dsh-read-url 1.4.0 → 1.6.1

package/CHANGELOG.md

@@ -1,5 +1,112 @@
 # Changelog
 
+## [1.6.1] - 2026-08-30
+
+> 迭代 3（效率与缓存轮）：缓存 FIFO → LRU、分页/跟随总时间预算、全量收官回归。零 schema 变更。
+
+### 改进
+
+- **成功/失败缓存 FIFO → LRU**：原实现按插入序驱逐，32 条容量下「老但热」的条目（长会话中反复读取的同一页）会被后来者挤掉触发重复 fetch。命中即重插为最新槽（Map delete+set），热点页面在 TTL 内永不被驱逐；failCache 同样处理（反复失败的坏 URL 也不会在 TTL 内被挤出导致重复请求）。
+- **meta-refresh 跟随链总预算**：每跳自带 fetch 超时，3 跳最坏 3×timeout——此前未计入工具超时公式（该公式只预留 paginateMax+1 次 fetch + 渲染余量）。整条跟随链共享一个 timeout 预算，超限即停止跟随（保留当前 HTML，fail-open）。
+- **分页链总时间预算**：while 循环增加墙钟守卫 `startedAt + timeoutMs × (paginateMax+1)`——首页或 SPA 渲染变慢时，分页链自动让出预算而不是顶穿工具超时（顶穿意味着整个调用被宿主杀掉、结果全丢；现在至少返回已拼好的部分）。
+
+### 验证
+
+- 单元断言 **175 → 176**（+1：LRU 驱逐语义——填满 3 条、命中最老条 a 提升、加入 d 驱逐次老条 b、b 重新 fetch 而 a 仍命中；测试内通过插件自身 dispose 清空共享缓存保证确定性）+ 12 SPA + probe 47 PASS 零 THROW；
+- 收官全量回归：152 站扫描 + site-types 语义探针 + 对抗 probe（见下轮备注/日志）。
+
+## [1.6.0] - 2026-08-30
+
+> 迭代 2（站点类型扩展轮）：AMP/响应式图片、纯文本家族原生路径、非标准 JSON 消毒。零 schema 变更。
+
+### 新增
+
+- **AMP 页面图片**：AMP HTML 用 `<amp-img>` 代替 `<img>`，markdown 模式图片此前全丢——图片预处理同等对待 amp-img（alt/src/data-src 同一套规则）。
+- **`<picture>/<source srcset>` 回退**：响应式布局把真实图片 URL 放在 `<source srcset>`、`<img>` 只留 data: 占位时，首个 source URL 注入 img 兜底；img 自带可用 src 的 picture 不动。
+- **纯文本家族原生路径**：`text/plain` / `text/markdown` / `text/csv` 且前 5000 字符无 HTML 标签形状时不再走 HTML 管线（那会把每行压平成一段）——行结构原样保留（CRLF 归一、行尾空白清理、3+ 空行折叠）。.txt 日志、.md 文档、robots.txt、CSV 数据集逐行可读。被错误标成 text/plain 的真 HTML 页仍走 HTML 管线。
+- **JSON NaN/Infinity 消毒**：JS 手写序列化产物含 `NaN`/`Infinity`（非法 JSON）时 `JSON.parse` 整页失败、退化成 HTML 管线输出原始转储。解析失败后以字面量置 null 重试一次（仅在已非法的 JSON 上应用）。
+
+### 验证
+
+- 单元断言 **170 → 175**（+5：amp-img、picture 注入、picture 不动、text/plain 行结构 + 误标 HTML 回归、NaN JSON）+ 12 SPA + probe 47 全绿；
+- site-types.mjs **12 类 10/12 语义通过**（2 例境外连接超时）：text/markdown（155 行结构保留）、robots.txt（164 行）实网验证；WordPress feed / `<time>` 博客 / 懒加载图片复验通过。
+
+## [1.5.2] - 2026-08-30
+
+> 迭代 1（底层管线审查轮）：fetchPage/重试/竞速/解码/spa/爬虫链路的例行审查，2 个真 bug + 2 处健壮性，零 schema 变更。
+
+### 修复
+
+- **双次 429 形状崩溃（P0）**：`directFetchOnce` 在任意一次 429/503+Retry-After 都返回 `{ retryAfterMs }` 形状；`directFetch` 对第二次结果不检查原样返回——无代理路径绕过 raceFirstSuccess 的形状过滤，`decodeBuffer(undefined)` 在 `buffer[0]` 处裸抛 TypeError。现在重试预算耗尽时返回明确的 rate-limited error；等待 Retry-After 期间被 abort 则返回 cancelled（原先仍会补打一次请求）。`directFetch` 导出供测试直测（本机 Windows 系统代理无法通过环境变量屏蔽，无代理路径无法端到端复现）。
+- **爬虫不跟随 meta-refresh 壳页（P0）**：`crawlSite` 直连 fetchPage——read_url 已跟随跳板壳，read_url_site 却把壳页当正文爬（无标题跳板文本 + 永远到不了真实内容）。爬虫页处理接入 `followMetaRefresh`（fail-open），提取与出站链接均基于跟随后的文档。
+
+### 加固
+
+- **spa.js 渲染轮询**：`evaluate` 连续失败两次（页面已崩）即放弃 DOM 稳定等待，不再对崩溃页烧满 10s；移除 playwright `page.goto` 不支持的 `signal` 选项（静默忽略、误导读者），abort 语义由循环内检查与 catch 承担。
+
+### 验证
+
+- 单元断言 **168 → 170**（+2：directFetch 双 429 error 形状（本地 429 服务器直测，确认重试恰一次）；read_url_site 跟随壳页落在真实页）+ 12 SPA 断言 + probe 47 PASS 零 THROW 全绿。
+
+## [1.5.1] - 2026-08-30
+
+> 例行审查轮：3 个实测坐实的 bug 修复 + 3 项健壮性加固 + 2 项站点适配，零 schema 变更。
+
+### 修复（P0，均有复现用例）
+
+- **锚点读 × ld+json 兜底越权**：`extract()` 的 articleBody 兜底不感知锚点语义——定位到 <200 字符的小节时整篇 LD 全文覆盖小节内容（`anchored: true` 却返回全文）。现在锚点命中时跳过该兜底（与 readability 升级的既有守卫对齐）；无锚点的薄页仍正常兜底。
