Update OpenClaw handoff with keyword governance

.gitignore

@@ -0,0 +1,12 @@
+.claude/
+.codex/
+__pycache__/
+*.pyc
+*.egg-info/
+.env.local
+findings.md
+progress.md
+task_plan.md
+outputs/freshrss/
+data/term_index/
+outputs/term_index/

README.md

@@ -1,6 +1,6 @@
 # Content Extract MCP
 
-Python MCP scaffold for article content extraction.
+Python MCP scaffold for article content extraction, structured summary validation, deterministic filtering, and Markdown sink output.
 
 ## Run
 
@@ -9,22 +9,208 @@ pip install -e .
 summary-mcp
 ```
 
-The server exposes two tools:
+The server exposes four tools:
 
 - `extract_url_content`
 - `extract_item_content`
+- `filter_summary_result`
+- `run_freshrss_openclaw_pipeline`
 
 Validate an LLM summary result:
 
 ```bash
-validate-llm-result outputs/result.json --extracted outputs/read-flow-2026.extracted.json
+validate-llm-result outputs/reference/summary/result.json --extracted outputs/reference/extracted/read-flow-2026.extracted.json
 ```
 
 Run the minimal extraction-to-summary loop:
 
 ```bash
 python scripts/run_summary_loop.py ^
-  --extracted outputs/read-flow-2026.extracted.json ^
-  --prompt outputs/llm-summary-prompt.txt ^
-  --output outputs/result.json
+  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
+  --prompt outputs/prompts/llm-summary-prompt.txt ^
+  --output outputs/reference/summary/result.loop.json
 ```
+
+Pull FreshRSS entries and map them into normalized `item` objects:
+
+```bash
+set FRESHRSS_API_BASE_URL=http://127.0.0.1:8081/api/greader.php
+set FRESHRSS_USERNAME=bot
+set FRESHRSS_API_PASSWORD=your-api-password
+python scripts/pull_freshrss_items.py --limit 5 --mark-read
+```
+
+By default the script excludes entries already tagged as `read`. Add `--include-read` if you want the full reading list.
+When `--mark-read` is enabled, fetched entries are marked as read after the script finishes successfully.
+
+The script writes:
+
+- `outputs/freshrss/raw/freshrss.raw.json`
+- `outputs/freshrss/items/freshrss.items.json`
+
+Pull FreshRSS entries and run content extraction for each mapped item:
+
+```bash
+set FRESHRSS_API_BASE_URL=http://127.0.0.1:8081/api/greader.php
+set FRESHRSS_USERNAME=osiman
+set FRESHRSS_API_PASSWORD=your-api-password
+python scripts/run_freshrss_extract.py --limit 1 --mark-read
+```
+
+By default the script excludes entries already tagged as `read`. When `--mark-read` is enabled, only entries with successful extraction are marked as read.
+
+The script writes:
+
+- `outputs/freshrss/raw/freshrss.raw.json`
+- `outputs/freshrss/items/freshrss.items.json`
+- `outputs/freshrss/extracted/freshrss.extracted.json`
+
+Run the full FreshRSS pipeline and mark items as read only after the final OpenClaw delivery payload is written:
+
+```bash
+set FRESHRSS_API_BASE_URL=http://127.0.0.1:8081/api/greader.php
+set FRESHRSS_USERNAME=osiman
+set FRESHRSS_API_PASSWORD=your-api-password
+set LLM_API_URL=https://api.deepseek.com
+set LLM_API_KEY=your-llm-api-key
+set LLM_MODEL=deepseek-chat
+python scripts/run_freshrss_pipeline.py --limit 5 --mark-read
+```
+
+If you want to apply your personal engineering and AI-agent interest profile during filtering, pass a context file:
+
+```bash
+python scripts/run_freshrss_pipeline.py ^
+  --limit 5 ^
+  --context configs/filter_context.personal.json ^
+  --mark-read
+```
+
+This is the recommended production entrypoint. By default it writes only:
+
+- `outputs/freshrss/rerun/<timestamp>/raw/freshrss.raw.json`
+- `outputs/freshrss/rerun/<timestamp>/candidates/openclaw-delivery-payload.json`
+- `outputs/freshrss/rerun/<timestamp>/run-report.json`
+
+It also updates the daily keyword index runtime data:
+
+- `data/term_index/daily/YYYY-MM-DD.json`
+- `data/term_index/term_stats.json`
+
+If you need per-item intermediates, add `--debug-artifacts`.
+
+When OpenClaw is connected to the MCP server, it should call `run_freshrss_openclaw_pipeline` for the same behavior directly through MCP. The tool also supports `debug_artifacts=true` when deeper inspection is needed.
+
+Run deterministic filter rules against a structured summary result:
+
+```bash
+python scripts/run_filter_rules.py ^
+  --summary outputs/reference/summary/result.loop.json ^
+  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
+  --output outputs/reference/filter/filter-decision.json
+```
+
+You can optionally pass a context file to inject interest topics or source tags:
+
+```bash
+python scripts/run_filter_rules.py ^
+  --summary outputs/reference/summary/result.loop.json ^
+  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
+  --context outputs/reference/filter/filter-context.json ^
+  --output outputs/reference/filter/filter-decision.with-context.json
+```
+
+Rule engine details and rule authoring guidance live in:
+
+- `docs/design/filter-rule-engine-design.md`
+- `docs/design/filter-rule-engine-usage.md`
+
+Write a filtered result into the Markdown sink:
+
+```bash
+python scripts/run_markdown_sink.py ^
+  --summary outputs/reference/summary/result.loop.json ^
+  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
+  --filter outputs/reference/filter/filter-decision.json
+```
+
+The script writes markdown notes under `knowledge-base/`.
+
+Build an internal `ArticleCandidateRecord` and a slim `OpenClawCandidateInput`:
+
+```bash
+python scripts/run_article_candidate.py ^
+  --summary outputs/reference/summary/result.loop.json ^
+  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
+  --filter outputs/reference/filter/filter-decision.json ^
+  --section-hint tools_and_workflows
+```
+
+The script writes by default:
+
+- `outputs/reference/candidates/article-candidate-record.json`
+- `outputs/reference/candidates/openclaw-candidate-input.json`
+
+Build a batch OpenClaw delivery payload:
+
+```bash
+python scripts/build_openclaw_delivery.py ^
+  --input-dir outputs/freshrss/candidates/batch ^
+  --sort-by-rank ^
+  --date 2026-03-25
+```
+
+The script writes by default:
+
+- `outputs/reference/candidates/openclaw-delivery-payload.json`
+
+Output layout details live in `outputs/README.md`.
+
+Keyword index defaults live in:
+
+- `configs/term_aliases.json`
+- `configs/term_stopwords.json`
+- `configs/term_cleanup_policy.json`
+- `configs/term_watchlist.json`
+- `configs/term_change_log.json`
+
+You can also rebuild the keyword index from an existing delivery payload:
+
+```bash
+python scripts/build_keyword_index.py ^
+  --input outputs/reference/candidates/openclaw-delivery-payload.json
+```
+
+Runtime keyword data is stored under `data/term_index/`.
+
+The keyword cleanup review skill lives in:
+
+- `skills/keyword-cleanup-review/`
+
+To build a review bundle for the LLM skill:
+
+```bash
+python skills/keyword-cleanup-review/scripts/build_review_bundle.py ^
+  --days 7 ^
+  --top 50 ^
+  --output outputs/term_index/review/keyword-cleanup-bundle.json
+```
+
+The review bundle now also carries cleanup governance context:
+
+- cleanup thresholds from `configs/term_cleanup_policy.json`
+- the current watch list from `configs/term_watchlist.json`
+- recent applied changes from `configs/term_change_log.json`
+
+The skill only produces review inputs and suggestions. It does not modify `term_aliases`, `term_stopwords`, or `filter_context.personal.json` automatically.
+
+To preview accepted suggestions before writing any config files:
+
+```bash
+python scripts/apply_term_suggestions.py ^
+  --suggestions outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.json ^
+  --accept-watch Cron Heartbeat Memory ^
+  --dry-run
+```
+
+Remove `--dry-run` to write the accepted changes. The script can also apply accepted `alias`, `stopword`, and `interest keyword` suggestions through `--accept-alias`, `--accept-stopword`, and `--accept-interest`. Accepted watch terms are written into `configs/term_watchlist.json`, and every applied action is appended into `configs/term_change_log.json`.

TODO.md

@@ -1,73 +1,83 @@
-# TODO
+# TODO
 
 ## 当前状态
 
-项目当前处于 `Content Extract MCP` 的 MVP 完成阶段。
+项目当前已经进入“可交付给 OpenClaw 调用”的阶段。
 
-已完成：
+当前主链路：
 
-- [x] 明确整体阅读流链路
-- [x] 明确 `source -> item -> document/article` 的数据抽象
-- [x] 明确 `MCP 负责提取，LLM 负责摘要` 的职责边界
-- [x] 搭建 Python MCP 服务骨架
-- [x] 实现 `extract_url_content`
-- [x] 实现 `extract_item_content`
-- [x] 完成标题与正文提取
-- [x] 完成质量标记与结构化错误输出
-- [x] 用参考文章完成真实提取测试
-- [x] 输出结构化提取 JSON 文件
-- [x] 设计并迭代 LLM 摘要 prompt
-- [x] 验证 LLM 输出的 `result.json` 基本符合预期
-- [x] 已封装 LLM 摘要校验 workflow skill
+`FreshRSS 未读 -> RSS 内容提取 -> LLM 总结 -> 规则过滤 -> OpenClaw delivery payload`
+
+当前已经完成：
+
+- [x] FreshRSS `greader` API 接入
+- [x] `entry -> item` 标准化映射
+- [x] RSS-first 内容提取策略
+- [x] LLM 摘要与校验闭环
+- [x] 第一版规则引擎
+- [x] `ArticleCandidateRecord` / `OpenClawCandidateInput` 分层
+- [x] `OpenClawDeliveryPayload` 批量投递结构
+- [x] 最终 payload 成功后才标记 FreshRSS 已读
+- [x] MCP 工具 `run_freshrss_openclaw_pipeline`
+- [x] 默认精简输出模式
+- [x] 日报级 `keywords` 词元库与周期性词元清洗 skill 设计完成
+- [x] 日报级 `keywords` 词元库与全局词频统计实现完成
+- [x] `keyword-cleanup-review` skill 骨架与 review bundle 脚本实现完成
+- [x] 词元清洗低复杂治理层落地：`term_cleanup_policy` / `term_watchlist` / `term_change_log`
+- [x] 已支持人工确认采纳建议并写入 `term_watchlist` / `term_change_log`
 
 ---
 
-## P0 - 近期必须完成
+## P0 - 交接前后最优先
 
-- [x] 为 LLM 摘要结果定义正式 JSON Schema
-- [x] 增加一个本地校验脚本，自动校验 `result.json` 是否符合 schema
-- [x] 把“提取 JSON -> LLM 摘要 JSON”串成一个标准化本地流程
-- [x] 清理或移除当前已不再使用的 `src/summary_mcp/core/summarizer.py`
-- [x] 更新旧设计文档中仍然残留的 `summary mcp` 描述，避免和当前实现冲突
+- [x] 为 OpenClaw 补齐交接文档
+- [x] 将 MCP 工具作为统一生产入口
+- [x] 将默认输出收敛为最小必要文件
+- [ ] 设计 OpenClaw webhook / delivery payload 的主动推送方式
+- [ ] 明确 OpenClaw 侧如何注册和启动本 MCP 服务
 
 ---
 
 ## P1 - 下一阶段推进
 
-- [ ] 把上游 RSS 聚合结果映射成标准化 `item`
-- [ ] 补充 `item` 的实际样例文件
-- [ ] 定义规则过滤层的输入输出 schema
-- [ ] 设计知识库入库格式
-- [ ] 设计 webhook / 推送格式
-- [ ] 给提取结果增加更多正文清洗策略，比如尾部噪音清理
+- [ ] 设计“人工确认后再沉淀知识库”的状态流转
+- [ ] 收敛 `paywall` 误判规则，降低中文文本误报
+- [ ] 细化过滤规则并引入更多个性化上下文
+- [ ] 将 `keyword-cleanup-review` skill 接入周期性执行流程，产出别名/停用词/兴趣词建议
+- [ ] 增加清洗前后效果对比报告，验证配置调整是否真的改善过滤质量
+- [ ] 增加批量 run 的保留策略与历史清理策略
+- [ ] 为 OpenClaw 补一份更正式的 MCP 调用示例和接线说明
 
 ---
 
-## P2 - 后续增强项
+## P2 - 后续增强
 
-- [ ] 增加批量提取能力
-- [ ] 引入 Playwright 作为动态页面兜底抓取方案
-- [ ] 支持更多 `content_kind`，例如 `release`、`thread`、`video`
+- [ ] 将 `Markdown sink` 进一步降级为 debug / fallback 能力
+- [ ] 增加按天聚合 `ArticleCandidateRecord` 的批处理能力
+- [ ] 让 OpenClaw 聚合候选内容并生成日级摘要
+- [ ] 将日级摘要写入知识库，并同步生成面向用户的日报消息
+- [ ] 支持更多 `content_kind`
 - [ ] 增加提取缓存、重试和更细粒度日志
-- [ ] 重命名包和项目名，从 `summary_mcp` 调整为更符合当前职责的名称
-- [ ] 与 OpenClaw 做更正式的工作流编排整合
+- [ ] 整理历史 rerun 目录与调试产物保留策略
 
 ---
 
 ## 当前建议的下一步
 
+优先做这三件事：
 
-优先做这两件事：
-
-1. 把上游 RSS 聚合结果映射成标准化 `item`
-2. 补充一个真实 `item` 样例文件并开始设计规则过滤 schema
+1. 将 `keyword-cleanup-review` skill 接入周期性执行流程
+2. 设计“人工确认后再沉淀知识库”的状态流转
+3. 收敛规则误判，尤其是 `paywall` 相关启发式
 
 ---
 
-## 收束上下文后建议先看
-
-- docs/context-reset-brief.md
-- docs/content-extract-mcp-mvp-archive.md
-- TODO.md
-
+## 交接时优先阅读
 
+- `README.md`
+- `docs/openclaw/openclaw-handoff.md`
+- `docs/openclaw/openclaw-candidate-input-field-spec.md`
+- `docs/openclaw/openclaw-delivery-payload-spec.md`
+- `docs/design/daily-keyword-index-design.md`
+- `skills/keyword-cleanup-review/SKILL.md`
+- `docs/current/context-reset-brief.md`

configs/filter_context.personal.json

@@ -0,0 +1,56 @@
+{
+  "source_tags": [
+    "backend",
+    "ai-agent",
+    "frontier-tech"
+  ],
+  "interest_topics": [
+    "后端工程",
+    "系统设计",
+    "分布式系统",
+    "数据库",
+    "缓存",
+    "消息队列",
+    "性能优化",
+    "可观测性",
+    "云原生",
+    "Kubernetes",
+    "AI Agent",
+    "LLM",
+    "RAG",
+    "MCP",
+    "工作流自动化",
+    "前沿科技"
+  ],
+  "interest_keywords": [
+    "Java",
+    "Go",
+    "Python",
+    "Spring",
+    "FastAPI",
+    "Gin",
+    "gRPC",
+    "MySQL",
+    "PostgreSQL",
+    "Redis",
+    "Kafka",
+    "微服务",
+    "可观测性",
+    "Kubernetes",
+    "云原生",
+    "AI Agent",
+    "Agent",
+    "LLM",
+    "RAG",
+    "MCP",
+    "Prompt Engineering",
+    "Workflow",
+    "向量数据库",
+    "知识库",
+    "OpenAI",
+    "DeepSeek",
+    "OpenClaw",
+    "AliSQL",
+    "MySQL复制延迟"
+  ]
+}

configs/filter_rules.json

@@ -0,0 +1,138 @@
+[
+  {
+    "rule_id": "drop-low-content",
+    "enabled": true,
+    "stop_on_match": true,
+    "conditions_all": [
+      { "field": "article.quality_flags.is_low_content", "op": "eq", "value": true }
+    ],
+    "action": {
+      "decision": "drop",
+      "reason": "Extracted article quality is too low.",
+      "labels": ["quality", "low-content"],
+      "priority": 100
+    }
+  },
+  {
+    "rule_id": "review-truncated-content",
+    "enabled": true,
+    "stop_on_match": false,
+    "conditions_all": [
+      { "field": "article.quality_flags.is_truncated", "op": "eq", "value": true }
+    ],
+    "action": {
+      "decision": "review",
+      "reason": "Extracted article may be truncated.",
+      "labels": ["quality", "truncated"],
+      "priority": 95
+    }
+  },
+  {
+    "rule_id": "review-paywalled-content",
+    "enabled": true,
+    "stop_on_match": false,
+    "conditions_all": [
+      { "field": "article.quality_flags.is_paywalled", "op": "eq", "value": true }
+    ],
+    "action": {
+      "decision": "review",
+      "reason": "Article may be behind a paywall.",
+      "labels": ["quality", "paywall"],
+      "priority": 95
+    }
+  },
+  {
+    "rule_id": "drop-ephemeral-news",
+    "enabled": true,
+    "stop_on_match": true,
+    "conditions_all": [
+      { "field": "summary.category", "op": "eq", "value": "资讯" },
+      { "field": "summary.worth_keeping", "op": "eq", "value": false }
+    ],
+    "action": {
+      "decision": "drop",
+      "reason": "News-like content marked as not worth keeping.",
+      "labels": ["summary", "ephemeral-news"],
+      "priority": 90
+    }
+  },
+  {
+    "rule_id": "keep-worth-keeping-method",
+    "enabled": true,
+    "stop_on_match": false,
+    "conditions_all": [
+      { "field": "summary.worth_keeping", "op": "eq", "value": true },
+      { "field": "summary.category", "op": "in", "value": ["方法论", "工具实践"] }
+    ],
+    "action": {
+      "decision": "keep",
+      "reason": "Structured summary marked the content as worth keeping in a durable category.",
+      "labels": ["summary", "durable"],
+      "priority": 80
+    }
+  },
+  {
+    "rule_id": "keep-interest-topic",
+    "enabled": true,
+    "stop_on_match": false,
+    "conditions_all": [
+      { "field": "summary.worth_keeping", "op": "eq", "value": true },
+      { "field": "summary.category", "op": "not_in", "value": ["资讯"] },
+      { "field": "summary.topics", "op": "overlap", "value": { "from_field": "context.interest_topics" } }
+    ],
+    "action": {
+      "decision": "keep",
+      "reason": "Topics overlap with the backend and AI-agent interest profile on non-news content.",
+      "labels": ["interest", "topic-match"],
+      "priority": 78
+    }
+  },
+  {
+    "rule_id": "keep-interest-keyword",
+    "enabled": true,
+    "stop_on_match": false,
+    "conditions_all": [
+      { "field": "summary.worth_keeping", "op": "eq", "value": true },
+      { "field": "summary.category", "op": "not_in", "value": ["资讯"] },
+      { "field": "summary.keywords", "op": "overlap", "value": { "from_field": "context.interest_keywords" } }
+    ],
+    "action": {
+      "decision": "keep",
+      "reason": "Keywords match the backend engineering or AI-agent watchlist on durable content.",
+      "labels": ["interest", "keyword-match", "engineering"],
+      "priority": 77
+    }
+  },
+  {
+    "rule_id": "review-interest-news-watch",
+    "enabled": true,
+    "stop_on_match": false,
+    "conditions_all": [
+      { "field": "summary.category", "op": "eq", "value": "资讯" }
+    ],
+    "conditions_any": [
+      { "field": "summary.topics", "op": "overlap", "value": { "from_field": "context.interest_topics" } },
+      { "field": "summary.keywords", "op": "overlap", "value": { "from_field": "context.interest_keywords" } }
+    ],
+    "action": {
+      "decision": "review",
+      "reason": "Relevant frontier or engineering news should be watched first instead of auto-kept.",
+      "labels": ["interest", "news-watch"],
+      "priority": 70
+    }
+  },
+  {
+    "rule_id": "review-worth-keeping-other",
+    "enabled": true,
+    "stop_on_match": false,
+    "conditions_all": [
+      { "field": "summary.worth_keeping", "op": "eq", "value": true }
+    ],
+    "action": {
+      "decision": "review",
+      "reason": "Worth-keeping signal is positive but no stronger keep rule matched.",
+      "labels": ["summary", "needs-review"],
+      "priority": 60
+    }
+  }
+]

configs/term_aliases.json

@@ -0,0 +1,5 @@
+{
+  "AI助手": "AI Agent",
+  "图文RAG": "RAG",
+  "Prompt架构": "Prompt Engineering"
+}

configs/term_change_log.json

@@ -0,0 +1,4 @@
+{
+  "schema_version": "v1",
+  "entries": []
+}

configs/term_cleanup_policy.json

@@ -0,0 +1,25 @@
+{
+  "schema_version": "v1",
+  "interest_keyword_review": {
+    "min_total_count": 3,
+    "min_days_seen": 2
+  },
+  "watch_term_review": {
+    "min_total_count": 1,
+    "min_days_seen": 1,
+    "max_total_count": 2,
+    "max_days_seen": 2
+  },
+  "alias_review": {
+    "min_total_count": 2,
+    "min_days_seen": 2
+  },
+  "stopword_review": {
+    "max_total_count": 2,
+    "max_days_seen": 2
+  },
+  "notes": [
+    "当前阶段采用保守阈值，避免在低样本条件下直接扩充 interest_keywords。",
+    "watch_terms 先用于观察，后续再决定是否升格为 interest_keywords 或进入 alias/stopword 配置。"
+  ]
+}

configs/term_stopwords.json

@@ -0,0 +1,6 @@
+[
+  "小银",
+  "银行客户经理",
+  "飞盘物理",
+  "奋斗文化"
+]

configs/term_watchlist.json

@@ -0,0 +1,5 @@
+{
+  "schema_version": "v1",
+  "updated_at": "2026-03-27T00:00:00Z",
+  "terms": []
+}

docs/README.md

@@ -0,0 +1,112 @@
+# 文档索引
+
+## 当前目录结构
+
+- `docs/README.md`
+  - 文档总索引
+- `docs/current/`
+  - 当前状态、收束入口、阶段导航
+- `docs/design/`
+  - 当前实现的设计文档
+- `docs/openclaw/`
+  - OpenClaw 日报聚合与下游对象设计
+- `docs/notes/`
+  - 较上层的方案笔记与非最终设计
+- `docs/archive/`
+  - 历史归档，不作为最新事实来源
+
+## 当前推荐阅读顺序
+
+1. `docs/current/context-reset-brief.md`
+   - 当前真实进度与下一步入口
+2. `docs/openclaw/openclaw-handoff.md`
+   - 给 OpenClaw 的接手说明、环境变量、MCP 调用方式与已知限制
+3. `docs/openclaw/openclaw-candidate-input-field-spec.md`
+   - 提供给 OpenClaw 的单篇结构化输入字段说明
+4. `docs/openclaw/openclaw-delivery-payload-spec.md`
+   - 提供给 OpenClaw 的批量投递 envelope 说明
+5. `docs/design/summary-mcp-service-design.md`
+   - 当前 MCP 服务的职责、接口和边界
+6. `docs/design/filter-rule-engine-design.md`
+   - 过滤层的输入输出、规则结构与当前实现
+7. `docs/design/filter-rule-engine-usage.md`
+   - 规则怎么写、怎么跑、结果怎么解读的使用说明
+8. `docs/design/daily-keyword-index-design.md`
+   - 日报级词元库与周期性词元清洗 skill 设计
+9. `docs/design/markdown-sink-design.md`
+   - 第一版 Markdown sink 的输入输出、目录结构与落地方式
+10. `docs/openclaw/openclaw-daily-digest-refactor.md`
+   - 为什么要从单篇入库改成 OpenClaw 日报聚合链路
+11. `docs/openclaw/article-candidate-daily-digest-schema.md`
+   - `ArticleCandidateRecord`、`OpenClawCandidateInput` 与 `DailyDigest` 的正式设计
+12. `docs/design/source-schema-design.md`
+   - `source -> item -> document` 的对象设计
+13. `docs/notes/reading-pipeline-design-notes.md`
+   - 更上层的阅读流方案与阶段划分
+14. `docs/design/summary-loop-explained.md`
+   - 当前 LLM 摘要校验闭环的解释
+
+## 当前文档分层
+
+### 1. 当前状态与导航
+
+- `README.md`
+  - 仓库入口与脚本运行方式
+- `TODO.md`
+  - 当前优先级、已完成项、下一阶段任务
+- `docs/current/context-reset-brief.md`
+  - 当前阶段状态的最短摘要
+- `docs/README.md`
+  - 文档索引与阅读顺序
+
+### 2. 当前实现设计
+
+- `docs/design/summary-mcp-service-design.md`
+  - 当前内容提取 MCP 的真实设计
+- `docs/design/summary-core-interface-design.md`
+  - 摘要/提取内核的接口抽象
+- `docs/design/source-schema-design.md`
+  - `source`、`item`、`document` 三层 schema
+- `docs/design/summary-loop-explained.md`
+  - 提取 JSON -> LLM 摘要 JSON -> 校验 的闭环说明
+- `docs/design/filter-rule-engine-design.md`
+  - 第一版规则过滤引擎设计与落地位置
+- `docs/design/filter-rule-engine-usage.md`
+  - 规则配置、调用方式与结果解读
+- `docs/design/daily-keyword-index-design.md`
+  - 日报级词元库与周期性词元清洗 skill 设计
+- `scripts/apply_term_suggestions.py`
+  - 人工确认后将建议写回 watchlist / change log 的脚本入口（见 `README.md` 用法）
+- `docs/design/markdown-sink-design.md`
+  - 第一版 Markdown sink 设计与落地位置
+
+### 3. OpenClaw 与下游设计
+
+- `docs/openclaw/openclaw-handoff.md`
+  - OpenClaw 接手所需的运行说明、工具入口与已知限制
+- `docs/openclaw/openclaw-candidate-input-field-spec.md`
+  - 提供给 OpenClaw 的单篇结构化输入字段说明
+- `docs/openclaw/openclaw-delivery-payload-spec.md`
+  - 提供给 OpenClaw 的批量投递 envelope 说明
+- `docs/openclaw/openclaw-daily-digest-refactor.md`
+  - 改造为 OpenClaw 日报聚合链路的原因与目标结构
+- `docs/openclaw/article-candidate-daily-digest-schema.md`
+  - `ArticleCandidateRecord`、`OpenClawCandidateInput` 与 `DailyDigest` 的字段设计与对象关系
+
+### 4. 方案笔记
+
+- `docs/notes/reading-pipeline-design-notes.md`
+  - 整体阅读流、规则、sink、push 的方案笔记
+
+### 5. 历史归档
+
+- `docs/archive/content-extract-mcp-mvp-archive.md`
+  - MVP 阶段归档，部分状态已被后续进展覆盖
+
+## 当前文档维护原则
+
+- `docs/current/context-reset-brief.md` 记录当前最新状态
+- `TODO.md` 记录任务优先级与下一步
+- `outputs/README.md` 记录当前输出目录约定
+- `docs/archive/content-extract-mcp-mvp-archive.md` 只当历史快照，不再作为最新事实来源
