菜单
Conducts in-depth code research using a tiered tool strategy. Use when investigating codebases, researching libraries/APIs, debugging errors, or understanding unfamiliar code patterns.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/cercova-studios/terminal-agent-plugins --skill code-research命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
Routes code search requests to semantic, structural, repo-discovery, or text-search workflows. Use when the user wants to find code, inspect repository layout, or search by meaning or syntax.
Semantic code search via the `osgrep` CLI. Use when the user wants to find code by behavior, concept, or intent ("where do we validate JWTs?", "where is rate limiting handled?") rather than by exact text. Prefer ripgrep for literal matches and ast-grep for structural patterns.
Use the Doppler CLI to access and manage secrets safely in both ad hoc scripts and long-lived projects. Covers CLI install, env injection, project binding via doppler.yaml, environments, service tokens, Docker, Kubernetes, and CI/CD integration. Use when a task touches secrets and must avoid exposing them in logs, prompts, command history, or tool output.
| name | code-research |
| description | Conducts in-depth code research using a tiered tool strategy. Use when investigating codebases, researching libraries/APIs, debugging errors, or understanding unfamiliar code patterns. |
<quick_start>
Before any Bash command that uses ${AGENT_ROOT}, run:
source ./scripts/agent-env.sh
This resolves AGENT_ROOT to ./.agents in the current working directory.
When researching code, follow the tool escalation ladder:
code-search skill first; otherwise use Grep/Glob/Readgh-cli skill for GitHub, scripts for Stack OverflowStart simple. Escalate only when simpler tools fail. </quick_start>
<tool_hierarchy> Local codebase exploration:
Grep - Search for patterns, function names, error messages in codeGlob - Find files by pattern (e.g., **/*.ts, **/config.*)Read - Read specific files once you know what to look atTask with subagent_type=Explore - For open-ended codebase explorationWeb research:
WebSearch - General web search for docs, tutorials, discussionsWebFetch - Fetch and analyze specific URLs (works for most static sites)When to use: Always start here. These tools are fast, low-cost, and handle 80% of research tasks.
Fallback to POSIX tools:
grep, find, sed, awk, cut, sort, uniqTerminal web/doc workflows:
w3m / lynx for fast doc browsing!gh, !so, !npm, !pypi) to jump directly to sourcescurl + jq + rg for structured data and targeted extractionpup/htmlq/python -m bs4 for HTML parsing when neededreadability-lxml (or python -m readability) to clean article contentcsvkit / xsv for CSV docs and tablesfzf to interactively select snippets and URLspbcopy (macOS) / xclip -selection clipboard (Linux)Rate-limit hygiene:
curl --retry 3 --retry-delay 2 --compressed + backoff (sleep)ETag/If-Modified-Since to avoid refetching unchanged docsWhen to use: Quick web/CLI research before WebSearch, or when you need high-throughput data extraction.
GitHub CLI (gh) - Use the gh-cli skill for comprehensive GitHub operations:
Skill: gh-clistackoverflow-api.sh - Find solutions to errors
# Search Stack Overflow for solutions
${AGENT_ROOT}/skills/code-research/scripts/stackoverflow-api.sh "error message or question"
When to use: When you need structured GitHub data (issues, PRs, code across repos) that WebSearch can't provide cleanly.
Exa MCP - Semantic web search with AI understanding
Deepwiki MCP - Documentation and wiki content
Chrome Web Tools MCP - Headless browser automation
https://codewiki.google/github.com/anomalyco/opencodeWhen to use: When simpler tools fail—JS-rendered content, semantic search needs, or complex documentation sites. </tool_hierarchy>
<research_workflow> Classify the research type:
Glob: **/{keyword}.{ts,py,go}
Grep: "functionName|className|errorMessage"
Task(subagent_type=Explore): "Find how authentication is implemented"
</step>
<step name="2.5" label="Terminal Research">
Use fast terminal tools to browse docs and extract data:
WebSearch: "!gh repo:org/project authentication middleware"
curl -s https://docs.example.com/api | rg -n "endpoint" | sed -n '1,120p'
gh repo view org/repo --json description --jq '.description'
curl -s https://blog.example.com/post | python -m readability | rg -n "API" | sed -n '1,80p'
curl -s https://docs.example.com/table.csv | xsv select 1,3 | xsv table | sed -n '1,40p'
curl -s https://docs.example.com/api | rg -n "endpoint" | fzf
curl -s https://docs.example.com/api | pbcopy # or: xclip -selection clipboard
Prefer Rust utilities when available (rg/fd/bat/sd/xsv); fall back to standard Unix tools otherwise.
