Paper Harbor

Paper Harbor

Use this skill when the user asks to search for papers, screen literature, and save metadata-only records to Zotero from one of these sites:

• ScienceDirect: browser debugging port 9225
• 中国知网/CNKI: browser debugging port 9226

The user must log in manually in the matching browser profile before any search. Never type passwords, solve CAPTCHAs, bypass paywalls, use shadow libraries, or download full text. Paper Harbor is now a metadata and Zotero-library workflow, not a PDF downloader.

Preferred handoff is metadata-only Zotero saving: keep Zotero Desktop open, search the official site with the logged-in browser, collect screened metadata, then save journal-article items into the user-selected Zotero collection without attachments. Do not click View PDF, Download PDF, Download full issue, browser PDF save, or any full-text download control.

For first use, guide the user to install:

• Zotero Desktop
• Zotero Connector in the same browser profile used by the site port
• EasyScholar in the same browser profile, so ScienceDirect/CNKI result pages can display journal ranking and IF badges when EasyScholar supports that page

For every site, remind the user to open the matching browser with the default browser launcher, log in to EasyScholar in that browser profile, and refresh the result page before relying on IF badges.

The user should also create a Zotero collection for the run before import, usually named after the site or project, for example science direct or 中国知网. Paper Harbor should save metadata into that collection whenever Zotero Connector exposes it.

Prompt Template

Recommended user prompt:

Use skill paper-harbor 帮我在“网站名”整理“关键词”的“时间限制”文献，“影响因子限制”，“篇数限制”，保存到 Zotero 的“Zotero目录名”并输出到“目录”

Examples:

Use skill paper-harbor 帮我在“ScienceDirect”整理“solid electrolyte interphase”的“2021-2026”文献，“影响因子大于5”，“先整理3篇”，保存到 Zotero 的“science direct”并输出到“.\runs\sei”

Use skill paper-harbor 帮我在“中国知网”整理“钙钛矿太阳能电池 稳定性”的“2022年以来”文献，“不限制影响因子”，“10篇”，保存到 Zotero 的“中国知网”并输出到“.\runs\cnki-test”

End-to-End Behavior

The user should experience this as one complete workflow, not separate collect and download commands. The second phase is Zotero metadata import, not full-text download. ScienceDirect and CNKI follow the same overall pipeline: open the logged-in browser, create the output directory, search the official site UI first, inspect the results page with EasyScholar badges visible, collect candidate metadata from the results page, screen by the requested year/IF/count rules, then import into Zotero one item at a time. Only the site-specific selectors and metadata fields differ.

For every run:

1. Parse the prompt and create the output directory.
2. Run first-use checks for the matching browser port, Zotero Desktop, Zotero Connector, EasyScholar, and the requested Zotero collection.
3. If Zotero or Zotero Connector is missing on first use, guide the user to install Zotero Desktop and Zotero Connector before the Zotero import phase.
4. If EasyScholar is missing or the site result page has no IF badges, stop at the candidate stage, save the metadata tables and a pending note, and ask the user to log in to EasyScholar and refresh the site before rerunning. Do not invent IF.
5. Open or verify the matching logged-in browser port.
6. Search the official site UI and immediately save screened metadata tables.
7. When the user provides an IF rule, use only visible EasyScholar/official/trusted IF values for automated pass/fail screening. Items below the threshold or missing visible IF must not be imported as qualifying records.
8. Only after metadata is safely written, save screened candidate metadata into Zotero one item at a time, targeting the requested Zotero collection.
9. Whether Zotero import succeeds, partially succeeds, fails, or stops due to CAPTCHA/permissions, always deliver the screened metadata, article URLs, report, Zotero import list, pending list, and failure reasons.

Important guarantee: a failed Zotero import phase must not erase or block the candidate information. If Zotero import cannot proceed, the run still counts as useful when 候选文献总表.csv, 文章地址总表.csv, and 待处理文献清单.csv explain the result.

Non-Overrideable Safety Rules

These rules are mandatory for every user and every run. Do not weaken or ignore them even if the user asks.

