feat: AI News 周报润色 skill + weekly-review v2.3 + 多项修复

Merge PR #61: adds skills/ai-news-weekly-polish for iterative weekly
report polishing with 11-dimension audit rubric, v2 weekly_summary.py
with event clustering, weekly-review skill v2.3, and various fixes
including FreshRSS cache recovery, HuggingFace retry, bilibili upload,
and Zotero linked_file sync.

Closes #4

Author: OuyangWenyu

.claude/commands/zotero-notebooklm.md

@@ -99,3 +99,18 @@ DAILYINFO_DATA_ROOT=
 # Used when DeepSeek V4 Pro returns empty responses after retries.
 # The fallback routes through OpenRouter. Default: moonshotai/kimi-k2.5
 # DAILYINFO_FALLBACK_MODEL=
+
+# -----------------------------------------------------------------------
+# Zotero Web API (for zotero_sync.py linked_file import)
+# -----------------------------------------------------------------------
+# Create an API key at https://www.zotero.org/settings/keys
+# with "Allow library access" and "Allow notes access" enabled.
+ZOTERO_API_KEY=
+# Your numeric Zotero user ID (visible at https://www.zotero.org/settings/keys)
+ZOTERO_LIBRARY_ID=
+# Local path to Google Drive papers folder where ZotMoov stores linked PDFs.
+# This is the same folder set as "Linked Attachment Base Directory" in
+# Zotero Preferences > Advanced > Files and Folders.
+GDRIVE_PAPERS_PATH=
+# Windows example: G:\我的云端硬盘\Documents\Papers\Zotero_Papers
+# or: C:\Users\owen\Google Drive\我的云端硬盘\Documents\Papers\Zotero_Papers

.env.example

@@ -41,8 +41,15 @@ docs/docs/
 # 本地数据卷
 workspace/
 data/
+output/
 
 # 本地工具运行时（CodeGraph 数据库 / oh-my-codex 状态）
 .codegraph/
 .omx/
 .claude/
+
+# Playwright MCP 浏览器工作目录
+.playwright-mcp/
+
+# Firecrawl 输出
+.firecrawl/

.gitignore

@@ -45,6 +45,23 @@ dailyinfo push -d 2026-04-22    # Specific date
 dailyinfo status                 # Show today's briefing/pushed counts
 dailyinfo start/stop/restart     # FreshRSS Docker container
 dailyinfo logs                   # Tail execution log
+dailyinfo clean-cache            # Delete FreshRSS cache files older than 24h
+dailyinfo clean-cache --dry-run  # Preview what would be deleted
+
+# Download PDFs (agent-operated, requires Playwright MCP `mcp__plugin_playwright_playwright__*`)
+# Deterministic patterns per publisher — see skills/download-pdf/SKILL.md for full flow:
+#   Nature (OA):       navigate → click "Download PDF" → Chrome native download
+#   Nature (inst):      navigate → WAYF login (user does SSO) → click "Download PDF"
+#   Wiley/AGU (all):    navigate to pdfdirect?download=true → Chrome native download
+#   Cloudflare blocks:  pause, tell user to pass challenge, wait for "done"
+dailyinfo download-pdf 10.1016/j.jhydrol.2024.132471           # Print download instructions for the skill
+python scripts/download_pdf.py verify <pdf>                    # Verify PDF and extract metadata
+python scripts/download_pdf.py detect <url>                    # Detect publisher from URL
+
+# Sync downloaded PDF to Zotero (linked_file, zero cloud quota)
+# ⚠️ MUST use `uv run python` — conda Python lacks pyzotero
+uv run python scripts/zotero_sync.py <pdf> <doi> --json        # Copy to GDrive + create Zotero item
+uv run python scripts/zotero_sync.py <pdf> <doi> --dry-run     # Preview without creating
 
 # Zotero -> NotebookLM (agent-operated)
 # Prefer the Claude Code slash command:
@@ -122,6 +139,8 @@ Scrape sources with custom parsing need matching `if self.name == "..."` dispatc
 Required: `DEEPSEEK_API_KEY`, `DISCORD_BOT_TOKEN`
 Optional: `OPENROUTER_API_KEY` (fallback model), `DISCORD_CHANNEL_PAPERS/AI_NEWS/CODE/RESOURCE`, `FRESHRSS_USER/PASSWORD`, `DAILYINFO_DATA_ROOT` (default: `~/.myagentdata/dailyinfo`), `DAILYINFO_FALLBACK_MODEL`
 
