codeindexlisted

Answer code questions structurally instead of reading whole files. Use when you need to find where something is defined, read one function, see what calls a symbol, trace imports, judge what a change breaks, or orient yourself in an unfamiliar file or repository. Reach for this before Read, Grep or Glob on a code question — and equally before `grep`, `rg`, `cat`, `head`, `sed -n` or `find` in a shell, which is where the same question goes when a harness or permission mode routes file work through Bash. A grep finds a string; it cannot tell you what calls what. Backed by the codeindex MCP server (tree-sitter, 40+ languages).
munhq/codeindex · ★ 3 · AI & Automation · score 76

Install: claude install-skill munhq/codeindex

# Answering code questions with codeindex ## Why this matters `Read` is the single largest consumer of context in agent sessions. Measured on real transcripts: tool results dominate conversation text, `Read` dominates the tool-result tokens, and a single `Read` call averages **1,563 tokens**. Most Read calls answer a question that did not need the whole file. | tool | tokens per call | |---|---| | `Read` | 1,563 | | `get_outline` | 113 | | `find_symbol` | 102 | | `search` | 79 | | `find_word` | 41 | | `read_symbol` | **35** | The codeindex tools are named bare throughout this skill. Their live prefix depends on the install: `mcp__codeindex__` when the server is registered directly, `mcp__plugin_codeindex_codeindex__` when the plugin registers it. Take the exact name from your own tool list — a name written for the other install fails the call and sends you back to `Read`. Some harnesses defer MCP schemas: the names are in your tool list, but calling one before its schema loads fails validation. Load the schemas first, in one call — `ToolSearch` with `select:find_symbol,read_symbol,get_outline,find_word` — then call the tools. That is one extra call per session, against 1,563 tokens for every file you would otherwise read. A token you never put into context is never billed. A token you put in and later compress has already been re-sent on every intervening turn, and editing it invalidates the prompt cache. So the cheapest possible intervention is to not read the file in