← cyberMap 导航← cyberMap Home
每日 08:00 更新Daily 08:00 UTC+8

算法与数据口径

这份榜单的每个名次都应该能被复算、被质疑、被回溯。下面把口径全部摊开。

一、这个榜是什么,不是什么

它统计的是 在 GitHub 上有公开可核验仓库、且属于某个 AI Agent 平台插件生态的项目,按当日公开指标排序。它不是官方下载量榜,也不代表插件的实际使用人数或质量评审。

数据分两层:榜单页(本页下述评分口径,总榜与各分榜各取前 100)和 官方市场目录页(各平台本地市场清单的全量条目,不做分数排序)。

最重要的局限:Codex、WorkBuddy 等平台的官方市场里存在大量没有公开 GitHub 仓库的闭源技能、连接器与专家包。它们不参与总榜分数排序(否则「中性缺失」会把零数据实体抬进 Top 100),但会在官方市场目录页完整列出。这是数据可得性的边界,不是遗漏,也不会用估算值补齐。

二、评分公式

Score = 100 × Σ( wᵢ · norm(xᵢ) ) / Σ( wᵢ 可用 )

norm 用当日全样本分位(并列取平均名次),而不是除以最大值,避免头部仓库碾压长尾;stars / forks / 下载量这类长尾指标先取 log1p。四个维度:

维度原始权重构成数据来源
热度 P0.40stars 0.60 + forks 0.30 + npm 周下载 0.10(子项按可用性重归一)GitHub API / npm downloads API
活跃 A0.25exp(−Δt / 90),Δt 为最近 push 距今天数GitHub pushed_at
势能 M0.25与 ≤7 天前历史快照比较的 stars 增量分位本站 data/history/ 快照
质量 Q0.10license 0.25、描述完整度 0.20、主页 0.15、topics 0.15、wiki/pages 0.15、语言 0.10GitHub 仓库元数据

三、缺失数据怎么处理

四、去重与平台归属

五、准入与反作弊

六、本地官方市场清单(发现源)

GitHub topic 搜索只能发现「打了标签的仓库」,会漏掉各平台市场里已正式收录、但没打 topic 的插件。为此流水线只读本机下列清单文件,把它们作为权威发现源(绝不修改用户配置):

平台清单文件可得字段
Claude Code~/.claude/plugins/marketplaces/*/.claude-plugin/marketplace.json插件名、描述、作者、分类、版本 ref、source.url(多为 GitHub 仓库)、子目录路径
Codex~/.codex/plugins/cache/**/.codex-plugin/plugin.json名称、版本、描述、作者、homepage
WorkBuddyconnectors.json、各 marketplace.json、插件缓存、.skill-list-cache.json中英双语描述、类型、版本、可见范围、部分 repository
dshdsh --profile web --dump-config + profiles/*/package.jsonnpm 包名(可查周下载量)

三条口径必须分清:

七、数据源与证据链

只使用 api.github.com/search/repositories、api.github.com/repos/* 与 api.npmjs.org/downloads 的公开返回。第三方榜单的 stars / 安装量一概不采信,最多作为发现线索,进榜前必须自行调用 API 复核。

官方清单里指向的仓库先用 Search API 的多个 repo:owner/name 限定符批量取(走 search 独立配额);实测有一批真实仓库不在搜索索引中(返回 422「资源不可搜索」,例如 anthropics/claude-plugins-public),这些再用 GET /repos/{owner}/{repo} 二次复核,REST 仍 404 才判为不可核验。所有结果含「确认查不到」都写入 data/repo-metadata.json 缓存,7 天内复用,既不臆造也不每天重查死链。

每条记录保存采集 query、UTC 采集时间、API 端点与当时的原始字段,详情页「证据链」表格可直接对照回查。每日运行报告在 state/run-YYYY-MM-DD.json,含全部淘汰条目与理由。

八、跨日比较请注意

分数是当日相对分位,样本集与有效权重每天都可能变化,因此分数绝对值不可跨日比较。跨日只看排名与排名变化(▲/▼),详情页趋势曲线展示的是同一口径下的历史采样,供参考而非精确审计。

九、怎么自己复算

GITHUB_TOKEN=xxx python pipelines/run_daily.py
python pipelines/run_daily.py --dry-run # 只算不生成榜单产物
python pipelines/run_daily.py --no-local-market # 只跑 GitHub topic 采集(对照用)
python pipelines/run_daily.py --no-market-npm # 跳过 dsh 包名下载量查询

权重、query、准入词表在 pipelines/config.py,评分在 pipelines/score.py,官方清单解析与仓库核验在 pipelines/local_market.py。当日生效权重写进 data/top100.json 的 meta.scoring;官方目录在 data/marketplace.json;仓库/npm 元数据缓存在 data/repo-metadata.json 与 data/npm-metadata.json。改口径请同步改这一页的说明。

没有 GITHUB_TOKEN 时,GitHub 未认证限额为 Search 10 次/分钟、core 60 次/小时。首次运行需批量核验约 240 个仓库(约 3 分钟),之后靠缓存几乎零请求。建议设置 token,可显著提高限额并缩短首次运行时间。

Methodology & Data Policy

Every rank on this leaderboard should be reproducible, challengeable and traceable. Below we lay out the full methodology.