+- 新增阶段性进展，优先更新 `README.md`、`TODO.md`、`docs/current/context-reset-brief.md`

docs/context-reset-brief.md

@@ -1,65 +0,0 @@
-# 项目当前状态简报
-
-## 当前已完成
-
-- 已明确整体链路：`来源 -> 聚合池 -> 内容提取 MCP -> LLM 摘要 -> 校验 -> 过滤 -> 入库/推送`
-- 已确定当前 MCP 的职责边界：
-  - 只负责内容提取
-  - 不负责摘要、分类、价值判断
-- 已完成 Python MCP 骨架
-- 已实现两个 tool：
-  - `extract_url_content`
-  - `extract_item_content`
-- 已完成真实 URL 提取验证
-- 已完成 LLM 摘要 prompt
-- 已完成 LLM 摘要结果 schema 校验器
-- 已完成“提取 JSON -> LLM 摘要 JSON -> 校验”的最小闭环脚本
-- 已完成 LLM 摘要校验 skill 封装
-- 已清理旧的启发式 `summarizer.py`
-- 已将旧的 MCP 设计文档更新为当前“Content Extract MCP”语义
-
-## 当前关键文件
-
-- MCP 入口：
-  - `src/summary_mcp/server.py`
-- 提取主流程：
-  - `src/summary_mcp/core/pipeline.py`
-- LLM 结果模型：
-  - `src/summary_mcp/models/llm_result.py`
-- LLM 校验器：
-  - `src/summary_mcp/validators/llm_result.py`
-- 校验 CLI：
-  - `src/summary_mcp/validate_llm_result.py`
-- 最小闭环脚本：
-  - `scripts/run_summary_loop.py`
-- 当前提示词：
-  - `outputs/llm-summary-prompt.txt`
-- MVP 归档：
-  - `docs/content-extract-mcp-mvp-archive.md`
-- 当前 TODO：
-  - `TODO.md`
-
-## 当前已经验证通过
-
-- 参考文章 URL 可提取为结构化 JSON
-- LLM 可根据提取结果生成摘要 JSON
-- validator 可校验摘要 JSON
-- 最小闭环脚本可直接调用 LLM 接口并产出通过校验的结果
-
-## 当前未开始的下一阶段
-
-- 让 FreshRSS 作为主聚合池
-- 设计 FreshRSS entry -> `item` 的映射
-- 准备真实 `item` 样例
-- 开始定义规则过滤层 schema
-
-## 收束后建议从这里继续
-
-优先从这两个问题继续：
-
-1. FreshRSS 的 entry 字段如何映射成 `item`
-2. 先做一个真实 `item` 样例文件，再讨论规则过滤
-
-## 一句话结论
-
-当前 MVP 已完成，下一阶段不再是继续打磨 MCP，而是开始接入 FreshRSS 上游并建立 `item` 标准化入口。

docs/current/context-reset-brief.md

@@ -0,0 +1,153 @@
+# 项目当前状态简报
+
+## 当前结论
+
+当前仓库已经具备交付给 OpenClaw 的基础条件。
+
+当前主链路是：
+
+`FreshRSS 未读 -> RSS 内容提取 -> LLM 总结 -> 规则过滤 -> OpenClaw delivery payload`
+
+OpenClaw 应通过 MCP 工具 `run_freshrss_openclaw_pipeline` 调用这条链路，而不是自行拼接脚本。
+
+## 当前已完成
+
+- 已完成 FreshRSS `greader` API 接入与未读拉取
+- 已完成 FreshRSS 条目到标准化 `item` 的映射
+- 已完成 RSS-first 提取策略
+- 已完成 LLM 总结与校验闭环
+- 已完成规则引擎过滤
+- 已完成 `ArticleCandidateRecord` 与 `OpenClawCandidateInput` 分层
+- 已完成 `OpenClawDeliveryPayload` 批量投递结构
+- 已完成 FreshRSS 已读状态回写
+- 已完成“仅在最终 payload 成功写盘后再标记已读”的语义
+- 已完成 MCP 工具 `run_freshrss_openclaw_pipeline`
+- 已完成默认精简输出模式，减少中间文件
+- 已完成日报级 `keywords` 词元库与全局词频统计
+- 已完成 `keyword-cleanup-review` skill 骨架与 review bundle 脚本
+- 已完成低复杂治理层：`term_cleanup_policy` / `term_watchlist` / `term_change_log`
+- 已完成采纳建议写回脚本 `scripts/apply_term_suggestions.py`
+
+## 当前 MCP 工具
+
+当前服务入口：
+
+- `src/summary_mcp/server.py`
+
+当前暴露的 MCP 工具：
+
+- `extract_url_content`
+- `extract_item_content`
+- `filter_summary_result`
+- `run_freshrss_openclaw_pipeline`
+
+其中生产主入口是：
+
+- `run_freshrss_openclaw_pipeline`
+
+## 当前关键文件
+
+- MCP 服务入口
+  - `src/summary_mcp/server.py`
+- FreshRSS 统一工作流
+  - `src/summary_mcp/workflows/freshrss_pipeline.py`
+- 词元统计核心
+  - `src/summary_mcp/core/keyword_index.py`
+- 词元统计模型
+  - `src/summary_mcp/models/keyword_index.py`
+- 摘要循环
+  - `src/summary_mcp/core/summary_loop.py`
+- 提取主流程
+  - `src/summary_mcp/core/pipeline.py`
+- FreshRSS 集成
+  - `src/summary_mcp/integrations/freshrss.py`
+- 规则引擎
+  - `src/summary_mcp/filters/engine.py`
+- LLM 结果校验
+  - `src/summary_mcp/validators/llm_result.py`
+- OpenClaw candidate 模型
+  - `src/summary_mcp/models/article_candidate.py`
+- OpenClaw delivery 模型
+  - `src/summary_mcp/models/openclaw_delivery.py`
+- 生产脚本入口
+  - `scripts/run_freshrss_pipeline.py`
+- 词元统计重建脚本
+  - `scripts/build_keyword_index.py`
+- 词元清洗 skill
+  - `skills/keyword-cleanup-review/SKILL.md`
+- skill review bundle 脚本
+  - `skills/keyword-cleanup-review/scripts/build_review_bundle.py`
+- 采纳建议写回脚本
+  - `scripts/apply_term_suggestions.py`
+- 清洗治理配置
+  - `configs/term_cleanup_policy.json`
+  - `configs/term_watchlist.json`
+  - `configs/term_change_log.json`
+- OpenClaw 交接说明
+  - `docs/openclaw/openclaw-handoff.md`
+
+## 当前输出规则
+
+默认生产模式只输出：
+
+- `raw/freshrss.raw.json`
+- `candidates/openclaw-delivery-payload.json`
+- `run-report.json`
+
+同时会更新本地运行数据：
+
+- `data/term_index/daily/YYYY-MM-DD.json`
+- `data/term_index/term_stats.json`
+
+如果需要词元清洗审阅输入，可额外生成：
+
+- `outputs/term_index/review/keyword-cleanup-bundle.json`
+
+如果需要在人工确认后把建议正式写入 watchlist / change log，可使用：
+
+- `scripts/apply_term_suggestions.py`
+
+如果需要排障，可开启：
+
+- `debug_artifacts=true`
+- 或脚本参数 `--debug-artifacts`
+
+这样才会额外输出逐条中间文件。
+
+## 当前验证状态
+
+已经验证通过：
+
+- FreshRSS 未读拉取成功
+- 已读回写成功
+- MCP 工具入口可直接触发完整链路
+- 微信公众号样本可直接使用 RSS 提供的 `summary` 内容提取，不再回源抓网页
+- 精简输出模式已实际跑通
+- 日报级词元统计已通过离线样例验证，确认别名、停用词、非 `drop` 过滤和 rerun 覆盖逻辑正常
+- `keyword-cleanup-review` skill 已通过 `quick_validate.py` 结构校验
+- review bundle 脚本已实际跑通
+- `apply_term_suggestions.py` 已通过 dry-run 与临时副本写回验证
+
+## 当前已知限制
+
+- 当前对 FreshRSS 条目采用 RSS-first 策略，不再回源抓原网页
+- 如果 RSS 中没有足够正文内容，该条会直接跳过，不会进入后续总结
+- 某些规则仍偏保守，部分内容可能落到 `review`
+- `paywall` 相关启发式仍可能误判中文文本
+- Webhook / 主动投递到 OpenClaw 外部接口尚未实现，当前是由 OpenClaw 通过 MCP 主动调用
+- 词元清洗 skill 当前已支持“bundle 构建 -> 建议审阅 -> 人工确认写回 watchlist/change_log”，但尚未接入周期性调度
+- 当前词元统计仍以前置 `OpenClawDeliveryPayload` 作为日报前代理输入，真实 `DailyDigest` 接入后还需切换上游
+
+## 当前最建议的交接阅读顺序
+
+1. `README.md`
+2. `docs/openclaw/openclaw-handoff.md`
+3. `docs/openclaw/openclaw-candidate-input-field-spec.md`
+4. `docs/openclaw/openclaw-delivery-payload-spec.md`
+5. `docs/design/daily-keyword-index-design.md`
+6. `skills/keyword-cleanup-review/SKILL.md`
+7. `TODO.md`
+
+## 一句话结论
+
+当前仓库已经从“提取 MCP 原型”演进到“可供 OpenClaw 调用的 FreshRSS -> OpenClaw payload 上游处理器”，并已补上第一阶段的日报级词元统计能力和词元清洗 skill 骨架；后续重点转向 skill 周期调度、知识库状态流转和 webhook 接线。

docs/design/daily-keyword-index-design.md

@@ -0,0 +1,461 @@
+# 日报级词元库与词元清洗 Skill 设计
+
+## 1. 设计目标
+
+当前项目已经具备：
+
+`FreshRSS -> extraction -> LLM summary -> rule engine -> OpenClaw payload`
+
+下一阶段希望新增“词元库”能力，目标不是做全文级检索索引，而是解决两件事：
+
+1. 为后续规则配置提供稳定、轻量、可读的词元来源
+2. 为周期性的词元清洗、合并和兴趣词补充提供数据基础
+
+因此这套设计的核心原则是：
+
+- 词元来源轻量化
+- 数据粒度日报化
+- 主链路程序化维护
+- 清洗治理由独立 skill 周期性执行
+- LLM 不直接修改规则或兴趣词配置
+
+## 2. 为什么不用逐篇词元库
+
+逐篇记录每篇文章的词元事件，虽然可追溯，但当前阶段成本过高，收益不足：
+
+- 生成文件会很多
+- 存储和调试负担更大
+- 后续真正调规则时，用户更关心“最近日报里反复出现什么词”，而不是“某一篇文章具体抽到了什么词”
+- 当前目标是服务日报与规则配置，不是做文章级分析平台
+
+所以本项目不采用：
+
+`article -> term event -> term store`
+
+而采用：
+
+`daily digest -> keyword aggregate -> daily term index -> global term stats`
+
+## 3. 为什么只保留 `keywords`
+
+当前 LLM 摘要结果里已经有两个候选字段：
+
+- `keywords`
+- `topics`
+
+本设计只使用 `keywords` 进入词元库，不使用 `topics` 作为主来源。
+
+原因：
+
+- `keywords` 更具体，适合后续规则配置
+  - 例如 `Java`、`Go`、`Python`、`MCP`、`RAG`、`Kafka`
+- `topics` 更宽泛，适合摘要展示，不适合作为精确规则命中基础
+  - 例如“后端工程”“AI Agent”“前沿科技”过于宽泛
+- 只保留一个字段可以显著控制词元库规模
+- 当前项目的真实需求是“词元可配置”，不是“主题聚类”
+
+结论：
+
+- `keywords` 进入词元库
+- `topics` 继续保留在单篇摘要结果中，但不纳入词元统计主流程
+
+## 4. 词元库的上游边界
+
+词元库不直接读取所有候选文章，而只读取“最终进入日报”的内容。
+
+也就是说，真正的词元来源是：
+
+- `DailyDigest`
+- 或等价的“已被日报选中”的 candidate 集合
+
+不纳入词元库的内容：
+
+- 被 `drop` 的内容
+- 仅在中间候选层出现、但未进入日报的内容
+- 调试产物中的临时摘要结果
+
+这样做的好处是：
+
+- 词元库只反映真正进入日级产物的内容
+- 高频词更接近长期兴趣，而不是临时噪声
+- 数据量更可控
+
+## 5. 总体链路
+
+目标链路调整为：
+
+`DailyDigest -> keyword normalization -> daily term index -> global term stats -> cleaning skill review -> human confirm -> config update`
+
+职责拆分如下：
+
+### 5.1 程序负责
+
+- 从日报中提取 `keywords`
+- 归一化词元
+- 应用别名映射
+- 应用停用词过滤
+- 生成每日词频
+- 更新全局累计统计
+
+### 5.2 Skill 负责
+
+- 周期性读取词频结果
+- 识别重复词、近义词、大小写变体
+- 识别泛词、噪声词、低价值词
+- 建议哪些词应合并、停用、加入兴趣词配置
+
+### 5.3 人工负责
+
+- 审核 skill 输出的建议
+- 决定是否更新：
+  - `term_aliases`
+  - `term_stopwords`
+  - `filter_context.personal.json`
+
+## 6. 数据文件设计
+
+建议新增以下文件：
+
+- `data/term_index/daily/YYYY-MM-DD.json`
+  - 某一天日报的词元聚合结果
+- `data/term_index/term_stats.json`
+  - 全局累计词元统计
+- `configs/term_aliases.json`
+  - 词元别名归一配置
+- `configs/term_stopwords.json`
+  - 词元停用词配置
+- `configs/term_cleanup_policy.json`
+  - 清洗阈值与治理策略配置
+- `configs/term_watchlist.json`
+  - 当前处于观察状态的词元列表
+- `configs/term_change_log.json`
+  - 已确认生效的词元治理变更记录
+
+说明：
+
+- 不新增逐篇 `term_events.jsonl`
+- 不新增文章级明细文件
+- 默认只保留“日报聚合结果 + 全局统计结果”
+
+## 7. 每日词元文件结构
+
+文件路径示例：
+
+- `data/term_index/daily/2026-03-26.json`
+
+建议结构：
+
+```json
+{
+  "date": "2026-03-26",
+  "source": "daily_digest",
+  "digest_id": "digest-2026-03-26",
+  "generated_at": "2026-03-26T21:30:00+08:00",
+  "candidate_count": 8,
+  "terms": [
+    {
+      "term": "AI Agent",
+      "normalized_term": "AI Agent",
+      "count": 4
+    },
+    {
+      "term": "MCP",
+      "normalized_term": "MCP",
+      "count": 3
+    },
+    {
+      "term": "RAG",
+      "normalized_term": "RAG",
+      "count": 2
+    }
+  ]
+}
+```
+
+字段说明：
+
+- `date`
+  - 日报日期
+- `source`
+  - 固定标记为 `daily_digest`
+- `digest_id`
+  - 日报对象唯一标识
+- `generated_at`
+  - 该词元文件生成时间
+- `candidate_count`
+  - 当天日报包含的条目数
+- `terms[]`
+  - 当天词元聚合结果
+
+其中单条 `terms[]` 只保留：
+
+- `term`
+- `normalized_term`
+- `count`
+
+不保留：
+
+- 逐篇文章来源列表
+- 逐条命中明细
+- 本地文件路径
+
+## 8. 全局统计文件结构
+
+文件路径：
+
+- `data/term_index/term_stats.json`
+
+建议结构：
+
+```json
+{
+  "schema_version": "v1",
+  "generated_at": "2026-03-26T21:30:00+08:00",
+  "terms": [
+    {
+      "term": "AI Agent",
+      "total_count": 18,
+      "days_seen": 6,
+      "first_seen": "2026-03-20",
+      "last_seen": "2026-03-26"
+    },
+    {
+      "term": "MCP",
+      "total_count": 12,
+      "days_seen": 5,
+      "first_seen": "2026-03-21",
+      "last_seen": "2026-03-26"
+    }
+  ]
+}
+```
+
+字段说明：
+
+- `term`
+  - 最终归一后的词元
+- `total_count`
+  - 累计出现次数
+- `days_seen`
+  - 出现过的天数
+- `first_seen`
+  - 首次出现日期
+- `last_seen`
+  - 最近出现日期
+
+当前阶段不额外记录：
+
+- 各 category 分桶统计
+- keep/review/drop 分桶统计
+- 文章级来源列表
+
+原因是：日报级词元库的第一目标是轻量稳定，不是分析平台。
+
+## 9. 标准化与归一规则
+
+程序在写入日报词元前，应先做标准化。
+
+### 9.1 基础标准化
+
+- 去除首尾空白
+- 保留中英文大小写风格中的稳定写法
+- 去重
+- 过滤空字符串
+
+### 9.2 别名映射
+
+通过 `configs/term_aliases.json` 做归一。
+
+示例：
+
+```json
+{
+  "Agent": "AI Agent",
+  "智能体": "AI Agent",
+  "Postgres": "PostgreSQL",
+  "Model Context Protocol": "MCP"
+}
+```
+
+### 9.3 停用词过滤
+
+通过 `configs/term_stopwords.json` 过滤过泛词和噪声词。
+
+示例：
+
+```json
+[
+  "技术",
+  "系统",
+  "方案",
+  "实践",
+  "文章"
+]
+```
+
+## 10. 为什么不让 LLM 直接维护词元库
+
+LLM 可以帮助做清洗建议，但不适合直接维护主词元库。
+
+原因：
+
+- 主词元库更新应该稳定、低成本、可复现
+- 词频统计属于纯程序逻辑，没必要消耗模型调用
+- 如果让 LLM 直接写词元库，会引入不稳定和难审计问题
+
+因此主流程固定为：
+
+- LLM 只负责在摘要结果里输出 `keywords`
+- 程序负责归一、聚合、统计
+
+## 11. 词元清洗 Skill 设计
+
+新增一个周期性清洗 skill，定位是“治理器”，不是“实时生产者”。
+
+### 11.1 Skill 输入
+
+建议输入：
+
+- `data/term_index/term_stats.json`
+- 最近 N 天的 `data/term_index/daily/*.json`
+- `configs/term_aliases.json`
+- `configs/term_stopwords.json`
+- `configs/filter_context.personal.json`
+
+### 11.2 Skill 输出
+
+skill 不直接修改配置文件，而是生成建议文件，例如：
+
+- `outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.md`
+- `outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.json`
+
+低复杂治理层建议补充三类输入：
+
+- `term_cleanup_policy`
+  - 用来定义 watch 和 interest 的最小证据阈值
+- `term_watchlist`
+  - 用来记录“先观察、暂不升级”的词元
+- `term_change_log`
+  - 用来记录已经确认落地的配置变更，避免后续遗忘上下文
+
+建议项包括：
+
+- 建议合并的别名词
+- 建议新增的停用词
+- 建议加入 `interest_keywords` 的候选词
+- 建议降权观察的热点词
+
+### 11.3 Skill 允许做什么
+
+- 发现重复词
+- 发现大小写变体
+- 发现中英文混用的近义词
+- 发现持续高频但尚未进入兴趣配置的词
+- 发现明显过泛的词
+
+### 11.4 Skill 不允许做什么
+
+- 直接改 `filter_rules.json`
+- 直接改 `filter_context.personal.json`
+- 直接覆盖 `term_stats.json`
+- 在无人工确认的情况下自动生效
+
+## 12. 建议的清洗建议文件结构
+
+建议 JSON 文件结构如下：
+
+```json
+{
+  "date": "2026-03-26",
+  "based_on_days": 7,
+  "alias_suggestions": [
+    {
+      "from": "Agent",
+      "to": "AI Agent",
+      "reason": "和现有高频词语义一致，建议归并。"
+    }
+  ],
+  "stopword_suggestions": [
+    {
+      "term": "系统",
+      "reason": "出现频繁但语义过泛，难以作为规则命中词。"
+    }
+  ],
+  "interest_keyword_suggestions": [
+    {
+      "term": "MCP",
+      "reason": "最近多日持续高频，且符合当前 AI Agent 学习方向。"
+    }
+  ]
+}
+```
+
+## 13. 周期与触发方式
+
+建议触发周期：
+
+- 词元统计更新：每天一次，跟随日报生成
+- skill 清洗：每周一次，或人工手动触发
+
+推荐流程：
+
+1. 当天日报生成完成
+2. 程序更新 `daily/YYYY-MM-DD.json`
+3. 程序更新 `term_stats.json`
+4. 每周或人工触发一次词元清洗 skill
+5. skill 输出建议
+6. 人工确认后再更新配置文件
+
+## 14. 与规则引擎的关系
+
+这套词元库设计不是规则引擎的替代品，而是规则配置的辅助层。
+
+关系如下：
+
+- `keywords`
+  - 是词元库来源
+- `term_stats`
+  - 是观察与调优依据
+- `filter_context.personal.json`
+  - 是真正给规则引擎使用的兴趣词配置
+- `filter_rules.json`
+  - 是最终裁决逻辑
+
+也就是说：
+
+`日报词元统计 -> 清洗建议 -> 人工确认 -> 更新 interest_keywords -> 规则引擎命中`
+
+而不是：
+
+`日报词元统计 -> 自动改规则`
+
+## 15. 分阶段落地建议
+
+### Phase 1
+
+先做最小可用版本：
+
+- 只读取日报中的 `keywords`
+- 生成每日词元文件
+- 生成全局累计词频文件
+- 支持 `term_aliases` 和 `term_stopwords`
+
+### Phase 2
+
+再补治理层：
+
+- 增加词元清洗 skill
+- 输出建议文件
+- 人工确认后更新配置
+
+### Phase 3
+
+最后再考虑增强：
+
+- 增加趋势分析
+- 增加最近 7 天热点词视图
+- 增加“建议加入兴趣词”的自动排序
+
+## 16. 一句话结论
+
+这套设计选择“只统计日报中的 `keywords`，由程序维护轻量词元库，再由独立 skill 周期性做清洗建议”，目的是在控制数据规模的前提下，为规则配置和长期兴趣演化提供稳定、可审计、可扩展的基础设施。

docs/design/filter-rule-engine-design.md

@@ -0,0 +1,225 @@
+# 规则过滤引擎设计
+
+## 1. 目标
+
+当前过滤层的定位是：
+
+- 不让 LLM 直接做最终过滤决策
+- 让 LLM 只产出结构化信号
+- 由规则引擎输出最终 `keep / drop / review` 决策
+
+当前链路是：
+
+`item -> content extraction -> llm summary -> filter rule engine`
+
+## 2. 输入输出
+
+### 2.1 FilterInput
+
+过滤层统一读取四类输入：
+
+- `item`
+- `article`
+- `summary`
+- `context`
+
+其中：
+
+- `item` 表示上游标准化候选条目
+- `article` 表示正文提取结果
+- `summary` 表示结构化 LLM 摘要结果
+- `context` 表示额外的用户偏好或运行时上下文
+
+### 2.2 FilterDecisionResult
+
+过滤结果统一输出：
+
+- `decision`
+  - `keep`
+  - `drop`
+  - `review`
+- `matched_rules`
+- `reasons`
+- `labels`
+- `priority`
+- `matches`
+
+这样后续的知识库 sink、推送层或人工审核都可以稳定消费。
+
+## 3. 规则结构
+
+当前规则文件位置：
+
+- `configs/filter_rules.json`
+
+单条规则结构为：
+
+```json
+{
+  "rule_id": "keep-worth-keeping-method",
+  "enabled": true,
+  "stop_on_match": false,
+  "conditions_all": [
+    { "field": "summary.worth_keeping", "op": "eq", "value": true },
+    { "field": "summary.category", "op": "in", "value": ["方法论", "工具实践"] }
+  ],
+  "action": {
+    "decision": "keep",
+    "reason": "Structured summary marked the content as worth keeping in a durable category.",
+    "labels": ["summary", "durable"],
+    "priority": 80
+  }
+}
+```
+
+## 4. 当前支持的操作符
+
+- `eq`
+- `ne`
+- `in`
+- `not_in`
+- `contains`
+- `overlap`
+- `gte`
+- `lte`
+- `exists`
+
+当前条件组合方式：
+
+- `conditions_all`
+- `conditions_any`
+
+## 5. 动态上下文字段
+
+规则支持从其他字段动态取值，例如：
+
+```json
+{ "field": "summary.topics", "op": "overlap", "value": { "from_field": "context.interest_topics" } }
+```
+
+这允许上层 Agent 在调用 `filter_summary_result` 时，把当前关注主题动态注入，而不必把偏好硬编码在规则文件里。
+
+## 6. 当前默认规则思路
+
+当前默认规则分为三类：
+
+- 质量拦截
+  - 低质量正文直接 `drop`
+  - 疑似截断或付费墙进入 `review`
+- 摘要价值判断
+  - `资讯 + worth_keeping=false` 直接 `drop`
+  - `方法论/工具实践 + worth_keeping=true` 直接 `keep`
+- 个性化补充信号
+  - `summary.topics` 与 `context.interest_topics` 重叠时提升为 `keep`
+  - 其他 `worth_keeping=true` 的结果默认进入 `review`
+
+## 7. 规则执行流程
+
+当前过滤流程按以下顺序执行：
+
+1. 上层先产出结构化 `summary`
+2. 可选附带 `item`、`article` 与 `context`
+3. 规则引擎按 `priority` 从高到低遍历规则
+4. 每条规则根据 `conditions_all` / `conditions_any` 判断是否命中
+5. 命中的规则被收集为 `matches`
+6. 最终根据命中结果收敛为 `keep / drop / review`
+
+当前决策收敛规则是：
+
+- 只要命中任意 `drop`，最终结果就是 `drop`
+- 否则只要命中任意 `keep`，最终结果就是 `keep`
+- 否则如果命中任意 `review`，最终结果就是 `review`
+- 如果没有任何规则命中，默认回落到 `review`
+
+这样设计的原因是：
+
+- `drop` 应该拥有最高约束力
+- `keep` 只在没有硬性淘汰时生效
+- 默认不自动放行未知内容
+
+## 8. 为什么让 LLM 调用规则 tool，而不是直接裁决
+
+当前架构故意拆成两层：
+
+- LLM 负责生成结构化信号
+- 规则引擎负责输出最终过滤决策
+
+原因是：
+
+- LLM 适合做语义理解、归类、摘要和价值信号提取
+- 规则引擎适合做稳定、可复现、可审计的最终判断
+
+因此推荐的调用方式是：
+
+1. LLM 先调用提取 tool
+2. LLM 或本地脚本产出 `summary_result`
+3. LLM 再调用 `filter_summary_result`
+4. 后续根据过滤结果决定是否入库、推送或人工审核
+
+这意味着：
+
+- LLM 是编排者
+- rule engine 是裁决器
+
+而不是让 LLM 在过滤阶段再次自由发挥。
+
+## 9. 调用示例
+
+### 9.1 本地脚本
+
+```bash
+python scripts/run_filter_rules.py ^
+  --summary outputs/reference/summary/result.loop.json ^
+  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
+  --output outputs/reference/filter/filter-decision.json
+```
+
+### 9.2 带上下文的调用
+
+```json
+{
+  "interest_topics": ["个人信息管理", "阅读工作流"]
+}
+```
+
+将这份上下文作为 `--context` 传入后，规则就可以根据当前关注主题做加权。
+
+### 9.3 MCP tool
+
+`filter_summary_result` 接收：
+
+- `summary_result`
+- 可选 `extracted_article`
+- 可选 `item`
+- 可选 `context`
+
+返回：
+
+- `decision`
+- `matched_rules`
+- `reasons`
+- `labels`
+- `priority`
+- `matches`
+
+## 10. 当前落地位置
+
+- 过滤模型：
+  - `src/summary_mcp/models/filtering.py`
+- 规则引擎：
+  - `src/summary_mcp/filters/engine.py`
+- 默认规则：
+  - `configs/filter_rules.json`
+- 本地运行脚本：
+  - `scripts/run_filter_rules.py`
+- MCP tool：
+  - `filter_summary_result`
+
+## 11. 当前阶段的设计结论
+
+第一版过滤层采用“规则主导，LLM 提供信号”的架构：
+
+- LLM 不直接做最终裁决
+- 上层 Agent 可以决定是否调用过滤 tool
+- 真正的过滤决策由规则引擎输出
+- 灰区内容后续再考虑是否引入第二层 LLM 辅助判断

docs/design/filter-rule-engine-usage.md

@@ -0,0 +1,328 @@
+# 规则引擎使用说明
+
+## 1. 它解决什么问题
+
+规则引擎负责把上游产出的结构化信号收敛成最终筛选决策：
+
+`item -> content extraction -> llm summary -> rule engine -> candidate/openclaw`
+
+这里有一个明确边界：
+
+- LLM 负责理解正文、生成结构化摘要信号
+- 规则引擎负责输出稳定、可复现、可审计的 `keep / drop / review`
+
+也就是说，规则引擎不是“再让 LLM 判断一遍”，而是用确定性规则做最后裁决。
+
+## 2. 相关文件
+
+- 规则模型: `src/summary_mcp/models/filtering.py`
+- 规则执行器: `src/summary_mcp/filters/engine.py`
+- 默认规则: `configs/filter_rules.json`
+- 本地脚本: `scripts/run_filter_rules.py`
+- MCP tool: `filter_summary_result`
+
+## 3. 输入与输出
+
+规则引擎统一读取一个 `FilterInput`，包含四部分：
+
+- `item`
+  - 标准化后的条目对象
+- `article`
+  - 正文提取结果
+- `summary`
+  - LLM 结构化摘要结果
+- `context`
+  - 运行时注入的偏好信息
+
+当前最常用的判断信号主要来自两类字段：
+
+- `article.quality_flags.*`
+  - 如 `is_low_content`、`is_truncated`、`is_paywalled`
+- `summary.*`
+  - 如 `category`、`worth_keeping`、`topics`
+
+输出是 `FilterDecisionResult`：
+
+- `decision`
+  - 最终决策，`keep / drop / review`
+- `matched_rules`
+  - 命中的规则 ID 列表
+- `reasons`
+  - 命中规则的原因说明
+- `labels`
+  - 聚合后的标签
+- `priority`
+  - 命中规则中的最高优先级
+- `matches`
+  - 每条命中规则的明细
+
+## 4. 规则文件怎么写
+
+规则文件位置是 `configs/filter_rules.json`，顶层必须是一个 JSON 数组。
+
+单条规则结构：
+
+```json
+{
+  "rule_id": "keep-worth-keeping-method",
+  "enabled": true,
+  "stop_on_match": false,
+  "conditions_all": [
+    { "field": "summary.worth_keeping", "op": "eq", "value": true },
+    { "field": "summary.category", "op": "in", "value": ["方法论", "工具实践"] }
+  ],
+  "action": {
+    "decision": "keep",
+    "reason": "Structured summary marked the content as worth keeping in a durable category.",
+    "labels": ["summary", "durable"],
+    "priority": 80
+  }
+}
+```
+
+字段说明：
+
+- `rule_id`
+  - 规则唯一标识，建议稳定命名
+- `enabled`
+  - 是否启用
+- `stop_on_match`
+  - 命中后是否立即停止继续匹配后续规则
+- `conditions_all`
+  - 全部命中才算命中
+- `conditions_any`
+  - 任意命中即可
+- `action.decision`
+  - 该规则命中时产出的规则级决策
+- `action.reason`
+  - 命中原因
+- `action.labels`
+  - 打到结果里的标签
+- `action.priority`
+  - 执行时排序优先级，越大越先执行
+
+## 5. 当前支持的操作符
+
+- `eq`
+- `ne`
+- `in`
+- `not_in`
+- `contains`
+- `overlap`
+- `gte`
+- `lte`
+- `exists`
+
+常见例子：
+
+```json
+{ "field": "summary.worth_keeping", "op": "eq", "value": true }
+```
+
+```json
+{ "field": "summary.category", "op": "in", "value": ["方法论", "工具实践"] }
+```
+
+```json
+{ "field": "summary.topics", "op": "overlap", "value": ["知识管理", "阅读工作流"] }
+```
+
+## 6. 执行顺序和收敛逻辑
+
+执行顺序不是按文件书写顺序，而是按 `action.priority` 从高到低排序。
+
+命中后会先收集所有规则，再做最终收敛：
+
+- 只要命中过任意 `drop`，最终就是 `drop`
+- 否则只要命中过任意 `keep`，最终就是 `keep`
+- 否则只要命中过任意 `review`，最终就是 `review`
+- 如果完全没有命中，默认 `review`
+
+这意味着：
+
+- `drop` 是硬拦截
+- `keep` 只能在没有更高约束的 `drop` 时生效
+- `review` 是默认灰区兜底
+
+如果你希望某条高优规则一旦命中就不再继续匹配，把 `stop_on_match` 设为 `true`。
+
+## 7. 动态上下文怎么用
+
+规则支持从 `context` 动态取值，不需要把用户偏好写死进规则文件。
+
+示例：
+
+```json
+{
+  "field": "summary.topics",
+  "op": "overlap",
+  "value": { "from_field": "context.interest_topics" }
+}
+```
+
+对应的 `context` 可以是：
+
+```json
+{
+  "interest_topics": ["个人知识管理", "阅读工作流"]
+}
+```
+
+这样同一套规则就可以被不同用户、不同运行场景复用。
+
+## 8. 本地怎么跑
+
+最小调用：
+
+```bash
+python scripts/run_filter_rules.py ^
+  --summary outputs/reference/summary/result.loop.json ^
+  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
+  --output outputs/reference/filter/filter-decision.json
+```
+
+带上下文：
+
+```bash
+python scripts/run_filter_rules.py ^
+  --summary outputs/reference/summary/result.loop.json ^
+  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
+  --context outputs/reference/filter/filter-context.json ^
+  --output outputs/reference/filter/filter-decision.with-context.json
+```
+
+也可以显式指定另一份规则文件：
+
+```bash
+python scripts/run_filter_rules.py ^
+  --summary outputs/reference/summary/result.loop.json ^
+  --rules configs/filter_rules.json ^
+  --output outputs/reference/filter/filter-decision.json
+```
+
+## 9. MCP 怎么调用
+
+MCP tool 名称是 `filter_summary_result`。
+
+输入参数：
+
+- `summary_result`
+- `extracted_article`
+- `item`
+- `context`
+
+其中只有 `summary_result` 是必填，其余都是可选补充信号。
+
+示例：
+
+```json
+{
+  "summary_result": {
+    "title": "如何构建个人阅读工作流",
+    "url": "https://example.com/read-flow",
+    "summary": "文章介绍了从采集、提炼到沉淀的个人阅读工作流设计。",
+    "highlights": ["先采集再提炼", "用规则做稳定筛选", "日报再进入知识库"],
+    "keywords": ["阅读工作流", "知识管理", "RSS", "规则引擎", "日报"],
+    "topics": ["阅读工作流", "知识管理", "信息筛选"],
+    "category": "方法论",
+    "worth_keeping": true,
+    "reason": "提供了可复用的方法框架。"
+  },
+  "context": {
+    "interest_topics": ["阅读工作流", "知识管理"]
+  }
+}
+```
+
+## 10. 结果怎么看
+
+一个典型结果会像这样：
+
+```json
+{
+  "decision": "keep",
+  "matched_rules": [
+    "keep-worth-keeping-method",
+    "keep-interest-topic"
+  ],
+  "reasons": [
+    "Structured summary marked the content as worth keeping in a durable category.",
+    "Topics overlap with current interest profile."