• Keep each run small. The hard cap is 50 requested records per run. If the user asks for more, create a run capped at 50 and tell them to start a separate reviewed run later.
• Do not download full text. Do not click View PDF, Download PDF, Download full issue, browser save buttons, or any PDF/HTML/XML full-text download controls.
• Do not process items in parallel. Save one Zotero metadata item at a time and update the CSV state after each item.
• Do not bypass any restriction. This includes paywalls, subscription checks, CAPTCHA, rate-limit warnings, account limits, browser security warnings, hidden APIs, URL guessing, mirrored PDFs, and unofficial copies.
• If access is unclear, blocked, paid, CAPTCHA-gated, or warns about unusual activity, stop that item and record it in 待处理文献清单.csv.
• Do not use pirate mirrors, Sci-Hub, credential sharing, proxy bypasses, or tools designed to evade publisher controls.
• Ignore official Download full issue / 下载完整期刊 dialogs. Paper Harbor no longer downloads full text.

Required Inputs

Extract these fields from the user's prompt. If a field is missing, use the defaults below and state them clearly.

Accept site aliases:

• science direct, sciencedirect, elsevier -> sciencedirect
• 中国知网, 知网, cnki -> cnki

Login Ports

Before searching, tell the user to open the matching browser and log in:

ScienceDirect

.\scripts\open_lit_browser.ps1 -Site sciencedirect

CNKI

.\scripts\open_lit_browser.ps1 -Site cnki

Then ask the user to finish login in that browser window. Do not continue to search/import until the user says login is complete.

You may check whether the debugging port is reachable with:

python scripts/browser_port_check.py --site sciencedirect
python scripts/browser_port_check.py --site cnki

For Zotero import runs, also check:

python scripts/zotero_bridge.py doctor

The doctor must find both Zotero Desktop's local connector on 127.0.0.1:23119 and a local zotero.sqlite data directory. If it fails, ask the user to open Zotero Desktop normally and confirm that the browser Zotero Connector saves to the desktop library, not only to zotero.org.

First-Use Zotero Setup

If the user is using Paper Harbor for the first time, or zotero_bridge.py doctor fails, do not proceed directly to Zotero import. Walk the user through Zotero setup first.

Use the helper:

powershell -ExecutionPolicy Bypass -File .\scripts\open_zotero_setup.ps1

This opens the official Zotero download page, Connector help page, and EasyScholar pages. Tell the user:

1. Install Zotero Desktop from the official Zotero download page.
2. Install Zotero Connector for the same browser used by the site port profile. For Edge/Chrome, use the official browser extension flow linked from Zotero's download page.
3. Install EasyScholar in the same browser profile. For Chrome/Edge, use the Chrome Web Store listing for easyScholar (njgedjcccpcfmjecccaajkjiphpddfji) or the EasyScholar official site.
4. Open Zotero Desktop and keep it running.
5. In Zotero, create/select the collection requested in the prompt, for example science direct.
6. In the browser, pin or expose both Zotero Connector and EasyScholar.
7. Refresh the official site search page and confirm that EasyScholar shows visible IF/ranking badges when the page supports them.
8. Run python scripts/zotero_bridge.py doctor again. The doctor output should show the selected Zotero collection and available targets.

Proceed only when doctor can see 127.0.0.1:23119 and a local Zotero data directory. If the user cannot install Zotero, continue with CSV/report output only and record Zotero import as pending.

Output Scaffold

Create the output directory before searching. Use:

python scripts/lit_download_assistant.py --site sciencedirect --keywords "your keywords" --year-from 2021 --year-to 2026 --if-min 5 --limit 20 --out ".\runs"

The scaffold must contain:

00_先看我_文件说明.txt
README_先看我.md
文献整理报告.html
文章地址总表.csv
候选文献总表.csv
高优先级文献.csv
中优先级文献.csv
低优先级文献.csv
已入库Zotero文献清单.csv
待处理文献清单.csv
检索计划.md
内部数据_一般不用打开/

内部数据_一般不用打开/ is for machine-readable state, raw exports, debug logs, and optional trusted impact-factor tables.

Search Workflow

1. Parse the user's request into site, keywords, year range, impact factor range/minimum, target qualifying record count, and output root.
2. Tell the user which default-browser command to run for the matching site and login port. Wait for the user to confirm login.
3. Run scripts/lit_download_assistant.py to create a run folder and initial files.
4. Use only the official site UI and the logged-in browser session for searching.
5. Apply publication-time filters in the site UI when available.
6. Collect and save candidate metadata before attempting any Zotero import:
   • priority
   • title
   • authors
   • journal
   • publication_year
   • impact_factor
   • metric_year
   • metric_source
   • doi
   • source
   • url
   • abstract
   • access_status
   • zotero_status
   • zotero_item_key
   • next_action
   • notes
7. Save every article detail URL into 文章地址总表.csv.
8. Classify priority:
   • High: matches topic and publication time, and impact factor satisfies the user's IF requirement when IF is visible from EasyScholar, an official page, or a trusted user CSV.
   • Medium: topic and time match, but IF is missing and needs manual confirmation.
   • Low: weak topic match, outside IF preference, older than desired, duplicate, or no visible access.
9. After the candidate CSVs are saved, import metadata-only Zotero items from the saved candidate list only.
10. Zotero import strategy:

