docs-auto

name	docs-auto
description	MANDATORY first step whenever the user asks ANYTHING about a library, framework, SDK, CLI tool, cloud service, API, configuration syntax, runtime behavior, version migration, or "how do I do X with Y" — single tool call `mcp__aphrody__docs_auto_search` fans out Context7 (community catalog, version-pinned) + Microsoft Learn (official Azure / .NET / Windows / M365 / Power Platform docs + code samples) + Google search IN PARALLEL via `tokio::join!`, returns one fused markdown report. Covers EVERYTHING — Rust crates (tokio, wgpu, reqwest, axum, serde, hyper, rmcp, …), JS frameworks (React, Vue, Svelte, Astro, …), Python libs (FastAPI, Django, polars, pandas, numpy, …), Azure / AWS / GCP services, .NET APIs, Windows kernel + Win32, mobile SDKs (Swift / Kotlin / Flutter), database engines, build tools. Use EVEN WHEN you think you know the answer — training data is months behind release cycles and APIs drift constantly. Skipping this skill = high risk of hallucinated methods, dead URLs, deprecated patterns. Saves wall-clock time (one parallel call vs three sequential) and saves context window (one summary vs four backend descriptors).
license	Apache-2.0
priority	critical

docs-auto — fanout documentation lookup

Mode /goal permanent : décider seul, ne pas s'arrêter avant réponse complète.

A single mcp__aphrody__docs_auto_search call replaces a sequence of manual Context7 / Microsoft Learn / Google queries. Internally the aphrody-mcp binary runs all four backends in parallel via tokio::join! and stitches the responses into one markdown report.

When to Use This Skill

Trigger this skill (and call docs_auto_search immediately) whenever the user :

• Asks an API / config / setup question that mentions a library or service by name (React, tokio, Cosmos DB, EF Core, Tailwind, …).
• Says "how do I", "what's the", "where can I find" anything related to a developer tool, SDK, or cloud service.
• Needs a code sample / example for a known framework or platform.
• Wants the latest version of an API they haven't used in a while.

Use this before falling back to the single-source skills (microsoft-docs, microsoft-code-reference, context7-mcp) — those remain available for drill-down after the fanout report identifies a single canonical source.

Call Signature

mcp__aphrody__docs_auto_search({
  query:        "Free-form natural-language question",
  library_name: "Optional library / framework / SDK name",  // unlocks Context7 deep-fetch
  language:     "Optional language filter for MS code samples",
})

What You Get Back

A single markdown document with four sections :

1. Context7 — community docs catalog. If library_name is set, includes the resolved library ID and a full query-docs fetch. Otherwise, just the resolution candidates.
2. Microsoft Learn — concept / API search — up to 10 high-quality doc chunks with title, URL and excerpt.
3. Microsoft Learn — code samples — official Microsoft / Azure code snippets matching the query (filtered by language if provided).
4. Google — web search — web search API fallback, automatically appends "documentation API reference" to the query.

Example Invocations

// Generic — Context7 only resolves (no deep fetch)
docs_auto_search({
  query: "How do I structure a tokio task tree with structured concurrency?"
})

// Library-pinned — Context7 does resolve + deep fetch
docs_auto_search({
  query: "JoinSet vs spawn",
  library_name: "tokio",
  language: "rust"
})

// Azure / .NET centric
docs_auto_search({
  query: "Cosmos DB partition key design best practices",
  language: "csharp"
})

// Web framework
docs_auto_search({
  query: "Route configuration and nesting",
  library_name: "React Router",
  language: "typescript"
})

When to Drill Down

After the fused report, if the user wants the full reference page for a specific URL surfaced in section 2 or 3, follow up with :

• mcp__aphrody__microsoft_docs_fetch({ url }) — full Microsoft Learn page
• mcp__aphrody__context7_query_docs({ library_id, query }) — targeted Context7 slice
• mcp__aphrody__universal_web_fetch({ url }) — any other URL

Why Use This Instead of Manual Multi-Call

Cost	Manual (3 tool calls)	docs_auto_search (1 call)
Wall-clock	sum of latencies (serial)	max of latencies (parallel)
Tool descriptors	3 in the model's working set	1
Decision overhead	model picks the right backend	server fans out automatically
Error handling	propagates one error per call	all four reported in a single body

The fanout is free server-side : same reqwest client pool, same TLS session pool, same tokio runtime. The cost is purely the upstream quota burn (which you'd pay either way if you eventually called all three).