AI RADAR · CHANGELOG

Changelog

记录 AI Radar 每一次可见改进与运行能力变化。

周二

  • 前端第二轮视觉与体验打磨，重点是对齐、留白与信息层次。桌面端：日期分组标题现在与下方所有时间戳右侧对齐成一条线；正文区不再固定宽度上限，宽屏下可用宽度被充分利用；时间线的圆点之间新增一条贯穿整个日期分组的连接线，收藏某条后该条圆点变为琥珀色；侧栏图标换成更清晰的图标；搜索按钮改为实心主按钮；新增全站定制滚动条与按钮按下 / 禁用状态；主题切换器用一块滑块指示当前档位；热点浮层不再在鼠标移向它的途中消失。
  • 手机与网页缩放（≤960px）：正文列变宽、左右留白对称；此区间的时间戳字号偏大的问题已修正；推荐理由不再被截断成两行；顶部条改为随页面滚动而非固定吸顶。列表行变得更清爽——行首只保留时间与信源、分数移到行尾，不再挤入精选标记、收藏按钮与标签，信源名不再被压缩省略。首页与「全部 AI 动态」的卡片收藏按钮现在只在桌面显示；手机上仍可经「更多 → 收藏」查看、取消、导出与导入收藏（导入依然能新增收藏）。手机上日期分组不再提供折叠。话题标签现在只出现在「全部 AI 动态」页。
  • 更新日志页修复了手机与缩放档的排版错乱（正文此前会被挤成窄条且左右不对称）。日期标题改为中文长格式并附星期，标题下方新增一行说明。
  • AI 日报页正文末尾新增四项数据概览：今日事件 / 一手报道 / 新模型 / 信源。四项均为本站自有定义——今日事件为该期条目数，一手报道为该期中非推文类且信源等级为 T1 的条目数，新模型为该期中属「模型发布」分类的条目数，信源为该期去重信源数；不声称与任何外部站点口径一致。日报摘要里的行内代码现在显示为等宽字体的浅底圆角块。
  • 热点榜页在桌面、平板与手机三档的正文宽度分别调整到更合适的比例。
  • 首页首批内容现在由服务端直出全部约 40 条，避免浏览器加载后用完整首批替换 12 条预渲染内容所造成的布局位移。手机上品牌 + 日期紧凑条改为只在首页出现；「更多」、「收藏」、「关于」和「微信文章解读」页改用各自的页面标题头，微信详情与 404 页的返回列表入口保留。

周日

  • 前端全面改版（参照 aihot.virxact.com 的信息架构与视觉）：默认主题从深色科技风改为浅色简约风，新增暗色变体与侧栏底部三态主题切换（浅色/深色/跟随系统，localStorage 持久，head 内联脚本防闪烁）；精选与全部动态页从分页控件改为无限下拉（IntersectionObserver + 请求代际校验防筛选切换竞态；/all 搜索态改用页码分页以规避 timeline API 搜索时忽略 cursor 的语义）；卡片新增收藏按钮与 /bookmarks 收藏页（localStorage 快照 + 导出/导入 JSON，导入经字段与日期可解析校验；服务端同步接口已预留约定未实现）；首页新增"当前热点"榜（新端点 GET /api/v1/hot，近 48 小时按 加权分×10+关联讨论×5 排序，单次一致快照取样，纳入 90s 公共缓存白名单）；日期分组可折叠且追加加载继承折叠态；主站页面移除 Google Fonts 改用系统字体栈（消除跨境字体请求的首屏延迟），AI 日报页保留原深色报纸风（旧样式与 token 原样迁移）。微信详情页安全契约由"整页无 script"收窄为"正文容器无 script"（新增 head 主题引导与导航 module 两个可信脚本）。改版经 Codex 高档对抗审查两轮修复复核后合入。

周日

  • 将 same-host performance-probe 从 busy/idle 双轨 gate + busy rollup 改为 idle-only：pipeline 运行或负载不确定时不保存/评估，只有 idle 窗的 22 样本确认窗超预算才直接 page，不再降为 notice。probe 调度同步从 hourly :17 crontab 改为专属 per-file launchd（StartInterval=300，经 ./run.sh performance-probe 进入 external watchdog），pipeline 仍保留既有 */15 crontab；install/uninstall/status 现管理该 plist 与 legacy symlink 迁移。Playwright Chromium 是微信抓取与默认 probe 共用的显式部署前置，安装器不会自动下载或校验。2026-07-26 live 证明中，全 8 个旅程 cell 在 4.93 小时取得第 22 条 idle 样本，满足 6 小时硬门槛但仅余约 1 小时负载裕度。PERF 投递契约明确为 at-least-once，并依赖 im-notify 持久 signature dedup 抑制同一 crash retry 的重复可见消息。

