vellum-ai Threadkeeper review
Use the threadkeeper tool with list_open_loops and list_reviews to gather a bounded review set. Present proposal state and accepted state distinctly.
Review order
- Active or blocked open loops, especially due or aging items.
- Gardener proposals and contradictions.
- Serendipity connections with a clear structural relationship and evidence for both endpoints.
- Pending archaeology jobs and persistent processing failures.
Every item needs provenance
For each proposal or connection, show the source conversation, source artifact, date, and evidence label when available. Explain the pattern in plain language. Distinguish directly stated, inferred, conflicted, stale, and weak-hypothesis evidence.
Actions
- Accept a proposal only after the user confirms it; use
accept_proposal. - Reject or dismiss with the user's reason when provided; use
reject_proposal. Rejections create durable suppression and should not recur without materially new evidence. - For a connection, offer explore, save as hypothesis, create open loop, dismiss, or mark wrong. A connection never automatically becomes memory, a task, a notification, or an external action.
- Use
snooze,mark_done, orarchive_threadonly for the target the user named. - Do not expose destructive database purge through this workflow.
After an action, report the queued or accepted state honestly and point the user to the dashboard for provenance and job progress.