+  ],
+  "labels": ["durable", "interest", "summary", "topic-match"],
+  "priority": 80,
+  "matches": [
+    {
+      "rule_id": "keep-worth-keeping-method",
+      "decision": "keep",
+      "reason": "Structured summary marked the content as worth keeping in a durable category.",
+      "labels": ["summary", "durable"],
+      "priority": 80
+    }
+  ]
+}
+```
+
+解读方式：
+
+- 看 `decision`
+  - 最终裁决
+- 看 `matched_rules`
+  - 哪些规则生效了
+- 看 `reasons`
+  - 为什么做出这个判断
+- 看 `matches`
+  - 需要排查时看完整命中明细
+
+## 11. 在当前整条链路里的位置
+
+当前生产链路里，规则引擎已经集成在 `run_freshrss_openclaw_pipeline` 中。
+
+顺序是：
+
+1. 从 FreshRSS 拉未读
+2. 从 RSS 项目里读取正文
+3. 调用 LLM 生成结构化摘要
+4. 规则引擎输出 `keep / drop / review`
+5. 构建 `ArticleCandidateRecord`
+6. 压缩成 `OpenClawCandidateInput`
+7. 生成 `openclaw-delivery-payload.json`
+8. 如果开启 `mark_read`，最后再标记已读
+
+所以在生产模式下，一般不需要单独跑 `run_filter_rules.py`，只有在调规则或排查命中逻辑时才单独跑。
+
+## 12. 调规则时的建议
+
+- 把“硬性淘汰”规则放高优先级
+  - 比如低质量正文、明显噪音内容
+- 把“强 keep”规则放在中高优先级
+  - 比如 `worth_keeping=true` 且类别是方法论
+- 把“兜底 review”规则放低一些
+  - 避免过早收敛
+- 用户偏好尽量走 `context`
+  - 不要把临时兴趣直接硬编码到规则里
+- `rule_id` 保持稳定
+  - 方便后续审计、统计和排障
+
+## 13. 一个最常见的改法
+
+如果你想增加一条“命中关注主题就 keep”的规则，可以直接在 `configs/filter_rules.json` 里加：
+
+```json
+{
+  "rule_id": "keep-interest-topic",
+  "enabled": true,
+  "stop_on_match": false,
+  "conditions_all": [
+    { "field": "summary.topics", "op": "overlap", "value": { "from_field": "context.interest_topics" } }
+  ],
+  "action": {
+    "decision": "keep",
+    "reason": "Topics overlap with current interest profile.",
+    "labels": ["interest", "topic-match"],
+    "priority": 75
+  }
+}
+```
+
+如果只是临时停用某条规则，最简单的是把它的 `enabled` 改成 `false`，不要先删规则。

docs/design/markdown-sink-design.md

@@ -0,0 +1,105 @@
+# Markdown Sink 设计
+
+## 1. 目标
+
+第一版 sink 不直接绑定外部平台，而是先落一个稳定的 `Markdown sink`。
+
+原因：
+
+- 可审计
+- 可版本化
+- 易迁移
+- 可直接兼容 Obsidian / 通用 Markdown 知识库
+
+## 2. 统一输入
+
+Markdown sink 读取统一的 `SinkInput`：
+
+- `item`
+- `article`
+- `summary`
+- `filter_decision`
+- `metadata`
+
+这样未来扩展 Notion sink、Webhook sink 时，可以继续复用同一套上游输入结构。
+
+## 3. 目录结构
+
+当前建议目录结构：
+
+```text
+knowledge-base/
+  inbox/
+  review/
+  archive/
+```
+
+具体落盘规则：
+
+- `keep` -> `knowledge-base/inbox/YYYY/YYYY-MM/`
+- `review` -> `knowledge-base/review/YYYY/YYYY-MM/`
+- `drop` -> `knowledge-base/archive/YYYY/YYYY-MM/`
+
+## 4. 文件命名
+
+文件名格式：
+
+```text
+YYYY-MM-DD-title-slug.md
+```
+
+例如：
+
+```text
+2026-03-21-kimi-cursor.md
+```
+
+## 5. Markdown 内容结构
+
+输出采用：
+
+- YAML frontmatter
+- 摘要区
+- 要点区
+- 过滤决策区
+- 原文信息区
+- 正文摘录区
+
+frontmatter 中保留：
+
+- `title`
+- `url`
+- `source_id`
+- `item_id`
+- `extract_id`
+- `category`
+- `worth_keeping`
+- `decision`
+- `priority`
+- `topics`
+- `keywords`
+- `labels`
+- `published_at`
+- `saved_at`
+- `quality_flags`
+
+## 6. 当前落地位置
+
+- sink 模型：
+  - `src/summary_mcp/models/sink.py`
+- Markdown sink：
+  - `src/summary_mcp/sinks/markdown.py`
+- 本地运行脚本：
+  - `scripts/run_markdown_sink.py`
+
+## 7. 当前阶段结论
+
+第一版知识库不先绑定 Notion、Lumina 或数据库，而是先让过滤后的内容稳定进入 Markdown 知识库。
+
+只要这层格式稳定，后续可以继续扩展：
+
+- Obsidian Vault
+- Logseq
+- Notion
+- Webhook
+- 自定义数据库 sink

docs/design/summary-loop-explained.md

@@ -1,4 +1,4 @@
-# Summary Loop 工作说明
+# Summary Loop 工作说明
 
 ## 1. 目的
 
@@ -15,7 +15,7 @@
 这个闭环由三部分组成：
 
 - 提取结果
-  - 来自 `outputs/*.extracted.json`
+  - 来自 `outputs/reference/extracted/*.json` 或 `outputs/freshrss/extracted/*.json`
   - 提供标题、链接、正文、质量标记
 
 - LLM 摘要
@@ -53,8 +53,8 @@
 
 脚本主要使用两个输入文件：
 
-- 提取结果：`outputs/read-flow-2026.extracted.json`
-- 摘要 prompt：`outputs/llm-summary-prompt.txt`
+- 提取结果：`outputs/reference/extracted/read-flow-2026.extracted.json`
+- 摘要 prompt：`outputs/prompts/llm-summary-prompt.txt`
 
 提取结果不会整包无差别塞给 LLM，而是先裁剪成更小的摘要输入：
 
@@ -187,16 +187,16 @@ validator
 
 每次执行脚本都会在输出目录下留下调试痕迹：
 
-- `result.loop.attempt-1.raw.txt`
+- `outputs/reference/summary/result.loop.attempt-1.raw.txt`
   - 模型原始输出
 
-- `result.loop.attempt-1.json`
+- `outputs/reference/summary/result.loop.attempt-1.json`
   - 提取出的 JSON 结果
 
-- `result.loop.attempt-1.validation.json`
+- `outputs/reference/summary/result.loop.attempt-1.validation.json`
   - 这一轮的校验报告
 
-- `result.loop.json`
+- `outputs/reference/summary/result.loop.json`
   - 当前最终结果
 
 这些文件的价值是：

docs/openclaw/article-candidate-daily-digest-schema.md

@@ -0,0 +1,488 @@
+# Article Candidate / OpenClaw Input / Daily Digest 设计
+
+## 1. 文档目的
+
+本文档正式定义 OpenClaw 日报链路中的三层对象：
+
+- `ArticleCandidateRecord`
+- `OpenClawCandidateInput`
+- `DailyDigest`
+
+目标不是只定义字段，而是明确每一层对象的职责边界，避免把“内部记录对象”和“发给 OpenClaw 的下游输入对象”混成同一个 payload。
+
+---
+
+## 2. 这次为什么要改
+
+前一版设计里，`article_candidate` 同时承担了三种职责：
+
+1. 本地审计与追溯
+2. 规则过滤后的内部候选记录
+3. 发给 OpenClaw 的下游输入
+
+这会直接带来三个问题：
+
+- 字段过重
+  - `article.plain_text` 这类正文内容会显著增加 token 消耗
+- 字段过细
+  - `filter_decision.matches`、`matched_rules`、`labels` 这类规则证据链更适合本地审计，不适合发给下游聚合器
+- 字段过耦合
+  - `source_refs` 是本地文件系统路径，对 OpenClaw 没有意义，反而会让上下游绑定内部实现
+
+批量跑完 `FreshRSS -> extraction -> LLM summary -> filter -> article_candidate` 后，这个问题已经非常清楚：
+
+- 本地对象需要尽量保留信息，便于追溯误判
+- OpenClaw 输入需要尽量精简，便于聚合日报和控制 token
+
+因此，正确改法不是继续给一个 `article_candidate` 做加减法，而是把对象拆层。
+
+---
+
+## 3. 设计结论
+
+建议正式拆成三层：
+
+### 3.1 `ArticleCandidateRecord`
+
+定位：
+
+- 内部候选记录对象
+- 面向本地留档、审计、调试、复跑
+
+它回答的问题是：
+
+- 这条候选内容从哪里来
+- 提取、摘要、过滤各阶段到底产出了什么
+- 为什么规则引擎会给出当前决策
+- 后续如果要重放或排查，该去哪里追溯
+
+### 3.2 `OpenClawCandidateInput`
+
+定位：
+
+- 发给 OpenClaw 的精简输入对象
+- 面向日报聚合与编排
+
+它回答的问题是：
+
+- 这条内容是什么
+- 为什么值得放进候选池
+- 应该如何排序和栏目化
+
+它不负责保存本地调试信息，也不负责承载完整正文。
+
+### 3.3 `DailyDigest`
+
+定位：
+
+- OpenClaw 聚合后的日级主产物
+- 面向日报汇报、知识沉淀、人工 review
+
+它回答的问题是：
+
+- 今天最值得关注的内容是什么
+- 这些内容能提炼出什么结论
+- 哪些内容值得进入长期知识库
+
+---
+
+## 4. 推荐链路
+
+推荐把链路明确为：
+
+`item -> article -> summary -> filter_result -> ArticleCandidateRecord -> OpenClawCandidateInput -> DailyDigest`
+
+这里有两个关键转换：
+
+1. `summary + filter_result` 先生成 `ArticleCandidateRecord`
+   - 这是内部标准记录
+2. `ArticleCandidateRecord` 再投影成 `OpenClawCandidateInput`
+   - 这是跨系统传输对象
+
+这样做的本质是：
+
+- 内部对象追求可追溯
+- 下游对象追求低耦合、低 token、高可消费性
+
+---
+
+## 5. `ArticleCandidateRecord` 设计
+
+### 5.1 角色定义
+
+`ArticleCandidateRecord` 是当前项目内部的正式候选记录对象。
+
+它不是：
+
+- 发给 OpenClaw 的最终 payload
+- 发给用户的日报消息
+- 最终知识库对象
+
+它是下游所有再加工动作之前的“内部事实底稿”。
+
+### 5.2 设计原则
+
+- 保留上游对象，便于调试和重放
+- 保留规则证据链，便于解释误判
+- 保留本地引用，便于审计
+- 允许后续重新生成不同版本的下游 payload
+
+### 5.3 建议字段
+
+```json
+{
+  "candidate_id": "cand:sha256:xxx",
+  "item": {"...": "Item"},
+  "article": {"...": "ExtractedArticle"},
+  "summary": {"...": "LlmSummaryResult"},
+  "filter_result": {"...": "FilterDecisionResult"},
+  "digest_section_hint": "insights",
+  "digest_rank": 60,
+  "review_state": "pending",
+  "rendered_markdown": null,
+  "metadata": {
+    "generated_at": "2026-03-25T10:00:00Z",
+    "pipeline_version": "v2",
+    "producer": "summary_mcp",
+    "run_id": "candidate-2026-03-25-001"
+  },
+  "source_refs": {
+    "item_path": "outputs/freshrss/items/batch/item-01.item.json",
+    "extracted_path": "outputs/freshrss/extracted/batch/item-01.extracted.json",
+    "summary_path": "outputs/freshrss/summary/batch/item-01/result.loop.json",
+    "filter_path": "outputs/freshrss/filter/batch/item-01.filter.json"
+  }
+}
+```
+
+### 5.4 字段说明
+
+- `candidate_id`
+  - 候选记录唯一标识
+  - 建议基于 `item_id` 或 `extract_id` 派生
+
+- `item`
+  - 保留来源、标题、发布时间、作者等上游元数据
+
+- `article`
+  - 保留正文提取结果与质量标记
+  - 供本地调试与重放使用
+
+- `summary`
+  - 保留结构化摘要结果
+  - 是后续 OpenClaw 输入映射的主要来源
+
+- `filter_result`
+  - 保留规则引擎的完整裁决与命中细节
+  - 主要用于解释和排查
+
+- `digest_section_hint`
+  - 给 OpenClaw 的栏目建议
+
+- `digest_rank`
+  - 给 OpenClaw 的排序信号
+
+- `review_state`
+  - 记录人工复核状态
+
+- `rendered_markdown`
+  - 可选的人类可读卡片
+  - 用于调试或 fallback 展示
+
+- `metadata`
+  - 记录运行时元信息
+
+- `source_refs`
+  - 仅用于本地追溯
+  - 不应进入跨系统 payload
+
+### 5.5 为什么内部对象要保留正文和完整过滤结果
+
+因为这层对象不是给 OpenClaw 直接消费的，而是给系统内部留档的。
+
+如果这里过早裁掉字段，会丢失两种关键能力：
+
+1. 误判排查能力
+   - 例如文章被误判为 `paywall` 时，必须能回看 `article` 与 `filter_result`
+2. 下游重放能力
+   - 后面如果调整了 OpenClaw payload 或摘要策略，可以从这层重新投影，而不必重新抓取全文
+
+---
+
+## 6. `OpenClawCandidateInput` 设计
+
+### 6.1 角色定义
+
+`OpenClawCandidateInput` 是当前项目发给 OpenClaw 的正式输入对象。
+
+它应当是：
+
+- 扁平化的
+- 精简的
+- 稳定的
+- 不依赖本地文件路径的
+
+### 6.2 设计原则
+
+- 不携带正文全文
+- 不携带规则引擎的完整证据链
+- 不携带本地 `source_refs`
+- 只保留 OpenClaw 聚合日报真正需要的字段
+
+### 6.3 建议字段
+
+```json
+{
+  "candidate_id": "cand:sha256:xxx",
+  "title": "套壳中国大模型撑起500亿美元估值？扒一扒 Cursor 的套壳疑云",
+  "url": "http://www.ruanyifeng.com/blog/2026/03/kimi-cursor.html",
+  "published_at": "2026-03-21T10:19:11Z",
+  "author": "阮一峰",
+  "source_name": "阮一峰博客",
+  "summary": "2 到 3 句话摘要",
+  "highlights": ["...", "...", "..."],
+  "keywords": ["Cursor", "Kimi K2.5"],
+  "topics": ["人工智能", "商业伦理"],
+  "category": "观点评论",
+  "worth_keeping": true,
+  "worth_reason": "有持续性的行业洞察价值",
+  "selection_decision": "review",
+  "selection_reason": "worth_keeping=true but no stronger keep rule matched",
+  "digest_section_hint": "insights",
+  "digest_rank": 60
+}
+```
+
+### 6.4 字段说明
+
+- `candidate_id`
+  - 与内部记录保持同一主键，便于回溯
+
+- `title` / `url` / `published_at` / `author` / `source_name`
+  - OpenClaw 做日报聚合所需的最小来源信息
+
+- `summary` / `highlights` / `keywords` / `topics` / `category`
+  - OpenClaw 聚合时真正需要的语义材料
+
+- `worth_keeping` / `worth_reason`
+  - 摘要层提供的长期价值判断信号
+
+- `selection_decision` / `selection_reason`
+  - 规则层的压缩结果
+  - 用于告诉 OpenClaw：这条为什么进入候选池，以及应该怎样对待
+
+- `digest_section_hint` / `digest_rank`
+  - OpenClaw 做栏目化和排序时最直接的信号
+
+### 6.5 为什么不用正文全文
+
+因为 OpenClaw 当前承担的是“日报聚合”角色，不是“再次全文阅读器”。
+
+如果把 `article.plain_text` 一并发给 OpenClaw，会出现三个问题：
+
+1. token 浪费
+   - 每条都携带全文，批量聚合时成本会迅速膨胀
+2. 角色混乱
+   - OpenClaw 会被迫重新阅读原文，而不是消费上游已经压缩好的摘要结果
+3. 输入不稳定
+   - 正文长度差异很大，会让聚合阶段的上下文更难控制
+
+因此，正文应保留在 `ArticleCandidateRecord`，而不是进入 `OpenClawCandidateInput`。
+
+### 6.6 为什么过滤结果要压缩
+
+完整的 `filter_result` 适合本地审计，不适合跨系统传输。
+
+OpenClaw 通常只需要知道：
+
+- 这条是 `keep / review / drop` 中的哪一种
+- 为什么会进入候选池
+- 优先级大概是多少
+
+它通常不需要知道：
+
+- 具体命中了哪几条规则
+- 每条规则携带了哪些 labels
+- 完整的 `matches` 证据链
+
+所以建议把规则层压缩为：
+
+- `selection_decision`
+- `selection_reason`
+- `digest_rank`
+
+### 6.7 为什么不要 `source_refs`
+
+因为 `source_refs` 是本地实现细节，不是业务语义。
+
+把它发给 OpenClaw 的问题有两个：
+
+1. 没价值
+   - OpenClaw 无法消费本地磁盘路径
+2. 高耦合
+   - 这会让 OpenClaw 输入隐式依赖当前项目的目录结构与输出命名方式
+
+因此，`source_refs` 应只保留在 `ArticleCandidateRecord`。
+
+---
+
+## 7. 两个对象之间的映射关系
+
+建议映射规则如下：
+
+- `OpenClawCandidateInput.candidate_id`
+  - 来自 `ArticleCandidateRecord.candidate_id`
+- `title` / `url` / `published_at` / `author`
+  - 优先来自 `item`
+- `source_name`
+  - 优先来自 `item.source_title`，没有时可退化为域名或作者来源
+- `summary` / `highlights` / `keywords` / `topics` / `category`
+  - 来自 `summary`
+- `worth_keeping` / `worth_reason`
+  - 来自 `summary.worth_keeping` 与 `summary.reason`
+- `selection_decision`
+  - 来自 `filter_result.decision`
+- `selection_reason`
+  - 可取 `filter_result.reasons[0]` 或压缩后的组合说明
+- `digest_section_hint`
+  - 来自 `ArticleCandidateRecord.digest_section_hint`
+- `digest_rank`
+  - 来自 `ArticleCandidateRecord.digest_rank`
+
+简化理解：
+
+- `ArticleCandidateRecord` 保留证据
+- `OpenClawCandidateInput` 保留结论
+
+---
+
+## 8. `DailyDigest` 设计
+
+### 8.1 角色定义
+
+`DailyDigest` 是 OpenClaw 聚合多个 `OpenClawCandidateInput` 后形成的日级主产物。
+
+它不应该再携带单篇全文，也不应该回退到内部调试结构。
+
+### 8.2 建议字段
+
+```json
+{
+  "digest_id": "digest:2026-03-25",
+  "date": "2026-03-25",
+  "title": "2026-03-25 阅读日报",
+  "summary": "今天的输入主要围绕 AI 工具、工程实践和软件产品策略展开。",
+  "sections": [
+    {
+      "section_id": "insights",
+      "title": "观察与判断",
+      "summary": "今天更值得关注的是 AI 对软件行业护城河和商业包装的影响。",
+      "items": [
+        {
+          "candidate_id": "cand:sha256:xxx",
+          "title": "文章标题",
+          "summary": "单篇摘要压缩版",
+          "why_it_matters": "为什么今天值得被放进这个栏目",
+          "action": "值得继续跟进"
+        }
+      ]
+    }
+  ],
+  "top_items": ["cand:sha256:xxx"],
+  "key_takeaways": [
+    "摘要层和知识沉淀层应分开。",
+    "规则证据链应保留在本地，而不是直接下发给聚合器。"
+  ],
+  "watchlist": [
+    "继续收敛 paywall 误判。"
+  ],
+  "candidate_ids": ["cand:sha256:xxx"],
+  "source_refs": [
+    {
+      "candidate_id": "cand:sha256:xxx",
+      "url": "https://example.com/post/1"
+    }
+  ],
+  "editor_notes": "今天的输入以观点和周刊类内容为主。",
+  "stats": {
+    "candidate_total": 12,
+    "kept_total": 4,
+    "review_total": 6,
+    "dropped_total": 2
+  },
+  "metadata": {
+    "generated_at": "2026-03-25T21:00:00Z",
+    "producer": "openclaw",
+    "pipeline_version": "v2"
+  }
+}
+```
+
+### 8.3 `sections.items[]` 子结构建议
+
+```json
+{
+  "candidate_id": "cand:sha256:xxx",
+  "title": "文章标题",
+  "summary": "单篇摘要压缩版",
+  "why_it_matters": "为什么这条内容今天值得看",
+  "action": "值得试用 / 值得跟进 / 值得收藏 / 待验证"
+}
+```
+
+这里最关键的不是原始摘要复述，而是 `why_it_matters`。
+
+因为日报的价值不在于再列一遍新闻，而在于表达“今天为什么值得看”。
+
+---
+
+## 9. 为什么这样设计更合理
+
+### 9.1 控制 token 成本
+
+把正文全文挡在 OpenClaw 之前，能显著降低聚合阶段的 token 开销。
+
+### 9.2 保持对象职责单一
+
+- `ArticleCandidateRecord` 负责本地审计
+- `OpenClawCandidateInput` 负责下游消费
+- `DailyDigest` 负责日级成品
+
+每层只做一件事，后续更容易演进。
+
+### 9.3 降低上下游耦合
+
+OpenClaw 不需要知道本地 `outputs/` 目录结构，也不应该依赖规则引擎的内部细节。
+
+### 9.4 保留调试与重放能力
+
+内部对象依然保留全文、质量标记、完整过滤证据链，因此不会因为“下游精简”而损失工程可维护性。
+
+### 9.5 为后续人审与知识库沉淀留出口
+
+当日报和知识库真正接起来时：
+
+- OpenClaw 基于精简输入做日报聚合
+- 人工基于日报成品做最终判断
+- 本地记录对象作为审计底稿长期保留
+
+这个边界是清晰且可持续的。
+
+---
+
+## 10. 当前阶段的实现建议
+
+建议按这个顺序推进：
+
+1. 在文档层正式采用这三个对象
+2. 将当前代码里的 `article_candidate` 重命名或重新定位为 `ArticleCandidateRecord`
+3. 新增 `OpenClawCandidateInput` 模型
+4. 增加 `ArticleCandidateRecord -> OpenClawCandidateInput` 的转换逻辑
+5. 让 OpenClaw 只消费精简输入对象
+6. 后续再补 `DailyDigest` 的真实生成逻辑
+
+---
+
+## 11. 一句话结论
+
+这次改动的本质，不是“删掉几个字段”，而是把“内部候选记录对象”和“发给 OpenClaw 的精简输入对象”彻底分层；只有这样，当前项目才能同时保留审计能力、控制 token 成本，并稳定服务 OpenClaw 的日报聚合。

docs/openclaw/openclaw-candidate-input-field-spec.md

@@ -0,0 +1,351 @@
+# OpenClaw Candidate Input 字段说明
+
+## 1. 文档目的
+
+本文档定义当前项目输出给 OpenClaw 的结构化输入对象：`OpenClawCandidateInput`。
+
+这是一份面向 OpenClaw 消费侧的接口说明文档，不要求了解本项目内部的 `item`、`article`、`filter_result` 等实现细节。
+
+当前定位是：
+
+- 单篇候选内容的精简输入
+- 用于 OpenClaw 做日报聚合、排序、栏目分配和后续汇报
+- 不承载正文全文、不承载本地文件路径、不承载规则引擎完整证据链
+
+---
+
+## 2. 设计原则
+
+这个 payload 的设计遵循以下原则：
+
+1. 足够轻
+   - 不发送正文全文，避免浪费 token
+2. 足够稳定
+   - 不暴露本地目录结构和内部中间文件路径
+3. 足够可消费
+   - 字段扁平化，避免 OpenClaw 反复解析嵌套对象
+4. 足够有判断信号
+   - 保留摘要、主题、价值判断、筛选结果、排序信号
+5. 足够可去重
+   - 同时提供原始 `url` 与归一化后的 `canonical_url`
+
+---
+
+## 3. 完整 JSON 示例
+
+```json
+{
+  "candidate_id": "cand:sha256:4139f277b8cb621b02b3398f8a7bd78e3f9eef7bcd320b70b8590e802540c6ca",
+  "title": "套壳中国大模型撑起500亿美元估值？扒一扒 Cursor 的\"套壳\"疑云",
+  "url": "http://www.ruanyifeng.com/blog/2026/03/kimi-cursor.html?utm_source=rss",
+  "canonical_url": "http://www.ruanyifeng.com/blog/2026/03/kimi-cursor.html",
+  "published_at": "2026-03-21T10:19:11Z",
+  "author": "阮一峰",
+  "source_name": "阮一峰的网络日志",
+  "language": "zh-CN",
+  "summary": "AI编程工具Cursor推出的Composer 2模型被证实套壳中国Kimi K2.5模型，引发侵权争议。Kimi官方确认Cursor通过Fireworks AI获得授权，不存在侵权。作者分析Cursor隐瞒事实是为了支撑其不断膨胀的估值，将其包装成大模型公司。",
+  "highlights": [
+    "Cursor的Composer 2模型被技术手段揭露实际调用的是Kimi K2.5模型。",
+    "Kimi官方确认Cursor通过Fireworks AI获得授权，因此不构成侵权。",
+    "Cursor隐瞒使用Kimi模型，被认为是为了支撑其高达500亿美元的估值。"
+  ],
+  "keywords": ["Cursor", "Composer 2", "Kimi K2.5", "Fireworks AI", "AI编程工具"],
+  "topics": ["人工智能", "大模型", "商业伦理"],
+  "category": "观点评论",
+  "worth_keeping": true,
+  "worth_reason": "文章深入剖析了AI行业的热点事件，涉及技术真相、商业动机和行业趋势，具有较高的参考价值。",
+  "selection_decision": "review",
+  "selection_reason": "Worth-keeping signal is positive but no stronger keep rule matched.",
+  "digest_section_hint": "insights",
+  "digest_rank": 60
+}
+```
+
+---
+
+## 4. 字段总览
+
+| 字段名 | 类型 | 必填 | 说明 |
+| --- | --- | --- | --- |
+| `candidate_id` | `string` | 是 | 单篇候选内容唯一标识 |
+| `title` | `string` | 是 | 文章或内容标题 |
+| `url` | `string` | 是 | 原始链接 |
+| `canonical_url` | `string \| null` | 否 | 归一化链接，用于更稳定的去重 |
+| `published_at` | `string \| null` | 否 | 发布时间，ISO 8601 格式 |
+| `author` | `string \| null` | 否 | 作者名 |
+| `source_name` | `string \| null` | 否 | 来源名称，例如博客名、站点名、Feed 名 |
+| `language` | `string \| null` | 否 | 语言标记，例如 `zh-CN`、`en` |
+| `summary` | `string` | 是 | 2 到 3 句话的精简摘要 |
+| `highlights` | `string[]` | 是 | 3 到 5 条单句要点 |
+| `keywords` | `string[]` | 是 | 5 到 8 个关键词 |
+| `topics` | `string[]` | 是 | 3 到 5 个高层主题标签 |
+| `category` | `string` | 是 | 内容分类 |
+| `worth_keeping` | `boolean` | 是 | 摘要层给出的长期保留判断 |
+| `worth_reason` | `string` | 是 | 对 `worth_keeping` 的解释 |
+| `selection_decision` | `string` | 是 | 规则层压缩后的筛选决策 |
+| `selection_reason` | `string \| null` | 否 | 对 `selection_decision` 的简短解释 |
+| `digest_section_hint` | `string \| null` | 否 | 建议栏目 |
+| `digest_rank` | `integer` | 是 | 建议排序分值，范围 `0-100` |
+
+---
+
+## 5. 字段详细说明
+
+### 5.1 `candidate_id`
+
+含义：
+- 单篇候选内容的稳定唯一标识
+
+约束：
+- 同一条上游内容在当前系统中应保持稳定
+- OpenClaw 可以把它作为去重、回溯、引用的主键
+
+### 5.2 `title`
+
+含义：
+- 候选内容标题
+
+来源：
+- 优先来自标准化 `item.title`
+- 如果上游标题缺失，则退回摘要结果中的 `summary.title`
+
+### 5.3 `url`
+
+含义：
+- 原始内容链接
+
+用途：
+- OpenClaw 在日报中回链原文
+- 后续人工 review 或知识库引用时回溯来源
+
+### 5.4 `canonical_url`
+
+含义：
+- 对原始 `url` 做轻量归一化后的链接
+
+当前归一化策略：
+- 去掉 fragment
+- 去掉常见追踪参数，例如 `utm_*`、`fbclid`、`gclid`、`ref` 等
+- 保留其他非追踪 query 参数
+
+用途：
+- 让 OpenClaw 在去重时更稳定
+- 避免同一篇文章因为带不同追踪参数被当成两篇
+
+说明：
+- `url` 保留原始值，`canonical_url` 提供去重辅助
+- OpenClaw 建议优先用 `canonical_url` 做去重键，不足时再结合 `candidate_id`
+
+### 5.5 `published_at`
+
+含义：
+- 内容发布时间
+
+格式：
+- ISO 8601，例如 `2026-03-21T10:19:11Z`
+
+### 5.6 `author`
+
+含义：
+- 作者名称
+
+### 5.7 `source_name`
+
+含义：
+- 内容来源名称
+
+常见示例：
+- `阮一峰的网络日志`
+- `FreshRSS`
+- 某个 Feed 标题
+- 某个站点域名
+
+### 5.8 `language`
+
+含义：
+- 语言标记
+
+常见示例：
+- `zh`
+- `zh-CN`
+- `en`
+
+用途：
+- OpenClaw 后续处理多语言输入时做风格和分类控制
+- 为知识库入库、多语言日报、翻译或过滤策略留出口
+
+说明：
+- 如果当前上游无法可靠识别语言，可以为 `null`
+
+### 5.9 `summary`
+
+含义：
+- 摘要主文本
+
+约束：
+- 当前上游约束是 2 到 3 句话
+- 长度不超过 140 个中文字符
+
+### 5.10 `highlights`
+
+含义：
+- 关键要点列表
+
+约束：
+- 3 到 5 条
+- 每条一条独立信息
+
+### 5.11 `keywords`
+
+含义：
+- 具体实体、工具名、方法名、关键概念
+
+约束：
+- 5 到 8 个
+
+### 5.12 `topics`
+
+含义：
+- 高层主题标签
+
+约束：
+- 3 到 5 个
+- 语义层级高于 `keywords`
+
+### 5.13 `category`
+
+含义：
+- 内容分类
+
+当前允许值：
+- `资讯`
+- `方法论`
+- `工具实践`
+- `观点评论`
+
+### 5.14 `worth_keeping`
+
+含义：
+- 这条内容是否具有持续保留价值
+
+说明：
+- 这是摘要层的价值判断，不是最终入库决定
+- OpenClaw 可以把它当作优先筛选信号，而不是绝对真值
+
+### 5.15 `worth_reason`
+
+含义：
+- 对 `worth_keeping` 的解释
+
+### 5.16 `selection_decision`
+
+含义：
+- 规则引擎压缩后的筛选决策
+
+当前允许值：
+- `keep`
+- `review`
+- `drop`
+
+### 5.17 `selection_reason`
+
+含义：
+- 对 `selection_decision` 的简短解释
+
+说明：
+- 当前一般取规则引擎 reasons 的第一条压缩说明
+- 不承诺包含全部规则命中细节
+
+### 5.18 `digest_section_hint`
+
+含义：
+- 建议日报栏目
+
+当前允许值：
+- `top_news`
+- `tools_and_workflows`
+- `risk_and_security`
+- `open_source`
+- `insights`
+- `deep_dive`
+- `null`
+
+### 5.19 `digest_rank`
+
+含义：
+- 建议排序分值
+
+范围：
+- `0` 到 `100`
+
+建议解释：
+- `80-100`：高优先级，建议优先关注
+- `50-79`：中优先级，建议正常进入候选池
+- `0-49`：低优先级，建议降权或仅保留参考
+
+---
+
+## 6. OpenClaw 消费建议
+
+推荐 OpenClaw 主要消费这几组信号：
+
+### 6.1 语义聚合信号
+
+- `summary`
+- `highlights`
+- `keywords`
+- `topics`
+- `category`
+
+### 6.2 选择与排序信号
+
+- `worth_keeping`
+- `worth_reason`
+- `selection_decision`
+- `selection_reason`
+- `digest_rank`
+- `digest_section_hint`
+
+### 6.3 来源展示信号
+
+- `title`
+- `url`
+- `canonical_url`
+- `published_at`
+- `author`
+- `source_name`
+- `language`
+
+---
+
+## 7. 当前不会提供的字段
+
+OpenClaw 当前不应期待以下字段：
+
+- 正文全文，例如 `plain_text`
+- 本地文件路径，例如 `source_refs`
+- 完整规则证据链，例如 `filter_result.matches`
+- 内部中间对象，例如完整的 `item`、`article`、`summary`、`filter_result`
+- 人工确认状态，例如 `review_status`、`knowledge_decision`
+
+最后一类字段不会放在这个对象中，原因是：
+
+- `OpenClawCandidateInput` 表示的是上游候选事实
+- 人工确认和知识沉淀属于下游流程状态
+- 两者混在一起会让输入对象逐渐失真、变脏
+
+---
+
+## 8. 兼容性约定
+
+当前版本约定：
+
+- 这是单篇输入对象，不是批量 envelope
+- 批量投递由上层 `OpenClawDeliveryPayload` 承载
+- 本文档只约束 `candidates[]` 中单条对象的字段
+
+---
+
+## 9. 一句话结论
+
+`OpenClawCandidateInput` 是“给 OpenClaw 聚合日报用的单篇精简对象”：它保留来源信息、摘要语义、价值判断、去重辅助和排序信号，但不会携带正文全文、本地路径、规则证据链或人工确认状态。

docs/openclaw/openclaw-daily-digest-refactor.md

@@ -0,0 +1,222 @@
+# OpenClaw 日报聚合改造说明
+
+## 1. 为什么要改
+
+当前项目已经跑通了：
+
+`FreshRSS -> item -> extraction -> LLM summary -> validator -> filter -> Markdown sink`
+
+这条链路证明了上游拉取、结构化提取、摘要校验和规则过滤都是可行的。
+
+但如果最终目标是：
+
+- 每日提炼总结进入知识库
+- OpenClaw 向用户汇报今天的日报
+
+那么当前“单篇过滤后直接写入本地 Markdown 知识库”的主路径就不够准确了。
+
+原因有四点：
+
+1. 单篇文章更像原材料，不是最终成品
+2. 用户要的是“日报级总结 + 知识沉淀”，不是“每篇都直接入库”
+3. 参考文章中真正的终态是 `Digest -> Daily Review -> 人工确认 -> Lumina`
+4. 当前 `article_candidate` 如果同时承担内部记录和下游投递，会导致 payload 过重、token 过高、边界混乱
+
+所以当前项目要从“单篇直接入库”切换为“单篇候选材料 -> 日报聚合 -> 人工确认 -> 知识沉淀”。
+
+## 2. 参考文章给出的真实结构
+
+参考文章里的关键分层是：
+
+1. `Digest`
+   - 去重、抓正文、质量检查、生成摘要
+   - 产出可判断的候选内容
+2. `Daily Review`
+   - 从候选内容中做栏目化精选
+   - 产出当天的日报
+3. `Human in the loop`
+   - 人工判断哪些内容值得长期保留
+4. `Lumina`
+   - 只承接真正值得长期保留的内容
+
+这意味着：
+
+- 日报不是知识库的替代品
+- 知识库也不应该承接所有单篇文章
+- 单篇内容更适合作为日报生成前的候选材料
+
+## 3. 当前设计哪里不够
+
+当前的不足主要在下游：
+
+### 3.1 `sink` 语义过重
+
+现在的 `Markdown sink` 更像“最终入库器”。
+
+但在新的目标里，它最多只应扮演：
+
+- debug sink
+- fallback sink
+- 审计/归档 sink
+
+主路径不应再把它视为最终知识库形态。
+
+### 3.2 缺少“日级对象”
+
+当前项目有：
+
+- `item`
+- `article`
+- `summary`
+- `filter_decision`
+
+但还没有：
+
+- `ArticleCandidateRecord`
+- `OpenClawCandidateInput`
+- `DailyDigest`
+
+如果没有这三个对象，就无法稳定表达：
+
+- 单篇内容如何作为内部候选材料存在
+- 单篇内容如何以低 token 的形式发给 OpenClaw
+- 一天的内容如何被聚合为一个正式产物
+
+### 3.3 缺少人工确认层
+
+规则引擎的 `keep` 只能表示“值得进入下一步”，不能直接等价于“正式入知识库”。
+
+如果没有人工确认层，系统就会退化成“规则命中即永久沉淀”，这和参考文章强调的 `Human in the loop` 不一致。
+
+### 3.4 下游输入边界不清楚
+
+如果把正文全文、完整规则证据链、本地文件路径都发给 OpenClaw，会出现三个直接问题：
+
+- token 浪费
+- 对象职责混乱
+- OpenClaw 输入与本地实现耦合
+
+所以必须把“内部记录对象”和“下游精简输入对象”分开。
+
+## 4. 改造后的主路径
+
+建议主路径调整为：
+
+`FreshRSS/RSS -> extraction -> LLM summary -> validator -> rule engine -> ArticleCandidateRecord -> OpenClawCandidateInput -> OpenClaw -> DailyDigest -> human review -> knowledge base`
+
+这里要注意几件事：
+
+1. 当前项目负责生成高质量候选材料与精简输入
+2. OpenClaw 负责按天聚合、生成日报、编排后续动作
+3. 知识库只接收日报和人工确认后的长期内容
+
+## 5. 新的对象分层
+
+### 5.1 `ArticleCandidateRecord`
+
+定位：
+
+- 单篇内容的内部候选记录
+- 用于留档、追溯、重放、审计
+
+最少应包含：
+
+- `item`
+- `article`
+- `summary`
+- `filter_result`
+- `metadata`
+- 可选 `source_refs`
+- 可选 `rendered_markdown`
+
+### 5.2 `OpenClawCandidateInput`
+
+定位：
+
+- 当前项目发给 OpenClaw 的标准投递对象
+- 用于日报聚合与排序
+
+最少应包含：
+
+- `candidate_id`
+- `title`
+- `url`
+- `published_at`
+- `author`
+- `summary`
+- `highlights`
+- `topics`
+- `category`
+- `worth_keeping`
+- `selection_decision`
+- `selection_reason`
+- `digest_section_hint`
+- `digest_rank`
+
+它不应包含：
+
+- `article.plain_text`
+- `filter_result.matches`
+- `source_refs`
+
+### 5.3 `DailyDigest`
+
+定位：
+
+- 一天的主产物
+- 同时服务于“知识沉淀”和“日报汇报”
+
+建议至少包含：
+
+- `date`
+- `sections`
+- `top_items`
+- `key_takeaways`
+- `watchlist`
+- `candidate_ids`
+- `source_refs`
+- `editor_notes`
+
+## 6. 为什么这样更合理
+
+这样改造有几个直接收益：
+
+1. 知识库不会被大量单篇摘要污染
+2. OpenClaw 不需要再次消费全文，token 更可控
+3. OpenClaw 的输入边界更清晰，不依赖本地目录结构和规则证据链
+4. 当前项目仍然保留完整候选记录，因此误判排查和重放能力不会丢
+5. 未来扩展周报、专题文章和反馈回流会更自然
+
+本质上，这是把当前系统从“单篇落库工具”调整为“日报生产链路中的候选记录生产器 + 精简投递器”。
+
+## 7. 对当前实现的影响
+
+不需要推翻已有能力，主要是重新定位：
+
+- `extraction` 保留
+- `LLM summary validator` 保留
+- `rule engine` 保留
+- `FreshRSS integration` 保留
+- `Markdown sink` 保留，但降级为调试/回退能力
+
+真正新增的是：
+
+- `ArticleCandidateRecord` schema
+- `OpenClawCandidateInput` schema
+- `ArticleCandidateRecord -> OpenClawCandidateInput` 映射逻辑
+- `DailyDigest` schema
+- 人工确认后的状态流转
+
+## 8. 下一步应该做什么
+
+建议按这个顺序推进：
+
+1. 在代码里把当前 `article_candidate` 重新定位为 `ArticleCandidateRecord`
+2. 新增 `OpenClawCandidateInput` 模型
+3. 增加精简 payload 的本地输出脚本或转换函数
+4. 让 OpenClaw 只消费精简输入对象
+5. 再设计批量聚合和 `DailyDigest` 生成
+
+## 9. 一句话结论
+
+如果最终目标是“每天有日报汇报，同时把日级提炼沉淀进知识库”，那么当前项目就不该继续围绕“单篇直接入库”演进，而应拆成“内部候选记录对象 + OpenClaw 精简输入对象 + 日级成品对象”三层结构。

docs/openclaw/openclaw-delivery-payload-spec.md

@@ -0,0 +1,195 @@
+# OpenClaw Delivery Payload 字段说明
+
+## 1. 文档目的
+
+本文档定义当前项目批量投递给 OpenClaw 的外层 envelope：`OpenClawDeliveryPayload`。
+
+它不是单篇 candidate 对象，而是承载一批 `OpenClawCandidateInput` 的批次对象。
+
+---
+
+## 2. 完整 JSON 示例
+
+```json
+{
+  "schema_version": "v1",
+  "generated_at": "2026-03-26T10:15:30Z",
+  "run_id": "openclaw-delivery-20260325-101530",
+  "date": "2026-03-25",
+  "candidates": [
+    {
+      "candidate_id": "cand:sha256:xxx",
+      "title": "文章标题",
+      "url": "https://example.com/post",
+      "canonical_url": "https://example.com/post",
+      "published_at": "2026-03-25T09:30:00Z",
+      "author": "作者",
+      "source_name": "来源",
+      "language": "zh-CN",
+      "summary": "2 到 3 句话摘要",
+      "highlights": ["要点1", "要点2", "要点3"],
+      "keywords": ["关键词1", "关键词2", "关键词3", "关键词4", "关键词5"],
+      "topics": ["主题1", "主题2", "主题3"],
+      "category": "观点评论",
+      "worth_keeping": true,
+      "worth_reason": "有长期参考价值",
+      "selection_decision": "review",
+      "selection_reason": "worth_keeping=true but no stronger keep rule matched",
+      "digest_section_hint": "insights",
+      "digest_rank": 60
+    }
+  ],
+  "stats": {
+    "total": 1,
+    "keep_total": 0,
+    "review_total": 1,
+    "drop_total": 0
+  }
+}
+```
+
+---
+
+## 3. 字段总览
+
+| 字段名 | 类型 | 必填 | 说明 |
+| --- | --- | --- | --- |
+| `schema_version` | `string` | 是 | 当前 envelope 的 schema 版本 |
+| `generated_at` | `string` | 是 | 本批数据的生成时间，ISO 8601 格式 |
+| `run_id` | `string` | 是 | 本次批量投递的唯一运行标识 |
+| `date` | `string` | 是 | 本批次所属日期，格式 `YYYY-MM-DD` |
+| `candidates` | `OpenClawCandidateInput[]` | 是 | 单篇候选内容列表 |
+| `stats` | `object` | 是 | 批次级统计信息 |
+
+---
+
+## 4. 字段详细说明
+
+### 4.1 `schema_version`
+
+含义：
+- 当前批量 envelope 的结构版本
+
+用途：
+- 未来字段调整时做兼容处理
+- 让 OpenClaw 能按版本选择不同解析逻辑
+
+当前值：
+- `v1`
+
+### 4.2 `generated_at`
+
+含义：
+- 本批数据的实际生成时间
+
+格式：
+- ISO 8601，例如 `2026-03-26T10:15:30Z`
+
+用途：
+- 排查数据新旧
+- 增量处理
+- 失败重跑和批次比对
+
+### 4.3 `run_id`
+
+含义：
+- 本次投递任务的唯一标识
+
+用途：
+- 失败重跑追踪
+- 批次去重
+- 运行日志关联
+
+建议：
+- 由上游在每次批量生成时唯一产生
+- 例如：`openclaw-delivery-20260325-101530`
+
+### 4.4 `date`
+
+含义：
+- 本批次所属日期
+
+格式：
+- `YYYY-MM-DD`
+
+用途：
+- OpenClaw 进行日报归档
+- 将多次投递映射到同一天的 digest 流程
+
+### 4.5 `candidates`
+
+含义：
+- 单篇候选输入对象列表
+
+说明：
+- 数组内每个元素都应满足 `OpenClawCandidateInput` 定义
+- 单篇字段说明请参考：
+  - `docs/openclaw/openclaw-candidate-input-field-spec.md`
+
+### 4.6 `stats`
+
+含义：
+- 本批次的统计信息
+
+当前字段包括：
+- `total`
+- `keep_total`
+- `review_total`
+- `drop_total`
+
+用途：
+- OpenClaw 在接收后快速感知这批输入的规模和分布
+- 便于记录运行状态和对账
+
+---
+
+## 5. `stats` 子结构说明
+
+| 字段名 | 类型 | 必填 | 说明 |
+| --- | --- | --- | --- |
+| `total` | `integer` | 是 | candidate 总数 |
+| `keep_total` | `integer` | 是 | `selection_decision=keep` 的数量 |
+| `review_total` | `integer` | 是 | `selection_decision=review` 的数量 |
+| `drop_total` | `integer` | 是 | `selection_decision=drop` 的数量 |
+
+说明：
+- 这些统计值由上游根据 `candidates[]` 自动汇总
+- OpenClaw 可以直接信任，也可以再次校验
+
+---
+
+## 6. 推荐消费方式
+
+建议 OpenClaw 的批处理逻辑按以下顺序进行：
+
+1. 先读取 `schema_version`
+   - 选择对应版本的解析逻辑
+2. 再读取 `generated_at`、`run_id` 和 `date`
+   - 建立本次批处理上下文
+3. 再读取 `stats`
+   - 快速判断这批输入规模与分布
+4. 最后逐条消费 `candidates[]`
+   - 去重
+   - 聚类
+   - 分栏目
+   - 排序
+   - 产生日报和 review 清单
+
+---
+
+## 7. 当前不会放入 envelope 的字段
+
+当前 envelope 不会额外放入：
+
+- 完整内部候选记录列表
+- 人工确认状态列表
+- 知识库入库状态列表
+- 执行日志文本
+
+这些都属于其他层的对象，不应混进上游投递协议。
+
+---
+
+## 8. 一句话结论
+
+`OpenClawDeliveryPayload` 是“给 OpenClaw 的批量投递对象”：它用 `schema_version`、`generated_at`、`run_id`、`date`、`stats` 管理批次上下文，用 `candidates[]` 承载真正的单篇候选输入。

docs/openclaw/openclaw-handoff.md

@@ -0,0 +1,209 @@
+# OpenClaw Handoff
+
+## Purpose
+
+This repository provides a FreshRSS-first reading pipeline for OpenClaw:
+
+`FreshRSS unread items -> RSS content extraction -> LLM summary -> rule engine -> OpenClaw delivery payload`
+
+OpenClaw should treat this repository as an MCP-backed upstream content processor.
