The Lark short-window debounce (MUL-2968, #3742) can land several user
messages in a chat session before a single agent run fires. But the
daemon claim built the agent prompt from only the *single most recent*
user message (walk history backward, take first user message, break).
So 「看上海天气」then「还有青岛」debounced into one run, and the agent
received only 「还有青岛」— it answered Qingdao and never saw Shanghai.
The session itself was correct (both messages persisted); the gap was in
what the run delivered to the agent. Before debouncing this was masked
because each message got its own run.
Build the prompt from the whole unanswered set instead: the trailing run
of user messages after the last assistant reply (every completed/failed
run writes an assistant row, so the anchor advances one turn at a time —
the full burst on the first turn, only the new message(s) after a reply).
Attachments are collected from each included message. Extracted the
selection into a pure trailingUserMessages helper with table-driven unit
tests, plus a DB-backed claim test asserting both messages reach the
agent and that a post-reply message delivers alone.
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
