What you're trying to do
When multiple agents (or one agent across sessions) propose knowledge to the same KB, we often get near-duplicate claim proposals — the same fact worded slightly differently. As a reviewer, I want to see that overlap at propose time, not only when I run vouch pending and read every row by hand.
Example: Agent A proposes "Auth uses JWTs in the Authorization header." Agent B proposes "Authentication uses JWT tokens in the Authorization header." Both land in proposed/. I only notice they're duplicates when I'm reviewing the queue, and I reject one with "duplicate of prop-XYZ" — after both proposals already consumed review attention.
I want the agent (and me) to get an immediate, non-blocking heads-up: "this looks like claim auth-uses-jwt or pending proposal 20260531-…" — so we can skip filing, merge rationale, or reject faster — without weakening the review gate.
What you've tried
- vouch pending + vouch show — works, but only after duplicates are already queued; no similarity signal.
- vouch dedup / kb.dedup_scan — finds near-duplicates via embeddings, but it's a separate audit step after artifacts exist; not wired into propose_claim.
- Manual workflow in multi-agent.md — "approve one, reject the other with reason duplicate of prop-XYZ"; accurate but entirely human-driven.
- Ingest-time check_and_log — logs to embedding_dupes on write; still post-hoc relative to the propose decision.
Suggested shape
Additive warnings on claim propose (v1 scope: claims only).
What you're trying to do
When multiple agents (or one agent across sessions) propose knowledge to the same KB, we often get near-duplicate claim proposals — the same fact worded slightly differently. As a reviewer, I want to see that overlap at propose time, not only when I run vouch pending and read every row by hand.
Example: Agent A proposes "Auth uses JWTs in the Authorization header." Agent B proposes "Authentication uses JWT tokens in the Authorization header." Both land in proposed/. I only notice they're duplicates when I'm reviewing the queue, and I reject one with "duplicate of prop-XYZ" — after both proposals already consumed review attention.
I want the agent (and me) to get an immediate, non-blocking heads-up: "this looks like claim auth-uses-jwt or pending proposal 20260531-…" — so we can skip filing, merge rationale, or reject faster — without weakening the review gate.
What you've tried
Suggested shape
Additive warnings on claim propose (v1 scope: claims only).