sophhub 0.2.3 → 0.2.4

package/skills/large-task-router/skill.json

@@ -0,0 +1,20 @@
+{
+  "name": "large-task-router",
+  "version": "1.0.0",
+  "types": [
+    "builtin"
+  ],
+  "displayName": "",
+  "description": "",
+  "changelog": [
+    {
+      "version": "1.0.0",
+      "date": "2026-04-16",
+      "changes": [
+        "初次提交"
+      ]
+    }
+  ],
+  "createdAt": "2026-04-16",
+  "updatedAt": "2026-04-16"
+}

package/skills/large-task-router/src/SKILL.md

@@ -0,0 +1,79 @@
+---
+name: large-task-router
+description: "大任务拆分规划。触发：单文件≥20MB、合计≥50MB、≥2 文件同类处理、用户提到并行/分批/拆分/subagent。命中时先 read SKILL.md 后按 Subagent Plan 执行。"
+---
+
+# Large Task Router
+
+## 目标
+
+把"大任务先拆分后执行"变成**强制**固定流程，避免直接串行处理导致慢、易中断、难恢复。本技能一旦被加载进上下文，触发条件匹配即自动生效，不再需要用户重申。
+
+## 何时启用（命中任一即必须启用）
+
+- 输入文件总大小 > 50MB
+- 任意单文件 > 20MB（即使只有一个文件）
+- 预计步骤 > 8
+- 预计耗时 > 10 分钟
+- 用户明确或隐含提到：并行 / 分批 / 拆分 / subagent / 多 agent / 同时处理 / 一起跑 / 并发
+
+> **关键**：以上条件是**客观触发器**，不是建议。不要用"任务简单"、"一个子任务就够"、"我直接跑更快"等理由跳过。触发条件成立即必须输出 Subagent Plan，即使 subagent 数量 = 1。
+
+## 固定流程（必须按顺序）
+
+1. **规模评估**
+   - 估算：文件大小、文件数量、步骤数、预估时长
+   - 输出一行结论：`任务规模=小|中|大`
+   - 若任一触发条件命中，`任务规模` 至少为 `中`，进入步骤 2
+2. **拆分计划**
+   - 先完整输出 "Subagent Plan" 模板（参考 `templates/plan.md`），再开始执行
+   - 子任务命名统一为 `subagent-1..n`
+   - **即使 n = 1，也要填写完整模板**，显式说明"无法进一步拆分的理由"
+3. **并行执行**
+   - 按计划并发执行
+   - 子任务失败可重试 1 次（同参数）
+4. **聚合汇总**
+   - 合并成功结果
+   - 标注失败项、重试结果、遗留风险
+5. **最终报告**
+   - 给出结论 + 可执行下一步
+
+## 并发策略
+
+- 默认并发数：`N = min(4, 子任务数)`
+- 若环境允许且任务明显可并行：可提高到 `N = 6~8`
+- 避免过度并发：当任务共享同一资源（同一大文件写入、同一数据库表）时改为分阶段并行
+
+## 拆分模板（按优先级选择）
+
+1. **按文件切块**：大文件按块/页/段/行数范围拆分（单文件 > 20MB 时首选此模式）
+2. **按阶段拆分**：采集 → 清洗 → 分析 → 汇总
+3. **按目标拆分**：A 子任务做提取，B 子任务做校验，C 子任务做格式化
+
+每个子任务必须包含：
+
+- `objective`：目标
+- `input`：输入范围
+- `output`：期望产出
+- `timeout`：超时限制
+- `success_criteria`：成功判定
+
+## 输出格式
+
+先使用 `templates/plan.md` 输出计划，再执行任务。
+
+最终报告必须包含：
+
+- `Summary`：总体结果
+- `Completed`：完成项
+- `Failed/Retried`：失败与重试
+- `Risks`：风险与边界
+- `Next Actions`：建议下一步
+
+## 护栏（硬性规则，不得绕过）
+
+- **禁止直接开跑**：触发条件命中的任务不得跳过 Subagent Plan 阶段，直接调用工具/读写文件。
+- **禁止自主降级**：不得以"任务简单 / 不需要拆分 / 我评估不用" 为由拒绝本流程。规模评估仅决定拆分粒度（subagent 数量），不决定是否输出计划。
+- **禁止静默失败**：子任务失败必须显式列入最终报告的 `Failed/Retried`，并说明重试结果。
+- **单文件大**：当触发器是"单文件 > 20MB"时，仍需输出 Subagent Plan；若该文件确实无法按块/行/段拆分，在计划中说明理由后可仅用 1 个 subagent 执行，但流程不变。
+- **关键词即触发**：用户提到"并行/分批/拆分"等关键词时，即使规模小，也要输出简化版 Subagent Plan（可 1~2 个 subagent）以确认并行意图，避免误解为串行。

