repo-radar-stinger
Scan GitHub for repos trending in the last 24 hours (vibe coding, AI coding, AI memory, vulnerabilities) via Exa/Firecrawl MCP, vet ONE, and output post options in Mario's voice with receipts.
legioncodeinc · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add legioncodeinc/vibe-coding-tools --skill repo-radar-stinger技能说明
name: repo-radar-stinger description: Scan GitHub for repos trending in the last 24 hours (vibe coding, AI coding, AI memory, vulnerabilities) via Exa/Firecrawl MCP, vet ONE, and output post options in Mario's voice with receipts. license: AGPL-3.0-or-later compatibility: Claude Cowork (Routines), Cowork desktop, Cursor, Codex metadata: hive-tier: standalone voice-skill: mario-voice-skill ledger: /areas/content-post-ledger.md in user memory paired-skill: viral-news-stinger (independent sibling, shared ledger) version: "1.0"
Repo Radar Stinger
Purpose
Runs on a cadence (Claude Routines) or on demand. Scans GitHub for repositories trending RIGHT NOW in the beats Mario cares about: vibe coding and AI-assisted development, AI coding tools, AI memory systems, newly disclosed vulnerabilities, and anything else on the board a builder audience would share. Vets the single best one, then produces three copy-paste-ready post options in Mario's voice with the repo link and citations underneath. If nothing on the board is worth his voice, it produces nothing.
When to use
- A scheduled Routine run whose prompt invokes this skill by name
- "What's trending on GitHub", "find me a repo to post about", "run the repo radar"
- Any request for a viral-ready post about a trending repository in Mario's voice
When not to use
- General news with no repository at the center: that is viral-news-stinger's beat
- Rewriting existing content in Mario's voice with no mining: use mario-voice-skill directly
- Reviewing or auditing a repo's code quality: that is a different job entirely
Hard gates (all of them, every run)
- Ledger first. Read the shared topic ledger before any scan. A repo already covered, by owner/repo, link, or substance, is dead on arrival. See
references/ledger-protocol.md. - 24 hours. The rule binds the SPIKE, not the repo's age. An old repo surging today is fair game; last week's darling is stale. Acceptable spike evidence is defined in
references/repo-vetting.md(today's trending board, stars-today, a release or disclosure inside 24 hours). - One repo. Pick ONE candidate from the scan, then vet only that one. Never deep-vet multiple repos; that burns tokens for nothing. If the chosen repo fails vetting, the run ends with a no-post report. It does not fall back to vetting a second repo.
- Receipts or nothing. Vetting per
references/repo-vetting.mdmust pass: real code, honest one-sentence description, exact captured numbers, and for security topics a mandatory second source. Never fabricate stars, capabilities, or severity. For vulnerability repos, post the disclosure as news; never write exploit usage instructions or frame the post as a how-to. - No forced output. If the board holds nothing worth posting, say exactly that in three sentences and stop.
- Editorial boundaries from
references/repo-vetting.md: no celebrity gossip crossovers, no Democrat-promoting briefs, sharp but never fabricated.
Procedure
- Load the memory tools if they are deferred (in Cowork, ToolSearch for
memory_list,memory_read,memory_write,memory_str_replace). Read the ledger perreferences/ledger-protocol.mdand hold the exclusion list. - Confirm the Exa and/or Firecrawl MCP tools are available (load via ToolSearch if deferred:
web_fetch_exa,web_search_exa,firecrawl_search,firecrawl_research_search_github). These are the required mining tools. If neither server is connected, report that and stop; only fall back to a built-in fetch/search tool if one exists, and say so in the output. - Scan, capped at three or four calls total:
- Fetch https://github.com/trending (daily board) with
web_fetch_exaor Firecrawl. Capture repo names, descriptions, languages, stars, and stars-today. - Optionally one topic-led pass:
firecrawl_searchwithcategories: ["github"]orfirecrawl_research_search_githubon the priority beats (agent skills, AI memory, CVE proof-of-concepts). - Optionally one momentum cross-check: OSS Insight trending API (
period=past_24_hours) or trendshift.io, perreferences/research/distilled-repo-radar.md.
- Fetch https://github.com/trending (daily board) with
- Filter candidates through the gates: current spike, not in the ledger, in a priority beat (order defined in
references/repo-vetting.md), and hookable. Skip star-farmed templates and README-only vaporware, unless the star-farming itself is the story. - Choose ONE repo. State the choice and the one-sentence description of what it actually does.
- Vet it per
references/repo-vetting.md: fetch the repo page, read the README, capture exact numbers, cross-check independent coverage, run the security-topic rules when they apply. - Load mario-voice-skill and follow its full workflow: pick one register (technical or hype grinder usually fit repo posts, but choose per its guide), read its guide for each length, calibrate against its corpus, and run its anti-AI gate on every option. Its hard rules override everything here, including no em dashes ever.
- Write the deliverable exactly per
references/output-template.md: micro (9 words or less), medium (10-60 words), long form (200+ words), the repository link, 2-4 hashtags actually moving in the niche, and the numbered receipts (repo page capture counts as receipt one). - Deliver the output as the run's final message. In an unattended Routine run, also send it through the user-facing message tool if one is available, so it reaches Mario verbatim.
- Append one ledger line (
repo | owner/repo | link) perreferences/ledger-protocol.md. A no-post run appends nothing.
References map
references/ledger-protocol.md: load at step 1 and step 10; the shared memory file location, format, and read/write rulesreferences/repo-vetting.md: load at steps 4-6; spike evidence, priority beats, the vetting checklist, security-topic rules, editorial boundariesreferences/output-template.md: load at step 8; the exact deliverable shape and the