+
+## Production Entrypoint
+
+OpenClaw should call the MCP tool:
+
+- `run_freshrss_openclaw_pipeline`
+
+This is the canonical entrypoint for production use.
+
+## Required Environment Variables
+
+The MCP server process must have these variables available:
+
+- `FRESHRSS_API_BASE_URL`
+- `FRESHRSS_USERNAME`
+- `FRESHRSS_API_PASSWORD`
+- `LLM_API_URL`
+- `LLM_API_KEY`
+- `LLM_MODEL`
+
+Example:
+
+```powershell
+set FRESHRSS_API_BASE_URL=http://127.0.0.1:8081/api/greader.php
+set FRESHRSS_USERNAME=osiman
+set FRESHRSS_API_PASSWORD=your-api-password
+set LLM_API_URL=https://api.deepseek.com
+set LLM_API_KEY=your-llm-api-key
+set LLM_MODEL=deepseek-chat
+```
+
+## Server Startup
+
+Install dependencies:
+
+```bash
+pip install -e .
+```
+
+Start the MCP server:
+
+```bash
+summary-mcp
+```
+
+## Recommended MCP Call
+
+Recommended default call:
+
+```json
+{
+  "limit": 5,
+  "mark_read": false,
+  "include_read": false,
+  "debug_artifacts": false,
+  "timeout_seconds": 60,
+  "max_retries": 2
+}
+```
+
+Recommended semantics:
+
+- Use `mark_read=false` while validating integration.
+- Use `mark_read=true` only after confirming OpenClaw will consume the returned payload successfully.
+- Keep `debug_artifacts=false` for routine production runs.
+- Set `debug_artifacts=true` only when troubleshooting a bad batch.
+
+## What The Tool Returns
+
+Primary return fields:
+
+- `run_id`
+- `output_dir`
+- `raw_output`
+- `delivery_output`
+- `report_output`
+- `pulled_count`
+- `delivered_count`
+- `marked_read_count`
+- `status_counts`
+- `delivery_payload`
+- `keyword_index`
+
+Optional:
+
+- `items`
+  - Returned only when `include_item_reports=true`
+
+## Minimal Output Files
+
+By default the pipeline writes only:
+
+- `outputs/freshrss/rerun/<run_id>/raw/freshrss.raw.json`
+- `outputs/freshrss/rerun/<run_id>/candidates/openclaw-delivery-payload.json`
+- `outputs/freshrss/rerun/<run_id>/run-report.json`
+
+It also updates local runtime keyword data:
+
+- `data/term_index/daily/YYYY-MM-DD.json`
+- `data/term_index/term_stats.json`
+
+If `debug_artifacts=true`, the pipeline additionally writes per-item intermediate files.
+
+## Payload Specs
+
+OpenClaw payload field specs live here:
+
+- `docs/openclaw/openclaw-candidate-input-field-spec.md`
+- `docs/openclaw/openclaw-delivery-payload-spec.md`
+
+## Read-State Semantics
+
+The pipeline reads from FreshRSS unread items by default.
+
+If `mark_read=true`:
+
+- items are marked as read only after the final `openclaw-delivery-payload.json` has been written successfully
+- only successfully delivered items are marked as read
+- failed or skipped items remain unread
+
+## FreshRSS Content Policy
+
+For FreshRSS items, the pipeline is RSS-first and does not re-crawl webpages.
+
+Behavior:
+
+- use `item.raw_content` first
+- if missing, use `item.raw_summary`
+- if neither contains usable content, skip the item
+- do not fetch the original webpage again for FreshRSS items
+
+This is intentional.
+
+## Keyword Cleanup Governance
+
+This repository also includes a lightweight keyword-governance flow for downstream review.
+
+Current pieces:
+
+- runtime keyword stats
+  - `data/term_index/daily/YYYY-MM-DD.json`
+  - `data/term_index/term_stats.json`
+- governance config
+  - `configs/term_cleanup_policy.json`
+  - `configs/term_watchlist.json`
+  - `configs/term_change_log.json`
+- review bundle builder
+  - `skills/keyword-cleanup-review/scripts/build_review_bundle.py`
+- accepted-suggestion writer
+  - `scripts/apply_term_suggestions.py`
+
+Current status:
+
+- OpenClaw can read the keyword review bundle as maintenance input
+- accepted suggestions still require explicit human confirmation
+- the repository can write accepted watch / alias / stopword / interest-keyword changes after confirmation
+- this governance flow is not yet wired into a periodic scheduler inside the repository
+
+Boundary:
+
+- keyword cleanup is a maintenance flow, not the production RSS ingestion path
+- the repository does not auto-apply cleanup suggestions without confirmation
+- current keyword stats are built from the delivered candidate payload, not yet from a final `DailyDigest`
+
+## Known Limitations
+
+- Some sources put only partial content in RSS; those items may be skipped if RSS content is insufficient.
+- WeChat articles often block direct crawling, but this pipeline now avoids that path for FreshRSS items and uses RSS-provided content when available.
+- Rule behavior is still conservative in some cases; many items may land in `review` depending on current rules.
+- Paywall heuristics may produce false positives for some Chinese text patterns.
+- Keyword cleanup governance is usable now, but periodic scheduling and before/after evaluation are not implemented yet.
+
+## Files OpenClaw Should Read First
+
+Recommended reading order for a new maintainer:
+
+1. `README.md`
+2. `docs/openclaw/openclaw-handoff.md`
+3. `docs/openclaw/openclaw-candidate-input-field-spec.md`
+4. `docs/openclaw/openclaw-delivery-payload-spec.md`
+5. `docs/design/daily-keyword-index-design.md`
+6. `skills/keyword-cleanup-review/SKILL.md`
+7. `docs/current/context-reset-brief.md`
+
+## Current Recommendation
+
+For integration handoff, the repository is usable now.
+
+The minimum you need to give OpenClaw is:
+
+- the repository code
+- the MCP server startup command
+- the required environment variables in the target environment
+- the instruction to call `run_freshrss_openclaw_pipeline`
+
+If OpenClaw will also participate in keyword-governance review, additionally point it to:
+
+- `docs/design/daily-keyword-index-design.md`
+- `skills/keyword-cleanup-review/SKILL.md`
+- `scripts/apply_term_suggestions.py`

knowledge-base/inbox/unknown-date/2026-03-25-信息过载时代-我的漏斗式阅读工作流.md

@@ -0,0 +1,282 @@
+---
+title: "信息过载时代，我的漏斗式阅读工作流"
+url: "https://shawnxie.top/blogs/tools/read-flow-2026.html"
+source_id: ""
+item_id: ""
+extract_id: "sha256:1ee177246bce46534c81718bee693fea43a4b491f545e61b3c41f22e20994057"
+category: "方法论"
+worth_keeping: true
+decision: "keep"
+priority: 80
+topics:
+  - "个人信息管理"
+  - "阅读工作流"
+  - "知识沉淀"
+keywords:
+  - "RSS"
+  - "FreshRSS"
+  - "OpenClaw"
+  - "Digest"
+  - "Daily Review"
+  - "Lumina"
+  - "信息漏斗"
+  - "人在回路"
+labels:
+  - "durable"
+  - "needs-review"
+  - "summary"
+published_at: ""
+saved_at: "2026-03-25T02:24:05.936615+00:00"
+quality_flags:
+  is_paywalled: false
+  is_truncated: false
+  is_low_content: false
+---
+
+# 摘要
+
+作者为解决信息过载问题，构建了一套以RSS为上游、FreshRSS为聚合池、OpenClaw为编排层的漏斗式工作流，通过分层筛选、AI精选和人工精读，将信息逐步沉淀至Lumina知识库，形成稳定可控的个人信息处理闭环。
+
+# 要点
+
+- 以RSS为主统一信息源，通过转换工具将公众号、社交动态等非RSS内容标准化接入。
+- Digest预处理负责去重、抓取正文和生成摘要，Daily Review结合AI将内容分类为日报栏目。
+- 人工精读是长期价值判断的核心环节，只有经过筛选的内容才进入Lumina知识库沉淀。
+- 通过轻量兴趣画像和反馈机制，系统能逐步优化筛选排序，同时避免形成信息茧房。
+- 沉淀后的内容可进一步生成周刊或主题文章，将信息流转化为长期资产。
+
+# 过滤决策
+
+- decision: `keep`
+- priority: `80`
+- matched rules: keep-worth-keeping-method, review-worth-keeping-other
+
+## reasons
+
+- Structured summary marked the content as worth keeping in a durable category.
+- Worth-keeping signal is positive but no stronger keep rule matched.
+
+# 原文信息
+
+- 标题: 信息过载时代，我的漏斗式阅读工作流
+- 链接: https://shawnxie.top/blogs/tools/read-flow-2026.html
+- 作者: unknown
+- 发布时间: unknown
+- 内容类型: article
+
+# 正文摘录
+
+信息过载时代，我的漏斗式阅读工作流
+这两年我越来越强烈地感觉到，信息问题早就不是获取不到，而是处理不过来。
+真正让我疲惫的，不是没东西看，而是每天都有太多东西值得看：公众号文章、技术博客、GitHub Release、AI 新闻、社区讨论、长文、短讯、碎片化观点等等，全都在争夺我的注意力。
+如果不做点什么，一个人的信息生活很容易退化成这样：
+- 收藏夹里躺满了文章，真正读完的却寥寥无几
+- 每天浏览大量内容，能留在脑子里的却屈指可数
+- 输入看似充实，输出却异常薄弱
+- 以为自己一直在吸收信息，实际上只是在不同窗口间疲于切换
+2026年，我开始认真整理自己的一套信息处理工作流。我不需要大而全的平台，也不想要那种号称能“一键替我读完互联网”的 AI 产品。
+我只需要一套围绕自己运转的信息吸收漏斗：既能尽可能广泛地捕捉信息，又能提前过滤掉噪音和重复内容，只把真正值得投入时间的内容呈现在我面前。更重要的是，这套系统还要能将我精读后的高价值内容沉淀下来，反过来优化下一轮的信息筛选。
+信息处理不追求看得更多，而是在于更稳定地吸收、判断和沉淀。
+这篇文章，我将把这套工作流思路完整写下来，包括：
+- 为什么我要这样做
+- 它是怎么搭起来的
+- 每一层分别承担什么职责
+- 未来优化方向
+如果你也在被信息过载折磨，或者你已经有不少输入工具，却还是觉得每天看了很多，脑子里没留下什么，那也许这篇文章会有点参考价值。
+为什么要做信息吸收漏斗？
+我以前也尝试过很多方法：
+- 订阅很多 RSS，然后每天刷
+- 把值得看的东西先扔进稍后阅读
+- 用收藏、标签、笔记软件把文章存起来
+- 靠搜索和回忆去找以前读过的内容
+这些工具各有用处，但通常只能解决信息处理中的某一个环节。而个人信息处理真正的难点，从来都不在于单点突破，而在于如何将这些分散的环节串联起来。信息系统如果不能同时解决这两个问题，就很容易失控：
+输入太多，内容太杂
+几乎每一个你关心的领域，都在持续产生新内容。技术突破、产品迭代、AI进展、工具更新、商业动态、研究成果、行业变化……
+而且在未经过滤之前，重要新闻、工具发布、经验复盘、广告软文、标题党、重复报道等往往都混在一起，让人难以分辨。
+如果每一条内容都需要你从零开始判断其价值，那日积月累下来的认知负担会非常高。很多时候，人真正感到疲惫的不是读一篇长文，而是反复做这些微小但无意义的判断：
+- 这篇文章值不值得点开？
+- 这条消息是不是和刚才那条重复？
+- 这个标题是不是在夸大其词？
+- 这条内容到底在说什么？
+- 这篇文章我只是需要了解一下，还是值得保存？
+缺少精读回路
+很多内容看完就过去了，既没有标记，也没有沉淀，更没有真正融入自己的知识体系。
+如果信息系统永远只知道最新内容是什么，却不知道什么内容真正对你有帮助，那最多只能算是一个信息输送管道，而不是一个能越来越懂你的个性化系统。
+所以我要的不是做一个更强的信息池，而是做一个更稳的信息漏斗。
+桶的思路是往里装更多，漏斗的思路是让信息在往下走的过程中不断收窄，最后留下真正值得进入大脑和长期记忆的部分。
+从实践效果来看，这套漏斗式信息处理工作流至少给我带来了几个显著的变化：
+- 我不再需要每天从海量标题中艰难地筛选重点内容
+- 低质量和重复的信息大部分在上游就被过滤掉了
+- 真正值得精读的内容，会在更靠后的环节中呈现在我面前
+- 我读过并认为有价值的内容，终于开始能够反过来优化后续的筛选逻辑
+不是追求更快地刷完信息，而是实现更稳定地吸收知识。
+为什么基于 OpenClaw 搭建？
+由于流程尚处探索阶段，选择先用 OpenClaw 将整条链路串联起来，让系统先运行起来，验证可行性，发现问题后再逐步优化和完善。
+OpenClaw 主要扮演编排层的角色，负责以下工作：
+- 定时触发各类任务
+- 串联不同的外部工具和系统
+- 与飞书文档进行交互
+- 与 Lumina 知识库进行联动
+- 在需要的环节引入 AI 进行结构化内容生成
+这种方式起步快、迭代快，非常适合边运行边优化，不需要一开始就投入大量精力构建完整的前后端系统。
+如果你也想搭建一套类似的信息处理流程，需要做这些准备：
+1. 可持续维护的信息源
+这是整个系统的上游基础，需要先想清楚几个问题：
+- 自己真正长期关注的主题领域是什么
+- 这些领域的主要信息来源有哪些
+- 哪些信息源值得长期订阅，哪些只是偶尔查看即可
+这一步不需要追求大而全，但要尽量保持稳定和高质量。
+如果你还没有构建自己的信息源，可以先参考我收集的 RSS 订阅源。
+2. 统一的聚合池
+我选择的聚合池是 FreshRSS，一个开源的RSS阅读器，支持多平台，支持自定义订阅，内容分类管理和API，非常适合用来进行内容存储和同步。
+它并非最终阅读工具，而是信息汇聚的中转站，让所有信息源先汇聚到同一个地方，形成可持续消费的内容候选池。
+3. 自动化编排层
+这是 OpenClaw 最能发挥价值的地方，极大地降低了自动化流程实现的难度，通过自然语言描述诉求，在对话中逐步落地想法。我主要用它来完成以下任务：
+- 定时执行 digest（自定义的预处理Skill）
+- 定时执行 daily-review（自定义的内容精选Skill）
+- 发布飞书文档
+- 调用 Lumina（个人开源的知识库项目）
+- 串联各个技能和脚本
+4. 长期知识沉淀层
+在我的工作流中，这一层使用 Lumina——个人开发的信息管理工作台。当然也可以使用别的笔记工具，如Notion、Obsidian等。
+Lumina 只承接我经过筛选后确认值得长期保留的优质内容，这一步直接决定了后续反馈机制的质量。
+一个系统最终能学到什么，很大程度上取决于你为它提供了什么样的正反馈样本。
+信息处理工作流
+一句话描述：用 RSS 尽可能广泛地捕捉信息，用 FreshRSS 进行稳定聚合，用 Digest 完成预处理，用 Daily Review 实现每日精选，通过人工精读判断内容的长期价值，用 Lumina 进行知识沉淀，最后将这些沉淀的价值反向转化为轻量的个性化信号。
+这套流程的核心优势不是追求全自动，而是在于分层处理的设计理念。这意味着：
+- 并非所有信息都值得投入时间精读
+- 并非所有信息都值得长期留存
+- 并非所有信息都需要进行个性化加权
+- 并非所有信息都适合直接作为输出内容
+一旦层次划分清晰，系统就不会退化为单一维度的推荐流，而会演变成一个真正的认知加工流程。漏斗的真正价值不在于让信息越来越少，而在于实现了这几个关键目标：
+- 上游宽广：确保不会错过真正重要的信息变化
+- 中游稳定：有效过滤噪音，避免其直接干扰注意力系统
+- 下游精准：让我不必在不值得精读的内容上浪费时间
+- 回流轻柔：通过轻量反馈机制，避免系统演变成封闭的信息茧房
+个人信息系统核心能力不是如何接入更多信息源，而是要明确：哪些信息值得进入系统，哪些信息值得投入时间精读，哪些内容值得长期留存，哪些知识最终真正融入了你的思考和行动？
+当能够清晰地回答这些问题时，就不再是互联网信息的被动接收者，你将拥有一套真正属于自己的认知处理系统。接下来将按顺序详细介绍每一层的工作原理。
+信息源：以 RSS 为主，但不局限于原生 RSS
+整个系统的最上游是信息源，主要依赖 RSS。
+RSS 是最被低估的个人信息基础设施。其天然符合个人信息系统最核心的几个要求：
+- 订阅权完全掌握在自己手中
+- 信息来源清晰明确
+- 更新内容结构化
+- 不受平台推荐算法直接支配
+- 便于程序接入和自动化处理
+然而在现实中，有一个不可避免的问题：并非所有值得关注的信息源都提供 RSS 订阅。例如：部分公众号内容、某些社区的特定栏目、垂直网站的更新页面和社交平台上的账号动态。
+因此，需要在信息源进入流程之前，尽可能将其统一转换为 RSS 或类似 feed 的格式。常见的方案有：
+- RSSHub / RSS-Bridge 等转换工具：能够将大量原本不提供 RSS 订阅的内容源，转换为可以被订阅和程序自动化消费的 feed 格式；
+- GitHub feed：项目的 Releases、Commits、Discussions 等，都天然提供了结构化的 feed 接口；
+- wewe-rss：能够将公众号文章转换成RSS订阅源；
+- nitter：将 Twitter/X 动态转换为RSS源；
+- 自定义抓取脚本：对于没有现成 RSS 解决方案的页面，也可以自行编写轻量级的抓取脚本，将其转换为内部可用的 feed 格式。
+这一步的原则很简单：上游信息源可以多种多样，但在进入系统之前，格式必须尽可能统一。只有这样，下游的预处理和筛选环节才能稳定可靠地运行。
+更多信息源归一处理方案可参考之前文章碎片时间刷文章！懒人阅读方案分享。
+聚合池：用 FreshRSS 打造稳定的“中间水库”
+所有订阅源最终都会汇聚到 FreshRSS 中，它并非"我每天真正坐下来阅读的地方"，而是流程的缓冲层，主要有以下作用：
+实现信息来源的统一管理
+无论内容来自博客、社区、GitHub 还是转换后的 feed，最终都以统一的格式呈现，成为可被消费的标准化对象。
+让下游环节无需直接对接互联网
+digest 模块无需再逐个访问各个网站抓取今日更新内容，只需从 FreshRSS 这个统一的内容池中获取未读候选即可。大大降低了系统各环节之间的耦合度。
+确保了信息处理的时间连续性
+信息系统最忌讳的是“今天临时查看一下、明天就忘了、后天又重新开始”的碎片化处理方式。FreshRSS 提供了一个稳定的时间窗口，使后续任务能够按照固定节奏有序运行。
+我的目标不是追求无边界的信息获取，而是实现有边界的信息处理。
+预处理：Digest 将海量候选内容转化为可判断对象
+如果把 FreshRSS 比作蓄水池，那么 Digest 就是这套系统中的第一道加工厂，专注于完成内容预处理任务。
+让人每天真正感到疲惫的，往往不是精读一篇高质量文章，而是反复判断大量低质量内容是否值得投入时间。
+核心处理任务
+- URL 精确去重
+- 相似内容去重
+- 正文抓取
+- 质量检查
+- 噪音过滤
+- 摘要生成
+- 初步排序与文档输出
+在真正的“阅读”开始之前，先将互联网上天然混乱的信息整理成一批更具可读性的候选内容。
+为什么重要？
+如果没有预处理环节，后续的所有精选工作都将建立在一堆未经整理的原始标题之上。而经过 Digest 处理后，情况会改善：
+- 标题党内容大幅减少
+- 重复报道得到有效过滤
+- 无法获取正文的无效数据明显减少
+- 每条候选内容至少附带一个可供快速判断的摘要
+这将显著降低后续阅读的认知成本。Digest 的职责并非编辑终稿，而是将候选内容池整理干净、结构化，为后续的精选环节奠定基础。
+AI 精选：Daily Review 为我呈现重点关注内容
+如果说 Digest 的作用是将原始候选内容处理得更具可读性，那么 Daily Review 则是从这些经过预处理的候选中，进一步提炼出当天真正值得关注的精华内容。
+这一步让流程从预处理迈向编辑阶段。不再追求内容的广度，而是致力于打造重点更突出、结构更清晰的阅读体验，让最终产物更像一份真正意义上的日报，而非简单堆砌的摘要集合。
+目前，我将 Daily Review 设计为几个固定栏目，将不同类型的信息分配到不同的认知槽位中，包含：
+- 今日大事：聚焦具有公共重要性的事件
+- 变更与实践：关注对个人有直接操作价值的内容
+- 安全与风险：警惕潜在的风险因素
+- 开源与工具：追踪工具生态的发展变化
+- 洞察与数据点：把握行业趋势和关键数据
+- 主题深挖：将单一新闻事件提升到趋势层面进行分析
+LLM 赋能
+从 Digest 输出的候选内容，到 Daily Review 最终的成稿层，中间有很多工作适合 AI 来完成：
+- 合并同一事件的多个信息来源
+- 提炼核心主题
+- 为内容分类并分配到对应栏目
+- 识别值得深入探讨的话题
+- 将零散的候选内容重组为人类可以快速阅读的日报结构
+但我对 AI 在这一环节的应用始终保持克制。不是让 AI 全自动生成日报，而是扮演结构化整理者的角色，从候选内容中筛选出当天和我相关的精华部分。
+精读留存：将 Human in the loop 置于系统核心
+前面所有流程，都是为将信息筛选到值得精读的阶段。真正让系统不至于退化为另一种自动化信息流的，正是这一关键环节：Human in the loop（人在回路中）。
+我坚信，在个人知识系统中，最不能完全外包的是长期价值判断能力。
+系统可以协助我完成许多任务，如信息收集、去重、摘要、聚类、排序和精选等。但它无法完全替我做出关键决策：
+- 哪些内容真正值得纳入长期知识库
+- 哪些内容在未来会持续发挥价值
+- 哪些内容会对我的写作和判断产生深远影响
+因此，在我的工作流中，Lumina 前面始终设有一道人工筛选门槛。我会从之前各个环节产生的内容中，挑选出真正值得精读和长期留存的内容。
+这是整套系统中最关键的价值确认环节。能够进入 Lumina 的内容，不仅代表我看过，更意味着我认为这篇内容值得在未来持续为我所用。
+个人画像：让系统学会识别对我真正有价值的内容
+如果流程到 Lumina 就戛然而止，那么它仍然只是一个单纯的过滤和沉淀系统。只有当沉淀的内容开始反过来影响上游的信息选择时，整个系统才真正形成了闭环。
+为此我加入了一层设计较为克制的兴趣画像逻辑。
+刻意控制了影响力，不希望整个系统演变成另一个猜你喜欢的推荐引擎。我只需要它能稍微更懂我，但又不过度迎合我的偏好，保留对公共重要性内容的关注和探索未知领域的空间。
+这层画像目前主要承担轻量 rerank 的功能，作为辅助信号，轻微影响 Digest 和 Daily Review 环节中候选内容的排序。主要从以下几个维度逐步学习我的偏好：
+- 长期精读的主题领域
+- 能稳定提供价值的信息源
+- 偏好的内容格式
+- 最终存入 Lumina 知识库的内容类型
+轻量引导的设计，旨在减少无效信息对注意力的浪费，同时避免构建封闭的信息茧房。个性化推荐应帮助我们减少无意义的判断，而非让我们躲进舒适区。
+沉淀：将信息流转化为长期内容资产
+信息漏斗的终点不是读完，而是沉淀。目前从两个方向拓展沉淀的价值：
+1. 周刊生成
+当系统积累了一周的高质量内容后，就不应再局限于每天生成一份日报。一周的时间跨度非常适合进行复盘总结。此时，系统已经拥有了丰富的素材：
+- 一周的 Digest 预处理结果
+- 一周的 Daily Review 精选内容
+- 若干经过价值确认的 Lumina 知识库内容
+- 一些开始反复出现的热门主题
+此时生成周刊，比单纯从网页上抓取热点更有价值。因为这些内容已经经过了个人筛选和沉淀，带有明显的个性化痕迹。周刊不应仅仅是本周发生了什么的简单罗列，更应该具备深度和价值：
+- 本周有哪些主题值得重点关注和记忆
+- 哪些变化只是短期噪音，哪些是值得重视的长期信号
+- 哪些内容具有长期参考价值，值得反复回看
+周刊合集👉🏻：肖恩技术周刊
+2. 主题文章生成
+另一个方向是让系统能够逐渐识别哪些主题已经积累了足够的素材，值得写成长篇文章。
+虽然信息流中的内容看起来是离散的，但如果拉长时间维度观察，就会发现很多内容其实都在指向同一个核心主题，例如：
+- AI Agent 工程的发展趋势
+- 开源工具链的演变
+- 内容平台分发机制的变革
+- 隐私保护、合规要求与数据治理
+一旦某个主题在一段时间内反复出现，并且我多次对相关内容进行精读、收藏和沉淀，那么它就不应再仅仅是多条零散的新闻，而应该逐渐发展成为一个可以深入挖掘和输出的长期主题。
+内容合集👉🏻：今日观察
+未来迭代方向
+尽管工作流目前已经能够稳定运行，但它远未达到最终完成的状态。可以继续完善的方面有：
+细化反馈机制
+目前系统中最强的反馈信号是这篇内容是否被存入 Lumina。但在理想状态下，我希望系统能够逐渐识别更多层次的用户行为：
+- 点开内容但未读完
+- 读完内容但未收藏
+- 内容值得精读
+- 内容值得长期沉淀
+- 内容最终影响了写作、决策或实际实现
+- 不喜欢的内容（负反馈）
+一旦这些层次的反馈机制得以完善，兴趣画像将变得更加精准，不再只是一个粗粒度的偏好集合。
+进一步抽象流程
+目前，我更倾向于继续在现有的 OpenClaw 体系内迭代优化，这是探索阶段最适合的方式。
+但如果这套流程能够变得更加稳定，职责边界也更加清晰，那么抽象出其中的信息处理内核，会更有利于后续工程化迭代。
+不过比起将其产品化，还是先让这套漏斗系统持续稳定地运转，越来越懂我。
+结语
+信息处理的关键，从来不是看到更多，而是让真正重要的信息被自己接住。
+当信息经过筛选、精读、沉淀和反馈，真正融入知识结构时，我们不再是信息的被动消费者，将真正成为自己信息环境的主人。
+感谢阅读
+微信公众号「肖恩聊技术」
+如果这篇文章对你有帮助，欢迎扫码关注，获取原创文章推送。

outputs/README.md

@@ -0,0 +1,83 @@
+# Outputs Layout
+
+`outputs/` 按来源和运行批次组织。当前推荐区分生产产物与调试产物。
+
+## 生产默认输出
+
+对于 FreshRSS 主流水线，默认只保留这 3 类文件：
+
+- `outputs/freshrss/rerun/<run_id>/raw/freshrss.raw.json`
+  - FreshRSS 原始响应快照
+- `outputs/freshrss/rerun/<run_id>/candidates/openclaw-delivery-payload.json`
+  - 发给 OpenClaw 的最终批量 payload
+- `outputs/freshrss/rerun/<run_id>/run-report.json`
+  - 本次批处理的运行报告
+
+这是当前推荐的生产模式。
+
+## 词元统计运行数据
+
+词元统计不放在 `outputs/` 下，而放在单独的运行数据目录：
+
+- `data/term_index/daily/YYYY-MM-DD.json`
+  - 某一天的日报级 `keywords` 聚合结果
+- `data/term_index/term_stats.json`
+  - 全局累计词频统计
+
+这两类文件属于可重建的本地运行数据，不作为仓库长期跟踪产物。
+
+## 调试扩展输出
+
+当启用 `debug_artifacts` 时，才会额外写出这些中间文件：
+
+- `outputs/freshrss/rerun/<run_id>/items/`
+  - 标准化 `item` 中间文件
+- `outputs/freshrss/rerun/<run_id>/extracted/`
+  - 提取结果
+- `outputs/freshrss/rerun/<run_id>/summary/`
+  - LLM 总结结果及调试尝试文件
+- `outputs/freshrss/rerun/<run_id>/filter/`
+  - 规则过滤结果
+- `outputs/freshrss/rerun/<run_id>/candidates/*.article-candidate-record.json`
+  - 内部候选记录
+- `outputs/freshrss/rerun/<run_id>/candidates/*.openclaw-candidate-input.json`
+  - 单篇 OpenClaw 输入对象
+
+## 何时启用调试产物
+
+只在这些场景启用：
+
+- 某一批次有异常，需要定位是提取、总结还是过滤阶段出错
+- 需要核对某条内容为什么被 `keep` / `review` / `drop`
+- 需要为规则调整收集样本
+
+平时不要默认开启。
+
+## 参考目录
+
+- `outputs/prompts/`
+  - Prompt 模板
+- `outputs/reference/`
+  - 手工样例、固定参考输入输出
+- `outputs/freshrss/`
+  - FreshRSS 实跑批次输出
+
+## 命名原则
+
+- 每次批处理使用独立目录：`outputs/freshrss/rerun/<timestamp>/`
+- 默认优先保留最终产物，不把所有中间文件都当成长期资产
+- 调试文件只在需要时生成
+- `reference/` 只放可复用样例，不放真实生产批次数据
+- `data/term_index/` 只保存本地词元统计运行数据，不纳入 git 跟踪
+
+## 当前建议
+
+如果目标是交给 OpenClaw 稳定消费，优先依赖：
+
+- `openclaw-delivery-payload.json`
+- `run-report.json`
+- `freshrss.raw.json`
+- `data/term_index/daily/YYYY-MM-DD.json`
+- `data/term_index/term_stats.json`
+
+其中前 3 个是本次批处理产物，后 2 个是持续累积的词元统计数据。

outputs/reference/candidates/openclaw-candidate-input.sample.json

@@ -0,0 +1,40 @@
+{
+  "candidate_id": "cand:sha256:4139f277b8cb621b02b3398f8a7bd78e3f9eef7bcd320b70b8590e802540c6ca",
+  "title": "套壳中国大模型撑起500亿美元估值？扒一扒 Cursor 的\"套壳\"疑云",
+  "url": "http://www.ruanyifeng.com/blog/2026/03/kimi-cursor.html",
+  "published_at": "2026-03-21T10:19:11Z",
+  "author": "阮一峰",
+  "source_name": "阮一峰的网络日志",
+  "summary": "AI编程工具Cursor推出的Composer 2模型被证实套壳中国Kimi K2.5模型，引发侵权争议。Kimi官方确认Cursor通过Fireworks AI获得授权，不存在侵权。作者分析Cursor隐瞒事实是为了支撑其不断膨胀的估值，将其包装成大模型公司。",
+  "highlights": [
+    "Cursor的Composer 2模型被技术手段揭露实际调用的是Kimi K2.5模型。",
+    "Kimi官方确认Cursor通过Fireworks AI获得授权，因此不构成侵权。",
+    "Cursor隐瞒使用Kimi模型，被认为是为了支撑其高达500亿美元的估值。",
+    "事件显示中国大模型技术已具备输出能力，国产模型与国外旗舰差距缩小。",
+    "Composer 2性能低于GPT-5.4但成本最低，生成速度较快。"
+  ],
+  "keywords": [
+    "Cursor",
+    "Composer 2",
+    "Kimi K2.5",
+    "Fireworks AI",
+    "套壳模型",
+    "模型授权",
+    "估值泡沫",
+    "AI编程工具"
+  ],
+  "topics": [
+    "人工智能",
+    "大模型",
+    "商业伦理",
+    "技术争议",
+    "创业融资"
+  ],
+  "category": "观点评论",
+  "worth_keeping": true,
+  "worth_reason": "文章深入剖析了AI行业的热点事件，涉及技术真相、商业动机和行业趋势，具有较高的参考价值。",
+  "selection_decision": "review",
+  "selection_reason": "Worth-keeping signal is positive but no stronger keep rule matched.",
+  "digest_section_hint": "insights",
+  "digest_rank": 60
+}

outputs/reference/filter/filter-decision.json

@@ -0,0 +1,39 @@
+{
+  "decision": "keep",
+  "matched_rules": [
+    "keep-worth-keeping-method",
+    "review-worth-keeping-other"
+  ],
+  "reasons": [
+    "Structured summary marked the content as worth keeping in a durable category.",
+    "Worth-keeping signal is positive but no stronger keep rule matched."
+  ],
+  "labels": [
+    "durable",
+    "needs-review",
+    "summary"
+  ],
+  "priority": 80,
+  "matches": [
+    {
+      "rule_id": "keep-worth-keeping-method",
+      "decision": "keep",
+      "reason": "Structured summary marked the content as worth keeping in a durable category.",
+      "labels": [
+        "summary",
+        "durable"
+      ],
+      "priority": 80
+    },
+    {
+      "rule_id": "review-worth-keeping-other",
+      "decision": "review",
+      "reason": "Worth-keeping signal is positive but no stronger keep rule matched.",
+      "labels": [
+        "summary",
+        "needs-review"
+      ],
+      "priority": 60
+    }
+  ]
+}

outputs/reference/filter/filter-decision.with-context.json

@@ -0,0 +1,53 @@
+{
+  "decision": "keep",
+  "matched_rules": [
+    "keep-worth-keeping-method",
+    "keep-interest-topic",
+    "review-worth-keeping-other"
+  ],
+  "reasons": [
+    "Structured summary marked the content as worth keeping in a durable category.",
+    "Topics overlap with current interest profile.",
+    "Worth-keeping signal is positive but no stronger keep rule matched."