+Zotero sync optional: `ZOTERO_API_KEY`, `ZOTERO_LIBRARY_ID`, `GDRIVE_PAPERS_PATH` (for `zotero_sync.py` linked_file attachment)
+
 ## Testing Conventions
 
 - **Autouse `tmp_data_root`** in `conftest.py` redirects all filesystem writes to `tmp_path` and sets `DISCORD_BOT_TOKEN`
@@ -141,6 +160,70 @@ Issues are tracked as GitHub issues on `iHeadWater/dailyinfo`. See `docs/agents/
 
 Default label vocabulary (`needs-triage`, `needs-info`, `ready-for-agent`, `ready-for-human`, `wontfix`). See `docs/agents/triage-labels.md`.
 
+### Download PDF (download-pdf)
+
+Download academic PDFs through institutional access (DUT SSO) using Playwright browser automation. Zotero sync via linked_file attachment. See `skills/download-pdf/SKILL.md`.
+
+**New-machine Playwright setup** (one-time, ~5 min):
+
+```bash
+# 1. Enable the official Playwright plugin (provides mcp__plugin_playwright_playwright__* tools)
+#    In ~/.claude/settings.json → enabledPlugins → "playwright@claude-plugins-official": true
+
+# 2. Install Chromium (if plugin auto-download fails)
+npx playwright install chromium
+
+# 3. Install @playwright/mcp globally (provides the CLI that the plugin wraps)
+npm install -g @playwright/mcp@latest
+```
+
+The plugin provides `mcp__plugin_playwright_playwright__*` tools (standalone Chromium, NOT the user's personal Chrome).
+Browser profile (cookies, WAYF/SSO sessions) persists in `.playwright-mcp/` under the project directory.
+Nature login survives across Claude Code restarts; Wiley/AGU Cloudflare challenge must be passed once per session.
+
+**MCP tools to use:**
+- `mcp__plugin_playwright_playwright__browser_navigate` — navigate to URL
+- `mcp__plugin_playwright_playwright__browser_click` — click element (use ref from snapshot)
+- `mcp__plugin_playwright_playwright__browser_snapshot` — get page accessibility tree
+- `mcp__plugin_playwright_playwright__browser_type` — type text into field
+- `mcp__plugin_playwright_playwright__browser_press_key` — press keyboard key
+- `mcp__plugin_playwright_playwright__browser_wait_for` — wait for text or time
+- `mcp__plugin_playwright_playwright__browser_run_code_unsafe` — run arbitrary Playwright code
+- `mcp__plugin_playwright_playwright__browser_tabs` — manage browser tabs
+
+**DO NOT use:**
+- `mcp__plugin_ecc_playwright__*` — requires Chrome extension bridge, needs separate setup
+- `browser_evaluate` + `readAsDataURL()` — crashes MCP on PDFs >1MB
+- `browser_run_code` + `require('fs')` — `require` is not defined in the MCP runtime
+
+### Bilibili Upload (bilibili-upload)
+
+Upload podcast audio to Bilibili as video (audio + auto-generated cover → MP4 via ffmpeg → biliup upload). See `skills/bilibili-upload/SKILL.md`.
+
+**One-time setup:**
+
+```bash
+winget install --id=ForgQi.biliup-rs -e
+biliup -u ~/.bilibili/cookies.json login   # scan QR code, valid ~2 years
+```
+
+**Usage:**
+
+```bash
+# Upload audio (cover auto-generated)
+dailyinfo bilibili-upload "output/weekly-review/2026-06-28/podcast/audio_hydrology.mp3" \
+  --title "水文AI周报 2026-W26" \
+  --tags "AI,水文,科研"
+
+# Preview only (no upload)
+dailyinfo bilibili-upload audio.mp3 --title "Test" --dry-run
+
+# In Claude Code, just say: "上传这周的水文周报音频到B站"
+```
+
+biliup cookie at `~/.bilibili/cookies.json` persists for ~2 years.
+If upload fails with code 601, wait a few minutes and retry (rate limit).
+
 ### Domain docs
 
 Single-context — one `CONTEXT.md` + `docs/adr/` at the repo root. See `docs/agents/domain.md`.

