Bug Description

System-generated messages — heartbeat prompts, pre-compaction memory flush prompts, and their corresponding NO_REPLY/HEARTBEAT_OK responses — are rendered as regular user and assistant messages in the TUI chat log.

These internal turns are injected by OpenClaw for maintenance purposes and should never appear in user-visible chat history.

Messages affected

1. Heartbeat prompts — e.g. "Read HEARTBEAT.md if it exists... reply HEARTBEAT_OK"
2. Memory flush prompts — e.g. "Pre-compaction memory flush. Store durable memories now." or custom prompts like "Extract key decisions, state changes..."
3. Silent responses — NO_REPLY, HEARTBEAT_OK, NO_FLUSH
4. Post-compaction audit messages — "Post-Compaction Audit: The following files were not read..."
5. Compaction status messages — "Compacted • Context 147k/200k (74%)"

Steps to Reproduce

1. Open the TUI (openclaw)
2. Wait for a heartbeat or memory flush cycle to run
3. Observe that the heartbeat/flush prompt appears as a "you" message and the NO_REPLY/HEARTBEAT_OK response appears as an assistant message

Expected Behavior

System-generated messages should be filtered from the TUI chat history. Only user-initiated messages and meaningful assistant responses should be displayed.

Actual Behavior

All transcript entries are displayed verbatim, including internal maintenance turns.

Environment

• OpenClaw 2026.2.21 (TUI)
• Gateway 2026.2.22
• Affects: TUI history loading and real-time event streaming

Related

• Google Chat: conversation metadata and cron announce payloads rendered as visible messages #23971 (Google Chat: similar issue with metadata leaking into chat)