openscite

OpenScite

Analyze papers citing one user-provided scholarly PDF. The uploaded PDF is the target, not a source of references to follow. Preserve user filenames and write the report in the user's language while keeping titles and quoted passages in their source language. Use keyless OpenAlex.

Use the resumable runner

Resolve this skill's directory as <skill-dir>. Run commands from the user's workspace so default artifacts are created there. Do not generate a substitute pipeline script.

python "<skill-dir>/scripts/openscite.py" prepare "<target.pdf>" --language <zh-TW|en> [--n N] [--run-dir <dir>]

Make prepare the first PDF operation. It runs the latest local Anydoc release, caches the output, matches arbitrary inbox filenames by DOI/title, and falls back only when its artifacts record an extraction failure. Add --require-page-aware only when exact page locators matter; this intentionally uses the page-aware fallback. Use --rule-triage only for diagnostics.

Reuse the same target, run directory, N, mode, and language on every retry. Follow the returned status:

• needs_user_selection: create target-claims.json first if needed, then read references/user-interaction.md, ask only for shortlist size, and rerun with --n N. Suggest 20; reuse an existing positive N without asking.
• needs_target_claims: read target-analysis-packet.json, extract discrete claims from the target paper, write a JSON array to target-claims.json, and rerun. Each item must include claim_id, claim, and source; include population, intervention/exposure, outcome, direction, and page when supported. Do not ask the user for this internal analysis.
• needs_abstract_triage: read references/ranking.md, load triage-context.json once, and analyze only triage-pending.jsonl. When sub-agents are available, dispatch up to four disjoint batches in one parallel wave. Append results to triage-results.jsonl, preserving each ID/hash, then rerun prepare.
• needs_analysis: read references/labeling.md, load analysis-context.json once, and analyze only analysis-pending.jsonl. Keep each paper group on one worker; when sub-agents are available, dispatch up to four paper-group batches in one parallel wave. Append one result per statement to analysis-results.jsonl, preserving IDs/hashes and setting label_source: model, then run:

python "<skill-dir>/scripts/openscite.py" finalize --run-dir "<run-dir>"

• needs_user_files: read references/user-interaction.md. When an interactive browser capability is available, ask once whether to use the user's current browser session (including institutional/VPN access) to download the authorized PDFs listed in fulltext-requests.md. On opt-in, use the bounded parallel browser procedure; otherwise present the links and accept attachments or files in fulltext/inbox/. Keep original filenames and rerun the same prepare command after files arrive.
• partial or complete: deliver report.md and briefly state coverage. partial is a valid evidence report when unavailable full text or unbound contexts are explicitly unknown.

Only pending rows are retried; valid result rows remain reusable. If finalize returns needs_analysis, process the regenerated pending file and rerun it.

If fulltext-manifest.json contains an ambiguous match, read references/user-interaction.md and ask one mapping question. Store the confirmed mapping in file-mappings.json:

{"mappings":[{"file_name":"original-name.pdf","citing_work_id":"W123"}]}

Then rerun prepare. Sub-agents return records to the parent and do not ask users or edit shared artifacts.

Evidence and access rules

Read references/providers.md for provider/OA issues, references/fulltext-processing.md for parsing or matching failures, and references/output-schema.md only when extending artifacts.

Use user-provided files and legitimate OA URLs. Validate PDF magic; obtain explicit consent before hosted OCR. Keep diagnostics and ranking internals out of report.md.

Assign stance per target-bound citation passage. Bibliography-only/topic-only hits and missing evidence stay unknown; rule output never becomes a model label.

Completion check

Require run.json, target.json, selection.json, citing-works.json, citations.json, fulltext-requests.md, and report.md. Every non-unknown statement needs exact context, immutable hash, confidence, reason, and label_source: model|human. Report unique papers first and passages second, in the user's language and without debugging metadata.