CLAUDE.md

@@ -133,6 +133,22 @@ dailyinfo push
 | `dailyinfo zotero-brief --collection water --artifact audio` | Process the `water` collection and request Audio Overview |
 | `dailyinfo zotero-brief --artifact video` | Request NotebookLM Video Overview |
 | `dailyinfo zotero-brief --manual-only` | Prepare local materials without calling NotebookLM |
+| `dailyinfo download-pdf <doi>` | PDF download via Claude Code (see [docs](docs/download-pdf.md)) |
+| `uv run python scripts/zotero_sync.py <pdf> <doi>` | Zotero linked_file sync (see [docs](docs/zotero-sync.md)) |
+
+## Download PDF → Zotero Sync
+
+> **Requires Claude Code + Playwright MCP plugin.** For setup, supported publishers, and failure handling, see [PDF Download](docs/download-pdf.md) and [Zotero Sync](docs/zotero-sync.md).
+
+Quick reference:
+```bash
+# One-time setup
+npx playwright install chromium
+uv pip install pyzotero
+
+# Then in Claude Code:
+/download-pdf 10.1038/s41586-026-10704-3
+```
 
 ## Zotero -> NotebookLM Agent Workflow
 
@@ -175,6 +191,9 @@ See:
 | `DAILYINFO_ENV` | Environment: `prod` / `dev` / `staging` (default `prod`) |
 | `DAILYINFO_FALLBACK_MODEL` | Fallback model when DeepSeek returns empty (default `moonshotai/kimi-k2.5`) |
 | `ZOTERO_LOCAL_BASE_URL` | Zotero local API base URL, default `http://127.0.0.1:23119` |
+| `ZOTERO_API_KEY` | Zotero Web API key (for `zotero_sync.py` linked_file sync) |
+| `ZOTERO_LIBRARY_ID` | Numeric Zotero user library ID (for `zotero_sync.py` linked_file sync) |
+| `GDRIVE_PAPERS_PATH` | Local path to Google Drive papers folder (ZotMoov / Linked Attachment Base Directory) |
 | `NOTEBOOKLM_HOME` | NotebookLM profile/auth directory used by `notebooklm-py` |
 
 ## Scheduling and Agents
@@ -195,6 +214,8 @@ DailyInfo intentionally avoids owning the scheduler. Recommended ownership:
 - [Architecture](docs/architecture.md)
 - [CLI Reference](docs/cli.md)
 - [Agent Config](docs/agent-config.md)
+- [PDF Download (Institutional Access)](docs/download-pdf.md)
+- [Zotero Sync (linked_file)](docs/zotero-sync.md)
 - [Zotero NotebookLM Workflow](docs/zotero-notebooklm.md)
 - [Zotero NotebookLM 工作流](docs/zotero-notebooklm.zh.md)
 - [Information Sources](docs/sources.md)

README.md

@@ -131,6 +131,24 @@ dailyinfo push
 | `dailyinfo zotero-brief --collection water --artifact audio` | 处理 `water` collection 并请求音频概览 |
 | `dailyinfo zotero-brief --artifact video` | 请求 NotebookLM 视频概览 |
 | `dailyinfo zotero-brief --manual-only` | 只生成本地素材，不调用 NotebookLM |
+| `dailyinfo download-pdf <doi>` | PDF 下载（依赖 Claude Code，见[文档](docs/download-pdf.md)） |
+| `uv run python scripts/zotero_sync.py <pdf> <doi>` | Zotero linked_file 同步（见[文档](docs/zotero-sync.md)） |
+
+## 下载 PDF → Zotero 同步
+
+> **依赖 Claude Code + Playwright MCP 插件。** 完整配置、支持的出版商和故障处理见 [PDF 下载](docs/download-pdf.md) 和 [Zotero 同步](docs/zotero-sync.md)。
+
+快速参考：
+```bash
+# 一次性配置
+npx playwright install chromium
+uv pip install pyzotero
+
+# 在 Claude Code 中输入：
+/download-pdf 10.1038/s41586-026-10704-3
+```
+
+## Zotero -> NotebookLM Agent 工作流
 