+- **feed `content:encoded` 全文丢失**：闭标签正则要求字面 `</content>`，`</content:encoded>` 永不匹配——WordPress 等博客 feed 的全文字段整段丢失只剩标题。改为捕获字段名 + 反向引用配对（`<content:encoded>` 必须配 `</content:encoded>`），顺带杜绝与 `</description>` 跨配对吞错内容。实测 WPJAM/WP学院 feed 每条恢复 200 字符实质摘要。
+- **隐藏/consent 嵌套泄漏**：容器内有先闭合的嵌套 `<div>` 时，非贪婪正则止步于第一个内层闭标签，其余横幅文本泄漏进正文（真实 consent 横幅普遍深嵌套）。改为深度计数平衡扫描（复用 tagBlockAt 思路）；扫描窗口 100KB 上限防「大量未闭合开标签」退化；未闭合 fail-open 保留内容。
+
+### 加固（P1）
+
+- **batch 单页异常隔离**：`mapLimit` 的 `Promise.all` 无守卫，单页意外 throw 会炸整批。每页 try/catch 转 `{error}`。
+- **batch 列表去重**：同一 URL 多次出现会并发双 fetch。按精确字符串去重（带 `#fragment` 的锚点读语义不同，不与裸 URL 合并）。
+- **`<time datetime>` 作用域收窄**：优先 `itemprop="datePublished"`/`pubdate` 标记的标签；扫描范围从全文档改为剥离噪声后的正文区——页脚/评论区的 time 标签不再劫持发布时间。
+
+### 站点适配（P2）
+
+- **noscript SEO 兜底**：SPA 壳把正文放 `<noscript>` 时（stripNoise 剥离的块），正文过薄则取最长的 noscript 块兜底（≥150 字符阈值；"请启用 JS" 类短横幅不触发）。
+- **懒加载图片**：markdown 模式 `<img src>` 缺失或为 data: 占位图时回退 `data-src`/`data-original`/`data-lazy-src`，描述性图片不再丢失。
+
+### 验证
+
+- 单元断言 **158 → 168**（+10：锚点守卫正反例 2、content:encoded + 跨配对 1、嵌套隐藏/consent 平衡移除 1、未闭合 fail-open 1、noscript 兜底正反例 2、懒加载回退 1、time 标记优先级 1；batch 上限断言改用 15 个不同 URL 并新增去重断言）
+- probe.mjs **39 → 47**（+8 对抗项：3000 个分散未闭合隐藏 div（窗口有界 408ms）、3000 层 consent 深嵌套完整移除 0.9ms、2000 未闭合 consent、隐藏 table、50k articleBody、畸形 LDJSON、3000 noscript 块、未闭合 noscript——全部无 THROW）
+- 新增 `site-types.mjs` 站点类型语义探针（10 类，按类型断言而非字符数）：**8/10 语义通过**，2 例境外连接超时（环境边界）。关键实锤——WordPress feed（content:encoded）摘要恢复、`<time datetime>` 博客 published 提取（2026-08-28）、懒加载站点 markdown 图片 15 张、MediaWiki 变体百科正文干净
+- 152 站全量回归：**103 OK / 25 THIN+EMPTY / 24 ERR / 0 THREW**，OK 站零噪声；5 个首窗字符下降站点经 v1.3.1/v1.4.0 模块 A/B 比对确认均为段落化重排效应，内容完整（总字符量普遍增加）
+
+## [1.5.0] - 2026-08-29
+
+> 元数据补全与 token 边角打磨轮。152 站回归确认 1.4.0 无内容性回归后，聚焦三处已核实的缺口（响应大小上限 3MB、RSS 摘要、nav 剔除等此前已就位的项不再重复）。
+
+### JSON-LD `articleBody` 正文兜底
+
+不少新闻/门户站把**全文**放在 `<script type="application/ld+json">` 的 `articleBody` 字段，可见 HTML 却是反爬壳或薄片段——DOM 提取器够不到。现在正文提取不足 200 字符时自动尝试 ld+json 正文兜底（strip 标签 → 段落化，块扫描沿用有界 LDJSON_RE，上限 200K 字符；DOM 正文充分时不触发，绝不用兜底覆盖真实正文）。
+
+### `<time datetime>` 元数据源 + 日期归一化
+
+- 发布时间合并链扩展为：显式 `article:`/`og:` meta → JSON-LD datePublished → **`<time datetime>`** → 通用 date meta → byline 文本。语义化 `<time>` 标记通常标示发布时间，优先于常为 last-modified 的通用 date meta。
+- 可识别格式的日期统一归一为 ISO（`2026年8月24日` → `2026-08-24`），覆盖 meta/JSON-LD/time/byline 全部来源；不匹配的格式原样透传。
+
+### markdown 表格行数帽
+
+兼容性对照表（40+ 行 × 多列）是 markdown 模式最大的单点 token 开销。表格超过 25 行时只保留前 25 行并追加一行 `…+N rows` 提示；text 模式不受影响。
+
+### 零宽字符剥离
+
+零宽空格/连接符（U+200B–200D、U+2060、BOM）对读者不可见却计 token，且 152 站实测中确有泄漏（如 vuejs-doc 输出）。textOnly / textLines / blockMd 三条渲染路径统一剥离。
+
+### 评估后保持不变：batch 默认 maxChars = 3000
+
+实测对比：ruanyifeng 全文不足 2000（两窗口无差异）；solidot 的 2000–3000 段是第二篇文章的真实正文——下调到 2000 会在聚合页切掉实际内容。且 batch schema 描述含字面 `default 3000`，改动必须动 schema 文本，违背静态 schema 承诺。**结论：维持 3000。**
+
+### 验证
+
+- 单元断言 **150 → 158**（+8：articleBody 兜底 2、无日期源不编造 1、time datetime 1、CJK 归一 1、表格帽 1、零宽 2；既有 byline 日期断言同步归一化行为更新）+ 12 SPA 断言全绿；
+- probe.mjs 39 PASS 无回归；
+- 实网：MDN markdown 模式表格静态 HTML 未超帽（兼容表为 JS 渲染），帽由单测覆盖；ruanyifeng/solidot 窗口评估见上。
+
 ## [1.4.0] - 2026-08-28
 