package/skills/large-task-router/src/templates/plan.md

@@ -0,0 +1,74 @@
+# Subagent Plan
+
+## 0) Task Size Assessment
+
+- Input size:
+- File count:
+- Estimated steps:
+- Estimated runtime:
+- Decision: `small | medium | large`
+- Split required: `yes | no` (if `no`, explain)
+
+## 1) Execution Strategy
+
+- Split mode: `by-file | by-stage | by-objective`
+- Planned subagents:
+- Planned concurrency (`N`):
+- Retry policy: `failed subtask retry once`
+
+## 2) Subtasks
+
+### subagent-1
+
+- objective:
+- input:
+- output:
+- timeout:
+- success_criteria:
+
+### subagent-2
+
+- objective:
+- input:
+- output:
+- timeout:
+- success_criteria:
+
+<!-- Add subagent-3..n when needed -->
+
+## 3) Merge Rules
+
+- Dedup strategy:
+- Conflict resolution:
+- Output ordering:
+
+---
+
+# Final Report
+
+## Summary
+
+- Overall result:
+- Time spent:
+- Coverage:
+
+## Completed
+
+- item 1
+- item 2
+
+## Failed/Retried
+
+- failed item:
+  - reason:
+  - retry result:
+
+## Risks
+
+- risk 1
+- risk 2
+
+## Next Actions
+
+- next step 1
+- next step 2

package/skills/skillhub/skill.json

@@ -1,12 +1,19 @@
 {
   "name": "skillhub",
-  "version": "2.0.0",
+  "version": "2.0.1",
   "types": [
     "builtin"
   ],
-  "displayName": "",
-  "description": "",
+  "displayName": "技能商店与安装",
+  "description": "在 Sophnet/Sophclaw 上搜索、列出、安装 skill（含 UI 设计等关键词）；SophHub 优先，ClawHub 兜底。",
   "changelog": [
+    {
+      "version": "2.0.1",
+      "date": "2026-04-16",
+      "changes": [
+        "扩充 SKILL.md 触发描述：覆盖「Sophnet UI 设计 skill」「再找一下」等口语检索场景"
+      ]
+    },
     {
       "version": "2.0.0",
       "date": "2026-04-09",
@@ -16,5 +23,5 @@
     }
   ],
   "createdAt": "2026-04-09",
-  "updatedAt": "2026-04-09"
+  "updatedAt": "2026-04-16"
 }

package/skills/skillhub/src/SKILL.md

@@ -1,12 +1,22 @@
 ---
 name: skillhub
-description: 当用户要求搜索、安装 skill （包括Sophnet/Sophclaw平台和clawhub社区） 时使用此 skill。
+description: 当用户要在 Sophnet / Sophclaw / OpenClaw 上查找、搜索、列出或安装 skill（技能、插件）时使用；包括「商店里有没有某某 skill」「Sophnet 有没有 UI 设计 / 前端 / 图表 / 自动化 相关 skill」「再去 SophHub 找一下」「用 npx sophhub 搜 skill」等说法。也适用于 clawhub、社区技能商店、下载 skill 到本地。与具体业务实现无关、仅涉及「发现与安装技能」时选本 skill。
 ---
 
 # SkillHub
 
 统一的技能安装入口，整合 Sophnet 平台 Sophclaw 的技能商店（SophHub，优先）和 **ClawHub**（社区商店，兜底）。
 
+## 何时应使用本 Skill（触发提示）
+
+用户可能用口语或中英文混说，只要意图是 **在商店里找 skill / 安装 skill**，即应走本流程，例如：
+
+- 「Sophnet / sophnet 有没有 **UI 设计** / UX / 界面 / 前端 相关的 skill？」「商店里有没有做海报 / 图表 / 视频的 skill？」
+- 「你再去 **找一下** skill」「在 SophHub / 技能商店 **搜一下**」「**列出** 商店里能装的技能」
+- 「**安装** 某个 skill」「从 Sophnet 平台 **下载** skill 到本地」「npx sophhub 怎么用」
+
+**不要**把「已经知道 skill 名称、只想写业务代码」误判为本 skill；但若用户明确要先在商店里 **检索、确认、安装** 再使用，仍用本 skill。
+
 ## 安装流程
 
