zhangzhuosjtu/youreadiread/paper-curate

paper-curate

Overview

Ingestion skill. Flows: /add, /track-group, /track-topic, /update (sweep). State: ~/.you-read-i-read/data/; never write to the plugin tree.

Hard rules

• Two-tier summarization. Deep (7-section) only on /add and [r]ead-now. /update uses abstracts — never fetch PDFs in a sweep.
• Confirm before persist. Show resolved sources (S2 IDs, arXiv queries) before saving any tracked group/topic.
• Watermarks don't advance for failing sources. On error, last_checked stays put; retried next sweep.
• Log every accept/reject. manage_data.py signal-log --event update_accepted|update_rejected.

Source-tier selection

Pick the highest tier available; tell the user which ran.

*Requires APIFY_TOKEN + budget.

Per-flow playbooks

/add <url>

1. Deep-summarize via paper-summarizer subagent. The subagent persists

raw fetched content via manage_data.py paper-set-raw <id> --ext ... (txt for grounding, plus pdf when arXiv). Skipped if summary.store_raw_content is false.

1. manage_data.py paper-add --url ... --title ... --authors ...
2. echo "$summary" | manage_data.py paper-set-summary <id>
3. Confirm (id, title, one-liner, tags). Log signal-log --event add_via_url.

/track-group <name>

Resolve to S2 author IDs. Confirm. manage_data.py group-add. state-set-last-checked --kind groups --id <id>.

/track-topic <topic>

Distill to per-source queries. Confirm. manage_data.py topic-add. Init watermark.

/update

Per source: (1) get watermark; (2) discover via highest tier; (3) dedupe against papers/index.json; (4) rank from signals.jsonl tail (~100) + preference-prompt.md — score accept_likelihood ∈ [0,1], blend recency + trending weights; (5) triage:

[k/N] <id> rel=0.82 like=0.74 rec=0.93 trend=0.85
<Title> · <authors> · via: <topic>
One-liner: …  Why: <rationale>
[a]dd [s]kip [r]ead-now [?]info [q]uit →

Log on a/s. ? fetches context, no persist. r → /add then /read. Advance watermark after full triage; q advances finished sources only.

Subagent dispatch

The paper-summarizer subagent always produces a deep 7-section summary AND persists raw content. Dispatch only when that's the desired output:

• `/add`: dispatch one summarizer per URL.
• `/update` `[r]ead-now`: dispatch one summarizer for the chosen candidate, then drop into /read.
• `/update` sweeps themselves: do NOT dispatch the summarizer for one-liners. Source adapters return abstracts; derive the per-card one-liner in-context from the abstract. The /update sweep is two-tier-quick by definition.

When dispatching for parallel /add (rare in v1), cap at 8 concurrent.

Common mistakes

• Deep-summarizing during /update (token-bombs sweep).
• Auto-adding candidates in interactive mode.
• Advancing a watermark for a failing source.
• Ranking from preference-state.yaml not signals.jsonl (digest lacks tag/exemplar data).
• Resolving an ambiguous group name without confirming.