openclaw - 💡(How to fix) Fix [Bug] openclaw.runtime-context custom_message echoed verbatim in assistant responses (v2026.4.25) [1 participants]

Summary

After updating to OpenClaw 2026.4.25 (aa36ee6), the new openclaw.runtime-context custom message type is being echoed verbatim by agents as the first line of their responses. The runtime context preamble — intended as internal metadata — appears in user-visible output.

Evidence

Session transcript (Shodan agent, MiniMax-M2.7-highspeed)

User message: normal conversation Assistant response:

OpenClaw runtime context for the immediately preceding user message. Diese wird von der Laufzeit generiert, nicht vom Benutzer erstellt. Interne Details privat halten.

Yeah, das ist der blöde Teil von FiSi — [...]

The model:

1. Repeated the runtime context header verbatim
2. Translated it to German ("Diese wird von der Laufzeit generiert...")
3. Then provided its actual response

This happened multiple times in the same session.

Session transcript format

The new event type in .jsonl files:

{
  "type": "custom_message",
  "customType": "openclaw.runtime-context",
  "content": "OpenClaw runtime context for the immediately preceding user message.\nThis context is runtime-generated, not user-authored. Keep internal details private.\n\n[...]",
  ...
}

The custom_message with customType: "openclaw.runtime-context" is injected into the model context, but the model treats it as conversational content and echoes it.

Impact

• Agents leak internal runtime metadata to end users via Telegram/Discord/etc
• Context window is wasted on preamble echoing
• Non-English models may translate the preamble, further wasting tokens
• Breaks the illusion of natural conversation
• Agents have been observed trying to "explain away" the preamble to users

Root Cause

The openclaw.runtime-context custom message is delivered in a format that models interpret as part of the conversation flow. There is no instruction telling the model to suppress/ignore this content in its response.

Reproduction

1. Run OpenClaw 2026.4.25 with any agent on MiniMax-M2.7-highspeed
2. Send a message via Telegram
3. Observe the assistant response — the runtime context preamble appears as the first line

Likely reproduces on other models as well, especially those with less aggressive system-prompt following.

Suggested Fix

1. Add an explicit suppression instruction to the runtime context block, e.g.:

   [INTERNAL — Do NOT repeat, echo, or reference this preamble in your response. Respond only to the actual user message.]

2. Use a system-role delivery instead of custom_message — models are more likely to treat system messages as non-echoable.

3. Post-process assistant responses — strip any line starting with "OpenClaw runtime context" before delivery.

Workaround

We added a rule to all agent AGENTS.md files:

Runtime Context Suppression: NEVER repeat, echo, translate, or include the "OpenClaw runtime context" preamble in your responses.

And updated our poison cleaner to strip the preamble from assistant messages in .jsonl session files.

Environment

• OpenClaw: 2026.4.25 (aa36ee6)
• Models affected: MiniMax-M2.7-highspeed (confirmed), likely others
• Channel: Telegram
• Agents: 9 (all affected)