Skip to content

README.md

Latest commit

 

History

History
124 lines (85 loc) · 6.32 KB

README.md

File metadata and controls

124 lines (85 loc) · 6.32 KB
Raw

qdev (retired)

Retired marketplace plugin. qdev is not listed in .claude-plugin/marketplace.json and cannot be installed from this marketplace. Its v2.0.6 source remains for reference and local inspection only.

Deep web research before you design or build.

Summary

The source implements one focused command: /qdev:research runs a structured, dual-source web sweep and persists a cited report for design, planning, or review. It is user-initiated and never fires contextually.

Principles

[P1] Research Before Analysis: Live sources beat training data. /research is a first-class command meant to run before design work begins; no decision rests on recall alone when a current source can be consulted.

[P2] Explicit Invocation Only: /research fires only when you call it. There is no auto-trigger and no background search — routine, automatic lookups are a separate concern owned by the web-search skill.

[P3] Persisted, Deduplicated Knowledge: Every sweep writes a frontmatter-tagged report under docs/research/ and regenerates docs/research/index.md; the dedup cycle updates, links, or supersedes overlapping prior reports rather than piling up duplicates.

Requirements

  • Claude Code (any recent version)
  • tavily MCP server (primary recall + JS-rendered page extraction; the agent fails soft to Brave/Serper when absent)
  • brave-search MCP server (cross-check recall source)
  • serper-search MCP server (Google-operator queries: site:, filetype:)
  • context7 MCP server (recommended; library/framework documentation gating)

Local inspection

The retired plugin can be loaded from this source tree for inspection or development:

claude --plugin-dir ./plugins/qdev

How It Works

flowchart TD
    A["/qdev:research [topic]"] --> B["Infer topic<br/>(args or context)"]
    B --> C["Dispatch qdev-researcher (Sonnet)"]
    C --> D["Topic kind?"]
    D -->|library| E["Context7<br/>resolve-library-id + query-docs"]
    D -->|pattern/topic| F["Skip Context7"]
    D -->|mixed| E
    D -->|mixed| F
    E --> G["Tavily-first recall<br/>6-8 queries"]
    F --> G
    G --> G2["Cross-check claims<br/>brave + serper (Google operators)"]
    G2 --> H["Tavily extract<br/>3-5 highest-signal pages"]
    H --> I["Corroboration check<br/>(footguns: 2+ sources)"]
    I --> J{"Coverage thin<br/>or open Qs?"}
    J -->|Yes| K["One follow-up pass"]
    J -->|No| L["Synthesize + grade sources"]
    K --> L
    L --> M["Persist report<br/>docs/research/[date]-[slug].md"]
    M --> N["Return structured report"]
Loading

Usage

# Research a technology or topic before starting design work
/qdev:research "Redis pub/sub with Python"

# Research from mid-session context (infers topic automatically)
/qdev:research

Commands

Command Description
/qdev:research Dual-source research sweep covering docs, practices, footguns, existing tools, security, and recent changes (dispatches qdev-researcher); persists a report under docs/research/

Agents

Agent Model Purpose
qdev-researcher Sonnet Tavily-first research with Brave/Serper cross-checks, Context7 docs gating, footgun corroboration (2+ sources), and a single follow-up pass for thin angles. Persists a structured report under docs/research/.

/qdev:research [topic]

Research a topic, technology, or problem space before designing or building, by dispatching the qdev-researcher subagent. Pass the topic as an argument, or invoke without arguments to have it inferred from project context and conversation history.

Coverage:

  • Official documentation (current API, recent changes)
  • Community best practices (established patterns, what has replaced older approaches)
  • Footguns and gotchas (2+ source corroboration required; single-source items demoted)
  • Existing tools (alternatives and prior art; avoid building what already exists)
  • Security and compatibility (CVEs, deprecations, advisories)
  • Recent changes (breaking changes, ecosystem shifts since the model's cutoff)

Output: A structured Markdown report persisted to docs/research/<YYYY-MM-DD>-<slug>.md. The file starts with project-standards research frontmatter, and the returned header includes the canonical path; downstream skills consume the artifact by reading that path rather than re-running the sweep.

Depth tiers: quick (3-4 queries), standard (6-8, default), thorough (12-15). For library/framework topics, the agent routes documentation queries through Context7 before falling back to web search.

Research reporting cycle

qdev-researcher treats docs/research/ as a small knowledge base, not a loose artifact pile. Reports carry project-standards research frontmatter; docs/research/index.md is regenerated from that frontmatter by scripts/build_research_index.py; scripts/validate_research_frontmatter.py checks the scoped corpus. Before writing a new report, the agent preflights the index, uses scripts/dedup.py to choose update vs new-with-related vs supersede, writes/validates the report, and regenerates the index.

When to use /qdev:research vs other tools

You want to Use
Research before design — output feeds superpowers:brainstorming /qdev:research
Look up a specific library API quickly Context7 directly

/qdev:research is opinionated for development decisions: fixed coverage angles, footgun corroboration, and frontmatter/index-backed persistence under docs/research/.

Handoff protocol

qdev-researcher writes its report to docs/research/<YYYY-MM-DD>-<slug>.md. Downstream skills consume the artifact by referencing that path:

  • superpowers:brainstorming: feed the report's Open Questions into the design conversation.
  • feature-dev:feature-dev: start architecture work with the report linked from the brief.

Reports are not auto-cleaned. The dedup cycle updates, relates, or supersedes overlapping research; stale reports can still be removed manually when they are no longer useful.

Links