 当用户需要安装一个技能时，按以下顺序操作：

package/skills/sophnet-dailynews/skill.json

@@ -0,0 +1,20 @@
+{
+  "name": "sophnet-dailynews",
+  "version": "1.0.0",
+  "types": [
+    "store"
+  ],
+  "displayName": "每日新闻助手",
+  "description": "支持实时新闻检索、分类、摘要与分析，以及基于 AI 的个性化新闻推荐和深度报道。",
+  "changelog": [
+    {
+      "version": "1.0.0",
+      "date": "2026-04-14",
+      "changes": [
+        "初次提交"
+      ]
+    }
+  ],
+  "createdAt": "2026-04-14",
+  "updatedAt": "2026-04-14"
+}

package/skills/sophnet-dailynews/src/SKILL.md

@@ -0,0 +1,179 @@
+---
+name: sophnet-dailynews
+description: Generate a daily, high-signal Markdown news report by scraping preset sources, filtering and deduplicating items. Use when the user asks for daily news, tech digest, today's links, or requests an automated curated report.
+---
+
+# Daily news report (Markdown)
+
+## Quick start
+
+1. Resolve the target date.
+   - Accept `YYYY-MM-DD` from the user.
+   - Default to local today.
+2. Read `sources.json` and `cache.json` **with explicit paths**.
+   - Always call the read tool with a `path` field (not `file_path`).
+   - Use these exact paths:
+     - `read(path="/Data/shutong.shan/clawd/skills/sophnet-dailynews/sources.json")`
+     - `read(path="/Data/shutong.shan/clawd/skills/sophnet-dailynews/cache.json")`
+3. Collect items in waves (Tier 1 → Tier 2 → Tier 3/browser) until the report has enough **high-quality** items.
+4. Write `NewsReport/YYYY-MM-DD-news-report.md`.
+5. Update `cache.json` (dedupe + historical stats).
+
+## Files in this skill
+
+- `sources.json`: tiers, batches, URLs, fetch method (`webfetch`/`browser`), extraction hint, enable/disable flags, and quality thresholds.
+- `cache.json`: last run metadata, per-source stats, URL/content dedupe caches, and per-day article history.
+
+## Common failure: `web_fetch` → `fetch failed`
+
+This usually means the fetch tool hit one of these:
+
+- a redirect it doesn't follow (302)
+- a method mismatch (some sites don't like `HEAD`)
+- bot protection / CDN quirks
+- response too large / slow for the tool limits
+
+This skill is tuned to avoid common failures:
+
+- Prefer **HN RSS** over HTML (`https://news.ycombinator.com/rss`).
+- Pin **HuggingFace Papers** to a concrete date URL to avoid redirects:
+  - `https://huggingface.co/papers/date/{{date}}`
+  - Replace `{{date}}` with the target date (`YYYY-MM-DD`) before fetching.
+
+## Output contract
+
+Produce exactly one Markdown file:
+
+- Path: `NewsReport/YYYY-MM-DD-news-report.md`
+- Target: `quality_thresholds.target_items` items (default 20)
+- Include only items with `quality_score >= quality_thresholds.min_score_to_include` (default 3)
+
+Each item must include:
+
+- `title`
+- `summary` (2–4 sentences, concrete and non-hypey)
+- `key_points` (max 3)
+- `url` (canonical if possible)
+- `source_id`
+- `keywords` (2–6)
+- `quality_score` (1–5, integer)
+
+## Collection workflow
+
+### 1) Initialize
+
+- Load `sources.json` → treat it as the source of truth.
+- Load `cache.json` → use it for dedupe and stats.
+- Create `NewsReport/` if missing.
+- If `NewsReport/YYYY-MM-DD-news-report.md` already exists, either:
+  - regenerate from scratch (preferred), or
+  - append only if explicitly requested.
+
+### 2) Fetch in waves (early stop)
+
+Follow the tier order in `sources.json`:
+
+- **Wave A (Tier 1 / batch_a + batch_b)**: high hit-rate sources first.
+- If you still have fewer than ~15 included items after filtering, continue.
+- **Wave B (Tier 2 / batch_a + batch_b)**: supplemental sources.
+- If you still have fewer than `target_items`, continue.
+- **Wave C (Tier 3 / browser sources)**: JS-rendered / blocked sources.
+
+Stop fetching when you have:
+