 > 省token主打轮。经实测复核，原方案中的「nav 导航块剔除」与「RSS/Atom 条目摘要」在 1.3.1 已就位（python-docs 实测 235K 全文 → 20K 内干净正文，导航零泄漏；feed 输出已含每条 ≤200 字符摘要），本轮聚焦真实缺口：隐藏元素剥离、URL 片段定位阅读、段落化正文。

package/README.md

@@ -213,12 +213,13 @@ const results = await Promise.all([
 2. **Paragraph-aligned truncation + offset continuation** — 6,000 chars by default (~3,000 tokens), cut at paragraph boundaries to keep semantics (text mode separates paragraphs with `\n\n`, so the cut lands on a real boundary, not mid-sentence); output notes a single line `(chars 6000/12990 — truncated, continue via offset)`; resume starts at the given offset, sliced from cache — **no repetition of already-read text** (measured 0+500 → 500+500, no overlap); offset past the end returns empty instead of repeating the head;
 3. **URL fragment reading (v1.4.0)** — `url#section-anchor` jumps straight to the target section of a long reference page: container anchors slice out exactly that block (depth-counted balanced extraction), heading anchors read from the heading onward; the cache key keeps the fragment so `#a`/`#b` never cross-pollute, and offset continues relative to the section start. Measured on python-docs `#str.startswith`: **235,393 → 111,587 chars (-52.6%)**, the default 6000-char window landing on the target method instead of the document head;
 4. **Hidden-element & consent-banner stripping (v1.4.0)** — `display:none`/`visibility:hidden` decoration trees, the bare `hidden` attribute, `aria-hidden="true"`, and OneTrust/Cookiebot/GDPR banners mounted by id no longer leak into the body; a bounded attribute-separator guard keeps compound names (`data-hidden`/`data-id`) from false matches;
-5. **`text` mode first** — Markdown structure is opt-in;
-6. **Compact text render** — the model sees a `title:` header + body directly, no JSON parsing; `siteName` is omitted when identical to the hostname; every status hint is one short line (truncated / cached / rendered), no verbose paragraphs;
-7. **Two-tier cache** — successful results cached per URL for 5 minutes (repeat reads hit cache: fewer network calls and fewer model retries); **failed results cached for 30 seconds** so a broken URL never triggers a re-fetch loop;
-8. **KV-cache friendly (DeepSeek cost tuning)** — tool schema/description stay **static text** (no config values embedded), so changing config never invalidates the reusable prompt prefix and KV cache keeps hitting. DeepSeek's cache-hit tokens cost about 1/10 of misses — the more stable the prefix, the cheaper the run (same analysis as the official `tool-web` docs);
-9. **Batch shares the cache** — `read_url_batch` reuses the same cache (repeat batches hit it directly) and caps each page at 3,000 chars (below the single-page 6,000) to bound total output;
-10. **Compact fixed cost** — the four tool descriptions total ~900 chars (audited by a budget assertion in tests, kept static for KV-cache); extended HTML-entity decoding (45 named entities) means leftovers like `—`/`…` never waste tokens or render as mojibake.
+5. **Three-way metadata fallback (v1.5.0)** — the published-time merge chain runs `article:`/`og:` meta → JSON-LD → `<time datetime>` → generic date meta → byline, with recognized formats normalized to ISO; when an anti-bot shell leaves the body too thin, the full ld+json `articleBody` serves as a body fallback; markdown tables beyond 25 rows truncate with a `…+N rows` hint; zero-width characters (U+200B etc.) are stripped uniformly;
+6. **`text` mode first** — Markdown structure is opt-in;
+7. **Compact text render** — the model sees a `title:` header + body directly, no JSON parsing; `siteName` is omitted when identical to the hostname; every status hint is one short line (truncated / cached / rendered), no verbose paragraphs;
+8. **Two-tier cache** — successful results cached per URL for 5 minutes (repeat reads hit cache: fewer network calls and fewer model retries); **failed results cached for 30 seconds** so a broken URL never triggers a re-fetch loop;
+9. **KV-cache friendly (DeepSeek cost tuning)** — tool schema/description stay **static text** (no config values embedded), so changing config never invalidates the reusable prompt prefix and KV cache keeps hitting. DeepSeek's cache-hit tokens cost about 1/10 of misses — the more stable the prefix, the cheaper the run (same analysis as the official `tool-web` docs);
+10. **Batch shares the cache** — `read_url_batch` reuses the same cache (repeat batches hit it directly) and caps each page at 3,000 chars (below the single-page 6,000) to bound total output;
+11. **Compact fixed cost** — the four tool descriptions total ~900 chars (audited by a budget assertion in tests, kept static for KV-cache); extended HTML-entity decoding (45 named entities) means leftovers like `&mdash;`/`&hellip;` never waste tokens or render as mojibake.
 
 ## Technical notes
 

package/README.zh.md

@@ -213,12 +213,13 @@ const results = await Promise.all([
 2. **段落级智能截断 + offset 续读**——默认 6000 字符（约 3000 token），在段落边界截断保证语义完整（text 模式段落以 `\n\n` 分隔，切口落在真实段落边界而非句中），输出行仅一行 `(chars 6000/12990 — 截断，offset 续读)` 引导；续读从指定偏移开始、命中缓存切片，**不重复返回已读前文**（实测 0+500 → 500+500，无重复）；offset 越界返回空而非重复开头；
 3. **URL 片段定位阅读（v1.4.0）**——`url#section-anchor` 直接定位长文档目标小节：容器锚点精确切块（深度计数平衡提取）、标题锚点从该处起读；cache key 保留 fragment 保证 `#a`/`#b` 互不串页，offset 相对小节起点。实测 python-docs `#str.startswith`：**235,393 → 111,587 字符（-52.6%）**，默认 6000 字符窗口直接落在目标方法而非文档头；
 4. **隐藏元素与同意弹窗剥离（v1.4.0）**——`display:none`/`visibility:hidden` 装饰树、`hidden` 属性、`aria-hidden="true"`、OneTrust/Cookiebot/GDPR id 挂载的横幅不再混入正文；复合属性名（`data-hidden`/`data-id`）有界卫兵不误伤；
-5. **text 模式优先**——Markdown 结构按需开启；
-6. **紧凑文本 render**——模型直接看到 `title:` 头部 + 正文，无需解析 JSON；`siteName` 与域名相同时省略；状态提示全部一行内（截断/续读/缓存/渲染标记），无长段落废话；
-7. **双层缓存**——成功结果按 URL 缓存 5 分钟（重复读取直接命中，省网络也省模型重试）；**失败结果缓存 30 秒**（坏 URL 不会触发重复 fetch 循环）；
-8. **KV Cache 友好（DeepSeek 成本特调）**——工具 schema/description 保持**静态文本**（不嵌入配置值），配置变更不会使可复用的 prompt 前缀失效，KV 缓存持续命中。DeepSeek 缓存命中 token 价格约为未命中的 1/10，前缀越稳定越省钱（官方 `tool-web` 文档同款分析）；
-9. **批量共用缓存**——`read_url_batch` 内部复用同一套缓存，重复批量读直接命中，且每页默认 3000 字符（低于单页 6000）控制总量；
-10. **固定开销压缩**——4 个工具 description 合计约 900 字符（有断言守卫，保持静态利于 KV 缓存）；HTML 实体解码扩展至 45 个命名实体，`—`/`…` 等残留不再浪费 token 或显示为乱码。
+5. **元数据三路兜底（v1.5.0）**——发布时间合并链 `article:`/`og:` meta → JSON-LD → `<time datetime>` → 通用 date meta → byline，可识别格式归一为 ISO；反爬壳页正文过薄时自动用 ld+json `articleBody` 全文兜底；markdown 表格超 25 行截断并附 `…+N rows` 提示；零宽字符（U+200B 等）统一剥离；
+6. **text 模式优先**——Markdown 结构按需开启；
+7. **紧凑文本 render**——模型直接看到 `title:` 头部 + 正文，无需解析 JSON；`siteName` 与域名相同时省略；状态提示全部一行内（截断/续读/缓存/渲染标记），无长段落废话；
+8. **双层缓存**——成功结果按 URL 缓存 5 分钟（重复读取直接命中，省网络也省模型重试）；**失败结果缓存 30 秒**（坏 URL 不会触发重复 fetch 循环）；
+9. **KV Cache 友好（DeepSeek 成本特调）**——工具 schema/description 保持**静态文本**（不嵌入配置值），配置变更不会使可复用的 prompt 前缀失效，KV 缓存持续命中。DeepSeek 缓存命中 token 价格约为未命中的 1/10，前缀越稳定越省钱（官方 `tool-web` 文档同款分析）；
+10. **批量共用缓存**——`read_url_batch` 内部复用同一套缓存，重复批量读直接命中，且每页默认 3000 字符（低于单页 6000）控制总量；
+11. **固定开销压缩**——4 个工具 description 合计约 900 字符（有断言守卫，保持静态利于 KV 缓存）；HTML 实体解码扩展至 45 个命名实体，`&mdash;`/`&hellip;` 等残留不再浪费 token 或显示为乱码。
 
