KingInYellows/yellow-plugins

最近记录的来源活动: 2026年11月5日 23:19

SkillsMP 收录数据更新: 2026年11月9日 14:22

已收集 skills: 114

GitHub 星标: 0
GitHub Forks: 0

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/KingInYellows/yellow-plugins.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# yellow-plugins Personal Claude Code plugin marketplace — 22 plugins for Git workflows, code review, CI, research, testing, documentation, code editing, security remediation, and cross-lineage code council. ## Requirements - Node.js `22.22.0` or later and below `25.0.0` - pnpm `8.0.0` or later ## Install Add the marketplace, then install individual plugins: ``` /plugin marketplace add KingInYellows/yellow-plugins /plugin install gt-workflow@yellow-plugins ``` ## Plugins | Plugin | Description | Components | | --------------------- | -------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------- | | `gt-workflow` | Graphite-native workflow commands for stacked PRs, smart commits, sync, and stack navigation | 7 commands, 2 hooks, 1 MCP | | `github-workflow` | GitHub-native stacked-PR provider — full command surface (setup, status, plan, submit, amend, sync, nav, cleanup, merge) | 9 commands, 9 skills, 2 hooks | | `yellow-agy` | Local delegate on the Antigravity CLI (`agy`) under your Google AI Pro sign-in — yellow-jules successor (experimental) | 8 commands | | `yellow-browser-test` | Autonomous web app testing with agent-browser — auto-discovery, structured flows, and bug reporting | 3 agents, 4 commands, 2 skills | | `yellow-ci` | CI failure diagnosis, workflow linting, and runner health management for self-hosted GitHub Actions runners | 4 agents, 9 commands, 8 skills, 1 hook | | `yellow-claudemod` | Claude Code mod: records the model each subagent runs on and can downshift listed Opus reviewers near a usage limit (experimental) | 1 command, 1 hooks module | | `yellow-codex` | OpenAI Codex CLI wrapper with review, rescue, and analysis agents for workflow integration | 3 agents, 4 commands, 2 skills | | `yellow-composio` | Composio MCP integration with usage tracking and budget guardrails | 2 commands, 1 skill, 1 MCP | | `yellow-core` | Dev toolkit with review agents, research agents, and workflow commands for TS/Py/Rust/Go | 21 agents, 19 commands, 23 skills | | `yellow-council` | On-demand cross-lineage code review fanning out to an in-process Claude reviewer plus the Codex, Gemini, and OpenCode CLIs in parallel | 3 agents, 2 commands, 1 skill | | `yellow-cursor` | Cursor Cloud Agent delegation — launch, track, and manage remote coding agents via a typed CLI (pilot Cursor distribution target) | 10 commands, 2 skills | | `yellow-debt` | Technical debt audit and remediation with parallel scanner agents for AI-generated code patterns | 7 agents, 6 commands, 2 skills, 1 hook | | `yellow-devin` | Devin.AI V3 API integration — delegate tasks, manage sessions, orchestrate plan-implement-review chains (legacy — see yellow-cursor) | 1 agent, 9 commands, 1 skill, 1 MCP | | `yellow-jules` | Google Jules integration (deprecated: read-only after 2026-11-11, off 2026-12-31) — delegate, supervise, review sessions under grants | 10 commands, 2 skills | | `yellow-docs` | Documentation audit, generation, and Mermaid diagram creation for any repository | 10 agents, 6 commands, 2 skills | | `yellow-goal` | Process bridge to the yellow-goal `goal-gen` engine (setup/request, stub run, approval-gated real run that may spend) | 4 commands | | `yellow-linear` | Linear MCP integration with PM workflows for issues, projects, initiatives, cycles, and documents | 3 agents, 9 commands, 1 skill, 1 MCP | | `yellow-morph` | Intelligent code editing and search via Morph Fast Apply and WarpGrep | 2 commands, 1 MCP | | `yellow-research` | Deep research with