+- at least `target_items` included items, and
+- at least `quality_thresholds.early_stop_threshold` total candidates evaluated (default 25), or you have exhausted all enabled sources.
+
+### 3) Extract and normalize
+
+For each enabled source entry:
+
+- Fetch the page content using the configured `fetch_method`:
+  - `webfetch`: normal HTTP fetch.
+  - `browser`: render with a headless browser when available; otherwise skip with a recorded error.
+- Apply the `extract` hint from the source entry (examples: `top_10`, `latest_5`, `latest_issue`, `today_top_5`).
+- For each candidate item, normalize:
+  - clean title (no site suffix noise)
+  - canonical URL (strip tracking params when safe)
+  - short summary + key points
+
+**RSS-first rule (important):**
+
+- For RSS sources, **do not** `web_fetch` the item URLs.
+- Use the RSS item's title + description/summary as the source of truth.
+- If the RSS description is too thin to summarize, either:
+  - drop the item, or
+  - keep it with a one-sentence, non-speculative summary based only on the RSS description.
+- Only fetch item URLs if the source entry explicitly says `extract: fetch_items` (none do by default).
+
+If a `webfetch` attempt fails:
+
+- retry once (respect `fetch_config.webfetch.timeout_ms`)
+- then mark the source as failed and continue (do not abort the whole report)
+
+### 4) Filter, score, and dedupe
+
+Apply these rules in order:
+
+1. Reject obvious low-signal items (marketing fluff, generic science, job posts, thin announcements).
+2. Score each remaining item (`quality_score` 1–5) with a consistent rubric:
+   - 5: deeply useful + specific + actionable/insightful
+   - 4: strong signal, worth reading
+   - 3: decent, include only if you need more items
+   - 1–2: exclude
+3. Deduplicate:
+   - exact URL match (including `cache.json:url_cache`)
+   - near-duplicate title (treat ~80% similarity as duplicate; keep the higher-scored one)
+   - optional content hash match (when you have the full text)
+
+### 5) Select and sort
+
+- Keep only items with `quality_score >= min_score_to_include`.
+- Sort by `quality_score` descending.
+- Break ties by source credibility (Tier 1 > Tier 2 > Tier 3) and recency.
+- Take the top `target_items`.
+
+## Markdown template
+
+Use this structure:
+
+```markdown
+# Daily News Report (YYYY-MM-DD)
+
+> Sources used: N | Candidates evaluated: M | Included: K
+> Generated at: <local timestamp> | Skill: sophnet-dailynews
+
+---
+
+## 1. <Title>
+
+- **Summary**: ...
+- **Key Points**:
+  1. ...
+  2. ...
+  3. ...
+- **Source**: <source_id> — <url>
+- **Keywords**: `k1` `k2` `k3`
+- **Score**: 4/5
+```
+
+## Cache update rules (`cache.json`)
+
+Update these fields every run:
+
+- `last_run`: date, duration, items_collected, items_published, sources_used
+- `source_stats[source_id]`: total_fetches, success_count, avg_items_per_fetch, avg_quality_score, last_fetch, last_success
+- `url_cache.entries`: add included URLs (store timestamps; respect `_ttl_hours`)
+- `content_hashes.entries`: add hashes when available (respect `_ttl_hours`)
+- `article_history[YYYY-MM-DD]`: record the final included item list (at minimum: title + url + source_id + score)
+
+## Editing sources (`sources.json`)
+
+- Disable a flaky/low-quality source by setting `enabled: false` or moving it to `disabled`.
+- Prefer fixing extraction hints before adding new sources.
+- Keep Tier 1 small and high-signal; use Tier 2 for "fill".
+
+## Failure handling
+
+- If a source 403s on `webfetch`, try `browser` (if available) or skip and record the error.
+- If all enabled sources fail, still write a report header and an explicit "no items" section; do not silently succeed.