 ## 技术说明
 

package/index.js

@@ -100,10 +100,11 @@ export function decodeBuffer(buffer, contentType) {
   return { text, charset: enc }
 }
 
-// Content types this plugin can consume. Beyond HTML: JSON (data APIs) and
-// XML (RSS/Atom feeds) are read natively — see readUrl's type dispatch.
+// Content types this plugin can consume. Beyond HTML: JSON (data APIs),
+// XML (RSS/Atom feeds) and plain-text families (txt/md/csv — logs, docs,
+// datasets) are read natively — see readUrl's type dispatch.
 // Note "application/rss+xml": the separator before xml is '+', not '/'.
-const FETCHABLE_CT = /text\/html|application\/xhtml|text\/plain|\/json|[+/]xml/i
+const FETCHABLE_CT = /text\/html|application\/xhtml|text\/(?:plain|markdown|csv)|\/json|[+/]xml/i
 
 // Retry-After header (seconds form) → ms, capped; undefined when absent/invalid.
 function retryAfterMs(res) {
@@ -168,14 +169,24 @@ async function directFetchOnce(url, signal, cfg) {
 }
 
 // Direct-connect fetch with a single Retry-After-aware retry on 429/503.
-async function directFetch(url, signal, cfg) {
+// Exported for tests: the double-throttle path must always produce an { error }
+// shape — a second `{ retryAfterMs }` would otherwise leak past the race
+// filter into decodeBuffer as a bufferless "success" (bare TypeError there).
+export async function directFetch(url, signal, cfg) {
   const first = await directFetchOnce(url, signal, cfg)
   if (first && first.retryAfterMs !== undefined && !(signal && signal.aborted)) {
     await new Promise((resolve) => {
       const t = setTimeout(resolve, first.retryAfterMs)
       if (signal) signal.addEventListener('abort', () => { clearTimeout(t); resolve() }, { once: true })
     })
-    return directFetchOnce(url, signal, cfg)
+    if (signal && signal.aborted) return { error: 'cancelled' }
+    const second = await directFetchOnce(url, signal, cfg)
+    // The retry budget is spent — report the rate limit instead of returning
+    // a shapeless { retryAfterMs } (no buffer, no error) to the caller.
+    if (second && second.retryAfterMs !== undefined) {
+      return { error: 'HTTP 429/503 rate limited (retried once after Retry-After, still throttled)' }
+    }
+    return second
   }
   return first
 }
@@ -329,6 +340,10 @@ function defuseLt(html) {
   return html.slice(0, from) + html.slice(from).replace(/</g, ' ')
 }
 
+// Invisible-to-readers joiner characters (zero-width space/joiners, word
+// joiner, BOM) cost tokens and leak into copy/paste text — strip at render.
+const ZW_RE = /[\u200b-\u200d\u2060\ufeff]/g
+
 function textOnly(html) {
   return decodeTextEntities(defuseLt(html)
     .replace(/<!--[\s\S]*?-->/g, ' ')
@@ -338,6 +353,7 @@ function textOnly(html) {
     .replace(/<noscript[\s>][\s\S]*?<\/noscript>/gi, ' ')
     .replace(/<[^>]{1,1000}>/g, ' ')
     .replace(/[ \t\r\n]+/g, ' ')
+    .replace(ZW_RE, '')
     .trim())
 }
 
@@ -357,7 +373,7 @@ function textLines(html) {
     .replace(/<\/?(?:p|div|section|article|main|h[1-6]|li|ul|ol|table|tbody|thead|tfoot|tr|blockquote|pre|figcaption|dt|dd|dl|figure|fieldset|option)\b[^>]{0,1000}>/gi, '\n')
     .replace(/<[^>]{1,1000}>/g, ' '))
   const lines = []
-  for (const rawLine of cleaned.split('\n')) {
+  for (const rawLine of cleaned.replace(ZW_RE, '').split('\n')) {
     let line = rawLine.replace(/[ \t\r\f\v]+/g, ' ').trim()
     // A heading becomes its own short paragraph in this layout; a trailing
     // permalink glyph on one is anchor decoration, not content.
@@ -400,7 +416,10 @@ function parseFeed(xml, limit) {
     const linkTag = /<link[^>]*>([\s\S]*?)<\/link>/i.exec(it)
     const linkHref = /<link[^>]+href=["']([^"']+)["']/i.exec(it)
     const link = (linkTag && xmlText(linkTag[1])) || (linkHref && linkHref[1]) || ''
-    const desc = xmlText(/<(?:description|summary|content)[^>]*>([\s\S]*?)<\/(?:description|summary|content)>/i.exec(it)?.[1] || '')
+    // Namespaced fields (WordPress <content:encoded>) pair via a captured
+    // name + backreference: the closer must match the SAME name, so a
+    // <content:encoded> opener can never pair with a distant </description>.
+    const desc = xmlText(/<((?:description|summary|content)(?::[a-zA-Z0-9]+)?)(?:\s[^>]*)?>([\s\S]*?)<\/\1\s*>/i.exec(it)?.[2] || '')
     if (!title && !link && !desc) continue
     items.push({ title: title.slice(0, 120), url: link, summary: desc.slice(0, 200) })
   }
@@ -453,13 +472,16 @@ export function densityFilter(html) {
 
 // Depth-counted block extraction: returns html from the opening tag at
 // openIdx through its BALANCED closing tag. A non-greedy regex would stop at
-// the first nested </div>, truncating heavily nested container divs.
-function tagBlockAt(html, openIdx, tagName) {
+// the first nested </div>, truncating heavily nested container divs. maxScan
+// bounds the search window — a closer further out than that is treated as
+// absent (keeps hostile many-unclosed-opens inputs linear).
+function tagBlockAt(html, openIdx, tagName, maxScan = Infinity) {
   const re = new RegExp(`</?${tagName}(?=[\\s>])`, 'gi')
   re.lastIndex = openIdx
   let depth = 0
   let m
   while ((m = re.exec(html))) {
+    if (m.index - openIdx > maxScan) return null
     depth += m[0][1] === '/' ? -1 : 1
     if (depth === 0) {
       const close = html.indexOf('>', m.index)
@@ -469,6 +491,29 @@ function tagBlockAt(html, openIdx, tagName) {
   return null
 }
 