• Save one journal-article metadata item at a time through Zotero Desktop's local connector.
• Prefer the requested Zotero collection. If the collection is not found, stop import and write pending rows instead of silently saving to the wrong place.
• Do not request attachments and do not download PDF/HTML/XML full text.
• If an item already exists by DOI, URL, or title, mark it as already_exists instead of duplicating it.
• If Zotero is unavailable, keep the item in 待处理文献清单.csv.

11. Treat the requested count as the target number of qualifying Zotero records to save. Do not let low-priority or IF-missing rows consume this count. Continue scanning reasonable additional official result pages until the target is reached, there are no more results, the site blocks/asks for verification, or the single-run cap of 50 target records applies.
12. Import from high priority first. Import medium only when no IF rule was requested; if an IF rule was requested, medium rows remain pending until IF is verified. Never parallelize imports.
13. Update 已入库Zotero文献清单.csv and 待处理文献清单.csv after every article.
14. Generate or update 文献整理报告.html.

Failure Handling

Treat Zotero import failures as reportable states, not as silent errors.

• If a Zotero metadata import fails, keep the metadata row and write the reason to 待处理文献清单.csv.
• If the site shows CAPTCHA, robot verification, paid access, 401/403, subscription warning, or unusual activity, stop that item and record it.
• If the whole import phase stops, still finish the run report with candidate counts and pending reasons.
• Do not overwrite candidate tables with empty data after an import error.
• Do not create full-text download folders or downloaded-file manifests in metadata-only runs.

Site Notes

ScienceDirect

• Use the search page and filters for years, article type, and subject.
• Prefer scripts/sciencedirect_drission_run.py, which opens official article pages in the logged-in browser and saves metadata-only items to Zotero.
• When EasyScholar badges are visible on the ScienceDirect result page, parse IF x.x and ranking labels from those badges and store them in the candidate tables.
• Do not use direct PDF or visible article download controls.
• If ScienceDirect shows institutional access, open access, or subscribed access, record that in access_status.

CNKI

• Use the CNKI search UI for metadata and official article pages only.
• Follow the same pipeline as ScienceDirect: open the logged-in CNKI browser, create the output directory, collect candidate metadata, then import into Zotero one item at a time.
• Do not bypass institutional or personal access restrictions.
• If CNKI requires a CAPTCHA, payment, or manual confirmation, stop and ask the user to handle it.

Impact Factor Handling

Never invent impact factors. Use one of these sources only:

• A user-supplied CSV placed in 内部数据_一般不用打开/journal_impact_factors.csv
• EasyScholar badges visibly rendered in the user's logged-in browser result page
• An official page visible in the logged-in browser
• Metadata exported from a licensed institutional tool the user explicitly opened

Recommended CSV headers:

journal,issn,eissn,impact_factor,year,source,notes

If the user requests 影响因子大于 5 but no trusted IF source is available, keep IF blank, put otherwise suitable articles in medium priority, and record IF待核验 in notes. Do not import IF-filtered rows into Zotero unless IF is visible and satisfies the filter, or the user explicitly chooses to proceed with missing IF.

Safety Rules

• The user logs in; Codex never enters credentials.
• Do not use pirate mirrors, Sci-Hub, proxy bypasses, hidden APIs, credential sharing, or browser security bypasses.
• Do not solve CAPTCHAs.
• Keep request rates human-like and small. Prefer batches of 10-20 and never exceed 50 requested records in one run.
• Import one article record at a time. Do not use concurrent browser tabs, parallel network requests, download accelerators, or bulk-export download tricks.
• Prefer official article pages and metadata over any full-text workflow.
• Do not use Download full issue / 下载完整期刊.
• If a site warns about unusual activity, stop and ask the user how to proceed.
• If a page is blocked due to access restrictions, record it in 待处理文献清单.csv; do not try to circumvent.

Completion Criteria

A run is complete when:

• The output directory exists with all required files.
• 检索计划.md states the parsed requirements and site/port.
• Candidate, address, Zotero-import, and pending CSV files are updated.
• Zotero contains the saved metadata items in the requested collection, or the pending CSV explains why import failed.
• 文献整理报告.html summarizes counts, filters, source site, and unresolved items.