</step>
<step name="3" label="Expand to Web">
If local exploration isn't enough:
WebSearch: "library-name how to implement X"
WebFetch: "https://docs.library.com/guide"
</step>
<step name="4" label="Use Skills & Scripts for Structured Data">
When you need GitHub/StackOverflow data:
Skill: gh-cli
gh issue list -R facebook/react --search "useEffect cleanup" gh search code "useEffect cleanup" --repo facebook/react
```bash
# Find error solutions
${AGENT_ROOT}/skills/code-research/scripts/stackoverflow-api.sh "React useEffect memory leak"
mcp__web_search_exa: "best practices for React state management 2024" mcp__get_code_context_exa: "langgraph deepagent cli" mcp__crawling_exa: "https://codewiki.google/github.com/anomalyco/opencode"
mcp__deepwiki__read_wiki_contents: "react/react" mcp__deepwiki__read_wiki_structure: "anomalyco/opencode" mcp__deepwiki__ask_question: "What are the context engineering strategies used by anomalyco/opencode?"
Use chrome-devtools-mcp to open up webpages and crawl through them for deeper information extraction tasks.
</step>
<step name="6" label="Save High-Value Sources to Readwise">
When a source is highly relevant (authoritative docs, key blog posts, insightful discussions), persist it for future retrieval:
```bash
# Save a URL to Reader with research tags
readwise reader-create-document --url "https://..." --tags "code-research,<topic>"
# Create a highlight for a key finding or quote
readwise readwise-create-highlights --highlights '[{"text": "key finding text", "title": "Source Title", "source_url": "https://...", "note": "Why this matters"}]'
Only save sources scoring 2-3 on relevance (directly answers the question or provides strong supporting context). Do not save noise.
<output_formats> Adapt output to the query:
Quick answer - For simple questions, respond inline with sources Structured summary - For broader research:
## Findings
- Key finding 1
- Key finding 2
## Recommendations
- Recommended approach with rationale
## Sources
- [Source 1](url)
- [Source 2](url)
Research report - For deep dives, create a markdown file:
Write: research-{topic}-{date}.md
</output_formats>
<tool_selection_heuristics>
| Situation | Tool Choice |
|---|---|
| "How does X work in this codebase?" | Grep → Read → Task/Explore |
| "Find where code lives in this repo" | code-search skill first, then continue research if needed |
| "What's the best library for X?" | WebSearch → Exa MCP |
| "How do I use library X?" | WebFetch (docs URL) → Deepwiki MCP |
| "Why am I getting error X?" | Grep (local) → stackoverflow-api.sh → WebSearch |
| "How do other projects do X?" | gh-cli skill → gh search |
| "Show me the docs for X" | WebFetch → Deepwiki → Chrome MCP |
| "Quickly skim docs" | w3m/lynx → curl + rg/sed |
| "Find issues related to X" | gh-cli skill → gh issue/search |
| "Understand large local codebase quickly" | Contextplus (get_context_tree/get_file_skeleton) with repo-root-relative paths |
| "Search code for pattern X" | rg (local) → Exa MCP |
| "Need fast CLI search" | rg/fd/bat (fallback: grep/find/cat) |
| "Navigate to JS-heavy site" | Chrome MCP (browser_navigate + browser_snapshot) |
| </tool_selection_heuristics> |
<anti_patterns> Don't reach for MCP servers immediately. Try built-in tools and scripts first—they're faster and use less context.
<success_criteria> Research is complete when:
code-search before broader research