+  ],
+  "labels": [
+    "durable",
+    "interest",
+    "needs-review",
+    "summary",
+    "topic-match"
+  ],
+  "priority": 80,
+  "matches": [
+    {
+      "rule_id": "keep-worth-keeping-method",
+      "decision": "keep",
+      "reason": "Structured summary marked the content as worth keeping in a durable category.",
+      "labels": [
+        "summary",
+        "durable"
+      ],
+      "priority": 80
+    },
+    {
+      "rule_id": "keep-interest-topic",
+      "decision": "keep",
+      "reason": "Topics overlap with current interest profile.",
+      "labels": [
+        "interest",
+        "topic-match"
+      ],
+      "priority": 75
+    },
+    {
+      "rule_id": "review-worth-keeping-other",
+      "decision": "review",
+      "reason": "Worth-keeping signal is positive but no stronger keep rule matched.",
+      "labels": [
+        "summary",
+        "needs-review"
+      ],
+      "priority": 60
+    }
+  ]
+}

scripts/apply_term_suggestions.py

@@ -0,0 +1,402 @@
+from __future__ import annotations
+
+import argparse
+import json
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_ALIASES_PATH = REPO_ROOT / "configs" / "term_aliases.json"
+DEFAULT_STOPWORDS_PATH = REPO_ROOT / "configs" / "term_stopwords.json"
+DEFAULT_CONTEXT_PATH = REPO_ROOT / "configs" / "filter_context.personal.json"
+DEFAULT_WATCHLIST_PATH = REPO_ROOT / "configs" / "term_watchlist.json"
+DEFAULT_CHANGE_LOG_PATH = REPO_ROOT / "configs" / "term_change_log.json"
+
+
+def _load_json(path: Path, default: Any) -> Any:
+    if not path.exists():
+        return default
+    return json.loads(path.read_text(encoding="utf-8-sig"))
+
+
+def _save_json(path: Path, payload: dict[str, Any] | list[Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+
+
+def _term_key(value: str) -> str:
+    return value.strip().casefold()
+
+
+def _utc_now() -> str:
+    return datetime.now(tz=UTC).isoformat().replace("+00:00", "Z")
+
+
+def _load_suggestions(path: Path) -> dict[str, Any]:
+    payload = _load_json(path, {})
+    if not isinstance(payload, dict):
+        raise RuntimeError("Suggestions file must contain a JSON object.")
+    return payload
+
+
+def _load_aliases(path: Path) -> dict[str, str]:
+    payload = _load_json(path, {})
+    if not isinstance(payload, dict):
+        raise RuntimeError("Aliases file must contain a JSON object.")
+    result: dict[str, str] = {}
+    for key, value in payload.items():
+        if isinstance(key, str) and isinstance(value, str) and key.strip() and value.strip():
+            result[key] = value
+    return result
+
+
+def _load_stopwords(path: Path) -> list[str]:
+    payload = _load_json(path, [])
+    if not isinstance(payload, list):
+        raise RuntimeError("Stopwords file must contain a JSON array.")
+    return [item for item in payload if isinstance(item, str) and item.strip()]
+
+
+def _load_context(path: Path) -> dict[str, Any]:
+    payload = _load_json(path, {})
+    if not isinstance(payload, dict):
+        raise RuntimeError("Context file must contain a JSON object.")
+    payload.setdefault("interest_keywords", [])
+    if not isinstance(payload["interest_keywords"], list):
+        raise RuntimeError("filter_context.personal.json interest_keywords must be a JSON array.")
+    payload["interest_keywords"] = [item for item in payload["interest_keywords"] if isinstance(item, str) and item.strip()]
+    return payload
+
+
+def _load_watchlist(path: Path) -> dict[str, Any]:
+    payload = _load_json(
+        path,
+        {
+            "schema_version": "v1",
+            "updated_at": _utc_now(),
+            "terms": [],
+        },
+    )
+    if not isinstance(payload, dict):
+        raise RuntimeError("Watchlist file must contain a JSON object.")
+    payload.setdefault("schema_version", "v1")
+    payload.setdefault("updated_at", _utc_now())
+    terms = payload.get("terms", [])
+    if not isinstance(terms, list):
+        raise RuntimeError("Watchlist terms must be a JSON array.")
+    payload["terms"] = [item for item in terms if isinstance(item, dict) and isinstance(item.get("term"), str)]
+    return payload
+
+
+def _load_change_log(path: Path) -> dict[str, Any]:
+    payload = _load_json(
+        path,
+        {
+            "schema_version": "v1",
+            "entries": [],
+        },
+    )
+    if not isinstance(payload, dict):
+        raise RuntimeError("Change log file must contain a JSON object.")
+    payload.setdefault("schema_version", "v1")
+    entries = payload.get("entries", [])
+    if not isinstance(entries, list):
+        raise RuntimeError("Change log entries must be a JSON array.")
+    payload["entries"] = [item for item in entries if isinstance(item, dict)]
+    return payload
+
+
+def _build_index(entries: list[dict[str, Any]], field: str) -> dict[str, dict[str, Any]]:
+    index: dict[str, dict[str, Any]] = {}
+    for entry in entries:
+        value = entry.get(field)
+        if isinstance(value, str) and value.strip():
+            index[_term_key(value)] = entry
+    return index
+
+
+def _append_change(
+    change_log: dict[str, Any],
+    *,
+    applied_at: str,
+    action: str,
+    term: str,
+    value: str | None,
+    reason: str,
+    suggestions_path: Path,
+    suggestion_date: str | None,
+    based_on_days: int | None,
+) -> None:
+    entry: dict[str, Any] = {
+        "applied_at": applied_at,
+        "action": action,
+        "term": term,
+        "reason": reason,
+        "suggestions_path": str(suggestions_path),
+    }
+    if value is not None:
+        entry["value"] = value
+    if suggestion_date is not None:
+        entry["suggestion_date"] = suggestion_date
+    if based_on_days is not None:
+        entry["based_on_days"] = based_on_days
+    change_log["entries"].append(entry)
+
+
+def _remove_from_watchlist(
+    watchlist: dict[str, Any],
+    *,
+    term: str,
+) -> bool:
+    folded = _term_key(term)
+    original_len = len(watchlist["terms"])
+    watchlist["terms"] = [item for item in watchlist["terms"] if _term_key(str(item.get("term", ""))) != folded]
+    return len(watchlist["terms"]) != original_len
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Apply accepted term cleanup suggestions into repo config files.")
+    parser.add_argument("--suggestions", type=Path, required=True, help="Suggestion JSON file")
+    parser.add_argument("--aliases", type=Path, default=DEFAULT_ALIASES_PATH, help="Term aliases JSON file")
+    parser.add_argument("--stopwords", type=Path, default=DEFAULT_STOPWORDS_PATH, help="Term stopwords JSON file")
+    parser.add_argument("--context", type=Path, default=DEFAULT_CONTEXT_PATH, help="Personal filter context JSON file")
+    parser.add_argument("--watchlist", type=Path, default=DEFAULT_WATCHLIST_PATH, help="Watchlist JSON file")
+    parser.add_argument("--change-log", type=Path, default=DEFAULT_CHANGE_LOG_PATH, help="Change log JSON file")
+    parser.add_argument("--accept-watch", nargs="*", default=[], help="Accept watch_terms by term name")
+    parser.add_argument(
+        "--accept-interest", nargs="*", default=[], help="Accept interest_keyword_suggestions by term name"
+    )
+    parser.add_argument("--accept-stopword", nargs="*", default=[], help="Accept stopword_suggestions by term name")
+    parser.add_argument("--accept-alias", nargs="*", default=[], help="Accept alias_suggestions by source term")
+    parser.add_argument("--dry-run", action="store_true", help="Preview changes without writing files")
+    args = parser.parse_args()
+
+    suggestions = _load_suggestions(args.suggestions)
+    aliases = _load_aliases(args.aliases)
+    stopwords = _load_stopwords(args.stopwords)
+    context = _load_context(args.context)
+    watchlist = _load_watchlist(args.watchlist)
+    change_log = _load_change_log(args.change_log)
+
+    alias_suggestions = suggestions.get("alias_suggestions", [])
+    stopword_suggestions = suggestions.get("stopword_suggestions", [])
+    interest_suggestions = suggestions.get("interest_keyword_suggestions", [])
+    watch_suggestions = suggestions.get("watch_terms", [])
+    if not all(isinstance(bucket, list) for bucket in [alias_suggestions, stopword_suggestions, interest_suggestions, watch_suggestions]):
+        raise RuntimeError("Suggestion JSON buckets must all be arrays.")
+
+    alias_index = _build_index([item for item in alias_suggestions if isinstance(item, dict)], "from")
+    stopword_index = _build_index([item for item in stopword_suggestions if isinstance(item, dict)], "term")
+    interest_index = _build_index([item for item in interest_suggestions if isinstance(item, dict)], "term")
+    watch_index = _build_index([item for item in watch_suggestions if isinstance(item, dict)], "term")
+
+    watch_map = {_term_key(str(item.get("term", ""))): item for item in watchlist["terms"]}
+    stopword_set = {_term_key(item) for item in stopwords}
+    interest_set = {_term_key(item) for item in context["interest_keywords"]}
+    alias_source_set = {_term_key(key) for key in aliases}
+
+    applied_at = _utc_now()
+    suggestion_date = suggestions.get("date") if isinstance(suggestions.get("date"), str) else None
+    based_on_days = suggestions.get("based_on_days") if isinstance(suggestions.get("based_on_days"), int) else None
+
+    applied: list[dict[str, Any]] = []
+    skipped: list[dict[str, Any]] = []
+
+    def skip(action: str, term: str, reason: str) -> None:
+        skipped.append({"action": action, "term": term, "reason": reason})
+
+    def applied_entry(action: str, term: str, reason: str, value: str | None = None) -> None:
+        item: dict[str, Any] = {"action": action, "term": term, "reason": reason}
+        if value is not None:
+            item["value"] = value
+        applied.append(item)
+
+    for raw_term in args.accept_watch:
+        term = raw_term.strip()
+        suggestion = watch_index.get(_term_key(term))
+        if suggestion is None:
+            skip("add_watch_term", term, "Term not found in watch_terms suggestions.")
+            continue
+        if _term_key(term) in watch_map:
+            skip("add_watch_term", term, "Term already exists in watchlist.")
+            continue
+        reason = str(suggestion.get("reason") or "Accepted from watch_terms suggestion.")
+        watch_item = {
+            "term": str(suggestion.get("term") or term),
+            "added_at": applied_at,
+            "source": str(args.suggestions),
+            "reason": reason,
+            "status": "watching",
+        }
+        watchlist["terms"].append(watch_item)
+        watch_map[_term_key(watch_item["term"])] = watch_item
+        _append_change(
+            change_log,
+            applied_at=applied_at,
+            action="add_watch_term",
+            term=watch_item["term"],
+            value=None,
+            reason=reason,
+            suggestions_path=args.suggestions,
+            suggestion_date=suggestion_date,
+            based_on_days=based_on_days,
+        )
+        applied_entry("add_watch_term", watch_item["term"], reason)
+
+    for raw_term in args.accept_interest:
+        term = raw_term.strip()
+        suggestion = interest_index.get(_term_key(term))
+        if suggestion is None:
+            skip("add_interest_keyword", term, "Term not found in interest_keyword_suggestions.")
+            continue
+        resolved_term = str(suggestion.get("term") or term)
+        if _term_key(resolved_term) in interest_set:
+            skip("add_interest_keyword", resolved_term, "Term already exists in interest_keywords.")
+            continue
+        reason = str(suggestion.get("reason") or "Accepted from interest_keyword_suggestions.")
+        context["interest_keywords"].append(resolved_term)
+        interest_set.add(_term_key(resolved_term))
+        _append_change(
+            change_log,
+            applied_at=applied_at,
+            action="add_interest_keyword",
+            term=resolved_term,
+            value=None,
+            reason=reason,
+            suggestions_path=args.suggestions,
+            suggestion_date=suggestion_date,
+            based_on_days=based_on_days,
+        )
+        if _remove_from_watchlist(watchlist, term=resolved_term):
+            _append_change(
+                change_log,
+                applied_at=applied_at,
+                action="remove_watch_term",
+                term=resolved_term,
+                value="promoted_to_interest_keyword",
+                reason="Removed from watchlist after promotion into interest_keywords.",
+                suggestions_path=args.suggestions,
+                suggestion_date=suggestion_date,
+                based_on_days=based_on_days,
+            )
+        applied_entry("add_interest_keyword", resolved_term, reason)
+
+    for raw_term in args.accept_stopword:
+        term = raw_term.strip()
+        suggestion = stopword_index.get(_term_key(term))
+        if suggestion is None:
+            skip("add_stopword", term, "Term not found in stopword_suggestions.")
+            continue
+        resolved_term = str(suggestion.get("term") or term)
+        if _term_key(resolved_term) in stopword_set:
+            skip("add_stopword", resolved_term, "Term already exists in stopwords.")
+            continue
+        reason = str(suggestion.get("reason") or "Accepted from stopword_suggestions.")
+        stopwords.append(resolved_term)
+        stopword_set.add(_term_key(resolved_term))
+        _append_change(
+            change_log,
+            applied_at=applied_at,
+            action="add_stopword",
+            term=resolved_term,
+            value=None,
+            reason=reason,
+            suggestions_path=args.suggestions,
+            suggestion_date=suggestion_date,
+            based_on_days=based_on_days,
+        )
+        if _remove_from_watchlist(watchlist, term=resolved_term):
+            _append_change(
+                change_log,
+                applied_at=applied_at,
+                action="remove_watch_term",
+                term=resolved_term,
+                value="promoted_to_stopword",
+                reason="Removed from watchlist after being added to stopwords.",
+                suggestions_path=args.suggestions,
+                suggestion_date=suggestion_date,
+                based_on_days=based_on_days,
+            )
+        applied_entry("add_stopword", resolved_term, reason)
+
+    for raw_term in args.accept_alias:
+        term = raw_term.strip()
+        suggestion = alias_index.get(_term_key(term))
+        if suggestion is None:
+            skip("add_alias", term, "Source term not found in alias_suggestions.")
+            continue
+        source_term = str(suggestion.get("from") or term)
+        target_term = str(suggestion.get("to") or "").strip()
+        if not target_term:
+            skip("add_alias", source_term, "Alias suggestion target is empty.")
+            continue
+        existing_target = aliases.get(source_term)
+        if existing_target == target_term:
+            skip("add_alias", source_term, "Alias already exists with the same target.")
+            continue
+        if _term_key(source_term) in alias_source_set and existing_target != target_term:
+            skip("add_alias", source_term, f"Alias source already exists with a different target: {existing_target}")
+            continue
+        reason = str(suggestion.get("reason") or "Accepted from alias_suggestions.")
+        aliases[source_term] = target_term
+        alias_source_set.add(_term_key(source_term))
+        _append_change(
+            change_log,
+            applied_at=applied_at,
+            action="add_alias",
+            term=source_term,
+            value=target_term,
+            reason=reason,
+            suggestions_path=args.suggestions,
+            suggestion_date=suggestion_date,
+            based_on_days=based_on_days,
+        )
+        if _remove_from_watchlist(watchlist, term=source_term):
+            _append_change(
+                change_log,
+                applied_at=applied_at,
+                action="remove_watch_term",
+                term=source_term,
+                value="resolved_as_alias_source",
+                reason="Removed from watchlist after alias mapping was accepted.",
+                suggestions_path=args.suggestions,
+                suggestion_date=suggestion_date,
+                based_on_days=based_on_days,
+            )
+        applied_entry("add_alias", source_term, reason, value=target_term)
+
+    watchlist["terms"] = sorted(watchlist["terms"], key=lambda item: (_term_key(str(item.get("term", ""))), str(item.get("term", ""))))
+    watchlist["updated_at"] = applied_at
+    stopwords = sorted(stopwords, key=lambda item: (_term_key(item), item))
+    context["interest_keywords"] = sorted(context["interest_keywords"], key=lambda item: (_term_key(item), item))
+
+    summary = {
+        "dry_run": args.dry_run,
+        "suggestions": str(args.suggestions),
+        "applied_count": len(applied),
+        "skipped_count": len(skipped),
+        "applied": applied,
+        "skipped": skipped,
+        "output_paths": {
+            "aliases": str(args.aliases),
+            "stopwords": str(args.stopwords),
+            "context": str(args.context),
+            "watchlist": str(args.watchlist),
+            "change_log": str(args.change_log),
+        },
+    }
+
+    if not args.dry_run:
+        _save_json(args.aliases, aliases)
+        _save_json(args.stopwords, stopwords)
+        _save_json(args.context, context)
+        _save_json(args.watchlist, watchlist)
+        _save_json(args.change_log, change_log)
+
+    print(json.dumps(summary, ensure_ascii=False, indent=2))
+
+
+if __name__ == "__main__":
+    main()

scripts/build_keyword_index.py

@@ -0,0 +1,76 @@
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = REPO_ROOT / "src"
+
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from summary_mcp.core.keyword_index import persist_keyword_indexes
+from summary_mcp.models.openclaw_delivery import OpenClawDeliveryPayload
+
+
+def _load_payload(path: Path) -> OpenClawDeliveryPayload:
+    return OpenClawDeliveryPayload.model_validate_json(path.read_text(encoding="utf-8-sig"))
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Build the daily keyword index and global keyword stats from an OpenClaw delivery payload."
+    )
+    parser.add_argument(
+        "--input",
+        type=Path,
+        required=True,
+        help="OpenClaw delivery payload JSON file",
+    )
+    parser.add_argument(
+        "--daily-dir",
+        type=Path,
+        default=REPO_ROOT / "data" / "term_index" / "daily",
+        help="Directory for daily keyword index files",
+    )
+    parser.add_argument(
+        "--stats-output",
+        type=Path,
+        default=REPO_ROOT / "data" / "term_index" / "term_stats.json",
+        help="Global keyword stats output path",
+    )
+    parser.add_argument(
+        "--aliases",
+        type=Path,
+        default=REPO_ROOT / "configs" / "term_aliases.json",
+        help="Keyword alias config JSON file",
+    )
+    parser.add_argument(
+        "--stopwords",
+        type=Path,
+        default=REPO_ROOT / "configs" / "term_stopwords.json",
+        help="Keyword stopwords config JSON file",
+    )
+    args = parser.parse_args()
+
+    payload = _load_payload(args.input)
+    result = persist_keyword_indexes(
+        payload.candidates,
+        for_date=payload.date,
+        digest_id=payload.run_id,
+        source="openclaw_delivery_payload",
+        daily_dir=args.daily_dir,
+        stats_path=args.stats_output,
+        aliases_path=args.aliases,
+        stopwords_path=args.stopwords,
+    )
+
+    print(f"Saved daily keyword index to {result['daily_output']}")
+    print(f"Saved keyword stats to {result['stats_output']}")
+    print(f"Indexed {result['candidate_count']} candidates and {result['term_count']} keywords")
+
+
+if __name__ == "__main__":
+    main()

scripts/build_openclaw_delivery.py

@@ -0,0 +1,85 @@
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from datetime import date
+from pathlib import Path
+from typing import Any
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = REPO_ROOT / "src"
+OUTPUT_ROOT = REPO_ROOT / "outputs"
+REFERENCE_OUTPUT_ROOT = OUTPUT_ROOT / "reference"
+
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from summary_mcp.models import OpenClawCandidateInput, build_openclaw_delivery_payload
+
+
+def _load_json(path: Path) -> dict[str, Any]:
+    return json.loads(path.read_text(encoding="utf-8-sig"))
+
+
+def _save_json(path: Path, payload: dict[str, Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def _collect_input_paths(inputs: list[Path], input_dir: Path | None, pattern: str) -> list[Path]:
+    resolved = list(inputs)
+    if input_dir is not None:
+        resolved.extend(sorted(input_dir.glob(pattern)))
+    seen: set[Path] = set()
+    unique_paths: list[Path] = []
+    for path in resolved:
+        normalized = path.resolve()
+        if normalized in seen:
+            continue
+        seen.add(normalized)
+        unique_paths.append(path)
+    return unique_paths
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Build an OpenClaw batch delivery payload from slim candidate JSON files.")
+    parser.add_argument("--inputs", nargs="*", type=Path, default=[], help="One or more OpenClaw candidate input JSON files")
+    parser.add_argument("--input-dir", type=Path, default=None, help="Directory containing OpenClaw candidate input JSON files")
+    parser.add_argument(
+        "--pattern",
+        type=str,
+        default="*.openclaw-candidate-input.json",
+        help="Glob pattern used together with --input-dir",
+    )
+    parser.add_argument("--run-id", type=str, default=None, help="Optional delivery run id")
+    parser.add_argument("--date", type=str, default=None, help="Optional delivery date in YYYY-MM-DD format")
+    parser.add_argument("--sort-by-rank", action="store_true", help="Sort candidates by digest_rank descending")
+    parser.add_argument(
+        "--output",
+        type=Path,
+        default=REFERENCE_OUTPUT_ROOT / "candidates" / "openclaw-delivery-payload.json",
+        help="Where to save the OpenClaw delivery payload",
+    )
+    args = parser.parse_args()
+
+    input_paths = _collect_input_paths(args.inputs, args.input_dir, args.pattern)
+    if not input_paths:
+        raise SystemExit("No candidate input files found. Pass --inputs or --input-dir.")
+
+    candidates = [OpenClawCandidateInput.model_validate(_load_json(path)) for path in input_paths]
+    if args.sort_by_rank:
+        candidates.sort(key=lambda candidate: candidate.digest_rank, reverse=True)
+
+    payload = build_openclaw_delivery_payload(
+        candidates,
+        run_id=args.run_id,
+        for_date=date.fromisoformat(args.date) if args.date else None,
+    )
+    _save_json(args.output, payload.model_dump(mode="json"))
+    print(f"Saved OpenClaw delivery payload with {len(candidates)} candidates to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

scripts/pull_freshrss_items.py

@@ -0,0 +1,114 @@
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+from typing import Any
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = REPO_ROOT / "src"
+OUTPUT_ROOT = REPO_ROOT / "outputs"
+FRESHRSS_OUTPUT_ROOT = OUTPUT_ROOT / "freshrss"
+
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from summary_mcp.integrations.freshrss import READ_TAG, FreshRSSClient, map_entry_to_item
+
+
+def _save_json(path: Path, payload: dict[str, Any] | list[Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def _load_required(name: str, value: str | None) -> str:
+    resolved = value or os.environ.get(name)
+    if not resolved:
+        cli_name = name.lower().replace("_", "-")
+        raise RuntimeError(f"Missing required value: pass --{cli_name} or set {name}.")
+    return resolved
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Pull FreshRSS entries and map them into normalized item objects.")
+    parser.add_argument("--api-base-url", type=str, default=None, help="FreshRSS greader API base URL")
+    parser.add_argument("--username", type=str, default=None, help="FreshRSS API username")
+    parser.add_argument("--api-password", type=str, default=None, help="FreshRSS API password")
+    parser.add_argument(
+        "--stream-id",
+        type=str,
+        default="user/-/state/com.google/reading-list",
+        help="Google Reader API stream id",
+    )
+    parser.add_argument("--limit", type=int, default=10, help="Maximum number of entries to request")
+    parser.add_argument("--continuation", type=str, default=None, help="Continuation token for paging")
+    parser.add_argument(
+        "--include-read",
+        action="store_true",
+        help="Do not exclude items already tagged as read.",
+    )
+    parser.add_argument(
+        "--mark-read",
+        action="store_true",
+        help="Mark fetched entries as read after this script finishes successfully.",
+    )
+    parser.add_argument(
+        "--raw-output",
+        type=Path,
+        default=FRESHRSS_OUTPUT_ROOT / "raw" / "freshrss.raw.json",
+        help="Where to save the raw FreshRSS response",
+    )
+    parser.add_argument(
+        "--items-output",
+        type=Path,
+        default=FRESHRSS_OUTPUT_ROOT / "items" / "freshrss.items.json",
+        help="Where to save the mapped item list",
+    )
+    parser.add_argument("--timeout", type=float, default=20.0, help="Request timeout in seconds")
+    args = parser.parse_args()
+
+    api_base_url = _load_required("FRESHRSS_API_BASE_URL", args.api_base_url)
+    username = _load_required("FRESHRSS_USERNAME", args.username)
+    api_password = _load_required("FRESHRSS_API_PASSWORD", args.api_password)
+
+    client = FreshRSSClient(
+        api_base_url=api_base_url,
+        username=username,
+        api_password=api_password,
+        timeout_seconds=args.timeout,
+    )
+    auth_token = client.client_login()
+    payload = client.fetch_stream_contents(
+        auth_token=auth_token,
+        stream_id=args.stream_id,
+        limit=args.limit,
+        continuation=args.continuation,
+        exclude_targets=[] if args.include_read else [READ_TAG],
+    )
+
+    entries = payload.get("items")
+    if not isinstance(entries, list):
+        raise RuntimeError("FreshRSS stream response does not contain an items array.")
+
+    mapped_items = [map_entry_to_item(entry).model_dump(mode="json") for entry in entries]
+
+    _save_json(args.raw_output, payload)
+    _save_json(args.items_output, mapped_items)
+
+    marked_count = 0
+    if args.mark_read:
+        item_ids = [entry.get("external_id") for entry in mapped_items if isinstance(entry.get("external_id"), str)]
+        if item_ids:
+            client.mark_items_as_read(auth_token=auth_token, item_ids=item_ids)
+            marked_count = len(item_ids)
+
+    print(f"Saved {len(mapped_items)} mapped items to {args.items_output}")
+    if args.mark_read:
+        print(f"Marked {marked_count} FreshRSS entries as read")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_article_candidate.py

@@ -0,0 +1,124 @@
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = REPO_ROOT / "src"
+OUTPUT_ROOT = REPO_ROOT / "outputs"
+REFERENCE_OUTPUT_ROOT = OUTPUT_ROOT / "reference"
+
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from summary_mcp.models.article_candidate import (
+    CandidateMetadata,
+    CandidateSourceRefs,
+    build_article_candidate_record,
+    build_openclaw_candidate_input,
+)
+from summary_mcp.models.document import ExtractedArticle
+from summary_mcp.models.filtering import FilterDecisionResult
+from summary_mcp.models.item import Item
+from summary_mcp.models.llm_result import LlmSummaryResult
+
+
+def _load_json(path: Path) -> dict[str, Any]:
+    return json.loads(path.read_text(encoding="utf-8-sig"))
+
+
+def _save_json(path: Path, payload: dict[str, Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Build an internal ArticleCandidateRecord and a slim OpenClawCandidateInput from pipeline outputs."
+    )
+    parser.add_argument("--summary", type=Path, required=True, help="Structured LLM summary JSON file")
+    parser.add_argument("--extracted", type=Path, required=True, help="Extracted article JSON file")
+    parser.add_argument("--filter", type=Path, required=True, help="Filter decision JSON file")
+    parser.add_argument("--item", type=Path, default=None, help="Optional normalized item JSON file")
+    parser.add_argument(
+        "--section-hint",
+        default=None,
+        choices=[
+            "top_news",
+            "tools_and_workflows",
+            "risk_and_security",
+            "open_source",
+            "insights",
+            "deep_dive",
+        ],
+        help="Optional digest section hint for downstream aggregation",
+    )
+    parser.add_argument(
+        "--rank",
+        type=int,
+        default=None,
+        help="Optional digest rank override. Defaults to filter priority when omitted.",
+    )
+    parser.add_argument(
+        "--rendered-markdown",
+        type=Path,
+        default=None,
+        help="Optional markdown file to embed as rendered_markdown",
+    )
+    parser.add_argument(
+        "--output",
+        dest="record_output",
+        type=Path,
+        default=REFERENCE_OUTPUT_ROOT / "candidates" / "article-candidate-record.json",
+        help="Where to save the internal ArticleCandidateRecord payload",
+    )
+    parser.add_argument(
+        "--openclaw-output",
+        type=Path,
+        default=REFERENCE_OUTPUT_ROOT / "candidates" / "openclaw-candidate-input.json",
+        help="Where to save the slim OpenClawCandidateInput payload",
+    )
+    args = parser.parse_args()
+
+    summary = LlmSummaryResult.model_validate(_load_json(args.summary))
+    extracted_payload = _load_json(args.extracted)
+    article = ExtractedArticle.model_validate(extracted_payload.get("article", extracted_payload))
+    decision = FilterDecisionResult.model_validate(_load_json(args.filter))
+    item = Item.model_validate(_load_json(args.item)) if args.item else None
+    rendered_markdown = args.rendered_markdown.read_text(encoding="utf-8-sig") if args.rendered_markdown else None
+
+    record = build_article_candidate_record(
+        summary=summary,
+        article=article,
+        filter_result=decision,
+        item=item,
+        digest_section_hint=args.section_hint,
+        digest_rank=args.rank,
+        rendered_markdown=rendered_markdown,
+        source_refs=CandidateSourceRefs(
+            item_path=str(args.item) if args.item else None,
+            extracted_path=str(args.extracted),
+            summary_path=str(args.summary),
+            filter_path=str(args.filter),
+        ),
+        metadata=CandidateMetadata(
+            generated_at=datetime.now(tz=UTC),
+            producer="run_article_candidate.py",
+            run_id=datetime.now(tz=UTC).strftime("candidate-%Y%m%d-%H%M%S"),
+        ),
+    )
+    openclaw_input = build_openclaw_candidate_input(record)
+
+    _save_json(args.record_output, record.model_dump(mode="json"))
+    _save_json(args.openclaw_output, openclaw_input.model_dump(mode="json"))
+    print(f"Saved article candidate record to {args.record_output}")
+    print(f"Saved OpenClaw candidate input to {args.openclaw_output}")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_filter_rules.py

@@ -0,0 +1,78 @@
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+from typing import Any
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = REPO_ROOT / "src"
+OUTPUT_ROOT = REPO_ROOT / "outputs"
+REFERENCE_OUTPUT_ROOT = OUTPUT_ROOT / "reference"
+
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from summary_mcp.filters.engine import load_filter_rules, evaluate_filter_rules
+from summary_mcp.models.document import ExtractedArticle
+from summary_mcp.models.filtering import FilterContext, FilterInput
+from summary_mcp.models.item import Item
+from summary_mcp.models.llm_result import LlmSummaryResult
+
+
+def _load_json(path: Path) -> dict[str, Any]:
+    return json.loads(path.read_text(encoding="utf-8-sig"))
+
+
+def _save_json(path: Path, payload: dict[str, Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Run deterministic filter rules against a structured summary result.")
+    parser.add_argument("--summary", type=Path, required=True, help="Structured LLM summary JSON file")
+    parser.add_argument("--extracted", type=Path, default=None, help="Extracted article JSON file")
+    parser.add_argument("--item", type=Path, default=None, help="Normalized item JSON file")
+    parser.add_argument("--context", type=Path, default=None, help="Optional filter context JSON file")
+    parser.add_argument(
+        "--rules",
+        type=Path,
+        default=REPO_ROOT / "configs" / "filter_rules.json",
+        help="Filter rule config JSON file",
+    )
+    parser.add_argument(
+        "--output",
+        type=Path,
+        default=REFERENCE_OUTPUT_ROOT / "filter" / "filter-decision.json",
+        help="Where to save the filter decision",
+    )
+    args = parser.parse_args()
+
+    summary = LlmSummaryResult.model_validate(_load_json(args.summary))
+
+    article = None
+    if args.extracted is not None:
+        extracted_payload = _load_json(args.extracted)
+        article_payload = extracted_payload.get("article", extracted_payload)
+        article = ExtractedArticle.model_validate(article_payload)
+
+    item = None
+    if args.item is not None:
+        item = Item.model_validate(_load_json(args.item))
+
+    context = FilterContext.model_validate(_load_json(args.context)) if args.context else FilterContext()
+    rules = load_filter_rules(args.rules)
+    decision = evaluate_filter_rules(
+        FilterInput(item=item, article=article, summary=summary, context=context),
+        rules,
+    )
+
+    _save_json(args.output, decision.model_dump(mode="json"))
+    print(f"Saved filter decision to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_freshrss_extract.py

@@ -0,0 +1,148 @@
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+from typing import Any
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = REPO_ROOT / "src"
+OUTPUT_ROOT = REPO_ROOT / "outputs"
+FRESHRSS_OUTPUT_ROOT = OUTPUT_ROOT / "freshrss"
+
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from summary_mcp.core.pipeline import extract_content
+from summary_mcp.integrations.freshrss import READ_TAG, FreshRSSClient, map_entry_to_item
+from summary_mcp.models.summary_io import ExtractionInput
+
+
+def _save_json(path: Path, payload: dict[str, Any] | list[Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def _load_required(name: str, value: str | None) -> str:
+    resolved = value or os.environ.get(name)
+    if not resolved:
+        cli_name = name.lower().replace("_", "-")
+        raise RuntimeError(f"Missing required value: pass --{cli_name} or set {name}.")
+    return resolved
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Pull FreshRSS entries and run content extraction for each mapped item.")
+    parser.add_argument("--api-base-url", type=str, default=None, help="FreshRSS greader API base URL")
+    parser.add_argument("--username", type=str, default=None, help="FreshRSS API username")
+    parser.add_argument("--api-password", type=str, default=None, help="FreshRSS API password")
+    parser.add_argument(
+        "--stream-id",
+        type=str,
+        default="user/-/state/com.google/reading-list",
+        help="Google Reader API stream id",
+    )
+    parser.add_argument("--limit", type=int, default=5, help="Maximum number of entries to request")
+    parser.add_argument("--continuation", type=str, default=None, help="Continuation token for paging")
+    parser.add_argument(
+        "--include-read",
+        action="store_true",
+        help="Do not exclude items already tagged as read.",
+    )
+    parser.add_argument(
+        "--mark-read",
+        action="store_true",
+        help="Mark entries as read after successful extraction for each item.",
+    )
+    parser.add_argument(
+        "--raw-output",
+        type=Path,
+        default=FRESHRSS_OUTPUT_ROOT / "raw" / "freshrss.raw.json",
+        help="Where to save the raw FreshRSS response",
+    )
+    parser.add_argument(
+        "--items-output",
+        type=Path,
+        default=FRESHRSS_OUTPUT_ROOT / "items" / "freshrss.items.json",
+        help="Where to save the mapped item list",
+    )
+    parser.add_argument(
+        "--extracted-output",
+        type=Path,
+        default=FRESHRSS_OUTPUT_ROOT / "extracted" / "freshrss.extracted.json",
+        help="Where to save the extraction results",
+    )
+    parser.add_argument("--timeout", type=float, default=20.0, help="Request timeout in seconds")
+    args = parser.parse_args()
+
+    api_base_url = _load_required("FRESHRSS_API_BASE_URL", args.api_base_url)
+    username = _load_required("FRESHRSS_USERNAME", args.username)
+    api_password = _load_required("FRESHRSS_API_PASSWORD", args.api_password)
+
+    client = FreshRSSClient(
+        api_base_url=api_base_url,
+        username=username,
+        api_password=api_password,
+        timeout_seconds=args.timeout,
+    )
+    auth_token = client.client_login()
+    payload = client.fetch_stream_contents(
+        auth_token=auth_token,
+        stream_id=args.stream_id,
+        limit=args.limit,
+        continuation=args.continuation,
+        exclude_targets=[] if args.include_read else [READ_TAG],
+    )
+
+    entries = payload.get("items")
+    if not isinstance(entries, list):
+        raise RuntimeError("FreshRSS stream response does not contain an items array.")
+
+    mapped_items = [map_entry_to_item(entry) for entry in entries]
+    extraction_results: list[dict[str, Any]] = []
+    success_count = 0
+    mark_read_ids: list[str] = []
+
+    for item in mapped_items:
+        extraction = extract_content(ExtractionInput(item=item))
+        extraction_results.append(
+            {
+                "item": item.model_dump(mode="json"),
+                "extraction": extraction.model_dump(mode="json"),
+            }
+        )
+        if extraction.success:
+            success_count += 1
+            if item.external_id:
+                mark_read_ids.append(item.external_id)
+
+    _save_json(args.raw_output, payload)
+    _save_json(args.items_output, [item.model_dump(mode="json") for item in mapped_items])
+    _save_json(
+        args.extracted_output,
+        {
+            "stream_id": args.stream_id,
+            "requested_limit": args.limit,
+            "entry_count": len(entries),
+            "extracted_success_count": success_count,
+            "results": extraction_results,
+        },
+    )
+
+    marked_count = 0
+    if args.mark_read and mark_read_ids:
+        client.mark_items_as_read(auth_token=auth_token, item_ids=mark_read_ids)
+        marked_count = len(mark_read_ids)
+
+    print(
+        f"Saved {len(mapped_items)} mapped items and {success_count} successful extractions to {args.extracted_output}"
+    )
+    if args.mark_read:
+        print(f"Marked {marked_count} FreshRSS entries as read")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_freshrss_pipeline.py

@@ -0,0 +1,100 @@
+from __future__ import annotations
+
+import argparse
+import sys
+from datetime import date
+from pathlib import Path
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = REPO_ROOT / "src"
+
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from summary_mcp.workflows import run_freshrss_pipeline
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Run the full FreshRSS -> extract -> LLM -> filter -> OpenClaw payload pipeline."
