CORE CONTENT

• Step 0 version gate is best-effort, cached for 24 hours, never blocks the task, and never downgrades a newer local Skill.
• Project-mode gate distinguishes greenfield architecture discovery from existing-project analysis before product search.
• The greenfield architecture interview records first-level Category context separately from Platform, Runtime, Deployment, and Ecosystem facets.
• Deep guidance (interview, architecture map, SkillHub protocol) loads on demand from public reference URLs.
• Required Capabilities and Preferred Capabilities are extracted only after architecture answers and constraints are confirmed.
• Candidates are ranked by required coverage, preferred coverage, relevance, and official primary evidence.
• Research uses staged retrieval (decision gate, capability resolution, candidate recall, relation expansion, and detail hydration); inspect meta.stages and preserve partial candidates when a later stage degrades.
• Capability coverage normalizes human requirements to canonical aliases (for example ssg, mdx, vue, image-optimization, and llms-txt) before ranking; inconsistent empty coverage triggers one bounded alias retry.
• A narrowly indexed sub-product is not treated as evidence that its parent framework is indexed; missing parents are flagged and sent to official external search.
• Candidate coverage is gated by the candidate's own matched capability tags; API coverage fields, keyword recall, and relevance scores cannot satisfy a capability, and false positives are explicitly rejected.
• Research and Search responses expose query alignment so unresolved requirements stay uncovered and product-specific features remain separate from reusable capabilities.
• Unknown AI-extracted capability terms are stored as review candidates instead of being promoted to the public verified taxonomy.
• Registry failures degrade to targeted npm keyword and optional GitHub topic fallback with an explicit registry-unavailable note.
• When an MCP host is available, the Skill can use the read-only reuseio-mcp tools while preserving the same skip conditions, evidence boundary, and decision rationale.
• Category is a first-level browse/navigation taxonomy; Platform, Runtime, Deployment, and Ecosystem remain parallel tag facets and never replace Capability evidence.
• Every recommendation follows Requirement → Capability → Evidence → Decision; small decisions may use a compact form.
• npm keyword, GitHub topic, and SkillHub results remain external discovery leads until verified.

CORE CONTENT

• researchTask accepts requiredCapabilities, preferredCapabilities, force, and external fallback options.
• Simple code changes are skipped locally; intentional ambiguous tasks can use force: true.
• Results expose required/preferred coverage, official source evidence, and decisionRationale.
• Targeted npm keyword and optional GitHub topic fallback searches run in parallel and are explicitly marked verified: false.
• The package exports compareVersions and getSyncDecision for the official version gate.
• MCP tool descriptions preserve the same Category versus parallel facet boundary used by the Skill.
• The read-only reuseio-mcp binary exposes registry search, research, manifests, sources, AI prompts, and official versions over MCP stdio.