+// Remove every element matched by an OPEN-TAG-ONLY pattern together with its
+// balanced content. The pattern must capture the tag name (group 1) so the
+// scanner counts only same-name open/close pairs. Unclosed matches (no closer
+// within the 100k scan window) are left in place — fail-open. Each match
+// continues from the removed block's end, so removed interiors are never
+// rescanned and total cost stays near-linear in the surviving input.
+const BALANCED_SCAN_WINDOW = 100000
+function stripBalanced(html, openTagRe) {
+  openTagRe.lastIndex = 0
+  let out = ''
+  let last = 0
+  let m
+  while ((m = openTagRe.exec(html))) {
+    const tag = m[1].toLowerCase()
+    const block = tagBlockAt(html, m.index, tag, BALANCED_SCAN_WINDOW)
+    if (!block) continue // unclosed container — keep the content
+    out += html.slice(last, m.index) + ' '
+    last = m.index + block.length
+    openTagRe.lastIndex = last
+  }
+  return out + html.slice(last)
+}
+
 export function pickMain(html) {
   // Aggregation pages (e.g. blog homepages) have one <article> per item:
   // collect all of them instead of only the first. A tiny <article> (e.g. a
@@ -540,11 +585,11 @@ function stripNoise(mainHtml) {
   // body. The [\s"'] guard keeps `hidden`/`style` from matching inside
   // compound attr names (data-hidden, aria-hidden). Consent/GDPR banners mount
   // by id (onetrust, cookiebot, cybot, gdpr) rather than class, so a second
-  // pass keys on the id attribute. Every scan run is bounded.
-  const hidden = /<(div|span|section|p|ul|table)\b[^>]{0,1000}[\s"'](?:hidden\b|aria-hidden\s*=\s*["']?true|style\s*=\s*["'][^"']{0,200}(?:display:\s*none|visibility:\s*hidden))[^>]{0,1000}>[\s\S]{0,50000}?<\/\1\s*>/gi
-  out = out.replace(hidden, ' ')
-  const consent = /<([a-z][a-z0-9]*)\s+id=["'][^"']{0,120}(?:onetrust|cookiebot|cybot|gdpr|consent|cookie-law|cookie_banner|cmp-)[^"']{0,120}["'][^>]{0,1000}>[\s\S]{0,50000}?<\/\1\s*>/gi
-  out = out.replace(consent, ' ')
+  // pass keys on the id attribute. Removal is a depth-counted balanced scan —
+  // banners are deeply nested and a lazy `[\s\S]*?` would stop at the first
+  // inner closer, leaking the rest of the banner text.
+  out = stripBalanced(out, /<(div|span|section|p|ul|table)\b[^>]{0,1000}[\s"'](?:hidden\b|aria-hidden\s*=\s*["']?true|style\s*=\s*["'][^"']{0,200}(?:display:\s*none|visibility:\s*hidden))[^>]{0,1000}>/gi)
+  out = stripBalanced(out, /<([a-z][a-z0-9]*)\s+id=["'][^"']{0,120}(?:onetrust|cookiebot|cybot|gdpr|consent|cookie-law|cookie_banner|cmp-)[^"']{0,120}["'][^>]{0,1000}>/gi)
   return out.replace(
     /<([a-z][a-z0-9]*)[^>]{1,1000}class=["'][^"']{0,300}(ad-|ads|advert|banner|sidebar|social|share|comment|popup|modal|cookie)[^"']{0,300}["'][^>]{0,1000}>[\s\S]*?<\/\1>/gi,
     ' ',
@@ -563,10 +608,29 @@ function escInline(s) {
 const IMG_SENTINEL = '\u0001'
 function imgsToMarkdown(html) {
   const store = []
-  const out = html.replace(/<img\b[^>]{0,1000}>/gi, (tag) => {
+  // AMP pages use <amp-img> instead of <img>; responsive layouts wrap images
+  // in <picture><source srcset>. When the <img> fallback carries no usable
+  // src (lazy placeholder), the first source URL is injected into it so the
+  // normal img pass below still emits the image.
+  html = html.replace(/<picture\b[^>]*>([\s\S]{0,50000}?)<\/picture\s*>/gi, (block) => {
+    const imgTag = /<img\b[^>]{0,1000}>/i.exec(block)
+    const imgSrc = imgTag && /(?:^|\s)src=["']([^"']+)["']/i.exec(imgTag[0])
+    if (imgTag && imgSrc && !/^data:/i.test(imgSrc[1])) return block
+    const source = /<source\b[^>]{0,1000}\bsrcset=["']([^"'\s>]+)/i.exec(block)
+    if (!source || !imgTag) return block
+    return block.replace(/<img\b[^>]{0,1000}>/i, (img0) => img0.replace(/<img/i, `<img src="${source[1]}"`))
+  })
+  const out = html.replace(/<(?:amp-)?img\b[^>]{0,1000}>/gi, (tag) => {
     const alt = /alt=["']([^"']*)["']/i.exec(tag)
-    const src = /src=["']([^"']+)["']/i.exec(tag)
-    store.push(src && alt && alt[1].trim() ? `![${alt[1].trim()}](${src[1]})` : '')
+    let src = /src=["']([^"']+)["']/i.exec(tag)
+    let srcVal = src && src[1]
+    // Lazy-load wrappers keep the real URL in data-* attributes; a data: URI
+    // src is an inline placeholder spacer, not a usable image reference.
+    if (!srcVal || /^data:/i.test(srcVal)) {
+      const lazy = /data-(?:src|original|lazy-src)=["']([^"']+)["']/i.exec(tag)
+      if (lazy) srcVal = lazy[1]
+    }
+    store.push(srcVal && alt && alt[1].trim() ? `![${alt[1].trim()}](${srcVal})` : '')
     return `${IMG_SENTINEL}${store.length - 1}${IMG_SENTINEL}`
   })
   return {
@@ -588,6 +652,7 @@ function imgsToMarkdown(html) {
 // the walker degrades to plain text.
 const OPEN_TAG_RE = /<([a-zA-Z0-9]+)((?:"[^"]*"|'[^']*'|[^'">]){0,1000})>/g
 const MD_MAX_DEPTH = 100
+const MD_TABLE_MAX_ROWS = 25
 function stripUnmatchedOpeners(html) {
   const closers = new Set()
   for (const m of html.matchAll(/<\/([a-zA-Z0-9]+)/g)) closers.add(m[1].toLowerCase())
@@ -713,7 +778,12 @@ export function blockMd(html, depth = 0) {
         // unescape cells' escaped pipes before deriving the separator row,
         // otherwise the --- line is one char wider per escaped pipe
         const sep = header.replace(/\\\|/g, '|').replace(/[^|]/g, '-')
-        out += `\n\n${rows.join('\n')}\n${sep}`
+        // Compatibility tables (40+ rows × many columns) are the single
+        // largest token sink in markdown mode; the hint keeps the model
+        // aware the table continues without repeating the cell noise.
+        const kept = rows.length > MD_TABLE_MAX_ROWS ? rows.slice(0, MD_TABLE_MAX_ROWS) : rows
+        const more = rows.length - kept.length
+        out += `\n\n${kept.join('\n')}\n${sep}${more > 0 ? `\n…+${more} rows` : ''}`
       }
     } else if (tag === 'a' || tag === 'strong' || tag === 'b' || tag === 'em' || tag === 'i') {
       const t = inlineMd(inner, depth + 1)
@@ -723,7 +793,7 @@ export function blockMd(html, depth = 0) {
       if (t) out += t
     }
   }
-  return img.restore(out)
+  return img.restore(out).replace(ZW_RE, '')
 }
 
