try again shortly")
   • Go silent (NO_REPLY) rather than forward the error
3. Auto-escalation (multi-agent setups): If an agent fails 2+ times consecutively, automatically notify the main/default agent so it can investigate and inform the user.
4. Session corruption detection: After updates, detect and flag sessions with incompatible thinking blocks before they cause repeated failures.

Current Workaround

Added ../shared/error-escalation.md protocol to all persona AGENTS.md files instructing agents to catch errors and escalate. But this only works if the agent can actually process the instruction — in the case above, the LLM call never succeeds, so the agent-level instructions are never read.

This needs to be handled at the framework level, before the LLM response reaches the messaging channel.

Environment

• OpenClaw 2026.3.7
• Multi-agent setup (9 personas)
• WhatsApp + Telegram channels
• Anthropic Claude as primary model