| name | tech-research |
| description | Технический ресёрч библиотек и кода через MCP-lookup (Context7/DeepWiki/Grep). Покрывает: исследуй интернет, documentation, best practices. Используй для: изучи библиотеку, исследуй фреймворк, найди официальную доку, найди production-паттерны кода, сравни реализации, разбери архитектуру, миграция версий, что изменилось. EN triggers: library API lookup, framework research, official docs, production code patterns, compare implementations, repo architecture, migration guide, breaking changes, since version, how does X library work. |
Tech Research
Auto-invoked when the user needs authoritative technical information before deciding on or implementing something. Routes the question to the right MCP transport rather than generic web search.
When to Use
- Library or framework API lookup ("how does X library handle Y in version Z?")
- Open-source repo architecture, design notes, module layout
- Real-world code patterns and idioms across GitHub at scale
- Comparing implementation approaches between several libraries
- Verifying potentially-stale documentation against current code
Skip when: question is news / current events (use web-research), or it's purely about local code (use Serena tools).
Workflow
- Pick the right transport for the question:
- Official docs / API reference -> Context7 (
context7_resolve-library-id then context7_get-library-docs)
- Open-source repo architecture, design rationale -> DeepWiki (
deepwiki_*)
- Real-world code patterns / production usage at scale -> Grep MCP (
grep_*)
- Query the chosen MCP first. Pull exact quotes, file paths, version numbers, section anchors.
- If the claim is critical or contested, cross-validate against at least one other source (often: docs claim vs code reality).
- Synthesize with citations: source name + version + section / file path.
Output Style
- Tag each finding with confidence: Confirmed / Strongly supported / Inferred
- Cite exact source -
<library>@<version>, repo path, doc section
- Distinguish documentation claims from real-world code observations
- Flag version-specific behavior with explicit version numbers
- Reply in Russian when the user wrote in Russian
Anti-patterns
- Single-source claims for critical decisions
- Mixing inferences with confirmed facts without explicit labels
- Ignoring contradictions between docs and code - surface them, don't hide
- Using
web-research when an MCP server has the authoritative answer
- Generic "best practice" lists without dated, named sources
