Problem
During curation, the agent makes editorial selections (which articles to include, what role each gets) silently in its reasoning. The user has no visibility into why articles were picked or dropped, and no chance to adjust the mix before 14 parallel subagents are spawned for summarization.
Context
The curation guide describes a structured scoring process (topic matching, source priority, diversification, feedback adjustment), but none of this is surfaced to the user. If the user disagrees with the selection, they only find out after summaries are already written — wasting time and tokens.
Suggested improvement
Add a step between article scoring and summarization that presents the user with:
- A ranked shortlist showing each selected article, its role (lead/lg/md/brief), and a one-line reason for inclusion
- A brief note on what was dropped and why (e.g., "Dropped SpaceX article — Bloomberg paywall, no content")
- An opportunity to swap articles or adjust roles before proceeding
Origin
- Date: 2026-06-05
- Working directory:
/home/sondre/openpaper
- Trigger: Reflection after first paper generation — no user checkpoint between scoring and parallel summarization
Problem
During curation, the agent makes editorial selections (which articles to include, what role each gets) silently in its reasoning. The user has no visibility into why articles were picked or dropped, and no chance to adjust the mix before 14 parallel subagents are spawned for summarization.
Context
The curation guide describes a structured scoring process (topic matching, source priority, diversification, feedback adjustment), but none of this is surfaced to the user. If the user disagrees with the selection, they only find out after summaries are already written — wasting time and tokens.
Suggested improvement
Add a step between article scoring and summarization that presents the user with:
Origin
/home/sondre/openpaper