.env.example

ARK_API_KEY=
ARK_MODEL=Doubao-Seed-1.6-lite
ARK_ENDPOINT_ID=
OPENAI_API_KEY=
OPENAI_BASE_URL=
OPENAI_MODEL=
DIGEST_WEBHOOK_URL=
GNEWS_API_KEY=
# Public deployment recommendation:
# - Do NOT put real personal keys in server-side .env for a public website.
# - Let visitors input their own keys in Streamlit sidebar (session-only, not persisted).
# - Keep this file as template only and never commit a real .env/.ENV.
# Default true: drop items that do not match intent keywords (reduces GNews noise). Set false to show unmatched items.
# STRICT_INTENT_MATCH=true
# 提速：跳过 Playwright 站点列表爬取，仅 RSS +（可选）GNews + LLM。需要 OpenAI Blog 等站点抓取时设 false。
# DIGEST_FAST_MODE=true
# 或显式关闭站点爬：SITE_CRAWLERS_ENABLED=false
# RSS 单源 HTTP 超时（秒），慢源早失败，默认 15；可设 10–12
# RSS_HTTP_TIMEOUT=12

# Streamlit 用户新闻偏好记忆：runs/user_news_memory.json（跑通日报后 LLM 自动合并）
# USER_NEWS_MEMORY_ENABLED=true
# USER_NEWS_MEMORY_UPDATE=true
# 检索模式（显式主题 / 泛化浏览）由代码根据 intent 自动分类；显式轮次不合并历史关键词，泛化轮次可加权重排。

# Outbound HTTP(S) for crawlers / RSS / APIs (NOT Ark LLM or webhooks).
#
# Webshare — same as youtube project: WEBSHARE_API_KEY pulls list via requests (mode=backbone, page_size=10).
# WEBSHARE_API_KEY=
# Optional: WEBSHARE_PROXY_LIST_PAGE_SIZE=10  WEBSHARE_PROXY_LIST_MODE=backbone
# Rotating gateway without API: WEBSHARE_PROXY_USERNAME=  WEBSHARE_PROXY_PASSWORD=
# If :80 is blocked: WEBSHARE_BACKBONE_PORT=1080
# Optional: WEBSHARE_PROXY_URL=http://USER:PASS@p.webshare.io:80/
# Same .env as youtube tool: YOUTUBE_PROXY_HTTPS=  YOUTUBE_PROXY_HTTP=  (generic; wins over Webshare)
# Optional: OUTBOUND_WEBSHARE_SCOPES=site_crawler,playwright
# HTTPS_PROXY=  OUTBOUND_HTTP_PROXY=

# Optional: extra HTTP API feeds (comma-separated). See project/ai_news_skill/integrations/public_api_feeds.py
# PUBLIC_API_FEEDS=hackernews,spaceflight,devto,arxiv,nasa,github
# PUBLIC_API_FEED_MAX=8
# NYTIMES_API_KEY=
# GUARDIAN_API_KEY=
# FMP_API_KEY=
# NASA_API_KEY=DEMO_KEY
# GITHUB_TOKEN=
# GITHUB_RELEASE_REPOS=microsoft/vscode,langchain-ai/langchain
# PRODUCT_HUNT_TOKEN=
# NYT_SEARCH_Q=artificial intelligence
# GUARDIAN_SEARCH_Q=technology OR artificial intelligence
# DEVTO_TAG=ai
# ARXIV_API_QUERY=cat:cs.AI
# Optional: LangGraph digest retries (0–5), default 2
# DIGEST_MAX_RETRIES=2
# Collect 后、Enrich 前：原始条目质检（多一次 LLM；默认关闭）。打开：PRE_ENRICH_REFLECTION_ENABLED=true
# 剔除单条：四维最低分 < PRE_REFLECT_DROP_ITEM_MIN（默认 2）则丢弃该条
# 仅对问题信源局部补拉 RSS（加宽窗口），非整轮重跑 collect：PRE_REFLECT_PARTIAL_RSS=true（默认）
# PRE_REFLECT_PARTIAL_MAX_SOURCES=5  PRE_REFLECT_PARTIAL_WINDOW_MULT=2  PRE_REFLECT_PARTIAL_EXTRA_PER_SOURCE=2
# PRE_REFLECT_PARTIAL_SOURCE_AVG_LT=3.5  PRE_REFLECT_PARTIAL_WORST_MIN_GREY_LO=2  PRE_REFLECT_PARTIAL_WORST_MIN_GREY_HI=3.25
# 质检前是否先拉正文摘录（更准但更慢）：PRE_ENRICH_REFLECT_FETCH_EXCERPTS=false
# 工具代理 + MCP：设为 true 且启用 LLM、存在仓库根目录 mcp_servers.json 时走 run_with_tool_agent（无侧栏开关）
# USE_TOOL_AGENT_MCP=false
# Site list crawlers (OpenAI news / Anthropic / HF / TechCrunch / VentureBeat): use visible browser if blocked
# SITE_CRAWLER_HEADLESS=true
# Streamlit defaults for item caps (optional)
# DIGEST_UI_ITEMS_CAP=100
# DIGEST_UI_GNEWS_MAX=100