code-tool-router

# Code Tool Router ## What it does Routes code work across the right tool: | Intent | First tool | Verification | | --- | --- | --- | | Exact string, config key, env var, error text, symbol spelling | `rg` | Open matching files directly | | Broad semantic discovery, "where is X implemented?", similar code | Semble | `rg` for exact terms, then file reads | | Definition, references, rename, signature change, type-aware navigation | Serena | `rg` for generated/dynamic callsites | | Callers, callees, complexity, dependency graph, architecture report | CGC | Serena or `rg` before edits | | Project decisions, product facts, user preferences, long-term notes | Existing project memory, when available | Source documents or current code | | Current library/API usage | Context7 or official docs | Local code examples | ## When to use Use this skill when a task requires finding code, understanding ownership, tracing behavior, planning a refactor, or comparing implementation patterns. The authored skill contract names Codex and Claude, and plugin targets mirror the same guidance into Cursor, Gemini, and OpenCode outputs. ## How to run Classify the question before searching, then use the cheapest reliable tool first. Exact search: ```bash rg -n "INTERVIEW_MEMORY_MODE" . rg -n "reset_session\\(" . rg --files | rg "memory|context" ``` Semantic discovery: ```bash uvx --from semble==0.2.0 semble search "where interview memory context is assembled" . uvx --from semble==0.2.0 semble search -k 10 "copy helper browser focus integration" . uvx --from semble==0.2.0 semble find-related src/path/file.ts 42 . ``` Graph and architecture: ```bash cgc stats . --context "$(basename "$PWD")" cgc find name reset_session --context "$(basename "$PWD")" cgc analyze callers reset_session --context "$(basename "$PWD")" cgc analyze calls assemble_stream_context_or_fallback --context "$(basename "$PWD")" cgc analyze complexity --limit 20 --context "$(basename "$PWD")" cgc report --output /tmp/CGC_REPORT.md --context "$(basename "$PWD")" ``` For Serena, use LSP-backed definitions, references, symbols, diagnostics, rename, and signature tools. If Serena exposes an initial-instructions tool, read it before using Serena in a new session. ## If optional tools are missing Do not block the task just because a specialized tool is unavailable. - If `rg` is missing, use the fastest available local text search and mention the fallback. - If Semble is missing, use `rg` to build a shortlist, then read the most relevant files directly. - If Serena is missing, use `rg` for definitions/references and avoid automated rename/refactor claims. - If CGC is missing or not indexed, use `rg --files`, source reads, and language tooling where available. - If project memory is missing, rely on source documents and current conversation context. Do not invent remembered facts. - If tool availability is unclear, run the read-only `code-intelligence-doctor` skill when available. ## Constraints - Use `rg` first for exact text. It is still the fastest and most trustworthy tool for literals. - Use Semble as a high-quality shortlist, not as complete impact analysis. - Use Serena for type-aware navigation and safe symbol edits. - Use CGC for graph questions, but confirm before changing behavior. - For HackInterview, keep CGC in tree-sitter mode by default. Previous local benchmarking showed SCIP mode missed important TypeScript/Rust caller and callee relationships. - Use memory for durable facts and decisions. Do not use it as a code index. If memory conflicts with source code, source code wins.