Problem

After context compaction, the assistant's outbound message dedup guard resets because it only tracks duplicates within a single assistantMessageIndex. When the model re-generates text from the compaction summary as new output, it gets delivered to Telegram as a duplicate message.

Reproduction

1. Have a long conversation that triggers compaction (compaction-safeguard logs: new content uses X% of context; dropped N older chunk(s))
2. The compaction summary includes the assistant's previous reply text verbatim
3. After compaction, the model sometimes re-outputs that same text as its new response
4. OpenClaw delivers it as a new Telegram message — user sees the same content twice

Root Cause

In src/agents/pi-embedded-subscribe.ts, shouldSkipAssistantText() only dedupes within the same assistantMessageIndex:

const shouldSkipAssistantText = (text: string) => {
    if (state.lastAssistantTextMessageIndex !== state.assistantMessageIndex) {
      return false;  // <-- resets on new turn, no cross-turn dedup
    }
    // ...
};

The messagingToolSentTexts dedup in filterMessagingToolDuplicates only covers messages sent via the message tool, not the main assistant reply path.

Suggested Fix

Add a small rolling hash cache (e.g., last 20 delivered text hashes, 1h TTL) that persists across turns within the same session. Before delivering an assistant reply chunk, check if a normalized hash of the first ~200 chars matches any recently delivered hash. If so, skip delivery.

This could live in rememberAssistantText / shouldSkipAssistantText as an additional cross-turn check, or as a lightweight filter in the Telegram send path.

Impact

• Affects all channels but most visible on Telegram (each chunk = separate message)
• Users report receiving the same reply twice after long conversations
• Workaround: system prompt instruction to not re-output summary text (behavioral, unreliable)

Environment

• OpenClaw 2026.3.3
• Compaction mode: safeguard
• Channel: Telegram (direct chat)