 // Read the content attribute of the FIRST <meta> whose property/name equals
@@ -818,6 +888,76 @@ function jsonLdMeta(html) {
   return { published: '', author: '' }
 }
 
+// News portals frequently ship the FULL article text inside ld+json
+// (articleBody) while the visible HTML is an anti-bot shell or a thin
+// fragment — a fallback body source the DOM extractors cannot reach. Returns
+// the first string articleBody found (array values joined), stripped to
+// paragraphs via textLines. Block scan is the same bounded LDJSON_RE.
+function jsonLdArticleBody(html) {
+  for (const m of html.matchAll(LDJSON_RE)) {
+    let json
+    try {
+      json = JSON.parse(m[1])
+    } catch {
+      continue
+    }
+    const nodes = []
+    const collect = (node, depth = 0) => {
+      if (!node || typeof node !== 'object' || depth > 20) return
+      if (Array.isArray(node)) {
+        for (const it of node) collect(it, depth + 1)
+        return
+      }
+      nodes.push(node)
+      for (const k of Object.keys(node)) {
+        const v = node[k]
+        if (v && typeof v === 'object') collect(v, depth + 1)
+      }
+    }
+    collect(json)
+    for (const n of nodes) {
+      const raw = Array.isArray(n.articleBody) ? n.articleBody.join('\n\n') : n.articleBody
+      if (typeof raw === 'string' && raw.trim()) return textLines(raw).slice(0, 200000)
+    }
+  }
+  return ''
+}
+
+// SPA shells ship their SEO copy inside <noscript> (the block the stripNoise
+// pass removes from main content). A thin body falls back to the LONGEST
+// noscript block when it carries substantive text; short ones are UI noise
+// ("please enable JavaScript"). Unclosed blocks simply never match — fail-open.
+function noscriptBody(html) {
+  let best = ''
+  for (const m of html.matchAll(/<noscript[^>]{0,500}>([\s\S]{0,50000}?)<\/noscript\s*>/gi)) {
+    const t = textLines(m[1])
+    if (t.length > best.length) best = t
+  }
+  return best.length >= 150 ? best : ''
+}
+
+// Semantic <time datetime="..."> — common on blogs/docs that ship no meta or
+// ld+json date. Tags explicitly marked as publication time (itemprop/pubdate)
+// win; otherwise the FIRST time tag in the scoped html is used. Callers pass
+// the noise-stripped main content so footer/comment timestamps (which carry
+// their own <time> tags) are out of scope. Empty attribute is a no-op.
+function timeTagMeta(html) {
+  if (!html) return ''
+  const marked = /<time\b[^>]{0,1000}(?:itemprop=["']datePublished["']|pubdate\b)[^>]{0,1000}datetime=["']([^"']{1,60})["']/i.exec(html)
+  if (marked) return marked[1].trim().slice(0, 40)
+  const m = /<time\b[^>]{0,1000}datetime=["']([^"']{1,60})["']/i.exec(html)
+  return m ? m[1].trim().slice(0, 40) : ''
+}
+
+// CJK / loose-variant dates normalize to ISO so the metadata line stays
+// comparable across sources ("2026年8月24日" → "2026-08-24"). Anything that
+// does not match a recognized shape passes through untouched.
+function normalizeDate(s) {
+  const m = /^(\d{4})\s*[年/\-.]\s*(\d{1,2})\s*[月/\-.]\s*(\d{1,2})\s*日?$/.exec((s || '').trim())
+  if (!m) return s
+  return `${m[1]}-${m[2].padStart(2, '0')}-${m[3].padStart(2, '0')}`
+}
+
 // Meta-less pages still print the byline inside the body head: "作者：阮一峰
 // 日期：2026年8月21日". Fill ONLY the missing fields, and only from the first
 // 600 chars of body text — deeper matches are content, not bylines.
@@ -833,7 +973,7 @@ function bylineFallback(text, meta) {
   }
   if (!out.published) {
     const m = BYLINE_DATE_RE.exec(head)
-    if (m) out.published = m[1].replace(/\s+/g, ' ').trim().slice(0, 40)
+    if (m) out.published = normalizeDate(m[1].replace(/\s+/g, ' ').trim().slice(0, 40))
   }
   return out
 }
@@ -897,13 +1037,20 @@ export function extract(html, mode, anchor = '') {
   }
   const meta = extractMeta(html)
   const ld = jsonLdMeta(html)
-  // Merge chain: <meta> → JSON-LD → byline. An explicit article:/og: meta date
-  // stays authoritative; otherwise a JSON-LD datePublished beats the generic
-  // date meta key (last-modified stamps are not publication times).
-  const published = ld.published
-    ? (meta.publishedExplicit ? meta.published : ld.published)
-    : meta.published
+  // Merge chain: explicit article:/og: meta date → JSON-LD datePublished →
+  // <time datetime> (semantic markup usually marks publication, unlike the
+  // generic date meta key which is often last-modified) → generic date meta →
+  // byline text. jsonLdArticleBody doubles as a thin-body fallback below.
+  const published = meta.publishedExplicit
+    ? meta.published
+    : normalizeDate(ld.published || timeTagMeta(main) || meta.published)
   const author = ld.author || meta.author
+  // Thin-body fallback: anti-bot shells and JS-only pages still carry the
+  // full article in ld+json — prefer it over the short og:description hint.
+  // An anchored read scopes the answer to one section: the ld+json body is
+  // the WHOLE article and would override the located section, so it applies
+  // only to full-document reads.
+  const ldBody = !scoped && bodyText.length < 200 ? jsonLdArticleBody(html) : ''
   // Empty-body pages (login walls, JS-only shells) still carry og:description
   // — surface it as a fallback hint instead of nothing.
   let text = bodyText
@@ -912,6 +1059,17 @@ export function extract(html, mode, anchor = '') {
     text = meta.description
     usedDescription = true
   }
+  if (ldBody && ldBody.length > text.length) {
+    text = ldBody
+    usedDescription = false
+  }
+  // Same thin-body shape, next fallback layer: SPA shells put their SEO copy
+  // in <noscript> (stripped from main) — surface it when substantive.
+  const nosBody = !scoped && text.length < 200 ? noscriptBody(html) : ''
+  if (nosBody && nosBody.length > text.length) {
+    text = nosBody
+    usedDescription = false
+  }
   const byline = bylineFallback(text, { published, author })
   return {
     title: titleTag ? textOnly(titleTag[1]) || titleTag[1].trim() : ogTitle,
@@ -1227,9 +1385,13 @@ async function followMetaRefresh(html, finalUrl, externalSignal, cfg) {
   let curUrl = finalUrl
   let curCharset = '' // set only when a hop actually happened
   const seen = new Set([normalizeUrl(finalUrl)])
+  // Chain budget: each hop carries its own fetch timeout, so 3 hops could
+  // triple the call's cost — the whole follow chain shares ONE timeout.
+  const t0 = Date.now()
   for (let i = 0; i < MAX_META_REFRESH_HOPS; i++) {
     const target = metaRefreshTarget(cur, curUrl)
     if (!target) break
+    if (Date.now() - t0 > cfg.timeoutMs) break
     const norm = normalizeUrl(target)
     if (!norm || seen.has(norm)) break
     seen.add(norm)
@@ -1251,6 +1413,10 @@ export async function readUrl(args, ctx, externalSignal, cfg = DEFAULTS) {
   const maxChars = Math.max(500, Math.min(20000, Number(args.maxChars) || cfg.maxChars))
   const mode = args.mode === 'markdown' ? 'markdown' : 'text'
   const offset = Math.max(0, Number(args.offset) || 0)
+  // Wall-clock anchor for the pagination budget below (the tool timeout
+  // formula reserves paginateMax+1 fetches + render slack; the in-code guard
+  // keeps a slow page from consuming the budget meant for the rest).
+  const startedAt = Date.now()
 