Ceramic, DeepWiki, Perplexity, Tavily, EXA, and Parallel Task MCPs | 2 agents, 4 commands, 3 skills, 6 MCPs | | `yellow-review` | Multi-agent PR review with adaptive agent selection, parallel comment resolution, and stack review | 16 agents, 9 commands, 2 skills | | `yellow-ruvector` | Persistent vector memory and semantic code search for Claude Code agents via ruvector | 2 agents, 8 commands, 3 skills, 4 hooks, 1 MCP | | `yellow-semgrep` | Semgrep security finding remediation — fetch, fix, and verify "to fix" findings from the Semgrep platform | 2 agents, 5 commands, 1 skill, 1 MCP | ## Codex Distribution The private/local Codex catalog selects ten plugins and 31 skills. The six new workflow slices are worktree inventory, documentation audit, complexity scan, public DeepWiki research, offline Cursor planning and local Codex readiness. See the exact support, setup, cache-refresh and unsupported states in [Codex distribution](docs/codex-distribution.md). WSL CLI evidence does not establish Windows desktop support. Claude command/setup behavior stays separate. ## Cursor Distribution (pilot) `yellow-cursor` is also generated as a native Cursor plugin (`.cursor-plugin/plugin.json` + a listing in the root `.cursor-plugin/marketplace.json`), alongside its normal Claude Code install path. `yellow-review` is the second Cursor-enabled plugin, exposing a single read-only skill (`yellow-thermonuclear-review`) and nothing else. Cursor support is **explicit opt-in per plugin** — this repo does not claim repository-wide Cursor support. See [docs/cursor-distribution.md](docs/cursor-distribution.md) for the full opt-in model, generated-artifact shape, and verification status. ## MCP Servers & Authentication Eight plugins bundle MCP servers. Authentication requirements vary by server. | Plugin | MCP Server | Auth | | ----------------- | ---------- | -------------------------------------------------------------------------------------------------------------------------------------------- | | `gt-workflow` | Graphite | Local stdio (`gt mcp`) — requires Graphite CLI login | | `yellow-composio` | Composio | Browser OAuth on `https://connect.composio.dev/mcp` (native HTTP, no API key). Headless `claude mcp add` with a consumer key is the fallback | | `yellow-devin` | Devin | `DEVIN_SERVICE_USER_TOKEN` & `DEVIN_ORG_ID` required | | `yellow-linear` | Linear | OAuth (browser popup on first use) | | `yellow-morph` | Morph | `MORPH_API_KEY` required | | `yellow-research` | Ceramic | OAuth (browser popup on first `ceramic_search` use) | | `yellow-research` | DeepWiki | None (public repos only) | | `yellow-research` | Perplexity | `PERPLEXITY_API_KEY` required | | `yellow-research` | Tavily | `TAVILY_API_KEY` required | | `yellow-research` | EXA | `EXA_API_KEY` required | | `yellow-research` | Parallel | No API key — auto-authenticated by Claude Code | | `yellow-ruvector` | ruvector | Local stdio — no auth required | | `yellow-semgrep` | semgrep | `SEMGREP_APP_TOKEN` required | `yellow-review` bundles no MCP server but can reach one. When the optional `yellow-linear` plugin is installed, its `save_issue` tool is discoverable, and the branch name carries a Linear ID, `/review:resolve` in `yellow-review` uses `yellow-linear`'s OAuth-backed Linear MCP server for reads, marker searches and follow-up issue creation. It writes to Linear only in that case, and authentication is the `yellow-linear` login above. If any condition is missing, or a Linear call fails, follow-ups are filed on GitHub instead. See "Optional integrations" in `plugins/yellow-review/README.md`. ### Context7 (user-level MCP) Provides up-to-date library documentation for LLMs. Works without an API key but with lower rate limits. **Free (no key):** Works out of the box. The plugin connects to `https://mcp.context7.com/mcp` with no configuration needed. **Free API key (higher rate limits):** Create an account at [context7.com/dashboard](https://context7.com/dashboard) and generate an API key (format: `ctx7sk_...