周三

  • 新增 curated_items.summary_json 精选 digest 缓存的常驻保留：每次 curate 后自动清空超过 keep_days（默认 7 天）且非最新 run 的可再生预计算缓存，使 radar.db 体量长期有界（此前约 8MB/天持续膨胀），生产库一次性瘦身实测由约 2.28GB 降到约 1.5GB（省约 785MB / 34%）。同时新增 ./run.sh admin db retain（只清列）与 ./run.sh admin db slim（清列 + VACUUM 回收磁盘、DB 同步前跑）子命令，slim 返回 retained/compacted 两阶段结果，--dry-run 零写只报待清量。唯一用户可感知的行为变化：/api/v1/curated?run_id=X 访问超窗口的历史 run 时，其 digest 改为 live 现算，内容反映当前 enrichment 而非 curation 时快照（TTL 语义）；所有 HTML 用户页只服务最新 run，字节一致、不受影响。
  • 将运维告警从单一 page 级别升级为 page/notice 分级：需立即处置的事故发往 ALERT，低打扰退化发往 NOTIFICATION。pipeline busy 期间但同视角 idle 正常的 PERF 超预算会合并成一条 PERF:rollup:busy notice，而真实 idle/公网退化仍保留 page。新增 data/alert-events.jsonl 已送达通知历史，可按规则、severity、firing/resolved 与通道查询最近 14 天的成功投递。

周日

  • 修复精选归档计数缓存的过度失效和微信 SSR 请求连接生命周期，使首页与微信页面在 pipeline 写入期间保持稳定响应，同时保留精确总数和分页语义。
  • 新增 performance-probe 用户旅程监控：从同机 origin/public 测量首页、微信列表、详情和翻页，区分 pipeline idle/busy，以 PERF:* 规则告警并保留 14 天诊断证据。结果明确标为 same-host provisional，不作为区域 SLO。
  • 新增 performance-remediate 候选修复 worker：confirmed 性能退化可触发单个、最长 60 分钟的隔离 Codex worktree，生成未进入主分支且未部署的本地候选 commit；越界配置会 fail closed。
  • 修复 performance-probe 首页浏览器探针把正常样本误判为 hard_failure 的缺陷。此前它拿完整渲染卡片列表与 12 项期望做全等比较，导致每个健康首页样本都被标记 hard_failure；现改为按前缀匹配期望，前缀不符仍会 hard-fail。同时收紧 performance-remediate 的启用门槛说明：缺陷虽已修复，运维仍须先确认 hard_failure=false 且首页 PERF:* 未告警，再安装 remediate cron。
  • 为公开分页路径（/、/wechat、/api/v1/curated、/api/v1/wechat）在安全分页变体下让 origin 发 Cache-Control: public, max-age=90, stale-while-revalidate=30，而带 q=（含空 q=）、分类/日期/未知参数或非 200 响应一律 private, no-store（fail-closed 白名单，已验证这些路由无 cookie/会话变量）；前端 app.js 在 SSR 预载后预取下一页 API 并在点击翻页时复用同一 promise，搜索/分类请求绕过该 90 秒前端缓存。头部代码需重启 serve 生效。配套在 Cloudflare zone aiplanet.live 手动加了 Cache Rule「AI Radar short public pagination TTL」——这是 Cloudflare dashboard 侧配置、非 repo 代码，Edge TTL 与 Browser TTL 均 respect origin 头——使这些路径改由 CF 边缘缓存，实测翻页 API 从约 3-5s 的 DYNAMIC 回源降到约 0.5-1.4s 的 CF HIT。

周一

  • A1-A4 告警的发送传输改为复用本机 im-notify --alert，不再由 AI Radar 直接调用飞书 webhook。原有 firing / resolved、debounce 与 30 分钟冷却状态机保持不变，且不叠加 --dedup-key；发送器失败会记录日志而不会终止周期告警检查。alert launchd 模板同时把 ~/.local/bin 加入 PATH。

周三

  • Split LLM usage accounting into data/llm_usage.db (AI_RADAR_LLM_USAGE_DB) so prefilter / score / enrich token writes no longer contend with the main radar.db writer. Existing radar.db.llm_usage history is copied into the dedicated DB on migration/first use, /admin/usage reads the dedicated DB while still showing item/source metadata from the main DB, and A2 prefilter P95 now uses a recent 2-hour sliding window so recovered latency incidents do not keep firing until midnight.
  • Added an internal /admin/usage page and /api/v1/admin/usage endpoint for maintainers to inspect LLM usage attribution. DeepSeek/ARK chat_json calls now persist one llm_usage row per prefilter / score / enrich LLM call, including model, input/output tokens from completion.usage, item attribution, and input size. The page uses the same Cloudflare Access / dev-only local bypass guard as /admin, is not linked from public navigation, and rolls up the last 30 days by day, model, and pipeline stage.

