morning-news-briefing

# 早新闻简报 v4.0 CC 设计审查驱动升级。三大修复：新闻数量保障（新增「今日要闻」摘要层）、TTS 语音版（Step 6）、Kanban v0.15 Swarm 双模式执行。 ## 🚨 Red Flags: DO NOT SKIP THIS SKILL | Excuse your brain will make | Why it's wrong | |------------------------------|----------------| | "I'll use yesterday's template, same CSS" | Style continuity gate: must auto-diff against last accepted baseline. Free-form styling = broken brand | | "I'll just run web_search for each section" | web-research-router picks the best engine per query type. Brave for broad coverage, Exa for semantic discovery, Tavily for fact grounding. Direct web_search skips all three | | "The analysis is thin today, I'll pad it" | Analysis format is locked: 前提→推理→结论 + 📈趋势 + 为什么重要. No hedging — grep for 一方面/另一方面/可能/或许 → auto-reject | | "10 articles is fine, they're all deep" | Sentinel #2 demands ≥20 independent headlines for normal runs and ≥25 for Mode B reruns. Deep analysis count does NOT substitute for 今日要闻 count. Assembly must not collapse multiple sources into one entry | | "I'll render as soon as assembly starts" | Render card MUST wait for assembly completion (parent dependency). Rendering before content = blank/stale PDF | | "TTS is optional, skip it" | v4.0 mandate: TTS generation (Step 6) is part of the pipeline. If text_to_speech tool unreachable, skip + annotate — never silently omit | ## 🔀 Execution Modes v4.0 supports two execution modes. The skill detects which environment it's running in and picks the right path. ### Mode A: Cron (scheduled, batch) Cron jobs run single-threaded without gateway dispatcher. Use shell background for search parallelism: ``` Trigger: cron scheduler (daily 08:00) ├── Step 0: Workspace setup │ mkdir -p ~/.hermes/workspaces/morning-news-{date}/{search,output,audit} │ ├── Step 1: Parallel search — PREFERRED: parallel tool calls in one turn │ 4 lanes × 3 engines = up to 12 queries, all in a single turn │ Each engine call is independent → they execute in parallel │ web_search (broad) + Exa (semantic) + Brave (cross-check) per lane │ Enrich with Tavily follow-ups where depth needed │ Direct markdown assembly from tool results — skip intermediate JSON │ Fallback: shell background jobs only if tool-call parallelism unavailable │ ├── Step 2: Assembly (sequential) │ Read search artifacts → deduplicate → structure with new layers │ Write: morning-news-{date}.md → workspace root │ ├── Step 2.5: de-slop (citation-protected) │ Run de-slop on non-citation blocks only │ ├── Step 3: Source verification │ Claim extraction → 2-path evidence → 4-level confidence on key claims │ ├── Step 4: Render PDF (two editions) │ Mobile: 430×932px + A4: 210×297mm │ ⚠️ 渲染后必须机械验证：`ls -lh output/*.pdf` 确认文件存在再进入 Step 5。 │ 不得只描述渲染结果——必须用 ls 检查文件系统。 │ ├── Step 5: Audit (7 sentinels + anti-hedging + source count) │ ├── Step 6: TTS Generation ← NEW │ Extract 摘要→分析→总结 from markdown → text_to_speech │ └── Step 7: Deliver (MEDIA: PDFs + audio to user) ``` > ⛔ **EXECUTION LAPSE GUARD**: v4.0 实测中 agent 多次读完 Mode A 后只描述 pipeline 而不执行工具调用（声称"Step 4 完成"但 PDF 未生成）。此为已知 Execution Lapse——skill 内容正确，agent 未遵循。**若 Mode A 连续失败，直接切 Mode B**。 ### Mode B: Interactive / Kanban Swarm (manual trigger) > 🚁 **PRE-FLIGHT**: 运行 `hermes kanban swarm` 前确保所有 worker（lane-zh/lane-en/lane-mixed/lane-tech）+ verifier（auditor）+ synthesizer（publisher）profile gateway 全部 running。检查：`hermes profile list`。任一 stopped → `hermes gateway start --profile <name>`。详见 `references/mode-b-operational-gotchas.md`。 > > 🎛️ **MODEL PINNING**: 若用户指定「协同 agent 都用某模型」（如 `deepseek-v4-pro`），先更新全部参与 profile 的 `model.default` 与 `fallback_providers`，再重启/kickstart gateways。不得让 fallback 滑回旧模型。DeepSeek Mode B 复跑细节见 `references/mode-b-deepseek-recovery-2026-06-05.md`。 **状态：✅ 已验证（2026-06-04）** — 并行搜索 + 验证通过，publisher gateway 需修端口配置。 When triggered interactively (not cron), use Kanban v0.15 Swarm for full multi-agent pipeline. ### ⚡ Gateway Pre-flight (MANDATORY before swarm) All worker/verifier/synthesizer profiles' gateways MUST be running. Check with `hermes profile list` — if any show `stopped`, start them: ```bash for p in lane-zh lane-en lane-mixed lane-tech auditor publisher; do hermes gateway start --profile "$p" done ``` Also verify each profile has `dispatch_in_gateway: true` in its `config.yaml` (or be prepared to run `hermes kanban dispatch` manually after swarm creation). ### Publisher profile skill setup Publisher profile needs these skill symlinks (one-time setup): ```bash mkdir -p ~/.hermes/profiles/publisher/skills/{productivity,creative,hermes} ln -sf ~/.hermes/skills/productivity/morning-news-briefing ~/.hermes/profiles/publisher/skills/productivity/ ln -sf ~/.hermes/skills/productivity/news-assembly ~/.hermes/profiles/publisher/skills/productivity/ ln -sf ~/.hermes/skills/creative/de-slop ~/.hermes/profiles/publisher/skills/creative/ ln -sf ~/.hermes/skills/hermes/tts-manager ~/.hermes/profiles/publisher/skills/hermes/ ``` ### Swarm command ⚠️ `--skill` flags MUST use full categorized paths, not short names. Short names cause `Unknown skill(s)` errors. ```bash # 实测语法 — 2026-06-04 验证（全路径 skill + gateway pre-flight 修复后） hermes kanban swarm \ --worker "lane-zh:中文搜索:productivity/morning-news-briefing,web-research-router" \ --worker "lane-en:英文搜索:productivity/morning-news-briefing,web-research-router" \ --worker "lane-mixed:市场数据:productivity/morning-news-briefing,web-research-router" \ --worker "lane-tech:科技新闻:productivity/morning-news-briefing,web-research-router" \ --verifier auditor \ --synthesizer publisher \ "生成 {date} 早新闻简报：四路并行搜索 → 汇编 → 深度分析 → 渲染 PDF → TTS → 交付" ``` After swarm creation, check task status. If publisher sits in `ready` (not auto-dispatched), run: ```bash hermes kanban dispatch ``` **实测结果（2026-06-04）**： | Worker | 产出 | 状态 | |--------|------|:--:| | lane-zh (flash) | 20 条中文新闻 | ✅ | | lane-en (flash) | 18 条英文新闻 6 板块 | ✅ | | lane-mixed (flash) | 8 板块市场数据 | ✅ | | lane-tech (flash) | 10 条科技新闻 | ✅ | | auditor (pro) | PASS — 仅 1 处轻微不一致 | ✅ | | publisher (pro) | 合成+渲染+TTS | ⚠️ 54 次崩溃 | **Publisher 崩溃根因**：`platforms.api_server.extra.port` 需为每个 profile 分配独立端口（当前默认为 8460，与 default 冲突）。 修复方法：在 publisher 的 `config.yaml` 中设置独立端口： ```yaml platforms: api_server: extra: port: 8461 # 必须与 default (8460) 不同 ``` 修复后 publisher 应正常完成合成。**此修复尚未应用到 publisher profile 配置中。** **🔄 Recovery: stuck workers**: 若 worker 显示 `running` 但 `hermes kanban log <task>` 卡在 "Initializing agent..." → reclaim + dispatch 救活：`hermes kanban reclaim <task> && hermes kanban dispatch`。