github-explore
面向 AI 编程代理的 gh CLI 探索与管理封装。以标准代理技能、Agent Plugins 1.0 插件（plugin.json + skills/）和 dsh 插件形式提供（dsh.bundle.patch 在运行时注册该技能）。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Fectivnfy112357/github-explore#e362c8c43b663083d496357c21ec5fb996fd5359Github Explore 以 github-explore 发布，当前版本为 3.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Discovery + management wrappers around the
ghCLI for AI coding agents.
github-explore is an agent skill that turns "search GitHub for X" into structured, deduplicated, relevance-scored output. It wraps gh search and gh repo view with smart filters, semantic multi-axis exploration, and layered output designed to keep an agent's context window small.
When you ask an agent "find multi-agent collaboration repos", you don't want a star-sorted dump of ollama, langchain, and a bunch of unrelated generic LLM frameworks. You want the canonical anchors (crewAI, autogen, MetaGPT, langgraph, camel, ChatDev, AutoGPT) surfaced first, with the protocol layer (A2A, ANP, ag-ui) as a separate axis, and awesome-* lists pushed to the bottom. That's what this skill does.
Plain gh search has three structural problems for agent-driven research:
"multi-agent" returns ollama (180k★) and langchain (140k★) on top because GitHub sorts by popularity, not topical fit.gh search repos --json returns full bodies, dates, and license objects per repo. Piping 50 of these into an LLM wastes thousands of tokens.github-explore addresses all three with a thin layer of Python around gh.
explore.py lets the agent define 2-4 semantic axes per topic, runs them in parallel, and unions results with a relevance score that combines cross-axis hits, canonical anchor recall, and awesome-list signals.fullName, and renders in a layered markdown summary (~3KB stdout).%TEMP%/gh-explore-{topic}-{ts}.md automatically; the agent reads the summary, and pulls the file only when it needs more detail. Default exploration drops your context from ~18KB to ~2KB.python scripts/<script>.py --schema prints the output JSON structure for the three scripts that support it (find_repos, explore, repo_summary), backed by the schema files in skills/github-explore/scripts/schemas/. The other scripts' JSON mirrors gh search's native camelCase fields (documented in skills/github-explore/references/commands-search-format.md).gh search or gh repo view. You can drop the skill and run the same gh commands by hand; the value is in the filter, dedup, and relevance scoring.# 1. Install the skill (Claude Code, Codex, Cursor, and 15+ agent CLIs)
npx skills add Fectivnfy112357/github-explore
# Hermes Agent users
hermes skills install https://raw.githubusercontent.com/Fectivnfy112357/github-explore/main/skills/github-explore/SKILL.md --force
# 2. Make sure gh CLI is authenticated
gh auth status
# 3. Try it — scripts live under the installed skill dir (e.g. ~/.claude/skills/github-explore/)
cd ~/.claude/skills/github-explore
python scripts/find_repos.py "vector database" --language python --min-stars 500
python scripts/explore.py "multi-agent" \
--axis "framework|multi-agent framework in:readme; collaborative agents in:readme" \
--axis "protocol|A2A agent protocol in:readme; agent-to-agent communication in:readme"
The same repo is also a dsh profile bundle: its package.json declares
dsh.bundle.patch, so it installs like any other dsh plugin and registers the
skill at runtime (no manual file copying). dsh plugin forwards its argument
to pnpm, so any pnpm spec works — shortest forms first:
# GitHub shorthand (no publish needed) — shortest
dsh plugin --profile web add Fectivnfy112357/github-explore
# Full git URL
dsh plugin --profile web add git+https://github.com/Fectivnfy112357/github-explore.git
# Bare npm name — after this package is published (npm publish)
dsh plugin --profile web add github-explore
# Local checkout / path / tarball (same flow)
dsh plugin --profile web add /path/to/github-explore
After a profile restart, the github-explore skill appears in the agent's
catalog — the plugin (lib/index.js) parses
skills/github-explore/SKILL.md and registers it with ctx.skills, with
resourceBase pointing at the skill directory so the skill body's
scripts/ / references/ paths keep working. Remove with
dsh plugin --profile web remove github-explore.
The repo is also an Agent Plugins 1.0 package:
plugin.json declares the manifest and the skill lives (self-contained, with
its scripts/ and references/) at the fixed skills/github-explore/
location. Any Agent Plugins 1.0-compatible client (ChatGPT, Codex, Cursor,
GitHub Copilot, Kiro, VS Code, …) can load it directly from the repo
(https://github.com/Fectivnfy112357/github-explore) or from a locally
checked-out copy — no extra build step:
git clone https://github.com/Fectivnfy112357/github-explore.git
# point your client at the repo root: plugin.json + skills/github-explore/SKILL.md
One repo, three install paths — npx skills add (standard skills),
dsh plugin --profile web add (DeepSeek Harness), and any Agent Plugins 1.0
client all read the same files.
Each command writes a layered markdown summary to stdout (~3KB) and a full report to a temp file. Pass --format json for machine-readable output (explicit; piping does not auto-switch).
