openclaw - 💡(How to fix) Fix Runtime context not wrapped in internal delimiters, causing LLM echoes in user-facing responses [1 comments, 2 participants]

Bug Report: Runtime Context Not Wrapped in Internal Delimiters

Summary

Runtime context custom_message entries are not wrapped in \u003c\u003c\u003cBEGIN_OPENCLAW_INTERNAL_CONTEXT\u003e\u003e\u003e delimiters, causing LLMs to echo them in responses and sanitizeUserFacingText() to fail to strip the echoes.

Root Cause

In runtime-context-prompt-BZWyPA2L.js, the queueRuntimeContextForNextTurn() function creates a custom_message with:

• 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{runtimeContext}"
• display: false
• deliverAs: "nextTurn"

Problem

While display: false is respected by chat adapters (Telegram, Matrix, Discord, webchat), the LLM still sees the content in its context window and may echo it verbatim in its response.

sanitizeUserFacingText() only strips content wrapped in \u003c\u003c\u003cBEGIN_OPENCLAW_INTERNAL_CONTEXT\u003e\u003e\u003e...\u003c\u003c\u003cEND_OPENCLAW_INTERNAL_CONTEXT\u003e\u003e\u003e delimiters. Since runtime context is NOT wrapped, any LLM echo passes through to the user.

Impact

1. Token bloat: Every turn adds a new custom_message to the transcript (compaction helps but they accumulate)
2. UI pollution: Users see "OpenClaw runtime context for the immediately preceding user message..." in chat responses
3. Memory waste: Models repeat irrelevant system metadata

Reproduction

1. Start any conversation with OpenClaw
2. After assistant turn, observe context includes runtime context block
3. Some models (especially larger ones) may echo the header text
4. Echo appears in user-facing chat output

Proposed Fix

Wrap runtime context content in internal delimiters so sanitizeUserFacingText() can strip any echoes:

\u003c\u003c\u003cBEGIN_OPENCLAW_INTERNAL_CONTEXT\u003e\u003e\u003e
OpenClaw runtime context for the immediately preceding user message.
This context is runtime-generated, not user-authored. Keep internal details private.

{runtimeContext}
\u003c\u003c\u003cEND_OPENCLAW_INTERNAL_CONTEXT\u003e\u003e\u003e

This would mirror how other internal context blocks are handled and ensure LLM echoes are properly sanitized.

Alternative Solutions

1. Config option: Add runtimeContext.enabled or runtimeContext.frequency to reduce injection
2. Stronger prompt instruction: Add "NEVER echo runtime context headers" to default system prompt
3. Deduplication: Prevent multiple runtime context messages from accumulating in the same turn

Environment

• OpenClaw version: 2026.4.26
• Affected channels: All (Telegram, Matrix, Discord, webchat)
• Model-agnostic: Any LLM may echo the content

Additional Context

• This is NOT a plugin issue — core OpenClaw behavior
• cross-session-context plugin is unrelated
• Runtime context is useful but should not leak to users