   // Cache stores the FULL extracted text keyed by url+mode+includeLinks, so
   // continuation reads (offset) and different maxChars hit the same entry —
@@ -1278,11 +1444,21 @@ export async function readUrl(args, ctx, externalSignal, cfg = DEFAULTS) {
   // broken URL in a loop (token + latency saver).
   const fh = failCache.get(cacheKey)
   if (fh) {
-    if (Date.now() - fh.time < 30000) return { error: fh.error, cached: true }
+    if (Date.now() - fh.time < 30000) {
+      // LRU touch: re-insert as the newest slot so repeated failures on a
+      // busy session don't evict (and re-fetch) this URL before its TTL ends.
+      failCache.delete(cacheKey)
+      failCache.set(cacheKey, fh)
+      return { error: fh.error, cached: true }
+    }
     failCache.delete(cacheKey)
   }
   const hit = cache.get(cacheKey)
   if (hit && Date.now() - hit.time < cfg.cacheTtlMs) {
+    // LRU touch: hot pages survive eviction even when cached long ago — the
+    // Map re-insertion moves the entry to the newest slot.
+    cache.delete(cacheKey)
+    cache.set(cacheKey, hit)
     const sliced = sliceFrom(hit.full, offset, maxChars)
     return { ...sliced, cached: true }
   }
@@ -1292,16 +1468,26 @@ export async function readUrl(args, ctx, externalSignal, cfg = DEFAULTS) {
     return { error }
   }
 
-  // Non-HTML payloads (JSON APIs, RSS/Atom feeds) are read natively: fetch,
-  // dispatch by content-type / payload shape, render compact. Falls back to
-  // the HTML pipeline for everything else.
+  // Non-HTML payloads (JSON APIs, RSS/Atom feeds, plain text files) are read
+  // natively: fetch, dispatch by content-type / payload shape, render compact.
+  // Falls back to the HTML pipeline for everything else.
+  // JSON with JS-only literals (NaN/Infinity — hand-rolled serializers) is
+  // not valid JSON: retry once with the literals nulled rather than dropping
+  // the whole page into the HTML pipeline as a raw-text dump.
+  const parseJsonLoose = (t) => {
+    try {
+      return JSON.parse(t)
+    } catch {
+      return JSON.parse(t.replace(/-?(?:NaN|Infinity)\b/g, 'null'))
+    }
+  }
   const dispatch = async () => {
     const viaSeam = await fetchViaWebSeam(ctx, url, externalSignal, cfg)
     if (viaSeam) {
       const t = viaSeam.html.trim()
       if (t.startsWith('{') || t.startsWith('[')) {
         try {
-          return { kind: 'json', text: compactJson(JSON.parse(t)) }
+          return { kind: 'json', text: compactJson(parseJsonLoose(t)) }
         } catch { /* not JSON — keep HTML pipeline */ }
       }
       return { kind: 'html', html: viaSeam.html, finalUrl: viaSeam.finalUrl, charset: 'provider-decoded' }
@@ -1312,7 +1498,7 @@ export async function readUrl(args, ctx, externalSignal, cfg = DEFAULTS) {
     const decoded = decodeBuffer(page.buffer, page.contentType)
     if (/json$/.test(ct)) {
       try {
-        return { kind: 'json', text: compactJson(JSON.parse(decoded.text)) }
+        return { kind: 'json', text: compactJson(parseJsonLoose(decoded.text)) }
       } catch { /* invalid JSON — serve through the HTML pipeline as text */ }
     }
     if (/[+/]xml$/.test(ct)) {
@@ -1323,11 +1509,30 @@ export async function readUrl(args, ctx, externalSignal, cfg = DEFAULTS) {
         return { kind: 'feed', feed: parseFeed(decoded.text, cfg.maxLinks) }
       }
     }
+    // Plain-text families (txt/md/csv): line structure IS the content — the
+    // HTML pipeline would flatten every newline into spaces. Pages that are
+    // actually HTML but mislabelled text/plain keep the HTML pipeline.
+    if (/^text\/(?:plain|markdown|csv)$/.test(ct)) {
+      const head = decoded.text.slice(0, 5000)
+      if (!/<(?:html|body|div|p|table)\b[\s>]/i.test(head)) {
+        const text = decoded.text
+          .replace(/\r\n?/g, '\n')
+          .replace(/[ \t]+\n/g, '\n')
+          .replace(/\n{4,}/g, '\n\n\n')
+          .trim()
+        return { kind: 'text', text, charset: decoded.charset }
+      }
+    }
     return { kind: 'html', html: decoded.text, finalUrl: page.finalUrl || url, charset: decoded.charset }
   }
   const payload = await dispatch()
   if (payload.error) return failWith(payload.error)
 
+  if (payload.kind === 'text') {
+    const full = { url, title: '', siteName: hostOf(url), lang: '', charset: payload.charset, mode: 'text', fullText: payload.text, paginated: 1 }
+    cacheStore(cacheKey, full, cfg.cacheMax)
+    return sliceFrom(full, offset, maxChars)
+  }
   if (payload.kind === 'json') {
     const full = { url, title: '', siteName: hostOf(url), lang: '', charset: 'json', mode: 'json', fullText: payload.text, paginated: 1 }
     cacheStore(cacheKey, full, cfg.cacheMax)
@@ -1378,8 +1583,8 @@ export async function readUrl(args, ctx, externalSignal, cfg = DEFAULTS) {
     }
     return target
   }
-  // Readability would replace the scoped section with the FULL article text;
-  // when an anchor matched, the scoped extraction IS the answer.
+  // Readability would replace the scoped section with the FULL article text;
+  // when an anchor matched, the scoped extraction IS the answer.
   extracted = extracted.anchored ? extracted : upgrade(extracted, html)
 