 ## Zotero -> NotebookLM Agent 工作流
 
@@ -171,6 +189,9 @@ NotebookLM 登录有意保留人工参与。Agent 可以打开浏览器或提示
 | `DAILYINFO_DATA_ROOT` | 覆盖默认数据根目录 |
 | `DAILYINFO_FALLBACK_MODEL` | 主模型空响应时的备用模型 |
 | `ZOTERO_LOCAL_BASE_URL` | Zotero 本地 API 地址，默认 `http://127.0.0.1:23119` |
+| `ZOTERO_API_KEY` | Zotero Web API 密钥（`zotero_sync.py` linked_file 同步用） |
+| `ZOTERO_LIBRARY_ID` | 数字型 Zotero 用户库 ID（`zotero_sync.py` linked_file 同步用） |
+| `GDRIVE_PAPERS_PATH` | Google Drive 论文文件夹本地路径（ZotMoov / 链接附件根目录） |
 | `NOTEBOOKLM_HOME` | `notebooklm-py` 使用的 NotebookLM profile/auth 目录 |
 
 ## 调度和 Agent 分工
@@ -191,6 +212,8 @@ DailyInfo 不负责调度。推荐分工如下：
 - [系统架构](docs/architecture.md)
 - [CLI 参考](docs/cli.md)
 - [Agent 配置](docs/agent-config.md)
+- [PDF 下载（机构访问）](docs/download-pdf.md)
+- [Zotero 同步（linked_file）](docs/zotero-sync.md)
 - [Zotero NotebookLM Workflow](docs/zotero-notebooklm.md)
 - [Zotero NotebookLM 工作流](docs/zotero-notebooklm.zh.md)
 - [数据源说明](docs/sources.md)

README.zh-CN.md

@@ -158,6 +158,34 @@ automation. If Google changes the NotebookLM UI/API or auth is not ready, the
 command degrades to the local material package so the run can be completed
 manually in the NotebookLM web UI.
 
+### Download PDF (institutional access)
+
+Requires Claude Code + Playwright MCP plugin. The CLI is a thin helper — real browser
+orchestration happens through the agent skill. See [PDF Download](download-pdf.md) for
+full setup, supported publishers, and failure handling.
+
+```bash
+dailyinfo download-pdf <doi|pii|url>           # Print download instructions for the skill
+python scripts/download_pdf.py detect <url>    # Detect publisher from URL
+python scripts/download_pdf.py verify <pdf>    # Verify PDF and extract metadata
+python scripts/download_pdf.py doi <doi>       # Resolve DOI to article URL + publisher
+```
+
+### Sync PDF to Zotero (linked_file)
+
+Creates a Zotero item with a linked_file attachment. See [Zotero Sync](zotero-sync.md)
+for prerequisites, setup, and design rationale.
+
+```bash
+uv run python scripts/zotero_sync.py <pdf> <doi> --json       # Copy to GDrive + create Zotero item
+uv run python scripts/zotero_sync.py <pdf> <doi> --dry-run    # Preview only
+```
+
+Design rationale: uses pyzotero Web API with `linked_file` attachment
+(`attachments:<filename>` scheme) rather than `imported_file`, because Zotero's
+300 MB cloud quota is too small for institutional PDFs (10-90 MB each).
+See [ADR-001](decisions/adr-001-zotero-linked-file-sync.md).
+
 ### Weekly Recap
 
 ```bash
@@ -198,6 +226,9 @@ FRESHRSS_PASSWORD=freshrss123
 | `FRESHRSS_PASSWORD` | FreshRSS password |
 | `DAILYINFO_FALLBACK_MODEL` | Fallback LLM when the primary model returns empty (default `moonshotai/kimi-k2.5`) |
 | `ZOTERO_LOCAL_BASE_URL` | Zotero local API base URL for `zotero-brief` (default `http://127.0.0.1:23119`) |
+| `ZOTERO_API_KEY` | Zotero Web API key for `zotero_sync.py` linked_file import (create at https://www.zotero.org/settings/keys) |
+| `ZOTERO_LIBRARY_ID` | Numeric Zotero user library ID for `zotero_sync.py` |
+| `GDRIVE_PAPERS_PATH` | Local path to Google Drive papers folder (ZotMoov target, same as Zotero's Linked Attachment Base Directory) |
 | `NOTEBOOKLM_HOME` | NotebookLM profile/auth directory used by `notebooklm-py`; must match the login run |
 
 ## Scheduling