What Problem This Solves

truncateReplyBody in extensions/matrix/src/matrix/monitor/reply-context.ts (reached via the exported summarizeMatrixReplyEvent) shortens long reply bodies with a raw value.slice(0, MAX_REPLY_BODY_LENGTH - 3). UTF-16 .slice cuts at a code-unit index, not a code-point boundary. When the cut index lands between the two halves of a surrogate pair (any emoji or other astral character), the slice keeps a lone high surrogate, which is appended directly before the ... ellipsis. The reply context surfaced to the agent then contains a dangling \uD83D that renders as a broken replacement glyph.

Repro: a plain m.room.message whose content.body is 496 a characters, then the emoji 😀 (U+1F600, a surrogate pair at UTF-16 indices 496–497), then bcd (total length 501):

summarizeMatrixReplyEvent({
  type: "m.room.message",
  content: { msgtype: "m.text", body: "a".repeat(496) + "😀" + "bcd" },
});

• Before (wrong): value.slice(0, 497) splits the emoji, returning "a".repeat(496) + "\uD83D" + "..." — a lone high surrogate (broken glyph) before the ellipsis.
• After (fixed): truncation falls back to the previous code-point boundary: "a".repeat(496) + "...". The half emoji is dropped; no orphaned surrogate.

The fix swaps the raw slice for sliceUtf16Safe (already exported from openclaw/plugin-sdk/text-utility-runtime and used by sibling extensions such as discord and memory-lancedb), which never cuts inside a surrogate pair. Bodies without astral characters are unaffected — a 600-char ASCII body still truncates to "x".repeat(497) + "..." exactly as before.

Evidence

Standalone Node red/green proof. It copies sliceUtf16Safe verbatim from src/utils.ts and runs both the buggy (.slice) and fixed (sliceUtf16Safe) versions of truncateReplyBody against the repro input plus an unaffected normal body:

PASS: repro input length is 501
PASS: emoji high surrogate at index 496
PASS: emoji low surrogate at index 497
PASS: BEFORE leaves a lone high surrogate (the bug)
PASS: BEFORE differs from expected (proves bug)
PASS: BEFORE is 'a'*496 + \uD83D + '...'
PASS: AFTER equals expected (a*496 + ...)
PASS: AFTER has no orphaned high surrogate
PASS: AFTER length <= 500
PASS: AFTER ends with ...
PASS: normal long ASCII body: BEFORE === AFTER (no behavior change)
PASS: normal long ASCII body truncates to x*497 + ...
PASS: short body unchanged BEFORE
PASS: short body unchanged AFTER

ALL CHECKS PASSED — red/green proof confirmed.

A regression test is added to reply-context.test.ts (truncates on a code-point boundary without orphaning a surrogate half) that fails on the raw-slice code and passes with the fix.