   // Optional SPA enhancement: if the page looks client-rendered and static
@@ -1407,7 +1612,7 @@ export async function readUrl(args, ctx, externalSignal, cfg = DEFAULTS) {
           ? '已渲染但被人机验证页拦截，保留静态提取结果'
           : '已渲染但被人机验证页拦截（无静态内容可用）'
       } else {
-        const r2raw = extract(rr.html, mode, anchor)
+        const r2raw = extract(rr.html, mode, anchor)
         const r2 = r2raw.anchored ? r2raw : upgrade(r2raw, rr.html, rr.finalUrl || finalUrl)
         // Accept the rendered text when it meaningfully beats the static one.
         // From an EMPTY static extraction even a short body is a real gain
@@ -1435,12 +1640,16 @@ export async function readUrl(args, ctx, externalSignal, cfg = DEFAULTS) {
   const paginateMax = cfg.paginate === false ? 1 : Math.max(1, Math.min(10, Number(cfg.paginateMax) || 3))
   const startHost = hostOf(finalUrl || url)
   const seen = new Set([normalizeUrl(finalUrl || url)])
-  // An anchored read is scoped to one section — joining the "next page" chain
-  // would append whole pages the fragment never asked for.
+  // An anchored read is scoped to one section — joining the "next page" chain
+  // would append whole pages the fragment never asked for.
   let nextUrl = paginateMax > 1 && !extracted.anchored ? findNextLink(renderedHtml || html, finalUrl || url) : null
   while (
     nextUrl && paginated < paginateMax &&
     sameHost(nextUrl, startHost) &&
+    // Budget guard: the whole read (initial fetch + renders + chain) must fit
+    // the tool timeout formula (paginateMax+1 fetches). A slow first page or
+    // SPA render eats into the chain's share here instead of overshooting.
+    Date.now() - startedAt < cfg.timeoutMs * (paginateMax + 1) &&
     !(externalSignal && externalSignal.aborted)
   ) {
     const nu = normalizeUrl(nextUrl)
@@ -1696,15 +1905,34 @@ function readUrlBatchTool(ctx, cfg) {
     },
     async execute(args, exec) {
       const a = args || {}
-      const urls = (Array.isArray(a.urls) ? a.urls : []).map((u) => String(u).trim()).filter(Boolean)
+      // Duplicate URLs collapse to one fetch (parallel copies would race the
+      // cache and fetch the same page twice); order is preserved. Exact-string
+      // identity: a #fragment read is an ANCHORED read with different
+      // semantics, so it must not merge with its bare URL.
+      const seenUrl = new Set()
+      const urls = (Array.isArray(a.urls) ? a.urls : [])
+        .map((u) => String(u).trim())
+        .filter(Boolean)
+        .filter((u) => {
+          if (seenUrl.has(u)) return false
+          seenUrl.add(u)
+          return true
+        })
       if (!urls.length) return { error: 'urls array is required (1-10 http(s) URLs)' }
       const list = urls.slice(0, 10)
       const perMax = Math.max(500, Math.min(20000, Number(a.maxChars) || 3000))
       const mode = a.mode === 'markdown' ? 'markdown' : 'text'
       const signal = exec && exec.signal
-      const pages = await mapLimit(list, 4, (u) =>
-        readUrl({ url: u, maxChars: perMax, mode, includeLinks: a.includeLinks === true }, ctx, signal, cfg),
-      )
+      // One page throwing must not reject the whole batch — readUrl returns
+      // error objects for expected failures, but any unexpected exception is
+      // converted to one here too.
+      const pages = await mapLimit(list, 4, async (u) => {
+        try {
+          return await readUrl({ url: u, maxChars: perMax, mode, includeLinks: a.includeLinks === true }, ctx, signal, cfg)
+        } catch (e) {
+          return { error: `Unexpected error: ${e && e.message ? e.message : e}` }
+        }
+      })
       const ok = pages.filter((p) => !p.error)
       return {
         total: list.length,
@@ -1777,11 +2005,19 @@ async function crawlSite(entryUrl, cfg, opts, externalSignal) {
         visited.add(norm)
         const page = await fetchPage(url, externalSignal, cfg)
         if (page.error) return { url, depth, error: page.error }
-        const html = decodeBuffer(page.buffer, page.contentType).text
+        let html = decodeBuffer(page.buffer, page.contentType).text
+        let pageUrl = page.finalUrl || url
+        // Same shell-following as read_url: a meta-refresh stub would otherwise
+        // be crawled AS the page (title-less hop text pollutes results and the
+        // real content is never reached). Fails open — fetch problems keep the
+        // stub's own HTML.
+        const followed = await followMetaRefresh(html, pageUrl, externalSignal, cfg)
+        html = followed.html
+        pageUrl = followed.finalUrl
         const ex = extract(html, 'text')
-        const links = extractLinks(html, 100, page.finalUrl || url)
+        const links = extractLinks(html, 100, pageUrl)
         return {
-          url: page.finalUrl || url,
+          url: pageUrl,
           depth,
           title: ex.title || '',
           chars: ex.text.length,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-read-url",
-  "version": "1.4.0",
+  "version": "1.6.1",
   "description": "DeepSeek Harness URL reader: fetch any webpage (HTML/JSON/RSS/Atom), auto-detect encoding (UTF-16 BOM/GBK/Big5/Shift-JIS), extract clean main content with auto-pagination, output compact text/Markdown to save tokens. Zero runtime dependencies, no API key, drop-in. Follows meta-refresh shells, reads JSON-LD metadata, honors base href.",
   "type": "module",
   "main": "index.js",

package/spa.js

@@ -58,13 +58,13 @@ export async function renderPage(url, externalSignal) {
     // 'domcontentloaded' instead of 'networkidle': heartbeat-polling sites
     // (qq-news, juejin) never go idle and would time out at 30s. Instead wait
     // for the DOM to stabilize (content stops growing) up to 10s — SPA paint
-    // usually lands within a couple of seconds of DOM-ready.
-    // The cooperative signal aborts both the navigation and the poll loop, so
-    // a cancelled/timed-out call releases the browser immediately instead of
-    // burning the full poll window.
-    await page.goto(url, { waitUntil: 'domcontentloaded', timeout: 30000, signal: externalSignal || undefined })
+    // usually lands within a couple of seconds of DOM-ready. Two consecutive
+    // evaluate failures mean the page crashed — stop waiting for growth that
+    // can never come.
+    await page.goto(url, { waitUntil: 'domcontentloaded', timeout: 30000 })
     const t0 = Date.now()
     let prevLen = -1
+    let evalFails = 0
     for (;;) {
       if (externalSignal && externalSignal.aborted) return { error: 'cancelled' }
       await page.waitForTimeout(500)
@@ -72,6 +72,7 @@ export async function renderPage(url, externalSignal) {
         .evaluate(() => (document.body ? document.body.innerHTML.length : 0))
         .catch(() => -1)
       if (Date.now() - t0 > 10000) break
+      if (len === -1 && ++evalFails >= 2) break
       // stop once two consecutive reads agree — including empty bodies (a
       // blank page should not burn the full 10s poll)
       if (len === prevLen && prevLen >= 0) break