+    )
+    parser.add_argument("--api-base-url", type=str, default=None, help="FreshRSS greader API base URL")
+    parser.add_argument("--username", type=str, default=None, help="FreshRSS API username")
+    parser.add_argument("--api-password", type=str, default=None, help="FreshRSS API password")
+    parser.add_argument(
+        "--stream-id",
+        type=str,
+        default="user/-/state/com.google/reading-list",
+        help="Google Reader API stream id",
+    )
+    parser.add_argument("--limit", type=int, default=5, help="Maximum number of entries to request")
+    parser.add_argument("--continuation", type=str, default=None, help="Continuation token for paging")
+    parser.add_argument(
+        "--include-read",
+        action="store_true",
+        help="Do not exclude items already tagged as read.",
+    )
+    parser.add_argument(
+        "--mark-read",
+        action="store_true",
+        help="Mark only successfully delivered items as read after the final payload is written.",
+    )
+    parser.add_argument(
+        "--debug-artifacts",
+        action="store_true",
+        help="Persist per-item intermediate files for debugging.",
+    )
+    parser.add_argument("--prompt", type=Path, default=None, help="LLM prompt template file")
+    parser.add_argument("--rules", type=Path, default=None, help="Filter rule config JSON file")
+    parser.add_argument("--context", type=Path, default=None, help="Optional filter context JSON file")
+    parser.add_argument("--max-retries", type=int, default=2, help="Number of repair retries after the initial attempt")
+    parser.add_argument("--timeout", type=float, default=60.0, help="Request timeout in seconds")
+    parser.add_argument("--llm-api-key", type=str, default=None, help="LLM API key")
+    parser.add_argument("--llm-model", type=str, default=None, help="LLM model name")
+    parser.add_argument("--llm-api-url", type=str, default=None, help="LLM chat completions API URL or base URL")
+    parser.add_argument("--run-id", type=str, default=None, help="Optional pipeline run id")
+    parser.add_argument("--date", type=str, default=None, help="Optional delivery date in YYYY-MM-DD format")
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=None,
+        help="Run output directory. Defaults to outputs/freshrss/rerun/<timestamp>",
+    )
+    args = parser.parse_args()
+
+    result = run_freshrss_pipeline(
+        api_base_url=args.api_base_url,
+        username=args.username,
+        api_password=args.api_password,
+        stream_id=args.stream_id,
+        limit=args.limit,
+        continuation=args.continuation,
+        include_read=args.include_read,
+        mark_read=args.mark_read,
+        debug_artifacts=args.debug_artifacts,
+        prompt=args.prompt,
+        rules=args.rules,
+        context_path=args.context,
+        max_retries=args.max_retries,
+        timeout_seconds=args.timeout,
+        llm_api_key=args.llm_api_key,
+        llm_model=args.llm_model,
+        llm_api_url=args.llm_api_url,
+        run_id=args.run_id,
+        delivery_date=date.fromisoformat(args.date) if args.date else None,
+        output_dir=args.output_dir,
+    )
+
+    print(f"Saved FreshRSS pipeline run to {result['output_dir']}")
+    print(f"Pulled {result['pulled_count']} items, delivered {result['delivered_count']} candidates")
+    print(
+        "Saved keyword index to "
+        f"{result['keyword_index']['daily_output']} and {result['keyword_index']['stats_output']}"
+    )
+    if args.mark_read:
+        print(f"Marked {result['marked_read_count']} FreshRSS entries as read")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_markdown_sink.py

@@ -0,0 +1,61 @@
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = REPO_ROOT / "src"
+
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from summary_mcp.models.document import ExtractedArticle
+from summary_mcp.models.filtering import FilterDecisionResult
+from summary_mcp.models.item import Item
+from summary_mcp.models.llm_result import LlmSummaryResult
+from summary_mcp.models.sink import SinkInput, SinkMetadata
+from summary_mcp.sinks.markdown import MarkdownSink
+
+
+def _load_json(path: Path) -> dict[str, Any]:
+    return json.loads(path.read_text(encoding="utf-8-sig"))
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Write a filtered result into the Markdown sink.")
+    parser.add_argument("--summary", type=Path, required=True, help="Structured LLM summary JSON file")
+    parser.add_argument("--extracted", type=Path, required=True, help="Extracted article JSON file")
+    parser.add_argument("--filter", type=Path, required=True, help="Filter decision JSON file")
+    parser.add_argument("--item", type=Path, default=None, help="Optional normalized item JSON file")
+    parser.add_argument(
+        "--base-dir",
+        type=Path,
+        default=REPO_ROOT / "knowledge-base",
+        help="Markdown sink base directory",
+    )
+    args = parser.parse_args()
+
+    summary = LlmSummaryResult.model_validate(_load_json(args.summary))
+    extracted_payload = _load_json(args.extracted)
+    article = ExtractedArticle.model_validate(extracted_payload.get("article", extracted_payload))
+    decision = FilterDecisionResult.model_validate(_load_json(args.filter))
+    item = Item.model_validate(_load_json(args.item)) if args.item else None
+
+    sink_input = SinkInput(
+        item=item,
+        article=article,
+        summary=summary,
+        filter_decision=decision,
+        metadata=SinkMetadata(generated_at=datetime.now(tz=UTC), source="run_markdown_sink.py"),
+    )
+    result = MarkdownSink(args.base_dir).write(sink_input)
+    print(json.dumps(result.model_dump(mode="json"), ensure_ascii=False, indent=2))
+
+
+if __name__ == "__main__":
+    main()

scripts/run_summary_loop.py

@@ -1,14 +1,8 @@
 from __future__ import annotations
 
 import argparse
-import json
-import os
-import re
 import sys
 from pathlib import Path
-from typing import Any
-
-import httpx
 
 
 REPO_ROOT = Path(__file__).resolve().parents[1]
@@ -17,192 +11,7 @@ SRC_ROOT = REPO_ROOT / "src"
 if str(SRC_ROOT) not in sys.path:
     sys.path.insert(0, str(SRC_ROOT))
 
-from summary_mcp.validators.llm_result import validate_llm_result
-
-
-JSON_BLOCK_RE = re.compile(r"```(?:json)?\s*(\{.*\})\s*```", re.DOTALL)
-
-
-def load_text(path: Path) -> str:
-    return path.read_text(encoding="utf-8")
-
-
-def load_json(path: Path) -> dict[str, Any]:
-    return json.loads(load_text(path))
-
-
-def save_json(path: Path, payload: dict[str, Any]) -> None:
-    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
-
-
-def build_summary_input(extracted: dict[str, Any]) -> dict[str, Any]:
-    article = extracted.get('article') or {}
-    return {
-        'article': {
-            'title': article.get('title'),
-            'url': article.get('url'),
-            'plain_text': article.get('plain_text'),
-            'quality_flags': article.get('quality_flags'),
-        },
-        'warnings': extracted.get('warnings', []),
-    }
-
-
-def build_initial_prompt(prompt_template: str, extracted: dict[str, Any]) -> str:
-    summary_input = build_summary_input(extracted)
-    return (
-        f"{prompt_template}\n\n"
-        "Below is the structured extracted article input. Generate the final summary JSON from it.\n\n"
-        f"{json.dumps(summary_input, ensure_ascii=False, indent=2)}"
-    )
-
-
-def build_repair_prompt(
-    errors: list[str],
-    extracted: dict[str, Any],
-    result_json: dict[str, Any],
-) -> str:
-    summary_input = build_summary_input(extracted)
-    return (
-        "Please repair the following invalid summary JSON.\n\n"
-        "Requirements:\n"
-        "- Output valid JSON only\n"
-        "- Keep fields that are already correct\n"
-        "- Fix only the validator-reported errors\n"
-        "- Do not add explanations\n\n"
-        f"validator errors:\n{json.dumps(errors, ensure_ascii=False, indent=2)}\n\n"
-        f"Extracted article input:\n{json.dumps(summary_input, ensure_ascii=False, indent=2)}\n\n"
-        f"Current summary JSON:\n{json.dumps(result_json, ensure_ascii=False, indent=2)}\n"
-    )
-
-
-def extract_json_text(raw_text: str) -> str:
-    fenced = JSON_BLOCK_RE.search(raw_text)
-    if fenced:
-        return fenced.group(1)
-
-    stripped = raw_text.strip()
-    start = stripped.find("{")
-    end = stripped.rfind("}")
-    if start == -1 or end == -1 or end <= start:
-        raise ValueError("Model output does not contain a JSON object.")
-    return stripped[start : end + 1]
-
-
-def call_llm(
-    prompt: str,
-    timeout_seconds: float,
-    api_key: str | None,
-    model: str | None,
-    api_url: str | None,
-) -> str:
-    api_key = api_key or os.environ.get("LLM_API_KEY") or os.environ.get("OPENAI_API_KEY")
-    model = model or os.environ.get("LLM_MODEL") or os.environ.get("OPENAI_MODEL")
-    api_url = api_url or os.environ.get("LLM_API_URL", "https://api.openai.com/v1/chat/completions")
-
-    if not api_key:
-        raise RuntimeError("Missing LLM_API_KEY or OPENAI_API_KEY, or pass --api-key.")
-    if not model:
-        raise RuntimeError("Missing LLM_MODEL or OPENAI_MODEL, or pass --model.")
-
-    headers = {
-        "Authorization": f"Bearer {api_key}",
-        "Content-Type": "application/json",
-    }
-    payload = {
-        "model": model,
-        "messages": [
-            {
-                "role": "system",
-                "content": "You are a precise JSON generator. Always output a single valid JSON object.",
-            },
-            {"role": "user", "content": prompt},
-        ],
-        "temperature": 0.2,
-    }
-
-    with httpx.Client(timeout=timeout_seconds) as client:
-        response = client.post(api_url, headers=headers, json=payload)
-        response.raise_for_status()
-        data = response.json()
-
-    try:
-        return data["choices"][0]["message"]["content"]
-    except (KeyError, IndexError, TypeError) as exc:
-        raise RuntimeError(f"Unexpected LLM response shape: {json.dumps(data, ensure_ascii=False)[:1000]}") from exc
-
-
-def run_loop(
-    extracted_path: Path,
-    prompt_path: Path,
-    output_path: Path,
-    max_retries: int,
-    timeout_seconds: float,
-    api_key: str | None,
-    model: str | None,
-    api_url: str | None,
-) -> int:
-    extracted = load_json(extracted_path)
-    prompt_template = load_text(prompt_path)
-    output_path.parent.mkdir(parents=True, exist_ok=True)
-
-    last_errors: list[str] = []
-    last_result: dict[str, Any] | None = None
-
-    for attempt in range(1, max_retries + 2):
-        if attempt == 1:
-            prompt = build_initial_prompt(prompt_template, extracted)
-        else:
-            assert last_result is not None
-            prompt = build_repair_prompt(last_errors, extracted, last_result)
-
-        raw_output = call_llm(prompt, timeout_seconds, api_key, model, api_url)
-        raw_path = output_path.with_name(f"{output_path.stem}.attempt-{attempt}.raw.txt")
-        raw_path.write_text(raw_output, encoding="utf-8")
-
-        try:
-            result_payload = json.loads(extract_json_text(raw_output))
-        except (json.JSONDecodeError, ValueError) as exc:
-            last_errors = [f"Model output is not valid JSON: {exc}"]
-            last_result = {"raw_output": raw_output}
-            validation_path = output_path.with_name(f"{output_path.stem}.attempt-{attempt}.validation.json")
-            save_json(
-                validation_path,
-                {
-                    "valid": False,
-                    "errors": last_errors,
-                    "warnings": [],
-                    "normalized_result": None,
-                },
-            )
-            if attempt > max_retries:
-                output_path.write_text(raw_output, encoding="utf-8")
-                return 1
-            continue
-
-        attempt_path = output_path.with_name(f"{output_path.stem}.attempt-{attempt}.json")
-        save_json(attempt_path, result_payload)
-        save_json(output_path, result_payload)
-
-        report = validate_llm_result(output_path, extracted_path)
-        validation_path = output_path.with_name(f"{output_path.stem}.attempt-{attempt}.validation.json")
-        save_json(
-            validation_path,
-            {
-                "valid": report.valid,
-                "errors": report.errors,
-                "warnings": report.warnings,
-                "normalized_result": report.normalized_result,
-            },
-        )
-
-        if report.valid:
-            return 0
-
-        last_errors = report.errors
-        last_result = result_payload
-
-    return 1
+from summary_mcp.core.summary_loop import run_loop
 
 
 def main() -> None:
@@ -214,7 +23,7 @@ def main() -> None:
     parser.add_argument("--timeout", type=float, default=60.0, help="LLM request timeout in seconds")
     parser.add_argument("--api-key", type=str, default=None, help="LLM API key")
     parser.add_argument("--model", type=str, default=None, help="LLM model name")
-    parser.add_argument("--api-url", type=str, default=None, help="LLM chat completions API URL")
+    parser.add_argument("--api-url", type=str, default=None, help="LLM chat completions API URL or base URL")
     args = parser.parse_args()
 
     raise SystemExit(

skills/keyword-cleanup-review/SKILL.md

@@ -0,0 +1,102 @@
+---
+name: keyword-cleanup-review
+description: Review and curate this repository's daily keyword index and frequency stats. Use when the user wants to inspect `data/term_index/term_stats.json`, recent `data/term_index/daily/*.json`, `configs/term_aliases.json`, `configs/term_stopwords.json`, or `configs/filter_context.personal.json` to propose alias merges, stopwords, watch terms, or `interest_keywords` updates without directly modifying configs.
+---
+
+# Keyword Cleanup Review
+
+Use this skill to turn the repository's keyword statistics into reviewable cleanup suggestions.
+
+## Workflow
+
+1. Build a compact review bundle:
+
+```bash
+python skills/keyword-cleanup-review/scripts/build_review_bundle.py
+```
+
+Optional knobs:
+
+- `--days 7`
+- `--top 50`
+- `--output outputs/term_index/review/keyword-cleanup-bundle.json`
+
+2. Read the generated bundle and the suggestion schema:
+
+- `outputs/term_index/review/keyword-cleanup-bundle.json`
+- `skills/keyword-cleanup-review/references/suggestion-schema.md`
+
+3. Produce two outputs:
+
+- A short Markdown review for humans
+- A JSON suggestion file matching the schema
+
+4. Keep the boundary strict:
+
+- Suggest changes to `configs/term_aliases.json`
+- Suggest changes to `configs/term_stopwords.json`
+- Suggest additions to `configs/filter_context.personal.json`
+- Do not directly edit these files unless the user explicitly asks
+- Do not suggest direct edits to `configs/filter_rules.json` unless the user asks for rule logic changes
+
+## Review Heuristics
+
+Prioritize these decisions:
+
+- Alias suggestion
+  - Same concept with different naming, casing, abbreviation, or Chinese/English variants
+- Stopword suggestion
+  - Too generic, too broad, or too noisy to help filtering
+- Interest keyword suggestion
+  - High-frequency and aligned with the user's backend engineering, AI-agent, and frontier-tech focus
+- Watch term
+  - Recent and potentially important, but evidence is still weak
+
+Prefer conservative suggestions. If confidence is low, put the term into `watch_terms`.
+
+## Inputs
+
+Primary inputs:
+
+- `data/term_index/term_stats.json`
+- `data/term_index/daily/*.json`
+- `configs/term_aliases.json`
+- `configs/term_stopwords.json`
+- `configs/filter_context.personal.json`
+- `configs/term_cleanup_policy.json`
+- `configs/term_watchlist.json`
+- `configs/term_change_log.json`
+
+The bundled script already compacts these into a single review bundle.
+
+## Output Expectations
+
+The Markdown output should:
+
+- Summarize the current state briefly
+- List the top terms worth acting on
+- Separate alias, stopword, interest-keyword, and watch-term recommendations
+- Explain reasoning in short, concrete sentences
+
+The JSON output should follow:
+
+- `references/suggestion-schema.md`
+
+## Repository Notes
+
+Current repository behavior:
+
+- Keyword stats are program-maintained, not LLM-maintained
+- Stats are built from `keywords`, not `topics`
+- Stats only include non-`drop` candidates
+- `data/term_index/term_stats.json` is rebuilt from daily files, so reruns overwrite the same day instead of double-counting
+- cleanup policy, watchlist, and change log are repository-managed governance inputs and should be respected during review
+
+Keep suggestions aligned with that design.
+
+## Resources
+
+- Script:
+  - `scripts/build_review_bundle.py`
+- Reference:
+  - `references/suggestion-schema.md`

skills/keyword-cleanup-review/agents/openai.yaml

@@ -0,0 +1,3 @@
+display_name: Keyword Cleanup Review
+short_description: Review keyword stats and suggest cleanup updates.
+default_prompt: Review the repository keyword index, identify duplicate or noisy terms, and produce alias, stopword, watch-term, and interest-keyword suggestions without modifying configs directly.

skills/keyword-cleanup-review/references/suggestion-schema.md

@@ -0,0 +1,47 @@
+# Suggestion Schema
+
+When this skill produces review output, prefer two files:
+
+- Markdown summary for humans
+- JSON suggestions for deterministic follow-up edits
+
+Recommended JSON shape:
+
+```json
+{
+  "date": "2026-03-27",
+  "based_on_days": 7,
+  "alias_suggestions": [
+    {
+      "from": "Agent",
+      "to": "AI Agent",
+      "reason": "High overlap with existing repository terminology."
+    }
+  ],
+  "stopword_suggestions": [
+    {
+      "term": "??",
+      "reason": "Too generic to be useful for filtering."
+    }
+  ],
+  "interest_keyword_suggestions": [
+    {
+      "term": "MCP",
+      "reason": "Repeated high-frequency term aligned with current learning focus."
+    }
+  ],
+  "watch_terms": [
+    {
+      "term": "OpenClaw",
+      "reason": "Trending recently but not enough history yet."
+    }
+  ]
+}
+```
+
+Rules:
+
+- Only output suggestions, never claim they are already applied.
+- Avoid suggesting changes that would directly edit `filter_rules.json`.
+- Prefer small, reviewable batches over large refactors.
+- If evidence is weak, put the term in `watch_terms` instead of alias or stopword suggestions.

skills/keyword-cleanup-review/references/trigger-notes.md

@@ -0,0 +1 @@
+Provide keyword cleanup guidance for this repository's daily keyword index. Use this skill when the user wants to review, merge, clean, or curate `data/term_index/term_stats.json`, recent `data/term_index/daily/*.json`, `configs/term_aliases.json`, `configs/term_stopwords.json`, or `configs/filter_context.personal.json`, especially to propose alias merges, stopwords, or `interest_keywords` updates without directly modifying configs.

skills/keyword-cleanup-review/scripts/build_review_bundle.py

@@ -0,0 +1,340 @@
+from __future__ import annotations
+
+import argparse
+import json
+from collections import defaultdict
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+
+DEFAULT_POLICY: dict[str, Any] = {
+    "schema_version": "v1",
+    "interest_keyword_review": {
+        "min_total_count": 3,
+        "min_days_seen": 2,
+    },
+    "watch_term_review": {
+        "min_total_count": 1,
+        "min_days_seen": 1,
+        "max_total_count": 2,
+        "max_days_seen": 2,
+    },
+    "alias_review": {
+        "min_total_count": 2,
+        "min_days_seen": 2,
+    },
+    "stopword_review": {
+        "max_total_count": 2,
+        "max_days_seen": 2,
+    },
+}
+
+
+def _load_json(path: Path) -> Any:
+    return json.loads(path.read_text(encoding="utf-8-sig"))
+
+
+def _save_json(path: Path, payload: dict[str, Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def _load_mapping(path: Path) -> dict[str, str]:
+    if not path.exists():
+        return {}
+    payload = _load_json(path)
+    return payload if isinstance(payload, dict) else {}
+
+
+def _load_list(path: Path) -> list[str]:
+    if not path.exists():
+        return []
+    payload = _load_json(path)
+    if isinstance(payload, list):
+        return [item for item in payload if isinstance(item, str)]
+    return []
+
+
+def _load_interest_keywords(path: Path) -> list[str]:
+    if not path.exists():
+        return []
+    payload = _load_json(path)
+    if not isinstance(payload, dict):
+        return []
+    values = payload.get("interest_keywords")
+    if not isinstance(values, list):
+        return []
+    return [item for item in values if isinstance(item, str)]
+
+
+def _casefold_set(values: list[str]) -> set[str]:
+    return {value.strip().casefold() for value in values if value.strip()}
+
+
+def _load_policy(path: Path) -> dict[str, Any]:
+    if not path.exists():
+        return DEFAULT_POLICY.copy()
+    payload = _load_json(path)
+    return payload if isinstance(payload, dict) else DEFAULT_POLICY.copy()
+
+
+def _load_watchlist(path: Path) -> list[dict[str, Any]]:
+    if not path.exists():
+        return []
+    payload = _load_json(path)
+    if not isinstance(payload, dict):
+        return []
+    terms = payload.get("terms")
+    if not isinstance(terms, list):
+        return []
+    return [item for item in terms if isinstance(item, dict) and isinstance(item.get("term"), str)]
+
+
+def _load_change_log(path: Path) -> list[dict[str, Any]]:
+    if not path.exists():
+        return []
+    payload = _load_json(path)
+    if not isinstance(payload, dict):
+        return []
+    entries = payload.get("entries")
+    if not isinstance(entries, list):
+        return []
+    return [item for item in entries if isinstance(item, dict)]
+
+
+def _meets_min_thresholds(item: dict[str, Any], thresholds: dict[str, Any]) -> bool:
+    min_total_count = int(thresholds.get("min_total_count", 1))
+    min_days_seen = int(thresholds.get("min_days_seen", 1))
+    total_count = int(item.get("total_count") or 0)
+    days_seen = int(item.get("days_seen") or 0)
+    return total_count >= min_total_count and days_seen >= min_days_seen
+
+
+def _within_watch_thresholds(item: dict[str, Any], thresholds: dict[str, Any]) -> bool:
+    min_total_count = int(thresholds.get("min_total_count", 1))
+    min_days_seen = int(thresholds.get("min_days_seen", 1))
+    max_total_count = int(thresholds.get("max_total_count", 999999))
+    max_days_seen = int(thresholds.get("max_days_seen", 999999))
+    total_count = int(item.get("total_count") or 0)
+    days_seen = int(item.get("days_seen") or 0)
+    return (
+        total_count >= min_total_count
+        and days_seen >= min_days_seen
+        and total_count <= max_total_count
+        and days_seen <= max_days_seen
+    )
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Build a compact review bundle for the keyword-cleanup-review skill."
+    )
+    repo_root = Path(__file__).resolve().parents[3]
+    parser.add_argument(
+        "--stats",
+        type=Path,
+        default=repo_root / "data" / "term_index" / "term_stats.json",
+        help="Global keyword stats file",
+    )
+    parser.add_argument(
+        "--daily-dir",
+        type=Path,
+        default=repo_root / "data" / "term_index" / "daily",
+        help="Directory containing daily keyword index JSON files",
+    )
+    parser.add_argument(
+        "--aliases",
+        type=Path,
+        default=repo_root / "configs" / "term_aliases.json",
+        help="Term aliases JSON file",
+    )
+    parser.add_argument(
+        "--stopwords",
+        type=Path,
+        default=repo_root / "configs" / "term_stopwords.json",
+        help="Term stopwords JSON file",
+    )
+    parser.add_argument(
+        "--context",
+        type=Path,
+        default=repo_root / "configs" / "filter_context.personal.json",
+        help="Personal filter context JSON file",
+    )
+    parser.add_argument(
+        "--policy",
+        type=Path,
+        default=repo_root / "configs" / "term_cleanup_policy.json",
+        help="Keyword cleanup policy JSON file",
+    )
+    parser.add_argument(
+        "--watchlist",
+        type=Path,
+        default=repo_root / "configs" / "term_watchlist.json",
+        help="Tracked watch terms JSON file",
+    )
+    parser.add_argument(
+        "--change-log",
+        type=Path,
+        default=repo_root / "configs" / "term_change_log.json",
+        help="Applied keyword cleanup change log JSON file",
+    )
+    parser.add_argument("--days", type=int, default=7, help="How many recent daily files to include")
+    parser.add_argument("--top", type=int, default=50, help="How many top global terms to include")
+    parser.add_argument(
+        "--output",
+        type=Path,
+        default=repo_root / "outputs" / "term_index" / "review" / "keyword-cleanup-bundle.json",
+        help="Output JSON file",
+    )
+    args = parser.parse_args()
+
+    stats_payload = _load_json(args.stats) if args.stats.exists() else {"terms": []}
+    stats_terms = stats_payload.get("terms") if isinstance(stats_payload, dict) else []
+    if not isinstance(stats_terms, list):
+        stats_terms = []
+
+    aliases = _load_mapping(args.aliases)
+    stopwords = _load_list(args.stopwords)
+    interest_keywords = _load_interest_keywords(args.context)
+    policy = _load_policy(args.policy)
+    watchlist = _load_watchlist(args.watchlist)
+    change_log = _load_change_log(args.change_log)
+
+    recent_daily_paths = sorted(args.daily_dir.glob("*.json"))[-args.days :] if args.daily_dir.exists() else []
+    recent_daily: list[dict[str, Any]] = []
+    recent_counter: defaultdict[str, int] = defaultdict(int)
+    for path in recent_daily_paths:
+        payload = _load_json(path)
+        if not isinstance(payload, dict):
+            continue
+        terms = payload.get("terms")
+        if not isinstance(terms, list):
+            terms = []
+        compact_terms = []
+        for term in terms:
+            if not isinstance(term, dict):
+                continue
+            normalized = term.get("normalized_term")
+            count = term.get("count")
+            if not isinstance(normalized, str) or not isinstance(count, int):
+                continue
+            compact_terms.append({"term": normalized, "count": count})
+            recent_counter[normalized] += count
+        recent_daily.append(
+            {
+                "date": payload.get("date"),
+                "candidate_count": payload.get("candidate_count"),
+                "top_terms": compact_terms[:20],
+            }
+        )
+
+    interest_set = _casefold_set(interest_keywords)
+    stopword_set = _casefold_set(stopwords)
+    alias_keys = _casefold_set(list(aliases.keys()))
+    alias_values = _casefold_set(list(aliases.values()))
+    watch_set = _casefold_set([str(item.get("term", "")) for item in watchlist])
+
+    top_global_terms = []
+    for item in stats_terms[: args.top]:
+        if not isinstance(item, dict):
+            continue
+        term = item.get("term")
+        if not isinstance(term, str):
+            continue
+        folded = term.strip().casefold()
+        top_global_terms.append(
+            {
+                "term": term,
+                "total_count": item.get("total_count"),
+                "days_seen": item.get("days_seen"),
+                "first_seen": item.get("first_seen"),
+                "last_seen": item.get("last_seen"),
+                "in_interest_keywords": folded in interest_set,
+                "is_stopword": folded in stopword_set,
+                "is_alias_source": folded in alias_keys,
+                "is_alias_target": folded in alias_values,
+                "in_watchlist": folded in watch_set,
+                "recent_count": recent_counter.get(term, 0),
+            }
+        )
+
+    uncovered_terms = [
+        item for item in top_global_terms if not item["in_interest_keywords"] and not item["is_stopword"]
+    ][:20]
+    interest_thresholds = policy.get("interest_keyword_review") if isinstance(policy, dict) else {}
+    watch_thresholds = policy.get("watch_term_review") if isinstance(policy, dict) else {}
+    interest_review_candidates = [
+        {
+            "term": item["term"],
+            "total_count": item["total_count"],
+            "days_seen": item["days_seen"],
+            "reason": (
+                "Meets the configured interest-keyword review threshold and is not yet covered "
+                "by interest keywords or stopwords."
+            ),
+        }
+        for item in uncovered_terms
+        if not item["in_watchlist"] and _meets_min_thresholds(item, interest_thresholds)
+    ][:20]
+    watch_review_candidates = [
+        {
+            "term": item["term"],
+            "total_count": item["total_count"],
+            "days_seen": item["days_seen"],
+            "reason": (
+                "Falls into the configured watch-term review range and should be observed "
+                "before promotion into interest keywords."
+            ),
+        }
+        for item in uncovered_terms
+        if not item["in_watchlist"]
+        and not _meets_min_thresholds(item, interest_thresholds)
+        and _within_watch_thresholds(item, watch_thresholds)
+    ][:20]
+    recent_hot_terms = sorted(
+        ({"term": term, "recent_count": count} for term, count in recent_counter.items()),
+        key=lambda item: (-item["recent_count"], item["term"].casefold(), item["term"]),
+    )[:20]
+
+    bundle = {
+        "generated_at": datetime.now(tz=UTC).isoformat(),
+        "days": args.days,
+        "top": args.top,
+        "sources": {
+            "stats": str(args.stats),
+            "daily_dir": str(args.daily_dir),
+            "aliases": str(args.aliases),
+            "stopwords": str(args.stopwords),
+            "context": str(args.context),
+            "policy": str(args.policy),
+            "watchlist": str(args.watchlist),
+            "change_log": str(args.change_log),
+        },
+        "policy": policy,
+        "current_config": {
+            "alias_count": len(aliases),
+            "stopword_count": len(stopwords),
+            "interest_keyword_count": len(interest_keywords),
+            "watch_term_count": len(watchlist),
+            "aliases": aliases,
+            "stopwords": stopwords,
+            "interest_keywords": interest_keywords,
+            "watchlist": watchlist,
+        },
+        "change_log_tail": change_log[-10:],
+        "top_global_terms": top_global_terms,
+        "recent_daily": recent_daily,
+        "recent_hot_terms": recent_hot_terms,
+        "uncovered_terms": uncovered_terms,
+        "governance_hints": {
+            "interest_review_candidates": interest_review_candidates,
+            "watch_review_candidates": watch_review_candidates,
+        },
+    }
+    _save_json(args.output, bundle)
+    print(f"Saved keyword cleanup bundle to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

src/summary_mcp.egg-info/PKG-INFO

@@ -1,27 +0,0 @@
-Metadata-Version: 2.4
-Name: summary-mcp
-Version: 0.1.0
-Summary: MCP service scaffold for article summarization
-Requires-Python: >=3.11
-Description-Content-Type: text/markdown
-Requires-Dist: beautifulsoup4>=4.12.0
-Requires-Dist: httpx>=0.27.0
-Requires-Dist: mcp>=1.17.0
-Requires-Dist: pydantic>=2.9.0
-Requires-Dist: trafilatura>=1.12.0
-
-# Summary MCP
-
-Python MCP scaffold for article summarization.
-
-## Run
-
-```bash
-pip install -e .
-summary-mcp
-```
-
-The server exposes two tools:
-
-- `summarize_url`
-- `summarize_item`

src/summary_mcp.egg-info/SOURCES.txt

@@ -1,23 +0,0 @@
-README.md
-pyproject.toml
-src/summary_mcp/__init__.py
-src/summary_mcp/server.py
-src/summary_mcp.egg-info/PKG-INFO
-src/summary_mcp.egg-info/SOURCES.txt
-src/summary_mcp.egg-info/dependency_links.txt
-src/summary_mcp.egg-info/entry_points.txt
-src/summary_mcp.egg-info/requires.txt
-src/summary_mcp.egg-info/top_level.txt
-src/summary_mcp/core/__init__.py
-src/summary_mcp/core/content_loader.py
-src/summary_mcp/core/errors.py
-src/summary_mcp/core/extractor.py
-src/summary_mcp/core/mapper.py
-src/summary_mcp/core/normalizer.py
-src/summary_mcp/core/pipeline.py
-src/summary_mcp/core/quality_checker.py
-src/summary_mcp/core/summarizer.py
-src/summary_mcp/models/__init__.py
-src/summary_mcp/models/document.py
-src/summary_mcp/models/item.py
-src/summary_mcp/models/summary_io.py

src/summary_mcp.egg-info/dependency_links.txt

@@ -1 +0,0 @@
-

src/summary_mcp.egg-info/entry_points.txt

@@ -1,2 +0,0 @@
-[console_scripts]
-summary-mcp = summary_mcp.server:main

src/summary_mcp.egg-info/requires.txt

@@ -1,5 +0,0 @@
-beautifulsoup4>=4.12.0
-httpx>=0.27.0
-mcp>=1.17.0
-pydantic>=2.9.0
-trafilatura>=1.12.0

src/summary_mcp.egg-info/top_level.txt

@@ -1 +0,0 @@
-summary_mcp

src/summary_mcp/core/content_loader.py

@@ -6,14 +6,24 @@ from summary_mcp.core.errors import SummaryError
 from summary_mcp.models.summary_io import ExtractionInput
 
 
+MIN_INLINE_CONTENT_LENGTH = 500
+
+
+def _has_usable_content(value: str | None) -> bool:
+    return bool(value and len(value.strip()) >= MIN_INLINE_CONTENT_LENGTH)
+
+
 def choose_inline_content(extraction_input: ExtractionInput) -> tuple[str | None, str]:
     if extraction_input.raw_html:
         return extraction_input.raw_html, "raw_html"
 
-    if extraction_input.item and extraction_input.item.raw_content and len(extraction_input.item.raw_content.strip()) >= 500:
+    if extraction_input.item and _has_usable_content(extraction_input.item.raw_content):
         return extraction_input.item.raw_content, "item.raw_content"
 
-    if extraction_input.rss_content and len(extraction_input.rss_content.strip()) >= 500:
+    if extraction_input.item and _has_usable_content(extraction_input.item.raw_summary):
+        return extraction_input.item.raw_summary, "item.raw_summary"
+
+    if _has_usable_content(extraction_input.rss_content):
         return extraction_input.rss_content, "rss_content"
 
     return None, "none"

src/summary_mcp/core/keyword_index.py

@@ -0,0 +1,228 @@
+from __future__ import annotations
+
+import json
+from collections import Counter
+from datetime import UTC, date, datetime
+from pathlib import Path
+from typing import Iterable
+
+from summary_mcp.models.article_candidate import OpenClawCandidateInput
+from summary_mcp.models.keyword_index import (
+    DailyKeywordIndex,
+    DailyKeywordTerm,
+    KeywordStat,
+    KeywordStatsIndex,
+)
+
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+DATA_ROOT = REPO_ROOT / "data" / "term_index"
+DEFAULT_DAILY_DIR = DATA_ROOT / "daily"
+DEFAULT_STATS_PATH = DATA_ROOT / "term_stats.json"
+DEFAULT_ALIASES_PATH = REPO_ROOT / "configs" / "term_aliases.json"
+DEFAULT_STOPWORDS_PATH = REPO_ROOT / "configs" / "term_stopwords.json"
+DEFAULT_INCLUDE_DECISIONS = {"keep", "review"}
+
+
+def _load_json(path: Path) -> object:
+    return json.loads(path.read_text(encoding="utf-8-sig"))
+
+
+def _save_json(path: Path, payload: dict | list) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def _term_key(value: str) -> str:
+    return value.strip().casefold()
+
+
+def load_term_aliases(path: Path | None = None) -> dict[str, str]:
+    aliases_path = path or DEFAULT_ALIASES_PATH
+    if not aliases_path.exists():
+        return {}
+
+    payload = _load_json(aliases_path)
+    if not isinstance(payload, dict):
+        raise RuntimeError("Term aliases file must contain a JSON object.")
+
+    aliases: dict[str, str] = {}
+    for raw_key, raw_value in payload.items():
+        if not isinstance(raw_key, str) or not isinstance(raw_value, str):
+            continue
+        normalized_key = _term_key(raw_key)
+        normalized_value = raw_value.strip()
+        if not normalized_key or not normalized_value:
+            continue
+        aliases[normalized_key] = normalized_value
+    return aliases
+
+
+def load_term_stopwords(path: Path | None = None) -> set[str]:
+    stopwords_path = path or DEFAULT_STOPWORDS_PATH
+    if not stopwords_path.exists():
+        return set()
+
+    payload = _load_json(stopwords_path)
+    if not isinstance(payload, list):
+        raise RuntimeError("Term stopwords file must contain a JSON array.")