`). Then configure it in your Claude Code settings: ```bash # In Claude Code, run /mcp → select context7 → edit config → add header: # "headers": { "CONTEXT7_API_KEY": "ctx7sk_your_key_here" } ``` ### DeepWiki (yellow-research) Works out of the box at `https://mcp.deepwiki.com/mcp`. No authentication needed — public repositories only, no private-repo path. `/devin:wiki` (yellow-devin) still queries it as a backward-compatible fallback, discovering the tool at its current location at runtime. ### Devin (yellow-devin) Devin sessions require a `DEVIN_SERVICE_USER_TOKEN` and `DEVIN_ORG_ID`. ```bash # Add to your shell profile (~/.zshrc, ~/.bashrc, etc.) export DEVIN_SERVICE_USER_TOKEN="cog_your_token_here" export DEVIN_ORG_ID="your-org-id" # Create a service user at: Enterprise Settings > Service Users # Find your org ID at: Enterprise Settings > Organizations ``` Never commit tokens to version control. ### Linear (OAuth) On first MCP tool call, Claude Code opens a browser popup to authenticate with your Linear account. Tokens are stored in your system keychain and refresh automatically. To re-authenticate or revoke access: run `/mcp` in Claude Code, select the server, and choose "Clear authentication". With Graphite's merge queue, Linear moves an issue to Done only through commit magic words, which needs a one-time GitHub push webhook from Linear's GitHub integration. The plugin never sees that webhook's secret; the setup steps are in `plugins/yellow-linear/README.md` "Graphite Merge Queue". These plugins require browser access and **will not work in headless SSH sessions**. ### Morph (yellow-morph) Morph runs as a local MCP server via `npx` and requires a `MORPH_API_KEY`. ```bash # Add to your shell profile (~/.zshrc, ~/.bashrc, etc.) export MORPH_API_KEY="your_key_here" ``` Restart Claude Code after setting the key, then run `/morph:setup` to verify API health and tool availability. ### yellow-research (API keys) Bundles six MCP servers for multi-source deep research. Three search providers require API keys. Ceramic and Parallel use OAuth managed by Claude Code (no API key), and DeepWiki needs none. The optional `ast-grep` CLI is a local binary, not an MCP server. ```bash # Add to your shell profile (~/.zshrc, ~/.bashrc, etc.) export PERPLEXITY_API_KEY="pplx-..." export TAVILY_API_KEY="tvly-..." export EXA_API_KEY="..." ``` Source or restart your shell, then restart Claude Code — MCP servers read environment variables at startup. - **Perplexity:** [perplexity.ai/settings/api](https://www.perplexity.ai/settings/api) - **Tavily:** [app.tavily.com](https://app.tavily.com) - **EXA:** [dashboard.exa.ai](https://dashboard.exa.ai) - **Ceramic:** No API key needed for the MCP — OAuth 2.1 (browser popup on first `ceramic_search` use) - **Parallel Task MCP:** No API key needed — Claude Code handles authentication automatically - **ast-grep CLI (optional):** No API key needed — install the `ast-grep` binary locally or run `/research:setup` Plugins degrade gracefully: if a key is missing, that provider is skipped and research continues with the remaining sources. ### Semgrep (yellow-semgrep) Semgrep runs as a local MCP server and also requires a Semgrep AppSec Platform API token: ```bash # Add to your shell profile (~/.zshrc, ~/.bashrc, etc.) export SEMGREP_APP_TOKEN="sgp_your_token_here" ``` The plugin also expects the `semgrep` CLI, `curl`, `jq`, and Graphite CLI (`gt`). Run `/semgrep:setup` to validate credentials, detect your deployment slug, and verify MCP tools. ### Jev shadow pre-filter (yellow-core, optional) yellow-core's Stop hook can ask TypeSafe's Jev model whether a finished session looks worth staging for knowledge capture, and log the answer without changing what is staged. It is off unless both variables are set in the environment Claude Code's hooks inherit, and it sends redacted session text to TypeSafe: ```bash export COMPOUND_JEV_PREFILTER=shadow export TYPESAFE_API_KEY="..." ``` See `docs/security.md` "Jev Shadow Pre-Filter" for exactly what is sent. ### ruvector (yellow-ruvector) Runs locally as a stdio MCP server through the plugin's own launcher, which installs the pinned ruvector into the plugin data directory on first use (no global or `npx` install; the first run needs network for the npm packages and the ~90MB embedding model). Requires Node.js 20+. No external services or API keys required. Run `/ruvector:setup` to install ahead of time and initialize the project's `.ruvector/` store. ### yellow-goal (process bridge) `/goal:setup`, `/goal:request`, and `/goal:run-stub` stay zero-spend. The fourth command, `/goal:run-real`, is user-only: it displays the engine manifest, forwards an operator-supplied approval path, and may spend. Default `--auth-mode subscription` uses the operator Claude Code login. `--auth-mode api-key` forwards `ANTHROPIC_API_KEY` from the environment. Never commit that key. Put the pinned `goal-gen` binary on PATH, then run `/goal:setup`. ## Usage After installing, use `/plugin install <name>@yellow-plugins` to activate individual plugins. Each plugin's commands are namespaced (e.g., `/ci:diagnose`, `/linear:create`, `/devin:delegate`, `/research:deep`, `/morph:status`, `/semgrep:fix`). Run `/plugin` to browse all available plugins in the Discover tab. ## Update, Disable, Remove ``` /plugin marketplace update yellow-plugins /plugin disable <plugin-name>@yellow-plugins /plugin uninstall <plugin-name>@yellow-plugins ``` ## Local Install (Development) Clone the repo and add it as a local marketplace: ```bash git clone https://github.com/KingInYellows/yellow-plugins.git cd yellow-plugins pnpm install ``` Then in Claude Code: ``` /plugin marketplace add ./ /plugin install gt-workflow@yellow-plugins ``` Verify `${CLAUDE_PLUGIN_ROOT}` resolves correctly after local install — local installs copy plugins to `~/.claude/plugins/cache/`. ## Create a New Plugin A new plugin lives under `plugins/`: ```text plugins/my-plugin/ .claude-plugin/ plugin.json # generated — do not create by hand commands/ my-command.md CLAUDE.md package.json ``` `package.json` (`"name"` + semver `version` + `"private": true`) and `catalog/plugins/my-plugin.json` are the two files you write by hand; `pnpm generate:manifests` emits `plugin.json` and the marketplace entry from them. The full numbered procedure (catalog fields, `pluginOrder`, `setup/all.md` wiring, the lockfile/changeset/validation steps) lives in [CONTRIBUTING.md "Adding a Plugin"](CONTRIBUTING.md#adding-a-plugin) — that is the canonical checklist; `docs/plugin-template.md` has the full worked example with concrete JSON. See each plugin's `CLAUDE.md` for conventions, component details, and usage guides. ## Project Structure ``` yellow-plugins/ ├── .claude-plugin/ │ └── marketplace.json # Plugin catalog ├── plugins/ │ ├── gt-workflow/ # Graphite workflow (7 commands, 2 hooks, 1 MCP) │ ├── github-workflow/ # GitHub-native stacked-PR provider (9 commands, 9 skills, 2 hooks) │ ├── yellow-agy/ # Antigravity CLI local delegate (8 commands) │ ├── yellow-browser-test/ # Browser testing (3 agents, 4 commands, 2 skills) │ ├── yellow-ci/ # CI toolkit (4 agents, 9 commands, 8 skills, 1 hook) │ ├── yellow-claudemod/ # Claude Code mod: spawn ledger + usage downshift (1 command, 1 hooks module) │ ├── yellow-codex/ # Codex CLI wrapper (3 agents, 