周五

  • Made fresh-clone setup degrade cleanly when optional private resources are missing. The Mp2RSS WeChat source now skips with a warning when MP2RSS_FEED_URL is unset or empty instead of aborting source loading, and .env.example no longer contains a fake Mp2RSS URL that would force a broken fetch. ./install.sh now checks each service before installing: serve always installs, pipeline needs one LLM API key, alert needs FEISHU_GENERAL_ALERT_WEBHOOK, and tunnel needs deploy/cloudflared/config.yml. Missing promptable values can be entered interactively and are appended to ./.env; non-interactive installs skip only the affected services and print a summary.

周二

  • Stopped A4 (article ingestion) from alerting on transient fetch flaps. All X/Twitter sources fetch through the single public nitter.net instance, which intermittently times out for one ~15-minute fetch round and then self-heals; each flap fired and resolved A4 within ~15 minutes as pure noise, while the daily ingestion count was never actually affected (a round skipped by the flap is backfilled by the next). A4 now debounces: a fetch-failure condition must persist past a 30-minute window (≈2 fetch rounds) before it notifies, and a flap that recovers within the window is absorbed silently — no firing and no resolved. The debounce is per-rule and configured only for A4 (a4.debounce_minutes); A1/A3 still notify immediately, so this does not delay genuinely urgent alerts. Also corrected A4's disposition text, which pointed at the retired wewe-rss/bridge for WeChat — WeChat now ingests via Mp2RSS, and a batch X-source failure is the more common trigger. The alert service must be running the updated code for this to take effect.

周一

  • Made /wechat search (and the shared timeline/curated search) whitespace-insensitive, so a query with extra internal spaces returns the same results as one without. Previously a stored title like 分享Claude Code was found by 分享Claude Code but not by 分享 Claude Code, because the query and the matched columns were compared with their spaces intact. Both the query patterns and the searched columns (title, author, abstract, tags) now have all whitespace — including full-width spaces — stripped before matching, and the longer-query FTS path handles spaced queries too. Simplified/Traditional matching is unchanged. The web layer must be restarted for this to take effect.
  • Fixed the missing avatar for the WeChat Official Account "赛博禅心", which showed a fallback initial instead of its real avatar. A single failed avatar scrape on 2026-06-02 had left its cache row empty and a 7-day negative cache prevented any retry. Added an admin wechat-avatar refresh --account <name> command that clears one account's cache row and re-scrapes immediately (used to repopulate 赛博禅心's avatar live), and shortened the failed-scrape negative-cache TTL from 7 days to 2 so a transient miss self-heals within days instead of a week.
  • Fixed the 日报 page so navigating to a past day now shows that day's curated articles instead of an empty report. Previously /daily/{date} (前一日/后一日 and direct date URLs) only ever populated for today, because a dated request was answered from the single latest curation run, which curates only about one day of fresh items. A dated daily report now aggregates curated items published on that date across all curation runs (deduplicated to each item's latest curation), the same cumulative-archive logic the home page / uses, so any past date with curated content renders a populated report. The admin explicit-run_id path and the / and /all archive pagination are unchanged.

周日

  • Reduced false alerts in the monitoring rules. A2 (pipeline health) no longer treats a long in-progress run as a fault: a SKIP log means "pipeline already running" (liveness), so it is no longer a standalone trigger, and the "no successful pipeline" heartbeat threshold was raised from 45 to 120 minutes — this eliminates the recurring fire/resolve flapping that produced the bulk of past alert noise. A3 (website) dropped its healthz dimension, which was a dead signal (hardcoded to never fire) that misleadingly displayed "healthz 连续失败 0 次"; A3 now reports only the real user-side 5xx rate. Each line in logs/alert-check.log is now timestamped (Asia/Shanghai) for incident forensics. A true active healthz probe and a time-aware A4 ingestion floor are tracked as follow-ups in docs/issues/general.md.
  • Prefixed every Feishu monitoring alert (both firing and resolved) with a 【AI Radar】 project label. Because the alert webhook (FEISHU_GENERAL_ALERT_WEBHOOK) is shared across projects, the prefix lets a recipient tell at a glance which project an alert came from.
  • Added an explicit "访问原文" (visit original) link to WeChat interpretations so readers can jump to the source Official Account article: a bordered button below the title on each /wechat/<slug> detail page, plus a compact "原文 ↗" link on every list card. Both open the original in a new tab and reuse the existing source URL. The shared frontend asset version was bumped so visitors receive the new behavior.
  • Added search to /wechat, scoped to interpretation card fields: Chinese title, Official Account author, abstract, and tags. Search URLs are shareable with ?q=, pagination and detail-page return links preserve the query, Simplified/Traditional Chinese variants match, and the shared frontend asset version was bumped so visitors receive the new behavior.

