Summary

Feishu direct messages inject metadata as two extra user-role blocks separate from the actual message text, causing 3 consecutive user messages per turn.

This breaks role alternation (Anthropic API rejects 400) and confuses OpenAI-compatible APIs. The model sees metadata blocks as user input and responds to them instead of the real message.

Actual payload (anonymized)

[
  {"role": "user", "content": "Conversation info (untrusted metadata): {sender details}"},
  {"role": "user", "content": "System: [timestamp] Feishu[main] DM | User [msg:id]", "cache_control": {"type": "ephemeral"}}
]

Key issue

Metadata blocks appear as separate user messages. When system events flush (heartbeat, cron), the model sees metadata as the freshest user input and responds with unrelated greetings instead of answering the real question.

OpenClaw version

2026.5.22 (a374c3a)

Environment

• OS: Linux (Docker, Ubuntu 24.04, x64)
• Channel: Feishu DM (websocket)
• Model: Claude-Opus-4.6/DeepSeek-V4-Pro

Root cause

1. buildInboundUserContextPrefix() in inbound-meta.ts generates metadata as separate user block
2. drainFormattedSystemEvents() in session-system-events.ts formats system events as System: lines with ephemeral cache control, injected as another user block
3. Neither is merged with the actual user message

Related issues

• Group chat context injection creates consecutive user-role messages, breaking Anthropic API compatibility #83419 - Group chat context injection (same root cause)
• Strip inbound metadata from past user turns on prompt replay #82337 - Strip inbound metadata from past turns
• Telegram inbound: openclaw.runtime-context envelope leaks into user message body #81241 - Telegram: runtime-context envelope leak
• [Bug]: Active Memory recall context uses full OpenClaw prompt envelope #88077 - Active Memory recall context envelope issue