+
+    values: set[str] = set()
+    for item in payload:
+        if not isinstance(item, str):
+            continue
+        normalized = _term_key(item)
+        if normalized:
+            values.add(normalized)
+    return values
+
+
+def normalize_keyword(
+    keyword: str,
+    *,
+    aliases: dict[str, str],
+    stopwords: set[str],
+) -> str | None:
+    raw_value = keyword.strip()
+    if not raw_value:
+        return None
+
+    aliased_value = aliases.get(_term_key(raw_value), raw_value).strip()
+    if not aliased_value:
+        return None
+    if _term_key(aliased_value) in stopwords:
+        return None
+    return aliased_value
+
+
+def build_daily_keyword_index(
+    candidates: Iterable[OpenClawCandidateInput],
+    *,
+    for_date: date,
+    digest_id: str,
+    source: str = "openclaw_delivery_payload",
+    include_decisions: set[str] | None = None,
+    aliases: dict[str, str] | None = None,
+    stopwords: set[str] | None = None,
+) -> DailyKeywordIndex:
+    allowed_decisions = include_decisions or DEFAULT_INCLUDE_DECISIONS
+    resolved_aliases = aliases or {}
+    resolved_stopwords = stopwords or set()
+
+    term_counter: Counter[str] = Counter()
+    candidate_count = 0
+
+    for candidate in candidates:
+        if candidate.selection_decision not in allowed_decisions:
+            continue
+
+        candidate_count += 1
+        seen_for_candidate: set[str] = set()
+        for keyword in candidate.keywords:
+            normalized = normalize_keyword(
+                keyword,
+                aliases=resolved_aliases,
+                stopwords=resolved_stopwords,
+            )
+            if normalized is None or normalized in seen_for_candidate:
+                continue
+            seen_for_candidate.add(normalized)
+            term_counter[normalized] += 1
+
+    terms = [
+        DailyKeywordTerm(term=term, normalized_term=term, count=count)
+        for term, count in sorted(term_counter.items(), key=lambda item: (-item[1], item[0].casefold(), item[0]))
+    ]
+
+    return DailyKeywordIndex(
+        schema_version="v1",
+        date=for_date,
+        source=source,
+        digest_id=digest_id,
+        generated_at=datetime.now(tz=UTC),
+        candidate_count=candidate_count,
+        terms=terms,
+    )
+
+
+def read_daily_keyword_index(path: Path) -> DailyKeywordIndex:
+    return DailyKeywordIndex.model_validate(_load_json(path))
+
+
+def rebuild_keyword_stats(*, daily_dir: Path = DEFAULT_DAILY_DIR) -> KeywordStatsIndex:
+    term_totals: dict[str, int] = {}
+    first_seen: dict[str, date] = {}
+    last_seen: dict[str, date] = {}
+    days_seen: dict[str, int] = {}
+
+    if daily_dir.exists():
+        for path in sorted(daily_dir.glob("*.json")):
+            daily_index = read_daily_keyword_index(path)
+            seen_today: set[str] = set()
+            for term in daily_index.terms:
+                normalized = term.normalized_term
+                term_totals[normalized] = term_totals.get(normalized, 0) + term.count
+                if normalized not in first_seen or daily_index.date < first_seen[normalized]:
+                    first_seen[normalized] = daily_index.date
+                if normalized not in last_seen or daily_index.date > last_seen[normalized]:
+                    last_seen[normalized] = daily_index.date
+                if normalized not in seen_today:
+                    days_seen[normalized] = days_seen.get(normalized, 0) + 1
+                    seen_today.add(normalized)
+
+    terms = [
+        KeywordStat(
+            term=term,
+            total_count=term_totals[term],
+            days_seen=days_seen[term],
+            first_seen=first_seen[term],
+            last_seen=last_seen[term],
+        )
+        for term in sorted(term_totals, key=lambda item: (-term_totals[item], item.casefold(), item))
+    ]
+
+    return KeywordStatsIndex(
+        schema_version="v1",
+        generated_at=datetime.now(tz=UTC),
+        terms=terms,
+    )
+
+
+def persist_keyword_indexes(
+    candidates: Iterable[OpenClawCandidateInput],
+    *,
+    for_date: date,
+    digest_id: str,
+    source: str = "openclaw_delivery_payload",
+    daily_dir: Path = DEFAULT_DAILY_DIR,
+    stats_path: Path = DEFAULT_STATS_PATH,
+    aliases_path: Path | None = None,
+    stopwords_path: Path | None = None,
+    include_decisions: set[str] | None = None,
+) -> dict[str, object]:
+    aliases = load_term_aliases(aliases_path)
+    stopwords = load_term_stopwords(stopwords_path)
+    daily_index = build_daily_keyword_index(
+        candidates,
+        for_date=for_date,
+        digest_id=digest_id,
+        source=source,
+        include_decisions=include_decisions,
+        aliases=aliases,
+        stopwords=stopwords,
+    )
+
+    daily_path = daily_dir / f"{for_date.isoformat()}.json"
+    _save_json(daily_path, daily_index.model_dump(mode="json"))
+
+    stats_index = rebuild_keyword_stats(daily_dir=daily_dir)
+    _save_json(stats_path, stats_index.model_dump(mode="json"))
+
+    return {
+        "daily_output": str(daily_path),
+        "stats_output": str(stats_path),
+        "source": source,
+        "candidate_count": daily_index.candidate_count,
+        "term_count": len(daily_index.terms),
+        "top_terms": [term.model_dump(mode="json") for term in daily_index.terms[:10]],
+    }

src/summary_mcp/core/pipeline.py

@@ -9,6 +9,19 @@ from summary_mcp.core.quality_checker import assess_quality
 from summary_mcp.models.summary_io import DebugInfo, ExtractionInput, ExtractionOutput
 
 
+RSS_ONLY_UPSTREAMS = {"freshrss"}
+
+
+def _should_skip_fetch(extraction_input: ExtractionInput) -> bool:
+    item = extraction_input.item
+    if item is None:
+        return False
+
+    metadata = item.metadata if isinstance(item.metadata, dict) else {}
+    upstream = metadata.get("upstream")
+    return isinstance(upstream, str) and upstream in RSS_ONLY_UPSTREAMS
+
+
 def extract_content(extraction_input: ExtractionInput) -> ExtractionOutput:
     try:
         normalized = normalize_input(extraction_input)
@@ -16,11 +29,20 @@ def extract_content(extraction_input: ExtractionInput) -> ExtractionOutput:
 
         inline_content, content_source = choose_inline_content(normalized)
         if inline_content is None:
+            if _should_skip_fetch(normalized):
+                raise SummaryError(
+                    code="RSS_CONTENT_MISSING",
+                    message="Skipping item because RSS content is unavailable.",
+                    retryable=False,
+                    stage="extract",
+                    details={"url": str(normalized.item.url), "upstream": normalized.item.metadata.get("upstream")},
+                )
             inline_content = fetch_html(str(normalized.item.url))
             content_source = "fetched_html"
 
         if normalized.item.title is None and (
-            content_source in {"raw_html", "fetched_html"} or inline_content.lstrip().startswith("<")
+            content_source in {"raw_html", "fetched_html", "item.raw_content", "item.raw_summary", "rss_content"}
+            or inline_content.lstrip().startswith("<")
         ):
             normalized.item.title = extract_title(inline_content)
 

src/summary_mcp/core/summary_loop.py

@@ -0,0 +1,273 @@
+from __future__ import annotations
+
+import json
+import os
+import re
+from pathlib import Path
+from typing import Any
+
+import httpx
+
+from summary_mcp.validators.llm_result import ValidationReport
+from summary_mcp.validators.llm_result import validate_llm_result as validate_llm_result_from_path
+from summary_mcp.validators.llm_result import validate_llm_result_payload
+
+
+JSON_BLOCK_RE = re.compile(r"```(?:json)?\s*(\{.*\})\s*```", re.DOTALL)
+DEFAULT_CHAT_COMPLETIONS_URL = "https://api.openai.com/v1/chat/completions"
+
+
+def load_text(path: Path) -> str:
+    return path.read_text(encoding="utf-8")
+
+
+def load_json(path: Path) -> dict[str, Any]:
+    return json.loads(load_text(path))
+
+
+def save_json(path: Path, payload: dict[str, Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def build_summary_input(extracted: dict[str, Any]) -> dict[str, Any]:
+    article = extracted.get("article") or {}
+    return {
+        "article": {
+            "title": article.get("title"),
+            "url": article.get("url"),
+            "plain_text": article.get("plain_text"),
+            "quality_flags": article.get("quality_flags"),
+        },
+        "warnings": extracted.get("warnings", []),
+    }
+
+
+def build_initial_prompt(prompt_template: str, extracted: dict[str, Any]) -> str:
+    summary_input = build_summary_input(extracted)
+    return (
+        f"{prompt_template}\n\n"
+        "Below is the structured extracted article input. Generate the final summary JSON from it.\n\n"
+        f"{json.dumps(summary_input, ensure_ascii=False, indent=2)}"
+    )
+
+
+def build_repair_prompt(
+    errors: list[str],
+    extracted: dict[str, Any],
+    result_json: dict[str, Any],
+) -> str:
+    summary_input = build_summary_input(extracted)
+    return (
+        "Please repair the following invalid summary JSON.\n\n"
+        "Requirements:\n"
+        "- Output valid JSON only\n"
+        "- Keep fields that are already correct\n"
+        "- Fix only the validator-reported errors\n"
+        "- Do not add explanations\n\n"
+        f"validator errors:\n{json.dumps(errors, ensure_ascii=False, indent=2)}\n\n"
+        f"Extracted article input:\n{json.dumps(summary_input, ensure_ascii=False, indent=2)}\n\n"
+        f"Current summary JSON:\n{json.dumps(result_json, ensure_ascii=False, indent=2)}\n"
+    )
+
+
+def extract_json_text(raw_text: str) -> str:
+    fenced = JSON_BLOCK_RE.search(raw_text)
+    if fenced:
+        return fenced.group(1)
+
+    stripped = raw_text.strip()
+    start = stripped.find("{")
+    end = stripped.rfind("}")
+    if start == -1 or end == -1 or end <= start:
+        raise ValueError("Model output does not contain a JSON object.")
+    return stripped[start : end + 1]
+
+
+def normalize_chat_completions_url(api_url: str | None) -> str | None:
+    if api_url is None:
+        return None
+
+    normalized = api_url.strip().rstrip("/")
+    if not normalized:
+        return None
+    if normalized.endswith("/chat/completions"):
+        return normalized
+    return f"{normalized}/chat/completions"
+
+
+def resolve_llm_settings(
+    *,
+    api_key: str | None = None,
+    model: str | None = None,
+    api_url: str | None = None,
+) -> tuple[str, str, str]:
+    resolved_api_key = api_key or os.environ.get("LLM_API_KEY") or os.environ.get("OPENAI_API_KEY")
+    if not resolved_api_key:
+        raise RuntimeError("Missing LLM_API_KEY or OPENAI_API_KEY, or pass an API key.")
+
+    resolved_model = model or os.environ.get("LLM_MODEL") or os.environ.get("OPENAI_MODEL")
+    if not resolved_model:
+        raise RuntimeError("Missing LLM_MODEL or OPENAI_MODEL, or pass a model.")
+
+    resolved_api_url = normalize_chat_completions_url(
+        api_url or os.environ.get("LLM_API_URL") or os.environ.get("OPENAI_API_URL") or DEFAULT_CHAT_COMPLETIONS_URL
+    )
+    if not resolved_api_url:
+        raise RuntimeError("Missing LLM_API_URL, OPENAI_API_URL, or pass an API URL.")
+
+    return resolved_api_key, resolved_model, resolved_api_url
+
+
+def call_llm(
+    prompt: str,
+    timeout_seconds: float,
+    api_key: str | None,
+    model: str | None,
+    api_url: str | None,
+) -> str:
+    resolved_api_key, resolved_model, resolved_api_url = resolve_llm_settings(
+        api_key=api_key,
+        model=model,
+        api_url=api_url,
+    )
+
+    headers = {
+        "Authorization": f"Bearer {resolved_api_key}",
+        "Content-Type": "application/json",
+    }
+    payload = {
+        "model": resolved_model,
+        "messages": [
+            {
+                "role": "system",
+                "content": "You are a precise JSON generator. Always output a single valid JSON object.",
+            },
+            {"role": "user", "content": prompt},
+        ],
+        "temperature": 0.2,
+    }
+
+    with httpx.Client(timeout=timeout_seconds) as client:
+        response = client.post(resolved_api_url, headers=headers, json=payload)
+        response.raise_for_status()
+        data = response.json()
+
+    try:
+        return data["choices"][0]["message"]["content"]
+    except (KeyError, IndexError, TypeError) as exc:
+        raise RuntimeError(f"Unexpected LLM response shape: {json.dumps(data, ensure_ascii=False)[:1000]}") from exc
+
+
+def _save_attempt_artifact(base_output_path: Path | None, suffix: str, payload: str | dict[str, Any]) -> None:
+    if base_output_path is None:
+        return
+
+    path = base_output_path.with_name(f"{base_output_path.stem}.{suffix}")
+    if isinstance(payload, str):
+        path.write_text(payload, encoding="utf-8")
+    else:
+        save_json(path, payload)
+
+
+def run_loop_payload(
+    *,
+    extracted_payload: dict[str, Any],
+    prompt_path: Path,
+    max_retries: int,
+    timeout_seconds: float,
+    api_key: str | None,
+    model: str | None,
+    api_url: str | None,
+    output_path: Path | None = None,
+) -> tuple[int, dict[str, Any] | None, ValidationReport | None]:
+    prompt_template = load_text(prompt_path)
+    if output_path is not None:
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    last_errors: list[str] = []
+    last_result: dict[str, Any] | None = None
+
+    for attempt in range(1, max_retries + 2):
+        if attempt == 1:
+            prompt = build_initial_prompt(prompt_template, extracted_payload)
+        else:
+            assert last_result is not None
+            prompt = build_repair_prompt(last_errors, extracted_payload, last_result)
+
+        raw_output = call_llm(prompt, timeout_seconds, api_key, model, api_url)
+        _save_attempt_artifact(output_path, f"attempt-{attempt}.raw.txt", raw_output)
+
+        try:
+            result_payload = json.loads(extract_json_text(raw_output))
+        except (json.JSONDecodeError, ValueError) as exc:
+            last_errors = [f"Model output is not valid JSON: {exc}"]
+            last_result = {"raw_output": raw_output}
+            report_payload = {
+                "valid": False,
+                "errors": last_errors,
+                "warnings": [],
+                "normalized_result": None,
+            }
+            _save_attempt_artifact(output_path, f"attempt-{attempt}.validation.json", report_payload)
+            if attempt > max_retries:
+                if output_path is not None:
+                    output_path.write_text(raw_output, encoding="utf-8")
+                return 1, None, ValidationReport(valid=False, errors=last_errors)
+            continue
+
+        _save_attempt_artifact(output_path, f"attempt-{attempt}.json", result_payload)
+        if output_path is not None:
+            save_json(output_path, result_payload)
+
+        report = validate_llm_result_payload(result_payload, extracted_payload)
+        _save_attempt_artifact(
+            output_path,
+            f"attempt-{attempt}.validation.json",
+            {
+                "valid": report.valid,
+                "errors": report.errors,
+                "warnings": report.warnings,
+                "normalized_result": report.normalized_result,
+            },
+        )
+
+        if report.valid:
+            return 0, result_payload, report
+
+        last_errors = report.errors
+        last_result = result_payload
+
+    return 1, None, ValidationReport(valid=False, errors=last_errors)
+
+
+def run_loop(
+    extracted_path: Path,
+    prompt_path: Path,
+    output_path: Path,
+    max_retries: int,
+    timeout_seconds: float,
+    api_key: str | None,
+    model: str | None,
+    api_url: str | None,
+) -> int:
+    extracted = load_json(extracted_path)
+    exit_code, _, report = run_loop_payload(
+        extracted_payload=extracted,
+        prompt_path=prompt_path,
+        output_path=output_path,
+        max_retries=max_retries,
+        timeout_seconds=timeout_seconds,
+        api_key=api_key,
+        model=model,
+        api_url=api_url,
+    )
+
+    if exit_code == 0:
+        return 0
+
+    if report is not None and output_path.exists() and extracted_path.exists():
+        fallback_report = validate_llm_result_from_path(output_path, extracted_path)
+        if fallback_report.valid:
+            return 0
+    return exit_code

src/summary_mcp/filters/__init__.py

@@ -0,0 +1,3 @@
+from .engine import DEFAULT_RULES_PATH, evaluate_filter_rules, load_filter_rules
+
+__all__ = ["DEFAULT_RULES_PATH", "evaluate_filter_rules", "load_filter_rules"]

src/summary_mcp/filters/engine.py

@@ -0,0 +1,136 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+from summary_mcp.models.filtering import (
+    FieldCondition,
+    FilterDecisionResult,
+    FilterInput,
+    FilterRule,
+    MatchedRule,
+)
+
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+DEFAULT_RULES_PATH = REPO_ROOT / "configs" / "filter_rules.json"
+
+
+def load_filter_rules(path: Path | None = None) -> list[FilterRule]:
+    rules_path = path or DEFAULT_RULES_PATH
+    payload = json.loads(rules_path.read_text(encoding="utf-8-sig"))
+    if not isinstance(payload, list):
+        raise RuntimeError("Filter rules file must contain a JSON array.")
+    return [FilterRule.model_validate(item) for item in payload]
+
+
+def _normalize_value(value: Any) -> Any:
+    if hasattr(value, "model_dump"):
+        return value.model_dump(mode="json")
+    return value
+
+
+def _resolve_field(filter_input: FilterInput, field_path: str) -> Any:
+    current: Any = filter_input
+    for part in field_path.split("."):
+        current = _normalize_value(current)
+        if isinstance(current, dict):
+            if part not in current:
+                return None
+            current = current[part]
+            continue
+        return None
+    return _normalize_value(current)
+
+
+def _expected_value(filter_input: FilterInput, value: Any) -> Any:
+    if isinstance(value, dict) and "from_field" in value:
+        field_name = value.get("from_field")
+        if isinstance(field_name, str):
+            return _resolve_field(filter_input, field_name)
+    return value
+
+
+def _match_condition(filter_input: FilterInput, condition: FieldCondition) -> bool:
+    current = _resolve_field(filter_input, condition.field)
+    expected = _expected_value(filter_input, condition.value)
+
+    if condition.op == "exists":
+        return (current is not None) if expected is not False else (current is None)
+    if condition.op == "eq":
+        return current == expected
+    if condition.op == "ne":
+        return current != expected
+    if condition.op == "in":
+        return current in expected if isinstance(expected, list) else False
+    if condition.op == "not_in":
+        return current not in expected if isinstance(expected, list) else False
+    if condition.op == "contains":
+        if isinstance(current, list):
+            return expected in current
+        if isinstance(current, str) and isinstance(expected, str):
+            return expected in current
+        return False
+    if condition.op == "overlap":
+        if isinstance(current, list) and isinstance(expected, list):
+            return bool(set(current) & set(expected))
+        return False
+    if condition.op == "gte":
+        return current is not None and expected is not None and current >= expected
+    if condition.op == "lte":
+        return current is not None and expected is not None and current <= expected
+    return False
+
+
+def _rule_matches(filter_input: FilterInput, rule: FilterRule) -> bool:
+    if not rule.enabled:
+        return False
+    if rule.conditions_all and not all(_match_condition(filter_input, condition) for condition in rule.conditions_all):
+        return False
+    if rule.conditions_any and not any(_match_condition(filter_input, condition) for condition in rule.conditions_any):
+        return False
+    return bool(rule.conditions_all or rule.conditions_any)
+
+
+def evaluate_filter_rules(filter_input: FilterInput, rules: list[FilterRule]) -> FilterDecisionResult:
+    matched: list[MatchedRule] = []
+
+    for rule in sorted(rules, key=lambda item: item.action.priority, reverse=True):
+        if not _rule_matches(filter_input, rule):
+            continue
+
+        matched_rule = MatchedRule(
+            rule_id=rule.rule_id,
+            decision=rule.action.decision,
+            reason=rule.action.reason,
+            labels=rule.action.labels,
+            priority=rule.action.priority,
+        )
+        matched.append(matched_rule)
+        if rule.stop_on_match:
+            break
+
+    if any(rule.decision == "drop" for rule in matched):
+        final_decision = "drop"
+    elif any(rule.decision == "keep" for rule in matched):
+        final_decision = "keep"
+    elif any(rule.decision == "review" for rule in matched):
+        final_decision = "review"
+    else:
+        final_decision = "review"
+
+    labels = sorted({label for rule in matched for label in rule.labels})
+    reasons = [rule.reason for rule in matched]
+    priorities = [rule.priority for rule in matched]
+
+    return FilterDecisionResult(
+        decision=final_decision,
+        matched_rules=[rule.rule_id for rule in matched],
+        reasons=reasons if reasons else ["No rule matched; defaulted to review."],
+        labels=labels,
+        priority=max(priorities, default=0),
+        matches=matched,
+    )
+
+

src/summary_mcp/integrations/__init__.py

@@ -0,0 +1 @@
+"""Integration helpers for upstream content sources."""

src/summary_mcp/integrations/freshrss.py

@@ -0,0 +1,292 @@
+from __future__ import annotations
+
+import hashlib
+from datetime import UTC, datetime
+from typing import Any
+from urllib.parse import urlencode, urlparse
+
+import httpx
+
+from summary_mcp.models.item import Item
+
+
+READ_TAG = "user/-/state/com.google/read"
+KEPT_UNREAD_TAG = "user/-/state/com.google/kept-unread"
+MIN_INLINE_CONTENT_LENGTH = 500
+
+
+def _trim_api_base_url(api_base_url: str) -> str:
+    return api_base_url.rstrip("/")
+
+
+def _build_item_id(source_id: str, external_id: str | None, url: str, published_at: datetime | None) -> str:
+    seed = "|".join(
+        [
+            source_id,
+            external_id or "",
+            url,
+            published_at.isoformat() if published_at else "",
+        ]
+    )
+    return f"sha256:{hashlib.sha256(seed.encode('utf-8')).hexdigest()}"
+
+
+def _build_source_id(entry: dict[str, Any], url: str) -> str:
+    origin = entry.get("origin") or {}
+    stream_id = origin.get("streamId")
+    if isinstance(stream_id, str) and stream_id.strip():
+        digest = hashlib.sha256(stream_id.encode("utf-8")).hexdigest()[:16]
+        return f"freshrss:{digest}"
+
+    host = urlparse(url).netloc or "unknown-source"
+    return f"freshrss:{host}"
+
+
+def _pick_entry_url(entry: dict[str, Any]) -> str:
+    candidates = [
+        entry.get("canonical"),
+        entry.get("alternate"),
+    ]
+
+    for candidate_list in candidates:
+        if not isinstance(candidate_list, list):
+            continue
+        for candidate in candidate_list:
+            href = (candidate or {}).get("href")
+            if isinstance(href, str) and href.strip():
+                return href.strip()
+
+    entry_id = entry.get("id")
+    if isinstance(entry_id, str) and entry_id.startswith("tag:"):
+        return entry_id
+
+    raise ValueError("FreshRSS entry does not contain a usable URL.")
+
+
+def _pick_content_block(entry: dict[str, Any], key: str) -> str | None:
+    block = entry.get(key)
+    if isinstance(block, dict):
+        content = block.get("content")
+        if isinstance(content, str) and content.strip():
+            return content
+    return None
+
+
+def _pick_categories(entry: dict[str, Any]) -> list[str]:
+    categories = entry.get("categories")
+    if not isinstance(categories, list):
+        return []
+
+    values: list[str] = []
+    for category in categories:
+        if not isinstance(category, str):
+            continue
+        if category.startswith("user/-/label/"):
+            values.append(category.removeprefix("user/-/label/"))
+        elif category.startswith("user/-/state/com.google/"):
+            continue
+        else:
+            values.append(category)
+    return values
+
+
+def _parse_datetime(timestamp: Any) -> datetime | None:
+    if timestamp is None:
+        return None
+    if isinstance(timestamp, (int, float)):
+        if timestamp > 10_000_000_000:
+            return datetime.fromtimestamp(timestamp / 1000, tz=UTC)
+        return datetime.fromtimestamp(timestamp, tz=UTC)
+    if isinstance(timestamp, str) and timestamp.isdigit():
+        value = int(timestamp)
+        if value > 10_000_000_000:
+            return datetime.fromtimestamp(value / 1000, tz=UTC)
+        return datetime.fromtimestamp(value, tz=UTC)
+    return None
+
+
+def _dedupe_non_empty(values: list[str] | None) -> list[str]:
+    if not values:
+        return []
+
+    seen: set[str] = set()
+    deduped: list[str] = []
+    for value in values:
+        if not value or value in seen:
+            continue
+        seen.add(value)
+        deduped.append(value)
+    return deduped
+
+
+def _has_usable_inline_content(value: str | None) -> bool:
+    return bool(value and len(value.strip()) >= MIN_INLINE_CONTENT_LENGTH)
+
+
+def map_entry_to_item(entry: dict[str, Any]) -> Item:
+    url = _pick_entry_url(entry)
+    summary_html = _pick_content_block(entry, "summary")
+    content_html = _pick_content_block(entry, "content")
+    published_at = _parse_datetime(entry.get("published"))
+    source_id = _build_source_id(entry, url)
+    external_id = entry.get("id") if isinstance(entry.get("id"), str) else None
+    fetch_state = "fetched" if _has_usable_inline_content(content_html) or _has_usable_inline_content(summary_html) else "pending"
+
+    metadata = {
+        "upstream": "freshrss",
+        "origin": entry.get("origin") or {},
+        "categories": _pick_categories(entry),
+        "crawled_at": _parse_datetime(entry.get("crawlTimeMsec")),
+        "published_epoch": entry.get("published"),
+    }
+
+    return Item(
+        item_id=_build_item_id(source_id, external_id, url, published_at),
+        source_id=source_id,
+        external_id=external_id,
+        title=entry.get("title"),
+        url=url,
+        author=entry.get("author"),
+        published_at=published_at,
+        discovered_at=datetime.now(tz=UTC),
+        content_kind="article",
+        language=None,
+        raw_summary=summary_html,
+        raw_content=content_html,
+        metadata=metadata,
+        fetch_state=fetch_state,
+    )
+
+
+class FreshRSSClient:
+    def __init__(
+        self,
+        api_base_url: str,
+        username: str,
+        api_password: str,
+        timeout_seconds: float = 20.0,
+    ) -> None:
+        self.api_base_url = _trim_api_base_url(api_base_url)
+        self.username = username
+        self.api_password = api_password
+        self.timeout_seconds = timeout_seconds
+
+    def _build_url(self, path: str) -> str:
+        return f"{self.api_base_url}/{path.lstrip('/')}"
+
+    def _build_auth_headers(self, auth_token: str) -> dict[str, str]:
+        return {"Authorization": f"GoogleLogin auth={auth_token}"}
+
+    def client_login(self) -> str:
+        data = {
+            "Email": self.username,
+            "Passwd": self.api_password,
+        }
+        with httpx.Client(timeout=self.timeout_seconds) as client:
+            response = client.post(self._build_url("accounts/ClientLogin"), data=data)
+            response.raise_for_status()
+
+        auth_token: str | None = None
+        for line in response.text.splitlines():
+            if line.startswith("Auth="):
+                auth_token = line.split("=", 1)[1].strip()
+                break
+
+        if not auth_token:
+            raise RuntimeError("FreshRSS ClientLogin succeeded but did not return an Auth token.")
+
+        return auth_token
+
+    def get_edit_token(self, auth_token: str) -> str:
+        with httpx.Client(
+            timeout=self.timeout_seconds,
+            headers=self._build_auth_headers(auth_token),
+        ) as client:
+            response = client.get(self._build_url("reader/api/0/token"))
+            response.raise_for_status()
+
+        edit_token = response.text.strip()
+        if not edit_token:
+            raise RuntimeError("FreshRSS token request succeeded but did not return an edit token.")
+        return edit_token
+
+    def fetch_stream_contents(
+        self,
+        auth_token: str,
+        stream_id: str = "user/-/state/com.google/reading-list",
+        limit: int = 20,
+        continuation: str | None = None,
+        exclude_targets: list[str] | None = None,
+        include_targets: list[str] | None = None,
+    ) -> dict[str, Any]:
+        params: list[tuple[str, str | int]] = [
+            ("output", "json"),
+            ("n", limit),
+        ]
+        if continuation:
+            params.append(("c", continuation))
+        for target in _dedupe_non_empty(exclude_targets):
+            params.append(("xt", target))
+        for target in _dedupe_non_empty(include_targets):
+            params.append(("it", target))
+
+        with httpx.Client(
+            timeout=self.timeout_seconds,
+            headers=self._build_auth_headers(auth_token),
+        ) as client:
+            response = client.get(
+                self._build_url(f"reader/api/0/stream/contents/{stream_id}"),
+                params=params,
+            )
+            response.raise_for_status()
+            return response.json()
+
+    def edit_tag(
+        self,
+        auth_token: str,
+        edit_token: str,
+        item_ids: list[str],
+        add_tags: list[str] | None = None,
+        remove_tags: list[str] | None = None,
+    ) -> str:
+        normalized_item_ids = _dedupe_non_empty(item_ids)
+        if not normalized_item_ids:
+            raise ValueError("FreshRSS edit-tag requires at least one item id.")
+
+        payload: list[tuple[str, str]] = [("ac", "edit"), ("T", edit_token)]
+        for item_id in normalized_item_ids:
+            payload.append(("i", item_id))
+        for tag in _dedupe_non_empty(add_tags):
+            payload.append(("a", tag))
+        for tag in _dedupe_non_empty(remove_tags):
+            payload.append(("r", tag))
+
+        headers = self._build_auth_headers(auth_token)
+        headers["Content-Type"] = "application/x-www-form-urlencoded"
+
+        with httpx.Client(
+            timeout=self.timeout_seconds,
+            headers=headers,
+        ) as client:
+            response = client.post(
+                self._build_url("reader/api/0/edit-tag"),
+                content=urlencode(payload).encode("utf-8"),
+            )
+            response.raise_for_status()
+
+        return response.text.strip()
+
+    def mark_items_as_read(
+        self,
+        auth_token: str,
+        item_ids: list[str],
+        edit_token: str | None = None,
+    ) -> str:
+        resolved_edit_token = edit_token or self.get_edit_token(auth_token)
+        return self.edit_tag(
+            auth_token=auth_token,
+            edit_token=resolved_edit_token,
+            item_ids=item_ids,
+            add_tags=[READ_TAG],
+            remove_tags=[KEPT_UNREAD_TAG],
+        )

src/summary_mcp/models/__init__.py

@@ -1 +1,60 @@
-"""Shared models for the summary MCP service."""
+"""Shared models for the summary MCP service."""
+
+from .article_candidate import (
+    ArticleCandidate,
+    ArticleCandidateRecord,
+    CandidateMetadata,
+    CandidateSourceRefs,
+    DigestSectionHint,
+    OpenClawCandidateInput,
+    ReviewState,
+    build_article_candidate_record,
+    build_openclaw_candidate_input,
+    candidate_id_for,
+    normalize_candidate_url,
+)
+from .candidate_review import CandidateReviewStatus, KnowledgeDecision, ReviewStatus
+from .daily_digest import (
+    DailyDigest,
+    DailyDigestItem,
+    DailyDigestMetadata,
+    DailyDigestSection,
+    DailyDigestSourceRef,
+    DailyDigestStats,
+)
+from .keyword_index import DailyKeywordIndex, DailyKeywordTerm, KeywordStat, KeywordStatsIndex
+from .openclaw_delivery import (
+    OpenClawDeliveryPayload,
+    OpenClawDeliveryStats,
+    build_openclaw_delivery_payload,
+)
+
+__all__ = [
+    "ArticleCandidate",
+    "ArticleCandidateRecord",
+    "CandidateMetadata",
+    "CandidateReviewStatus",
+    "CandidateSourceRefs",
+    "DailyDigest",
+    "DailyDigestItem",
+    "DailyDigestMetadata",
+    "DailyDigestSection",
+    "DailyDigestSourceRef",
+    "DailyDigestStats",
+    "DailyKeywordIndex",
+    "DailyKeywordTerm",
+    "DigestSectionHint",
+    "KeywordStat",
+    "KeywordStatsIndex",
+    "KnowledgeDecision",
+    "OpenClawCandidateInput",
+    "OpenClawDeliveryPayload",
+    "OpenClawDeliveryStats",
+    "ReviewState",
+    "ReviewStatus",
+    "build_article_candidate_record",
+    "build_openclaw_delivery_payload",
+    "build_openclaw_candidate_input",
+    "candidate_id_for",
+    "normalize_candidate_url",
+]

src/summary_mcp/models/article_candidate.py

@@ -0,0 +1,190 @@
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Literal
+from urllib.parse import parse_qsl, urlencode, urlparse, urlunparse
+
+from pydantic import BaseModel, Field, HttpUrl
+
+from .document import ExtractedArticle
+from .filtering import FilterDecision, FilterDecisionResult
+from .item import Item
+from .llm_result import Category, LlmSummaryResult
+
+
+ReviewState = Literal["pending", "accepted", "rejected", "deferred"]
+DigestSectionHint = Literal[
+    "top_news",
+    "tools_and_workflows",
+    "risk_and_security",
+    "open_source",
+    "insights",
+    "deep_dive",
+]
+
+_TRACKING_QUERY_PARAMS = {
+    "fbclid",
+    "gclid",
+    "igshid",
+    "mc_cid",
+    "mc_eid",
+    "mkt_tok",
+    "ref",
+    "ref_src",
+    "si",
+    "spm",
+}
+
+
+class CandidateSourceRefs(BaseModel):
+    item_path: str | None = None
+    extracted_path: str | None = None
+    summary_path: str | None = None
+    filter_path: str | None = None
+
+
+class CandidateMetadata(BaseModel):
+    generated_at: datetime | None = None
+    pipeline_version: str = "v1"
+    producer: str | None = None
+    run_id: str | None = None
+
+
+class ArticleCandidateRecord(BaseModel):
+    candidate_id: str
+    item: Item | None = None
+    article: ExtractedArticle
+    summary: LlmSummaryResult
+    filter_result: FilterDecisionResult
+    digest_section_hint: DigestSectionHint | None = None
+    digest_rank: int = Field(default=50, ge=0, le=100)
+    review_state: ReviewState = "pending"
+    source_refs: CandidateSourceRefs = Field(default_factory=CandidateSourceRefs)
+    rendered_markdown: str | None = None
+    metadata: CandidateMetadata = Field(default_factory=CandidateMetadata)
+
+
+class OpenClawCandidateInput(BaseModel):
+    candidate_id: str
+    title: str
+    url: HttpUrl
+    canonical_url: HttpUrl | None = None
+    published_at: datetime | None = None
+    author: str | None = None
+    source_name: str | None = None
+    language: str | None = None
+    summary: str
+    highlights: list[str] = Field(default_factory=list)
+    keywords: list[str] = Field(default_factory=list)
+    topics: list[str] = Field(default_factory=list)
+    category: Category
+    worth_keeping: bool
+    worth_reason: str
+    selection_decision: FilterDecision
+    selection_reason: str | None = None
+    digest_section_hint: DigestSectionHint | None = None
+    digest_rank: int = Field(default=50, ge=0, le=100)
+
+
+# Backward-compatible alias while the rest of the codebase migrates to the new name.