详见 `references/mode-b-operational-gotchas.md`。 ⚠️ **注意**： - `--worker` 格式是 `PROFILE:TITLE[:SKILL,SKILL]`，每个并行 worker 一个独立 flag - **没有** `--workers N`（数量 flag）、`--worker-model`、`--verifier-model` 这些 flag - Model 由各 profile 的 `config.yaml` 控制：搜索 worker 用便宜模型，verifier 用高质量模型 - 详见 `references/kanban-swarm-workflow.md` Swarm topology: ``` Root (Blackboard) ├── Worker 1: Lane ZH search (Brave → Tavily) ├── Worker 2: Lane EN search (Exa → Tavily) ├── Worker 3: Lane Mixed + Tech search (Brave → Exa → Tavily) └── Worker 4: Assembly + de-slop + source-verification ↓ Verifier: Audit (7 sentinels + anti-hedging) ↓ Synthesizer: Render PDFs + TTS + Deliver ``` Kanban advantages over cron: - Per-task model override: cheap models for search, expensive for verification - Artifact auto-delivery: `kanban_complete(artifacts=[...])` → Telegram - Audit trail: every worker run persisted in task_runs - Crash recovery: zombie detection + auto-reclaim See `references/kanban-swarm-workflow.md` for full Kanban integration spec. ## ⚡ Core Rules 1. **搜索必须三+路并行** — 不走单引擎。Brave 中文广撒网 + Exa 英文深度发现 + Tavily 事实校验。缺一路 = 缺信息维度。 2. **结果必须落盘持久 workspace** — 用 `~/.hermes/workspaces/morning-news-{date}/`。搜索完立即写 JSON。 3. **渲染必须等汇编完成** — 父子依赖不可跳。 4. **CSS 必须 diff-check** — 渲染前跑 `assets/diff-check.sh` 双版验证，偏离 baseline >5% 警告。 5. **交付前必须全量审计** — 7 sentinels × 2 editions，PyMuPDF 全量提取，反骑墙 grep，源数校验。任一未过 = 不得交付。 6. **搜索失败不阻塞整路** — 单源 404/单引擎超时 = 跳过 + 标注。整路失败 = 其他路填补。三路全败 = 中止。 7. **Workspace 持久化卫生** — 新建 workspace `chmod 700`，含 `.gitignore`（`*` 全忽略）。保留 7 天。 8. **TTS 属于 pipeline，不是可选项** — Text-to-speech 工具不可达时跳过并标注，但不得静默省略。 ## Content Specifications ### 📰 今日要闻 (Today's Headlines) — NEW in v4.0 - **Purpose**: Quick-scan summary before deep analysis. - **Count**: 普通运行 ≥20 条；Mode B 复跑/用户质检后重做 ≥25 条。每条 2-3 句完整描述，非标题碎片 - **Format**: 每条 2-3 句完整描述 + 信源编号 [sN] + 可选小时间线 - **Source**: 从四路搜索中提取，保留独立性，不在此层融合 - **Organization**: 按板块分组（🔥 中东/俄乌 / 🇺🇸 美国 / 🇨🇳 中国 / 🌍 国际 / 📊 市场+科技） ### 🕐 历史时间线 (v4.1 NEW) 每个有历史纵深的板块开头放**分区脉络框**（.sec-timeline），梳理事件来龙去脉。每条相关新闻上方放**小时间线**（.story-tl），紧凑箭头链。 格式： ``` 分区脉络框：灰底 + 左侧 IKB 粗线 + mono 日期 + 箭头链 小时间线： mono 小字 + accent 色日期 + → 箭头 ``` Example: ```markdown ## 📰 今日要闻 ### 🔥 中东·俄乌 - 美伊停火濒临破裂，伊朗导弹击中科威特机场，美军打击格什姆岛 [s1] - 俄罗斯发动战争以来最大规模空袭：73枚导弹+656架无人机，22死130伤 [s2] ### 🇺🇸 美国 - Trump-Xi北京峰会后续：$14B对台军售成最大悬念 [s3] - ... 总计: ≥15 items ``` ### 执行摘要 (Executive Summary) - Location: first page after cover - Format: 3-5 bullet points distilled from 今日要闻 ### 🔍 深度分析 (Deep Analysis) — v4.0 adjusted count Each analysis item follows the locked structure: ``` 🔍 分析：{标题} 前提：{1-2句事实陈述，引用具体数据/事件来源} 推理：{1-2句因果链，不骑墙，不含"可能/或许"} 结论：{1句明确判断} 趋势：📈/📉/⚠️ + 方向 为什么重要：{1句 impact statement} ``` - **Count**: 5-8 items (not the same as article count!) - **Selection**: pick top stories from 今日要闻 for deep dive - **Anti-hedging hard check**: grep for `一方面|另一方面|可能|或许|似乎`. Any hit → REJECT. ### 📌 今日总结 (Daily Summary) - Standalone card after all analysis - Core tension one-liner - 2-3 forward-looking observations ### TTS 语音稿 (Step 6 output) See §TTS Generation below for full specification. ## TTS Generation (Step 6) — ⛔ NOT OPTIONAL > **EXECUTION LAPSE GUARD**: 2026-06-03 测试中 agent 跳过了此步骤。