MCP-first code search and analysis (cross-validated)

Use this command to answer code/navigation/impact questions by prioritizing MCP server tools over default editor tooling, then cross-validating with text/LSP search to increase precision and recall. Inputs • Query: $ARGUMENTS • Optional scope hint (directory/module/package) • Optional mode: search | navigate | impact | arch • Optional server: server=<name> (normalized MCP server name from /mcp) • Optional prompt: prompt=<name> (specific MCP tool/prompt to prefer) Execution protocol 1) Discover and prioritize MCP tools • If server=<name> is provided, restrict MCP calls to that server; otherwise, iterate through all connected servers exposing dependency graph tools and prefer those indexing libraries/packages referenced by this workspace. • Prefer these tools (by intent): • Search/navigation: search_graph, get_node_by_name, find_references, find_definition, get_dependencies, get_dependents, get_call_chain • Impact/quality: analyze_impact, analyze_code_smells, analyze_circular_dependencies, analyze_architectural_layers, get_module_metrics, analyze_external_dependencies • Similarity/patterns: find_similar_nodes, find_similar_implementations, search_by_pattern • Use narrow scope if provided; otherwise start broad, then refine by directories/languages as indicated by early results. • If prompt=<name> is provided, prefer calling that MCP tool/prompt first when applicable. 2) Cross-validate with default/editor tools • Use IDE symbol search/LSP capabilities to corroborate MCP findings (definitions, references, symbols). • Run a conservative ripgrep for the query and any top MCP-derived identifiers to catch non-indexed text: • !rg -n --hidden --no-ignore -S "$ARGUMENTS" | head -200 • If the query looks like a symbol/path, also search likely variants (case, suffix/prefix, interface/impl patterns) inferred from MCP results. 3) Merge, rank, and reconcile • Deduplicate by canonical node/file identity; prefer MCP graph semantics over plain-text hits. • Rank by: • Direct MCP matches (exact node/file) > MCP structural proximity (deps/dependents/call chains) > editor/LSP > raw text hits. • Flag disagreements (present in text-only, absent in MCP) as potential extraction gaps or dead code. 4) Produce decision-ready output • Summary: 1–3 bullets with the most relevant findings and next action. • Evidence (MCP): key nodes, relationships, call chains, impact sets, and metrics. • Evidence (default): LSP/grep confirmations and any unique findings not in MCP. • Cross-validation: overlaps, conflicts, and confidence score. • If mode=impact, include affected files/modules and suggested test focus; if mode=arch, include layer/cycle findings. Heuristics • Prefer MCP graph relationships over text proximity when signals conflict. • When MCP tools return no results, expand scope and switch to text-first, then attempt to map text hits back to MCP nodes. • For library/package queries not in the local workspace, query MCP servers that index those artifacts first; only then fall back to local search. Response format (concise) • Summary • Findings (ranked list with brief rationale) • Cross-validation (MCP vs default) • Next steps Notes • This command assumes MCP prompts/tools are available via connected servers and will be invoked before editor defaults. • Keep output focused; surface only the top-most actionable results, link deeper evidence on demand. Examples • /Q "retry policy backoff" scope=shared/utils mode=search server=depgraph-mcp • /Q "transitive deps of FooService" mode=navigate server=depgraph-mcp prompt=get_dependencies • To discover server names and prompts exposed by MCP servers, use /mcp and look for commands like /mcp__<server-name>__<prompt-name>.