The meeting ends and the commitments show up in the thread

A finished meeting's commitments get posted into the thread they were promised in, and the overdue ones get one quiet nudge a day.

On-prem deployments can now turn on action item extraction. It is off until you do.

When a meeting finishes, its transcript gets handed to an extraction service that pulls out commitments, decisions, open questions and blockers. The commitments come back as a follow-up message in the meeting's own thread. If you leave the feature off you still get the transcript and the summary, you just get no action items.

Once a day, the overdue and nearly due commitments get a short message in the thread each one was promised in. Not a digest, not a separate view, and not a direct message. That choice also settles who can see what, because a message in a session is visible to exactly the people who could already see that session, and there is no second access model to configure.

The restraint is the part we spent the most time on. One message per thread, listing that meeting's outstanding items, so three overdue commitments from one meeting is one message with three lines. Once per commitment per deadline, so something that stays overdue is mentioned once rather than every morning. If the deadline moves, that is new information and it earns a new mention. A deadline nobody could read is logged instead of posted, because resolving by end of Friday to a date is reasonable and guessing at when the audit clears is not. Nothing is ever posted for a dictation.

Extraction runs against the inference host you point it at, over the same chat completions path ollama, llama.cpp and vLLM all expose. The full text of every finished meeting goes to that service, so its address is set in the deployment config and deliberately not in the list of servers an org admin can edit.

The goal is a meeting that leaves something behind in the place the work is already happening.