Step 6 不是 "如果时间允许就做"，是 v4.0 pipeline 的硬步骤。必须在 audit 通过后、交付前用 `text_to_speech` 工具实际调用，产出 `output/morning-news-{date}.mp3`。不调用 = 未完成。 Triggered after audit passes, before delivery. Produces `output/morning-news-{date}.mp3`. ### Content Structure (三段式) | Segment | Source | Length | Style | |---------|--------|--------|-------| | **开场摘要** | 执行摘要 → 口语化 | 30-45s | "早上好，今天是{date}。三件大事：…" | | **深度分析** | Top 3 分析条目 | 60-90s | 每条压缩为"事件 + 为什么重要"（≤50字/条）| | **收尾总结** | 📌 今日总结 | 20-30s | 核心张力 + 一句话前瞻 | **Total**: ≤5 分钟语音 ### Generation Flow ``` 1. 从 morning-news-{date}.md 提取三段内容 2. 改写为口语化脚本 → tts-script-{date}.txt 3. 调用 Hermes text_to_speech(text=script) 4. 保存 output/morning-news-{date}.mp3 5. 质检：试听 30s 片段，确认无吞字/断句 ``` ### Fallback If `text_to_speech` tool unreachable (CosyVoice down, no TTS provider): - Log: "TTS skipped — text_to_speech unavailable" - Continue delivery without audio - Do NOT silently omit the step ## Sections (Fixed) 1. 📰 **今日要闻** — 15-20 条快速摘要（NEW in v4.0） 2. 🔥 **深度分析** — 5-8 条 前提→推理→结论 分析 3. 🇺🇸 **美国** — domestic, economy, Congress, tech 4. 🇨🇳 **中国** — politics, economy, tech, diplomacy, society 5. 🌍 **国际** — Russia-Ukraine, Asia-Pacific, Africa, LatAm 6. 📊 **市场+科技** — oil, equities, forex, crypto, AI/tech 7. 📌 **今日总结** — core tension + outlook Note: 今日要闻 is a flat list, not sectional. The sections below it organize the deep analysis. ## 7 Sentinels (Missing Any = Rework) | # | Sentinel | Check Method | |---|----------|-------------| | 1 | **执行摘要** | 3-5 bullet points on first page | | 2 | **📰 今日要闻** | ≥15 独立新闻条目（来源计数以 📰 来源清单 S01-SNN 为准），按板块分组 | | 3 | **🔍 深度分析** | 5-8 items, each 前提→推理→结论 + 趋势 + 为什么重要 | | 4 | **📌 今日总结** | Standalone card with core tension one-liner | | 5 | **来源清单** | S01–SNN numbered list with outlet names + URLs | | 6 | **Alex Cai** | Cover/header attribution | | 7 | **日期** | Current date: YYYY年M月D日 format | ## Source Requirements - Managed by web-research-router confidence-based routing - Reference registry: `references/sources.json` - Target: ≥50 outlets, routed by locale (zh/en) - Cross-check: Tavily grounding + Brave verification for claims - Per-source error resilience: single source failure ≠ chain failure - **Article count**: 今日要闻 must contain ≥15 independent news items. Sources counted by 📰 来源清单 numbering (S01-SNN). ## Format Specifications **v4.1 起全部使用 Swiss Hybrid 设计系统。** 详见 `references/swiss-hybrid-design-system.md`。 模板：`assets/mobile-template.html`（手机）、`assets/standard-template.html`（A4）。均为 v7 Swiss IKB + serif 编辑体。 ### Mobile Edition — Swiss Hybrid (430×932px) | Property | Value | |----------|-------| | body font-size | 14px（不得低于此值！） | | content font-size | 12px | | section heading | 16px Inter sans-serif | | line-height | 1.65 | | @page margin | 14px 16px 14px | | body font | Noto Serif SC (serif) | | accent | #002FA7 IKB Blue | | paper | #fafaf8 | | layout | 自然流，不硬断页 | | timeline | 分区脉络框 + 每条新闻小时间线（见 design-system ref） | | KPI | 6 列紧凑数据条 | ### Key Layout Rules 1. **不硬断页。** 内容自然填满，不留大片底部空白。仅封面独立一页。 2. **正文 14px 底线。** 用户多次迭代确定。12px 以下不可接受。 3. **时间线必加。** 有历史纵深的板块加分区脉络框，每条相关新闻加小时间线。 4. **信源 50+。** 来源清单从 sources.json 的 50+ outlets 拉取。 ## References | File | Content | |------|---------|