A single repo serves all three packaging formats; the skill is self-contained under the Agent Plugins fixed location so it works identically no matter which installer copies it:
github-explore/
├── plugin.json # Agent Plugins 1.0 manifest ($schema + name required)
├── skills/
│ └── github-explore/ # the one skill, fully self-contained
│ ├── SKILL.md # skill body (frontmatter: name/description)
│ ├── scripts/ # 9 entry scripts + _lib.py + schemas/
│ └── references/ # gh command references (commands-*.md)
├── package.json # dsh plugin (dsh.bundle.patch) + npm metadata
├── cordis.patch.yml # dsh loader patch (inserts the skill entry)
├── lib/index.js # dsh plugin: registers the skill via ctx.skills
├── README.md / README_zh.md
└── LICENSE
plugin.json + skills/<name>/SKILL.md (+ optional mcp.json).package.json → dsh.bundle.patch → cordis.patch.yml → lib/index.js.npx skills add discovers skills/<name>/SKILL.md and installs the whole
skill directory (scripts + references included).| Script | Purpose | --schema | Notes |
|---|---|---|---|
find_repos.py | Smart repo search with multi-dimensional filters | ✅ | Default entry point. Multi-word free-text runs dual-scope (in:readme + default) for conceptual recall. |
explore.py | Multi-axis topic exploration | ✅ | Agent defines axes inline. Outputs canonical anchors + cross-axis hits + top 5/axis. |
discover.py | Auto-expand topics from a seed search | ❌ | Reads top seed results, extracts their topics, runs per-topic searches. Fast, opportunistic. |
trending.py | Time-windowed trending repos | ❌ | Default 7d window; supports --topic, --language, --min-stars. |
repo_summary.py | Deep dive on a single repo | ✅ | Topics, languages, recent activity, mentionable users, license. |
find_similar.py | Alternatives to a given repo | ❌ | Cross-language option (--no-language). |
code_search.py | GitHub code search by pattern | ❌ | --repo, --org, --owner, --extension, --filename. |
search_issues.py | Issue/PR search | ❌ | --state, --type, --label, --author, --assignee. |
org_landscape.py | Audit an entire org | ❌ | --group-by {language,topic,activity,stars}. |
_lib.py | Shared helpers | n/a | ensure_auth, gh_json, parse_since, print_schema. Not for direct use. |
__init__.py | Module docstring | n/a | Documents the scripts package. |
--schema gap: 6 of 9 scripts don't yet expose --schema as a CLI flag. The three scripts that do — find_repos, explore, repo_summary — cover the most-used paths, backed by repo.schema.json / explore.schema.json / repo_summary.schema.json.
┌─────────────────────────────────────────────┐
│ Agent (LLM, coder, etc.) │
│ - reads SKILL.md for trigger + protocol │
│ - decides which script + which axes │
└──────────────────┬──────────────────────────┘
│ python scripts/<name>.py [args]
▼
┌────────────────────────────────────────────────────┐
│ skills/github-explore/scripts/ │
│ (9 entry points + _lib + __init__) │
│ ─────────────────────────────────────────────────│
│ find_repos explore discover trending │
│ repo_summary find_similar code_search │
│ search_issues org_landscape │
│ │
│ shared: _lib.ensure_auth, _lib.gh_json, │
│ _lib.print_schema, _lib.parse_since │
└──────────────────┬─────────────────────────────────┘
│ subprocess.run(['gh', ...])
▼
┌────────────────────────────────────────────────────┐
│ gh CLI (search repos / repo view / search code) │
│ Authenticated via gh auth status. │
└──────────────────┬─────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────┐
│ GitHub REST + Search API │
│ ~5000/hr core / ~30/min search (authenticated) │
└────────────────────────────────────────────────────┘
Output:
- stdout: ~3KB layered markdown summary (default)
- stdout: full JSON when --format json (explicit)
- disk: %TEMP%/gh-explore-{topic}-{ts}.md (always)
Two layers, one mental model. Scripts handle discovery (search / dedup / score / render). Direct gh calls handle management (create / update / merge / label / workflow). The references/commands-*.md files document the management side without bloating SKILL.md.
| Task | Tool |
|---|---|
| Find repos about a topic | find_repos.py "<query>" |
| Map a field's full landscape | explore.py "<topic>" --axis ... |
| Auto-expand into related topics | discover.py "<seed>" |
| See what shipped recently | trending.py --window 7d |
| Read up on one repo | repo_summary.py owner/repo |
| Find alternatives | find_similar.py owner/repo |
| Where is this pattern used? | code_search.py "<pattern>" --org ... |
| Search issues / PRs | search_issues.py "<query>" |
| Audit an org | org_landscape.py <org> |
| Create a repo, open a PR, label, run CI | gh <command> (see references/commands-*.md) |
These are the non-obvious calls the skill makes, surfaced so you don't have to reverse-engineer them:
in:readme is the default for multi-word free text. Description is too short to disambiguate topics. find_repos runs two scopes and unions, with a relevance bonus for in:readme hits to keep canonical small projects above generic big repos.--exclude is a post-filter, not a query token. GitHub's -term exclusion is unreliable for awesome lists and tutorials; this skill filters at the merge stage on fullName / description substrings.awesome-* directories are tagged ☰list and heavily demoted, not deleted. They're a different artifact (curation vs. code) and should appear below real projects but still be findable.explore.py orders by a relevance score that combines canonical-anchor recall, cross-axis hits, and a log-scaled star count. A 100★ canonical anchor always beats a 200k★ repo that merely mentions the topic.Issues and pull requests are welcome. This is a personal skill that's been refactored over multiple real uses; the test surface is the scripts themselves, not a unit-test suite.
Before opening a PR:
python scripts/<name>.py --help and (where supported) --schema.skills/github-explore/scripts/schemas/.MIT. See LICENSE.
Built and maintained by 贾晓源 (@Fectivnfy112357).