package/skills/sophnet-dailynews/src/cache.json

@@ -0,0 +1,151 @@
+{
+  "schema_version": "1.0",
+  "description": "Daily News Report 缓存文件，用于避免重复抓取和跟踪历史表现",
+
+  "last_run": {
+    "date": "2026-02-09",
+    "duration_seconds": 30,
+    "items_collected": 3,
+    "items_published": 3,
+    "sources_used": ["fs_blog", "hackernoon_pm"]
+  },
+
+  "source_stats": {
+    "_comment": "记录每个源的历史表现，用于动态调整优先级",
+    "hn": {
+      "total_fetches": 2,
+      "success_count": 0,
+      "avg_items_per_fetch": 0,
+      "avg_quality_score": 0,
+      "last_fetch": "2026-02-09T04:44:19Z",
+      "last_success": null,
+      "last_error": "fetch failed"
+    },
+    "hf_papers": {
+      "total_fetches": 2,
+      "success_count": 1,
+      "avg_items_per_fetch": 0,
+      "avg_quality_score": 0,
+      "last_fetch": "2026-02-09T04:44:19Z",
+      "last_success": "2026-02-09T04:44:19Z",
+      "last_error": null,
+      "note": "No new papers for 2026-02-09 date page"
+    },
+    "one_useful_thing": {
+      "total_fetches": 2,
+      "success_count": 1,
+      "avg_items_per_fetch": 0,
+      "avg_quality_score": 0,
+      "last_fetch": "2026-02-09T04:44:24Z",
+      "last_success": "2026-02-09T04:44:24Z",
+      "last_error": null,
+      "note": "Landing page only, newsletter content not accessible"
+    },
+    "paul_graham": {
+      "total_fetches": 2,
+      "success_count": 1,
+      "avg_items_per_fetch": 0,
+      "avg_quality_score": 0,
+      "last_fetch": "2026-02-09T04:44:26Z",
+      "last_success": "2026-02-09T04:44:26Z",
+      "last_error": null,
+      "note": "Minimal content returned from articles page"
+    },
+    "james_clear": {
+      "total_fetches": 2,
+      "success_count": 1,
+      "avg_items_per_fetch": 0,
+      "avg_quality_score": 0,
+      "last_fetch": "2026-02-09T04:44:32Z",
+      "last_success": "2026-02-09T04:44:32Z",
+      "last_error": null,
+      "note": "Main site only, newsletter requires email signup"
+    },
+    "fs_blog": {
+      "total_fetches": 3,
+      "success_count": 3,
+      "avg_items_per_fetch": 1.0,
+      "avg_quality_score": 5.0,
+      "last_fetch": "2026-02-09T04:44:38Z",
+      "last_success": "2026-02-09T04:44:38Z",
+      "last_error": null
+    },
+    "hackernoon_pm": {
+      "total_fetches": 3,
+      "success_count": 3,
+      "avg_items_per_fetch": 1.0,
+      "avg_quality_score": 4.0,
+      "last_fetch": "2026-02-09T04:44:43Z",
+      "last_success": "2026-02-09T04:44:43Z",
+      "last_error": null,
+      "note": "Category page fetched, extracted 2 high-quality articles"
+    },
+    "scotthyoung": {
+      "total_fetches": 2,
+      "success_count": 1,
+      "avg_items_per_fetch": 0,
+      "avg_quality_score": 0,
+      "last_fetch": "2026-02-09T04:44:48Z",
+      "last_success": "2026-02-09T04:44:48Z",
+      "last_error": null,
+      "note": "Articles page only, book and course promo without article links"
+    }
+  },
+
+  "url_cache": {
+    "_comment": "已处理的 URL 缓存，避免重复收录",
+    "_ttl_hours": 168,
+    "entries": {
+      "https://fs.blog/brain-food/january-25-2026/": {
+        "timestamp": "2026-02-09T04:44:38Z",
+        "title": "The Spectrum of Independence"
+      },
+      "https://hackernoon.com/10-proven-ways-to-reduce-misalignment-between-stakeholders-in-product-teams": {
+        "timestamp": "2026-02-09T04:44:43Z",
+        "title": "10 Proven Ways to Reduce Misalignment Between Stakeholders in Product Teams"
+      },
+      "https://hackernoon.com/why-good-products-feel-broken": {
+        "timestamp": "2026-02-09T04:44:43Z",
+        "title": "Why Good Products Feel Broken"
+      }
+    }
+  },
+
+  "content_hashes": {
+    "_comment": "内容指纹，用于去重",
+    "_ttl_hours": 168,
+    "entries": {}
+  },
+
+  "article_history": {
+    "_comment": "已收录文章的简要记录",
+    "2026-02-07": [
+      {
+        "title": "The Spectrum of Independence",
+        "url": "https://fs.blog/brain-food/january-25-2026/",
+        "source_id": "fs_blog",
+        "score": 5
+      }
+    ],
+    "2026-02-09": [
+      {
+        "title": "The Spectrum of Independence",
+        "url": "https://fs.blog/brain-food/january-25-2026/",
+        "source_id": "fs_blog",
+        "score": 5
+      },
+      {
+        "title": "10 Proven Ways to Reduce Misalignment Between Stakeholders in Product Teams",
+        "url": "https://hackernoon.com/10-proven-ways-to-reduce-misalignment-between-stakeholders-in-product-teams",
+        "source_id": "hackernoon_pm",
+        "score": 4
+      },
+      {
+        "title": "Why Good Products Feel Broken",
+        "url": "https://hackernoon.com/why-good-products-feel-broken",
+        "source_id": "hackernoon_pm",
+        "score": 4
+      }
+    ]
+  }
+}