4 commands, 2 skills) │ ├── yellow-composio/ # Composio MCP (2 commands, 1 skill, 1 MCP) │ ├── yellow-core/ # Dev toolkit (21 agents, 19 commands, 23 skills) │ ├── yellow-council/ # Cross-lineage code council (3 agents, 2 commands, 1 skill) │ ├── yellow-cursor/ # Cursor Cloud Agent delegation, pilot target (10 commands, 2 skills) │ ├── yellow-debt/ # Debt audit (7 agents, 6 commands, 2 skills, 1 hook) │ ├── yellow-devin/ # Devin.AI, legacy (1 agent, 9 commands, 1 skill, 1 MCP) │ ├── yellow-docs/ # Documentation (10 agents, 6 commands, 2 skills) │ ├── yellow-goal/ # yellow-goal engine bridge (4 commands) │ ├── yellow-jules/ # Google Jules, deprecated, grant-gated (10 commands, 2 skills) │ ├── yellow-linear/ # Linear PM (3 agents, 9 commands, 1 skill, 1 MCP) │ ├── yellow-morph/ # Morph code editing and search (2 commands, 1 MCP) │ ├── yellow-research/ # Deep research (2 agents, 4 commands, 3 skills, 6 MCPs) │ ├── yellow-review/ # PR review (16 agents, 9 commands, 2 skills) │ ├── yellow-ruvector/ # Vector memory (2 agents, 8 commands, 3 skills, 4 hooks, 1 MCP) │ └── yellow-semgrep/ # Semgrep remediation (2 agents, 5 commands, 1 skill, 1 MCP) ├── packages/ # Validation tooling (domain, infrastructure, cli) ├── schemas/ # JSON schemas └── docs/ # Validation guides, operational docs, and solutions ``` ## Troubleshooting ### `claude doctor` says "descriptions dropped" If `claude doctor` shows a warning like `157 descriptions dropped (4.9%/1% of context)`, your skill listing budget is too small for the total skill descriptions installed across all your plugins. This is a **per-user Claude Code setting**, not a plugin problem — yellow-plugins skill descriptions are well under the official 1,536-char per-skill cap (per [code.claude.com/docs/en/skills](https://code.claude.com/docs/en/skills)). The default budget is 1% of the context window, with an 8,000-character fallback. Two ways to give Claude Code more room: 1. **Raise `skillListingBudgetFraction`** — edit `~/.claude/settings.json`: ```jsonc { "skillListingBudgetFraction": 0.04, } ``` `0.04` (4%) is enough to fit the full yellow-plugins marketplace plus typical adjacent installs, at the cost of roughly 8K extra characters reserved from the per-session context window. Note: this setting is community-discovered (decompiled from Claude Code 2.1.129) and not yet in the official docs — it may be renamed before public documentation. 2. **Set `SLASH_COMMAND_TOOL_CHAR_BUDGET`** as an environment variable for a one-off raise without editing settings. Default is 15,000 characters; set higher (e.g. `SLASH_COMMAND_TOOL_CHAR_BUDGET=40000`) to fit larger skill listings. (Disabling skills via `/skills` or `skillOverrides` is intentionally omitted: those knobs apply to user-scoped skills, not plugin-provided skills, and would not reduce yellow-plugins' contribution to the budget.) The warning is informational — Claude Code drops the lowest-priority descriptions first. Critical skills you use frequently stay listed. ## License MIT

正在加载仓库 README…

这个仓库中的 skills

软件开发工程师软件质量保证分析师与测试员其他计算机职业网络与计算机系统管理员

6 个职业分类 · 已分类 47%

未分类

原文语言：英语

2026-10-24

未分类

原文语言：英语

2026-10-23

ci-diagnose

未分类

原文语言：英语

2026-10-23

ci-runner-health

未分类

原文语言：英语

2026-10-23

jules-supervision

未分类

Host-neutral reference for supervising a Google Jules session with one bounded pass at a time: the decision table, what each decision allows, pause handling, and the correction limits. Use when you are asked to keep a Jules session moving inside a grant, or…

原文语言：英语

2026-10-23

未分类

Audit local repository documentation for missing coverage, stale claims and structure problems. Use when assessing doc health; return findings and recommendations without editing files.

原文语言：英语

2026-10-13

已展示 40 / 114 个已收集 Skill。