Recipe ergonomics for commodity agents: hybrid RRF returns unhydrated ids; descriptors lack when-to-use guidance

[fazpu]

Two findings from watching a commodity agent (gpt-4o-mini, per protocol) use the public recipes on a live deployment:

1. claims_hybrid_rrf returns only a ranking of claim UUIDs with scores — no hydrated text. For an agent, that is unusable without a follow-up the descriptor never mentions. The strongest search primitive went completely unused across all 8 questions, plausibly for this reason. Either hydrate top-k text into the envelope or make the descriptor state "returns ids — hydrate with explain".
2. Descriptors carry no when-to-use guidance. The agent answered a when-question with one claims_verbatim call and never touched entity_timeline; answered an identity question via identity_as_of's transcript-shaped envelope and extracted nothing, while a claims search would have surfaced the answering claim. One sentence per descriptor ("for questions about when something happened, prefer entity_timeline") is cheap and directly serves the D50 tool surface.

Note the interaction with #146: until valid-time is populated, entity_timeline guidance would steer agents into an empty room — sequence accordingly.

Found during RS-LoCoMo-Full-v3 trace forensics, 2026-07-26.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GKENhTLJg1HqhbdwCmmkbc