1. What this leaderboard is — and is not

It tracks projects with a public, verifiable GitHub repository that belong to an AI Agent platform's plugin ecosystem, ranked by the day's public metrics. It is not an official download chart, nor does it represent actual usage or a quality review.

The data has two layers: the leaderboard (the scoring policy described here, top 100 in the overall and each platform chart) and the official marketplace directory (all entries from each platform's local marketplace list, without score ranking).

The biggest limitation: platform marketplaces such as Codex and WorkBuddy contain many closed-source skills, connectors and expert packs that have no public GitHub repository. They do not take part in the overall score ranking (otherwise "neutral missing" would lift zero-data entities into the Top 100), but they are listed in full on the marketplace directory. This is a boundary of data availability, not an omission, and no estimated values are used to fill in.

2. Scoring formula

Score = 100 × Σ( wᵢ · norm(xᵢ) ) / Σ( wᵢ available )

norm uses the same-day percentile across the full sample (ties take the average rank) rather than dividing by the maximum, so top repositories don't crush the long tail; long-tail metrics such as stars / forks / downloads first take log1p. Four dimensions:

DimensionBase weightCompositionData source
Popularity P0.40stars 0.60 + forks 0.30 + npm weekly downloads 0.10 (sub-items renormalized by availability)GitHub API / npm downloads API
Activity A0.25exp(−Δt / 90), where Δt is days since the last pushGitHub pushed_at
Momentum M0.25Percentile of stars growth versus a historical snapshot from ≤7 days agoSite data/history/ snapshots
Quality Q0.10license 0.25, description completeness 0.20, homepage 0.15, topics 0.15, wiki/pages 0.15, language 0.10GitHub repository metadata

3. How missing data is handled

4. Deduplication and platform attribution

5. Inclusion and anti-gaming

6. Local official marketplace lists (discovery source)

GitHub topic search can only find "tagged repositories" and misses plugins officially listed in platform marketplaces but without a topic. To fix this, the pipeline reads the following local list files read-only as authoritative discovery sources (user configuration is never modified):

PlatformList fileAvailable fields
Claude Code~/.claude/plugins/marketplaces/*/.claude-plugin/marketplace.jsonplugin name, description, author, category, version ref, source.url (mostly GitHub repos), subdirectory path
Codex~/.codex/plugins/cache/**/.codex-plugin/plugin.jsonname, version, description, author, homepage
WorkBuddyconnectors.json, each marketplace.json, plugin cache, .skill-list-cache.jsonbilingual descriptions, type, version, visibility scope, partial repository
dshdsh --profile web --dump-config + profiles/*/package.jsonnpm package name (weekly downloads can be queried)

Three distinctions must be kept clear:

7. Data sources and evidence chain

Only public returns from api.github.com/search/repositories, api.github.com/repos/* and api.npmjs.org/downloads are used. stars / install counts from third-party leaderboards are never trusted, used at most as discovery leads and re-verified via our own API calls before inclusion.

Repositories referenced in official lists are first fetched in bulk with multiple repo:owner/name qualifiers through the Search API (separate search quota). In practice a number of real repositories are absent from the search index (HTTP 422 "resource not searchable", e.g. anthropics/claude-plugins-public); those are re-checked with GET /repos/{owner}/{repo}, and only a REST 404 marks them unverifiable. All results, including "confirmed not found", are cached in data/repo-metadata.json for 7 days — no fabrication and no daily re-checking of dead links.

Each record stores the collection query, UTC collection time, API endpoint and the original fields at that moment; the "Evidence chain" table on detail pages can be checked directly against them. Daily run reports live in state/run-YYYY-MM-DD.json, including all rejected entries and reasons.

8. A note on cross-day comparison

Scores are same-day relative percentiles; the sample set and effective weights can change every day, so absolute score values cannot be compared across days. Compare only ranks and rank changes (▲/▼). The trend curve on detail pages shows historical samples under the same methodology, for reference rather than precise auditing.

9. How to recompute it yourself

GITHUB_TOKEN=xxx python pipelines/run_daily.py
python pipelines/run_daily.py --dry-run # compute only, no leaderboard artifacts
python pipelines/run_daily.py --no-local-market # only run GitHub topic collection (for comparison)
python pipelines/run_daily.py --no-market-npm # skip the dsh package download query

Weights, queries and inclusion word lists are in pipelines/config.py, scoring in pipelines/score.py, official-list parsing and repo verification in pipelines/local_market.py. The effective weights for the day are written to data/top100.json under meta.scoring; the official directory is in data/marketplace.json; repo/npm metadata caches are in data/repo-metadata.json and data/npm-metadata.json. When changing the methodology, update this page accordingly.

Without GITHUB_TOKEN, GitHub's unauthenticated limits are Search 10/min and core 60/hour. The first run must verify roughly 240 repositories in bulk (about 3 minutes); afterwards the cache brings requests to nearly zero. Setting a token is recommended to significantly raise the limits and shorten the first run.