+ArticleCandidate = ArticleCandidateRecord
+
+
+def normalize_candidate_url(url: str) -> str:
+    parsed = urlparse(url)
+    filtered_query = [
+        (key, value)
+        for key, value in parse_qsl(parsed.query, keep_blank_values=True)
+        if key.lower() not in _TRACKING_QUERY_PARAMS and not key.lower().startswith("utm_")
+    ]
+    normalized = parsed._replace(query=urlencode(filtered_query, doseq=True), fragment="")
+    return urlunparse(normalized)
+
+
+def candidate_id_for(item: Item | None, article: ExtractedArticle) -> str:
+    if item is not None and item.item_id:
+        return f"cand:{item.item_id}"
+    if article.item_id:
+        return f"cand:{article.item_id}"
+    return f"cand:{article.extract_id}"
+
+
+def build_article_candidate_record(
+    *,
+    summary: LlmSummaryResult,
+    article: ExtractedArticle,
+    filter_result: FilterDecisionResult,
+    item: Item | None = None,
+    digest_section_hint: DigestSectionHint | None = None,
+    digest_rank: int | None = None,
+    rendered_markdown: str | None = None,
+    source_refs: CandidateSourceRefs | None = None,
+    metadata: CandidateMetadata | None = None,
+) -> ArticleCandidateRecord:
+    return ArticleCandidateRecord(
+        candidate_id=candidate_id_for(item, article),
+        item=item,
+        article=article,
+        summary=summary,
+        filter_result=filter_result,
+        digest_section_hint=digest_section_hint,
+        digest_rank=digest_rank if digest_rank is not None else filter_result.priority,
+        source_refs=source_refs or CandidateSourceRefs(),
+        rendered_markdown=rendered_markdown,
+        metadata=metadata or CandidateMetadata(),
+    )
+
+
+def _source_name_from_item(item: Item | None, article: ExtractedArticle) -> str | None:
+    if item is not None:
+        metadata = item.metadata if isinstance(item.metadata, dict) else {}
+        origin = metadata.get("origin")
+        if isinstance(origin, dict):
+            origin_title = origin.get("title")
+            if isinstance(origin_title, str) and origin_title.strip():
+                return origin_title.strip()
+
+        for key in ("source_title", "feed_title", "site_name", "source_name"):
+            value = metadata.get(key)
+            if isinstance(value, str) and value.strip():
+                return value.strip()
+
+    candidate_url = str(item.url) if item is not None else str(article.url)
+    hostname = urlparse(candidate_url).netloc
+    return hostname or None
+
+
+def build_openclaw_candidate_input(record: ArticleCandidateRecord) -> OpenClawCandidateInput:
+    item = record.item
+    article = record.article
+    summary = record.summary
+    filter_result = record.filter_result
+
+    title = item.title if item is not None and item.title else summary.title
+    raw_url = str(item.url) if item is not None else str(summary.url)
+    published_at = item.published_at if item is not None else article.published_at
+    author = item.author if item is not None and item.author else article.author
+    language = item.language if item is not None and item.language else article.language
+    selection_reason = filter_result.reasons[0] if filter_result.reasons else None
+
+    return OpenClawCandidateInput(
+        candidate_id=record.candidate_id,
+        title=title,
+        url=raw_url,
+        canonical_url=normalize_candidate_url(raw_url),
+        published_at=published_at,
+        author=author,
+        source_name=_source_name_from_item(item, article),
+        language=language,
+        summary=summary.summary,
+        highlights=summary.highlights,
+        keywords=summary.keywords,
+        topics=summary.topics,
+        category=summary.category,
+        worth_keeping=summary.worth_keeping,
+        worth_reason=summary.reason,
+        selection_decision=filter_result.decision,
+        selection_reason=selection_reason,
+        digest_section_hint=record.digest_section_hint,
+        digest_rank=record.digest_rank,
+    )

src/summary_mcp/models/candidate_review.py

@@ -0,0 +1,19 @@
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Literal
+
+from pydantic import BaseModel
+
+
+ReviewStatus = Literal["pending", "confirmed", "rejected", "deferred"]
+KnowledgeDecision = Literal["pending", "ingest", "skip"]
+
+
+class CandidateReviewStatus(BaseModel):
+    candidate_id: str
+    review_status: ReviewStatus = "pending"
+    reviewed_at: datetime | None = None
+    review_note: str | None = None
+    knowledge_decision: KnowledgeDecision = "pending"
+    knowledge_note: str | None = None

src/summary_mcp/models/daily_digest.py

@@ -0,0 +1,58 @@
+from __future__ import annotations
+
+from datetime import date, datetime
+
+from pydantic import BaseModel, Field, HttpUrl
+
+from .article_candidate import DigestSectionHint
+
+
+class DailyDigestItem(BaseModel):
+    candidate_id: str
+    title: str
+    summary: str | None = None
+    why_it_matters: str
+    action: str | None = None
+
+
+class DailyDigestSection(BaseModel):
+    section_id: DigestSectionHint
+    title: str
+    summary: str
+    items: list[DailyDigestItem] = Field(default_factory=list)
+
+
+class DailyDigestSourceRef(BaseModel):
+    candidate_id: str
+    url: HttpUrl | None = None
+    title: str | None = None
+
+
+class DailyDigestStats(BaseModel):
+    candidate_total: int = Field(default=0, ge=0)
+    kept_total: int = Field(default=0, ge=0)
+    review_total: int = Field(default=0, ge=0)
+    dropped_total: int = Field(default=0, ge=0)
+
+
+class DailyDigestMetadata(BaseModel):
+    generated_at: datetime | None = None
+    producer: str | None = None
+    pipeline_version: str = "v1"
+    run_id: str | None = None
+
+
+class DailyDigest(BaseModel):
+    digest_id: str
+    date: date
+    title: str
+    summary: str
+    sections: list[DailyDigestSection] = Field(default_factory=list)
+    top_items: list[str] = Field(default_factory=list)
+    key_takeaways: list[str] = Field(default_factory=list)
+    watchlist: list[str] = Field(default_factory=list)
+    candidate_ids: list[str] = Field(default_factory=list)
+    source_refs: list[DailyDigestSourceRef] = Field(default_factory=list)
+    editor_notes: str | None = None
+    stats: DailyDigestStats = Field(default_factory=DailyDigestStats)
+    metadata: DailyDigestMetadata = Field(default_factory=DailyDigestMetadata)

src/summary_mcp/models/filtering.py

@@ -0,0 +1,67 @@
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+from .document import ExtractedArticle
+from .item import Item
+from .llm_result import LlmSummaryResult
+
+
+FilterDecision = Literal["keep", "drop", "review"]
+ConditionOp = Literal["eq", "ne", "in", "not_in", "contains", "overlap", "gte", "lte", "exists"]
+
+
+class FilterContext(BaseModel):
+    source_tags: list[str] = Field(default_factory=list)
+    interest_topics: list[str] = Field(default_factory=list)
+    interest_keywords: list[str] = Field(default_factory=list)
+    now: datetime | None = None
+
+
+class FieldCondition(BaseModel):
+    field: str
+    op: ConditionOp
+    value: Any = None
+
+
+class FilterAction(BaseModel):
+    decision: FilterDecision
+    reason: str
+    labels: list[str] = Field(default_factory=list)
+    priority: int = 50
+
+
+class FilterRule(BaseModel):
+    rule_id: str
+    enabled: bool = True
+    stop_on_match: bool = False
+    conditions_all: list[FieldCondition] = Field(default_factory=list)
+    conditions_any: list[FieldCondition] = Field(default_factory=list)
+    action: FilterAction
+
+
+class FilterInput(BaseModel):
+    item: Item | None = None
+    article: ExtractedArticle | None = None
+    summary: LlmSummaryResult
+    context: FilterContext = Field(default_factory=FilterContext)
+
+
+class MatchedRule(BaseModel):
+    rule_id: str
+    decision: FilterDecision
+    reason: str
+    labels: list[str] = Field(default_factory=list)
+    priority: int = 0
+
+
+class FilterDecisionResult(BaseModel):
+    decision: FilterDecision
+    matched_rules: list[str] = Field(default_factory=list)
+    reasons: list[str] = Field(default_factory=list)
+    labels: list[str] = Field(default_factory=list)
+    priority: int = 0
+    matches: list[MatchedRule] = Field(default_factory=list)

src/summary_mcp/models/keyword_index.py

@@ -0,0 +1,35 @@
+from __future__ import annotations
+
+from datetime import date, datetime
+
+from pydantic import BaseModel, Field
+
+
+class DailyKeywordTerm(BaseModel):
+    term: str
+    normalized_term: str
+    count: int = Field(ge=1)
+
+
+class DailyKeywordIndex(BaseModel):
+    schema_version: str = "v1"
+    date: date
+    source: str
+    digest_id: str
+    generated_at: datetime
+    candidate_count: int = Field(default=0, ge=0)
+    terms: list[DailyKeywordTerm] = Field(default_factory=list)
+
+
+class KeywordStat(BaseModel):
+    term: str
+    total_count: int = Field(default=0, ge=0)
+    days_seen: int = Field(default=0, ge=0)
+    first_seen: date
+    last_seen: date
+
+
+class KeywordStatsIndex(BaseModel):
+    schema_version: str = "v1"
+    generated_at: datetime
+    terms: list[KeywordStat] = Field(default_factory=list)

src/summary_mcp/models/openclaw_delivery.py

@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+from datetime import UTC, date, datetime
+
+from pydantic import BaseModel, Field
+
+from .article_candidate import OpenClawCandidateInput
+
+
+class OpenClawDeliveryStats(BaseModel):
+    total: int = Field(default=0, ge=0)
+    keep_total: int = Field(default=0, ge=0)
+    review_total: int = Field(default=0, ge=0)
+    drop_total: int = Field(default=0, ge=0)
+
+
+class OpenClawDeliveryPayload(BaseModel):
+    schema_version: str = "v1"
+    generated_at: datetime
+    run_id: str
+    date: date
+    candidates: list[OpenClawCandidateInput] = Field(default_factory=list)
+    stats: OpenClawDeliveryStats = Field(default_factory=OpenClawDeliveryStats)
+
+
+def build_openclaw_delivery_payload(
+    candidates: list[OpenClawCandidateInput],
+    *,
+    run_id: str | None = None,
+    for_date: date | None = None,
+) -> OpenClawDeliveryPayload:
+    now = datetime.now(tz=UTC)
+    payload_date = for_date or now.date()
+    resolved_run_id = run_id or now.strftime("openclaw-delivery-%Y%m%d-%H%M%S")
+
+    stats = OpenClawDeliveryStats(
+        total=len(candidates),
+        keep_total=sum(1 for candidate in candidates if candidate.selection_decision == "keep"),
+        review_total=sum(1 for candidate in candidates if candidate.selection_decision == "review"),
+        drop_total=sum(1 for candidate in candidates if candidate.selection_decision == "drop"),
+    )
+    return OpenClawDeliveryPayload(
+        schema_version="v1",
+        generated_at=now,
+        run_id=resolved_run_id,
+        date=payload_date,
+        candidates=candidates,
+        stats=stats,
+    )

src/summary_mcp/models/sink.py

@@ -0,0 +1,35 @@
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+from .document import ExtractedArticle
+from .filtering import FilterDecisionResult
+from .item import Item
+from .llm_result import LlmSummaryResult
+
+
+SinkTarget = Literal["markdown"]
+
+
+class SinkMetadata(BaseModel):
+    generated_at: datetime | None = None
+    pipeline_version: str = "v1"
+    source: str | None = None
+
+
+class SinkInput(BaseModel):
+    item: Item | None = None
+    article: ExtractedArticle
+    summary: LlmSummaryResult
+    filter_decision: FilterDecisionResult
+    metadata: SinkMetadata = Field(default_factory=SinkMetadata)
+
+
+class SinkResult(BaseModel):
+    target: SinkTarget
+    success: bool
+    output_path: str | None = None
+    message: str | None = None

src/summary_mcp/server.py

@@ -1,10 +1,20 @@
 from __future__ import annotations
 
+import json
+import tempfile
+from datetime import date
+from pathlib import Path
+
 from mcp.server.fastmcp import FastMCP
 
 from summary_mcp.core.pipeline import extract_content
+from summary_mcp.filters.engine import evaluate_filter_rules, load_filter_rules
+from summary_mcp.models.document import ExtractedArticle
+from summary_mcp.models.filtering import FilterContext, FilterInput
 from summary_mcp.models.item import Item
+from summary_mcp.models.llm_result import LlmSummaryResult
 from summary_mcp.models.summary_io import ExtractionInput
+from summary_mcp.workflows import run_freshrss_pipeline
 
 
 mcp = FastMCP(name="content-extract-mcp")
@@ -26,12 +36,95 @@ def extract_url_content(url: str, language_hint: str | None = None) -> dict:
 def extract_item_content(item: dict) -> dict:
     """Extract structured article content from a normalized item object."""
     parsed_item = Item.model_validate(item)
-    result = extract_content(
-        ExtractionInput(item=parsed_item)
-    )
+    result = extract_content(ExtractionInput(item=parsed_item))
     return result.model_dump(mode="json")
 
 
+@mcp.tool()
+def filter_summary_result(
+    summary_result: dict,
+    extracted_article: dict | None = None,
+    item: dict | None = None,
+    context: dict | None = None,
+) -> dict:
+    """Apply deterministic filter rules to a structured summary result."""
+    parsed_summary = LlmSummaryResult.model_validate(summary_result)
+    parsed_article = ExtractedArticle.model_validate(extracted_article) if extracted_article else None
+    parsed_item = Item.model_validate(item) if item else None
+    parsed_context = FilterContext.model_validate(context or {})
+    rules = load_filter_rules()
+    decision = evaluate_filter_rules(
+        FilterInput(
+            item=parsed_item,
+            article=parsed_article,
+            summary=parsed_summary,
+            context=parsed_context,
+        ),
+        rules,
+    )
+    return decision.model_dump(mode="json")
+
+
+@mcp.tool()
+def run_freshrss_openclaw_pipeline(
+    limit: int = 5,
+    mark_read: bool = False,
+    include_read: bool = False,
+    debug_artifacts: bool = False,
+    continuation: str | None = None,
+    timeout_seconds: float = 60.0,
+    max_retries: int = 2,
+    stream_id: str = "user/-/state/com.google/reading-list",
+    api_base_url: str | None = None,
+    username: str | None = None,
+    api_password: str | None = None,
+    llm_api_key: str | None = None,
+    llm_model: str | None = None,
+    llm_api_url: str | None = None,
+    context: dict | None = None,
+    run_id: str | None = None,
+    date_value: str | None = None,
+    output_dir: str | None = None,
+    include_item_reports: bool = False,
+) -> dict:
+    """Run the full FreshRSS -> extract -> LLM -> filter -> OpenClaw payload pipeline."""
+    temp_context_path: Path | None = None
+
+    try:
+        if context is not None:
+            with tempfile.NamedTemporaryFile("w", encoding="utf-8", suffix=".json", delete=False) as handle:
+                json.dump(context, handle, ensure_ascii=False, indent=2)
+                temp_context_path = Path(handle.name)
+
+        result = run_freshrss_pipeline(
+            api_base_url=api_base_url,
+            username=username,
+            api_password=api_password,
+            stream_id=stream_id,
+            limit=limit,
+            continuation=continuation,
+            include_read=include_read,
+            mark_read=mark_read,
+            debug_artifacts=debug_artifacts,
+            context_path=temp_context_path,
+            max_retries=max_retries,
+            timeout_seconds=timeout_seconds,
+            llm_api_key=llm_api_key,
+            llm_model=llm_model,
+            llm_api_url=llm_api_url,
+            run_id=run_id,
+            delivery_date=date.fromisoformat(date_value) if date_value else None,
+            output_dir=Path(output_dir) if output_dir else None,
+        )
+    finally:
+        if temp_context_path and temp_context_path.exists():
+            temp_context_path.unlink()
+
+    if not include_item_reports:
+        result = {key: value for key, value in result.items() if key != "items"}
+    return result
+
+
 def main() -> None:
     mcp.run()
 

src/summary_mcp/sinks/__init__.py

@@ -0,0 +1,3 @@
+from .markdown import MarkdownSink
+
+__all__ = ["MarkdownSink"]

src/summary_mcp/sinks/markdown.py

@@ -0,0 +1,127 @@
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from pathlib import Path
+import re
+
+from summary_mcp.models.sink import SinkInput, SinkResult
+
+
+_DECISION_DIR = {
+    "keep": "inbox",
+    "review": "review",
+    "drop": "archive",
+}
+
+
+def _slugify(text: str) -> str:
+    slug = text.lower()
+    slug = re.sub(r"[^a-z0-9\u4e00-\u9fff]+", "-", slug)
+    slug = re.sub(r"-+", "-", slug).strip("-")
+    return slug or "untitled"
+
+
+def _yaml_scalar(value: str) -> str:
+    escaped = value.replace('"', '\\"')
+    return f'"{escaped}"'
+
+
+def _yaml_list(values: list[str]) -> str:
+    if not values:
+        return "[]"
+    return "\n".join(f"  - {_yaml_scalar(value)}" for value in values)
+
+
+def _frontmatter(payload: SinkInput) -> str:
+    article = payload.article
+    summary = payload.summary
+    decision = payload.filter_decision
+    generated_at = payload.metadata.generated_at or datetime.now(tz=UTC)
+    published_at = article.published_at.isoformat() if article.published_at else ""
+    source_id = article.source_id or (payload.item.source_id if payload.item else "")
+    item_id = article.item_id or (payload.item.item_id if payload.item else "")
+
+    return "\n".join(
+        [
+            "---",
+            f"title: {_yaml_scalar(summary.title)}",
+            f"url: {_yaml_scalar(str(summary.url))}",
+            f"source_id: {_yaml_scalar(source_id)}",
+            f"item_id: {_yaml_scalar(item_id)}",
+            f"extract_id: {_yaml_scalar(article.extract_id)}",
+            f"category: {_yaml_scalar(summary.category)}",
+            f"worth_keeping: {'true' if summary.worth_keeping else 'false'}",
+            f"decision: {_yaml_scalar(decision.decision)}",
+            f"priority: {decision.priority}",
+            "topics:",
+            _yaml_list(summary.topics),
+            "keywords:",
+            _yaml_list(summary.keywords),
+            "labels:",
+            _yaml_list(decision.labels),
+            f"published_at: {_yaml_scalar(published_at)}",
+            f"saved_at: {_yaml_scalar(generated_at.isoformat())}",
+            "quality_flags:",
+            f"  is_paywalled: {'true' if article.quality_flags.is_paywalled else 'false'}",
+            f"  is_truncated: {'true' if article.quality_flags.is_truncated else 'false'}",
+            f"  is_low_content: {'true' if article.quality_flags.is_low_content else 'false'}",
+            "---",
+        ]
+    )
+
+
+def _render_markdown(payload: SinkInput) -> str:
+    summary = payload.summary
+    article = payload.article
+    decision = payload.filter_decision
+    highlights = "\n".join(f"- {item}" for item in summary.highlights)
+    reasons = "\n".join(f"- {item}" for item in decision.reasons)
+
+    return (
+        f"{_frontmatter(payload)}\n\n"
+        f"# 摘要\n\n{summary.summary}\n\n"
+        f"# 要点\n\n{highlights}\n\n"
+        f"# 过滤决策\n\n"
+        f"- decision: `{decision.decision}`\n"
+        f"- priority: `{decision.priority}`\n"
+        f"- matched rules: {', '.join(decision.matched_rules) if decision.matched_rules else 'none'}\n\n"
+        f"## reasons\n\n{reasons}\n\n"
+        f"# 原文信息\n\n"
+        f"- 标题: {summary.title}\n"
+        f"- 链接: {summary.url}\n"
+        f"- 作者: {article.author or 'unknown'}\n"
+        f"- 发布时间: {article.published_at.isoformat() if article.published_at else 'unknown'}\n"
+        f"- 内容类型: {article.content_kind}\n\n"
+        f"# 正文摘录\n\n{article.plain_text.strip()}\n"
+    )
+
+
+class MarkdownSink:
+    def __init__(self, base_dir: Path) -> None:
+        self.base_dir = base_dir
+
+    def _target_dir(self, decision: str, published_at: datetime | None) -> Path:
+        bucket = _DECISION_DIR.get(decision, "archive")
+        if published_at is None:
+            return self.base_dir / bucket / "unknown-date"
+        return self.base_dir / bucket / published_at.strftime("%Y") / published_at.strftime("%Y-%m")
+
+    def _target_name(self, payload: SinkInput) -> str:
+        article = payload.article
+        published_at = article.published_at or datetime.now(tz=UTC)
+        title = payload.summary.title or article.title or "untitled"
+        return f"{published_at.strftime('%Y-%m-%d')}-{_slugify(title)}.md"
+
+    def write(self, payload: SinkInput) -> SinkResult:
+        target_dir = self._target_dir(payload.filter_decision.decision, payload.article.published_at)
+        target_dir.mkdir(parents=True, exist_ok=True)
+        output_path = target_dir / self._target_name(payload)
+        output_path.write_text(_render_markdown(payload), encoding="utf-8")
+        return SinkResult(
+            target="markdown",
+            success=True,
+            output_path=str(output_path),
+            message="Markdown note written successfully.",
+        )
+
+

src/summary_mcp/validators/llm_result.py

@@ -47,12 +47,31 @@ def _validate_business_rules(result: LlmSummaryResult, extracted: dict[str, Any]
         if extracted_url and str(result.url) != extracted_url:
             errors.append("`url` does not match extracted article url")
 
-    if result.category == "\u8d44\u8baf" and result.worth_keeping:
-        warnings.append("`??` category marked as worth keeping; check if this is intentional.")
+    if result.category == "资讯" and result.worth_keeping:
+        warnings.append("`资讯` category marked as worth keeping; check if this is intentional.")
 
     return errors, warnings
 
 
+def validate_llm_result_payload(
+    result_payload: dict[str, Any],
+    extracted_payload: dict[str, Any] | None = None,
+) -> ValidationReport:
+    try:
+        parsed = LlmSummaryResult.model_validate(result_payload)
+    except ValidationError as exc:
+        errors = [f"{'.'.join(str(part) for part in error['loc'])}: {error['msg']}" for error in exc.errors()]
+        return ValidationReport(valid=False, errors=errors)
+
+    errors, warnings = _validate_business_rules(parsed, extracted_payload)
+    return ValidationReport(
+        valid=not errors,
+        errors=errors,
+        warnings=warnings,
+        normalized_result=parsed.model_dump(mode="json"),
+    )
+
+
 def validate_llm_result(
     result_path: Path,
     extracted_path: Path | None = None,
@@ -69,16 +88,4 @@ def validate_llm_result(
         except json.JSONDecodeError as exc:
             return ValidationReport(valid=False, errors=[f"Invalid extracted JSON: {exc}"])
 
-    try:
-        parsed = LlmSummaryResult.model_validate(raw_result)
-    except ValidationError as exc:
-        errors = [f"{'.'.join(str(part) for part in error['loc'])}: {error['msg']}" for error in exc.errors()]
-        return ValidationReport(valid=False, errors=errors)
-
-    errors, warnings = _validate_business_rules(parsed, extracted)
-    return ValidationReport(
-        valid=not errors,
-        errors=errors,
-        warnings=warnings,
-        normalized_result=parsed.model_dump(mode="json"),
-    )
+    return validate_llm_result_payload(raw_result, extracted)

src/summary_mcp/workflows/__init__.py

@@ -0,0 +1,3 @@
+from .freshrss_pipeline import default_output_dir, read_delivery_payload, run_freshrss_pipeline
+
+__all__ = ["default_output_dir", "read_delivery_payload", "run_freshrss_pipeline"]

src/summary_mcp/workflows/freshrss_pipeline.py

@@ -0,0 +1,303 @@
+from __future__ import annotations
+
+import json
+import os
+from datetime import UTC, date, datetime
+from pathlib import Path
+from typing import Any
+
+from summary_mcp.core.keyword_index import persist_keyword_indexes
+from summary_mcp.core.pipeline import extract_content
+from summary_mcp.core.summary_loop import resolve_llm_settings, run_loop_payload
+from summary_mcp.filters.engine import evaluate_filter_rules, load_filter_rules
+from summary_mcp.integrations.freshrss import READ_TAG, FreshRSSClient, map_entry_to_item
+from summary_mcp.models.article_candidate import (
+    CandidateMetadata,
+    CandidateSourceRefs,
+    OpenClawCandidateInput,
+    build_article_candidate_record,
+    build_openclaw_candidate_input,
+)
+from summary_mcp.models.filtering import FilterContext, FilterInput
+from summary_mcp.models.llm_result import LlmSummaryResult
+from summary_mcp.models.openclaw_delivery import OpenClawDeliveryPayload, build_openclaw_delivery_payload
+from summary_mcp.models.summary_io import ExtractionInput
+
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+OUTPUT_ROOT = REPO_ROOT / "outputs"
+FRESHRSS_OUTPUT_ROOT = OUTPUT_ROOT / "freshrss"
+DATA_ROOT = REPO_ROOT / "data" / "term_index"
+DEFAULT_PROMPT_PATH = OUTPUT_ROOT / "prompts" / "llm-summary-prompt.txt"
+DEFAULT_RULES_PATH = REPO_ROOT / "configs" / "filter_rules.json"
+DEFAULT_TERM_ALIASES_PATH = REPO_ROOT / "configs" / "term_aliases.json"
+DEFAULT_TERM_STOPWORDS_PATH = REPO_ROOT / "configs" / "term_stopwords.json"
+DEFAULT_TERM_DAILY_DIR = DATA_ROOT / "daily"
+DEFAULT_TERM_STATS_PATH = DATA_ROOT / "term_stats.json"
+
+
+def _save_json(path: Path, payload: dict[str, Any] | list[Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def _load_json(path: Path) -> dict[str, Any]:
+    return json.loads(path.read_text(encoding="utf-8-sig"))
+
+
+def _load_required_env(name: str, value: str | None) -> str:
+    resolved = value or os.environ.get(name)
+    if not resolved:
+        raise RuntimeError(f"Missing required value: {name}.")
+    return resolved
+
+
+def default_output_dir() -> Path:
+    return FRESHRSS_OUTPUT_ROOT / "rerun" / datetime.now(tz=UTC).strftime("%Y%m%d-%H%M%S")
+
+
+def _maybe_path(enabled: bool, path: Path) -> Path | None:
+    return path if enabled else None
+
+
+def run_freshrss_pipeline(
+    *,
+    api_base_url: str | None = None,
+    username: str | None = None,
+    api_password: str | None = None,
+    stream_id: str = "user/-/state/com.google/reading-list",
+    limit: int = 5,
+    continuation: str | None = None,
+    include_read: bool = False,
+    mark_read: bool = False,
+    debug_artifacts: bool = False,
+    prompt: Path | None = None,
+    rules: Path | None = None,
+    context_path: Path | None = None,
+    max_retries: int = 2,
+    timeout_seconds: float = 60.0,
+    llm_api_key: str | None = None,
+    llm_model: str | None = None,
+    llm_api_url: str | None = None,
+    run_id: str | None = None,
+    delivery_date: date | None = None,
+    output_dir: Path | None = None,
+) -> dict[str, Any]:
+    started_at = datetime.now(tz=UTC)
+    resolved_output_dir = output_dir or default_output_dir()
+    run_stamp = started_at.strftime("%Y%m%d-%H%M%S")
+    resolved_run_id = run_id or f"freshrss-pipeline-{run_stamp}"
+    resolved_delivery_date = delivery_date or datetime.now(tz=UTC).date()
+
+    resolved_api_base_url = _load_required_env("FRESHRSS_API_BASE_URL", api_base_url)
+    resolved_username = _load_required_env("FRESHRSS_USERNAME", username)
+    resolved_api_password = _load_required_env("FRESHRSS_API_PASSWORD", api_password)
+    resolved_llm_api_key, resolved_llm_model, resolved_llm_api_url = resolve_llm_settings(
+        api_key=llm_api_key,
+        model=llm_model,
+        api_url=llm_api_url,
+    )
+
+    resolved_prompt_path = prompt or DEFAULT_PROMPT_PATH
+    resolved_rules_path = rules or DEFAULT_RULES_PATH
+    raw_output = resolved_output_dir / "raw" / "freshrss.raw.json"
+    delivery_output = resolved_output_dir / "candidates" / "openclaw-delivery-payload.json"
+    report_output = resolved_output_dir / "run-report.json"
+    items_list_output = _maybe_path(debug_artifacts, resolved_output_dir / "items" / "freshrss.items.json")
+
+    client = FreshRSSClient(
+        api_base_url=resolved_api_base_url,
+        username=resolved_username,
+        api_password=resolved_api_password,
+        timeout_seconds=timeout_seconds,
+    )
+    auth_token = client.client_login()
+    payload = client.fetch_stream_contents(
+        auth_token=auth_token,
+        stream_id=stream_id,
+        limit=limit,
+        continuation=continuation,
+        exclude_targets=[] if include_read else [READ_TAG],
+    )
+    entries = payload.get("items")
+    if not isinstance(entries, list):
+        raise RuntimeError("FreshRSS stream response does not contain an items array.")
+
+    _save_json(raw_output, payload)
+
+    items = [map_entry_to_item(entry) for entry in entries]
+    if items_list_output is not None:
+        _save_json(items_list_output, [item.model_dump(mode="json") for item in items])
+
+    loaded_rules = load_filter_rules(resolved_rules_path)
+    context = FilterContext.model_validate(_load_json(context_path)) if context_path else FilterContext()
+
+    delivered_candidates: list[OpenClawCandidateInput] = []
+    delivered_item_ids: list[str] = []
+    item_reports: list[dict[str, Any]] = []
+
+    for index, item in enumerate(items, start=1):
+        item_key = f"item-{index:02d}"
+        item_path = _maybe_path(debug_artifacts, resolved_output_dir / "items" / f"{item_key}.item.json")
+        extracted_path = _maybe_path(debug_artifacts, resolved_output_dir / "extracted" / f"{item_key}.extracted.json")
+        summary_output = _maybe_path(debug_artifacts, resolved_output_dir / "summary" / item_key / "result.loop.json")
+        filter_path = _maybe_path(debug_artifacts, resolved_output_dir / "filter" / f"{item_key}.filter.json")
+        record_path = _maybe_path(debug_artifacts, resolved_output_dir / "candidates" / f"{item_key}.article-candidate-record.json")
+        openclaw_path = _maybe_path(debug_artifacts, resolved_output_dir / "candidates" / f"{item_key}.openclaw-candidate-input.json")
+
+        if item_path is not None:
+            _save_json(item_path, item.model_dump(mode="json"))
+
+        item_report: dict[str, Any] = {
+            "item_key": item_key,
+            "item_id": item.item_id,
+            "external_id": item.external_id,
+            "url": str(item.url),
+            "title": item.title,
+            "status": "pulled",
+        }
+        if debug_artifacts:
+            item_report["paths"] = {
+                "item": str(item_path) if item_path else None,
+                "extracted": str(extracted_path) if extracted_path else None,
+                "summary": str(summary_output) if summary_output else None,
+                "filter": str(filter_path) if filter_path else None,
+                "article_candidate": str(record_path) if record_path else None,
+                "openclaw_candidate": str(openclaw_path) if openclaw_path else None,
+            }
+
+        extraction = extract_content(ExtractionInput(item=item))
+        extracted_payload = extraction.model_dump(mode="json")
+        if extracted_path is not None:
+            _save_json(extracted_path, extracted_payload)
+        if not extraction.success or extraction.article is None:
+            item_report["status"] = "extract_failed"
+            item_report["error"] = extraction.error.model_dump(mode="json") if extraction.error else None
+            item_reports.append(item_report)
+            continue
+
+        item_report["status"] = "extracted"
+        summary_exit_code, summary_payload, summary_report = run_loop_payload(
+            extracted_payload=extracted_payload,
+            prompt_path=resolved_prompt_path,
+            output_path=summary_output,
+            max_retries=max_retries,
+            timeout_seconds=timeout_seconds,
+            api_key=resolved_llm_api_key,
+            model=resolved_llm_model,
+            api_url=resolved_llm_api_url,
+        )
+        if summary_exit_code != 0 or summary_payload is None:
+            item_report["status"] = "summary_failed"
+            if summary_report is not None:
+                item_report["summary_errors"] = summary_report.errors
+            item_reports.append(item_report)
+            continue
+
+        summary = LlmSummaryResult.model_validate(summary_payload)
+        decision = evaluate_filter_rules(
+            FilterInput(item=item, article=extraction.article, summary=summary, context=context),
+            loaded_rules,
+        )
+        if filter_path is not None:
+            _save_json(filter_path, decision.model_dump(mode="json"))
+
+        record = build_article_candidate_record(
+            summary=summary,
+            article=extraction.article,
+            filter_result=decision,
+            item=item,
+            source_refs=CandidateSourceRefs(
+                item_path=str(item_path) if item_path else None,
+                extracted_path=str(extracted_path) if extracted_path else None,
+                summary_path=str(summary_output) if summary_output else None,
+                filter_path=str(filter_path) if filter_path else None,
+            ),
+            metadata=CandidateMetadata(
+                generated_at=datetime.now(tz=UTC),
+                producer="run_freshrss_pipeline",
+                run_id=resolved_run_id,
+            ),
+        )
+        openclaw_input = build_openclaw_candidate_input(record)
+
+        if record_path is not None:
+            _save_json(record_path, record.model_dump(mode="json"))
+        if openclaw_path is not None:
+            _save_json(openclaw_path, openclaw_input.model_dump(mode="json"))
+
+        delivered_candidates.append(openclaw_input)
+        if item.external_id:
+            delivered_item_ids.append(item.external_id)
+
+        item_report["status"] = "delivered"
+        item_report["selection_decision"] = decision.decision
+        item_report["candidate_id"] = openclaw_input.candidate_id
+        item_reports.append(item_report)
+
+    delivered_candidates.sort(key=lambda candidate: candidate.digest_rank, reverse=True)
+    delivery_payload = build_openclaw_delivery_payload(
+        delivered_candidates,
+        run_id=resolved_run_id,
+        for_date=resolved_delivery_date,
+    )
+    _save_json(delivery_output, delivery_payload.model_dump(mode="json"))
+
+    keyword_index_result = persist_keyword_indexes(
+        delivery_payload.candidates,
+        for_date=delivery_payload.date,
+        digest_id=delivery_payload.run_id,
+        source="openclaw_delivery_payload",
+        daily_dir=DEFAULT_TERM_DAILY_DIR,
+        stats_path=DEFAULT_TERM_STATS_PATH,
+        aliases_path=DEFAULT_TERM_ALIASES_PATH,
+        stopwords_path=DEFAULT_TERM_STOPWORDS_PATH,
+    )
+
+    marked_count = 0
+    if mark_read and delivered_item_ids:
+        client.mark_items_as_read(auth_token=auth_token, item_ids=delivered_item_ids)
+        marked_count = len({item_id for item_id in delivered_item_ids if item_id})
+
+    status_counts: dict[str, int] = {}
+    for item_report in item_reports:
+        status = str(item_report["status"])
+        status_counts[status] = status_counts.get(status, 0) + 1
+
+    report = {
+        "run_id": resolved_run_id,
+        "started_at": started_at.isoformat(),
+        "completed_at": datetime.now(tz=UTC).isoformat(),
+        "requested_limit": limit,
+        "pulled_count": len(items),
+        "delivered_count": len(delivered_candidates),
+        "marked_read_count": marked_count,
+        "mark_read_requested": mark_read,
+        "debug_artifacts": debug_artifacts,
+        "raw_output": str(raw_output),
+        "delivery_output": str(delivery_output),
+        "keyword_index": keyword_index_result,
+        "status_counts": status_counts,
+        "items": item_reports,
+    }
+    _save_json(report_output, report)
+
+    return {
+        "run_id": resolved_run_id,
+        "output_dir": str(resolved_output_dir),
+        "raw_output": str(raw_output),
+        "delivery_output": str(delivery_output),
+        "report_output": str(report_output),
+        "keyword_index": keyword_index_result,
+        "pulled_count": len(items),
+        "delivered_count": len(delivered_candidates),
+        "marked_read_count": marked_count,
+        "status_counts": status_counts,
+        "debug_artifacts": debug_artifacts,
+        "delivery_payload": delivery_payload.model_dump(mode="json"),
+        "items": item_reports,
+    }
+
+
+def read_delivery_payload(path: Path) -> OpenClawDeliveryPayload:
+    return OpenClawDeliveryPayload.model_validate(_load_json(path))