daily-report-lane-github-ai-projects

name	daily-report-lane-github-ai-projects
description	Generate the GitHub AI 项目 section of the AI Agent daily report from raw corpus.

GitHub AI 项目 Lane

Mission

主动发现 GitHub AI/coding-agent 项目，不能用旧 shared memory 文件作为集成点。

Input

Read the lane package provided by the master:

• input.md
• context.json
• raw/
• history/ if present

Do not use selected_items.json as primary judgment input. Raw corpus is the evidence source. Use recent reports only as reference-only dedupe context. Before selecting or writing, read yesterday and day-before-yesterday report files listed in context.json recent_report_paths or package history/. Reject exact repeats or substantially unchanged topics. Keep meaningful follow-ups with new facts and state what changed. Do not dedupe weather/current market items purely because yesterday had the same section. This is lane-agent judgment, not code-controlled filtering.

Selection Rules

Select only items that can be explained concretely for an AI/coding-agent reader. Prefer specific releases, workflows, repos, discussions, failures, tools, versions, or user-visible changes.

Hard GitHub repo floor: every selected repository must have verified stars >= 100. Reject repositories below 100 stars and repositories whose star count cannot be verified from raw corpus or GitHub metadata. Discovery/search queries for GitHub repositories must include stars:>=100; do not spend reader attention on tiny repos unless MT explicitly asks for an early-stage scan.

Rejection Rules

Reject items when:

• evidence is too thin;
• source link is missing;
• it is generic tech/news not relevant to AI/coding-agent workflows;
• it repeats another stronger item;
• it cannot be explained in concrete human terms.

Usefulness Bar（MT 反馈 2026-06-25）

长度不是问题；问题是“有用信息密度”。宁可少选，也不要用低信息量内容填满栏目。每个入选条目必须同时满足：

• 新信息：今天相对昨天/前天或常识，具体新增了什么；如果只是同一产品/观点反复出现，必须写出新增事实，否则拒绝。
• 可执行/可判断：读者看完能知道该试用、避坑、跟进、学习、忽略中的哪一种，以及理由。
• 证据锚点：至少包含一个可核验事实（版本号、repo/产品名、作者、数据、价格、接口/功能变化、错误现象、讨论分歧、链接来源）。
• 非空泛价值：禁止只写“值得关注/说明趋势/生态变化/更成熟/更清楚”；如果没有具体“为什么有用”，就不要选。

写作时优先保留事实与判断依据；不要为了变短删除最有用的细节。

Lane-Specific Judgment

Run internal discovery or evidence helper when useful. Never use the old shared memory output as integration point. Treat discovery output as evidence only; the subagent still selects/rejects and writes. Select only projects relevant to AI/coding-agent workflows.

Writing Style

Write Chinese-first, concrete, human-readable prose. Keep source links.

子条目要信息密度高但不要拖长：

• 每条默认 1-3 句；信息简单时 1 句，复杂更新最多 3 句。
• 单条正文最长不超过 200 个中文字符；链接不计入正文长度。
• 句式优先：发生了什么 + 关键事实/数字 + 对读者有什么用/风险，少写铺垫和评价。
• 删除“值得关注/值得跟踪/说明生态变化”等空泛判断，除非同句给出具体证据。

Forbidden Output

禁止使用：

• “趋势信息包含这些具体点”
• “这条原始信号给出的可核验信息集中在”
• “适合作为今日该栏目的迁移期素材”
• “具体变化见来源”
• “采集文本”
• “当前可作为”
• internal collector voice

Output

Write exactly:

• lane.md: reader-facing markdown section only.
• lane-meta.json: JSON metadata with lane, status, selected_count, rejected_count, sources, rejected, and notes. Allowed status: ok, empty, degraded, blocked. Use selected and rejected reasoning in metadata so the master can audit the lane without rewriting it.