Hi @YangManBOBO — dropping by from #12076 (Teknium closed mine as a duplicate of this one). Your fix shape matches what I landed on, and I'm happy to defer to your version. Two edge-case observations from my tests that might be worth adding to yours if you have cycles:

1. Non-mapping metadata values. If a caller passes metadata=[] or metadata="whatever" (theoretically out-of-contract but happens when callers stringify a payload), metadata.get("thread_id") would raise AttributeError. My fix handled it with an isinstance(metadata, dict) guard before the .get(). Small defensive pin:

   if not root_post_id and isinstance(metadata, dict):
       tid = metadata.get("thread_id")
       ...

   Test: test_send_invalid_metadata_type_ignored — asserts metadata={"bad"} (set) or metadata="str" doesn't crash.

2. Multi-chunk consistency. Long messages chunk into multiple posts via truncate_message(...). All chunks should carry the same root_id so the whole reply lands in-thread as one conversation, not with chunk 1 threaded and chunks 2+ at channel root. Your diff already handles this correctly (since root_post_id is computed once outside the chunk loop), but a test pinning that invariant guards against future refactors breaking it:

   # force chunking by posting > MAX_POST_LENGTH
   await adapter.send(chat_id, "x" * (MAX_POST_LENGTH + 100), metadata={"thread_id": "rid"})
   # assert every captured post payload has root_id == "rid"

Feel free to cherry-pick / ignore — whatever you think is useful. Cheers!

Related to #6617 and #12299 — competing Mattermost threading fixes. All address #12063.