周六

  • Removed the retired WeWe RSS bridge from the service layer. ./install.sh, ./uninstall.sh, and ./status.sh now manage four services (serve, tunnel, pipeline, alert) instead of five, and a bare ./install.sh no longer requires Docker or aborts when it is unavailable. WeChat ingestion continues through Mp2RSS; rollback material (deploy/wewe-rss/ + RUNBOOK) is retained, with the launchd wiring recoverable from git history.

周四

  • Changed the curated home page / from a single page of the latest curation round's top 40 into a cumulative archive of every item ever curated. It now aggregates all distinct items selected across past curation runs (deduplicated, currently about 1,793 items), ordered newest first, paginated about 40 per page (currently about 45 pages) using the same numbered page controls as /all. Page 1 still shows the latest curated picks, preserving the "skim in five minutes" use.
  • Upgraded pagination on the /all timeline and the /wechat interpretation list to numbered page controls: first and last pages are always visible, the current page shows two neighbors on each side, gaps collapse to an ellipsis, any page number is directly clickable, and previous/next arrows remain. Both pages share one pagination component.
  • Changed /all to show the true first and last pages. /api/v1/timeline now returns an exact total count instead of the previous forward-looking estimate, so the last page reflects real data (matching /wechat), and out-of-range requests such as ?page=9999 clamp to the real last page.

周二

  • Added the /wechat tab for WeChat Official Account article interpretations, with structured summaries, tags, worth-reading filtering, shareable detail pages, and ai-assistant knowledge-base writeback for saved articles.
  • Removed the disabled legacy WeWe source definitions wx_guizang and wx_crossing and deleted their historical item rows from the production DB after creating a verified backup.
  • Migrated WeChat Official Account ingestion from the self-hosted WeWe RSS bridge to the hosted Mp2RSS feed, removing the local Docker container and WeRead QR-login maintenance that frequently broke ingestion in production. The feed URL with its embedded key is read from the MP2RSS_FEED_URL environment variable and never committed.
  • Added real Official Account names and avatars to WeChat article cards. Cards now show each article's source account name and its avatar instead of the shared collection name, falling back to the WeChat icon when no avatar is cached.
  • Added the /admin operations dashboard for user traffic, ingestion, pipeline health, and active alert status.
  • Added the alert background service, which runs admin alert-check every five minutes and sends A1-A4 monitoring alerts through a Feishu custom-bot webhook.
  • Documented the monitoring runbook, including Cloudflare Access setup for /admin* and /api/v1/admin*, Feishu webhook setup, and daily service verification commands.

周一

  • Fixed Chinese-source search so recently fetched backfill articles are evaluated and visible even when their original publish date is older than the pipeline window.
  • Improved source-name searches by ranking source/author matches ahead of content-only matches and rotating same-name sources so a prolific source no longer hides lower-volume WeChat sources on the first page.
  • Added Simplified/Traditional Chinese query expansion, so searches such as 归藏 and 歸藏 find the same Chinese-source articles.

周六

  • Expanded /api/v1/timeline and /api/v1/curated search to match source names, authors, and Chinese titles in addition to article title/body. Searches of 3+ characters use FTS; 1-2 character queries fall back to short-field LIKE matching.
  • Changed search semantics intentionally: searching a source name can return all matching articles from that source, and scoring reasoning is no longer part of the search index.

周五

  • Added WeChat Official Account ingestion through a local WeWe RSS bridge, with Playwright-based full-text scraping for internal LLM processing and public cards limited to generated summaries plus original article links.

周四

  • Improved / and /all first-screen loading by serving SSR-preloaded feed HTML; production verification now shows no visible loading spinner and sub-1.5s median first content on the main feed URLs.
  • Fixed /all timeline entries without scoring data so every visible card renders a numeric score pill.
  • Fixed the About page repository link to point at your-org/ai-radar.

周日

  • Improved /api/v1/timeline and /all load performance by adding SQLite indexes, preloading enrichment data in the timeline query, and replacing the exact count query with a pagination-safe estimate.
  • Fixed timeline time display so visible times use the same Asia/Shanghai timezone as date grouping.
  • Updated the About page repository link to your-org/ai-radar.
AI RADAR · 持续更新
沪ICP备2026017013号