garbell

分类: 工具与效率 | 上传者: rberenguel | 下载: 0 | 版本: v1.0（最新）
Code indexing and search skill using the purely local, daemonless Cercle v2 ("garbell") tool. Use when you need to navigate a codebase, map out files, locate usages, or extract code blocks directly from the source using ripgrep under the hood.
更新日志: Source: GitHub https://github.com/rberenguel/garbell
目录结构

当前层级: skills/
📄 SKILL.md 9.0 KB
SKILL.md

---
name: garbell
description: Code indexing and search skill using the purely local, daemonless Cercle v2 ("garbell") tool. Use when you need to navigate a codebase, map out files, locate usages, or extract code blocks directly from the source using ripgrep under the hood.
allowed-tools: Bash(garbell *)
---

# Garbell Context Retrieval

`garbell` is a completely local, zero-dependency Go binary built around `ripgrep` (`rg`).
Output is pure, compact text—designed specifically for LLM contexts.
_Note on Reindexing: Reindexing is extremely fast. If the source code changes, you must re-run `index` instantly to avoid stale data._

## Tools / Commands

Invoke these via `~/.garbell/garbell <command> [args...]` from the root of a project.

- `index` — Traverses the codebase (respecting `.gitignore`), parses Go, Python, JS/TS, C++, CSS, HTML, Markdown (`.md`/`.mdx`), and Protobuf (`.proto`), and generates JSON chunk map shards in `~/.garbell/indexes/`. Also builds a **PPMI thesaurus** (`ppmi.json`) — a repository-specific map of co-occurring terms — used by `search-related`. Re-run this whenever you modify the codebase.
- `search-lexical <query> [--file <regex>]` — Full-text search using `rg`. Evaluates the query implicitly as **PCRE/Rust regex** (so `generate|tint` works! **Do NOT escape pipes like `\|`**). Returns a **compact grouped-by-file list** of matching chunks (`start-end: sig`). **If more than 80 chunks match, returns a directory-grouped overview instead** — narrow your query or use `file-skeleton` to drill in. Use `read-chunk` to read the body of any listed chunk. `--file` restricts results to files whose path matches the regex (e.g. `--file '.*_test\.go'` for test files only).
- `search-fuzzy <signature>` — Fuzzy searches across the entire vocabulary of the chunk map using pure-Go Levenshtein distance. Capped at **100 lines** of output. Use this if you know a symbol name but aren't sure of the exact spelling or capitalization.
- `file-skeleton <filepath|dir> [--file <regex>]` — Returns a structural view of a file (line numbers + signatures) or, if given a directory, the skeletons of all files inside. **If the output would exceed the line threshold, returns a directory summary with symbol counts instead** — use that to pick a subdirectory to drill into. In directory mode, `--file` restricts results to files whose path matches the regex.
- `read-chunk <filepath> <line_number> [--unsafe]` — Reads exactly the code block enclosing a specific line number. Capped at **1000 lines**. Use this after `search-lexical`, `search-related`, or `file-skeleton` to read the full body of a specific chunk. Pass **`-1` as the line number** to read the entire file: if the file exceeds the safe line limit, a warning is returned instead — add `--unsafe` to force the full read (but consider your native file-reading tool for very large files).
- `peek <filepath> <line_number> [radius]` — Shows `radius` lines above and below `line_number` from the raw source file (default radius: **5**). The target line is marked with `>>`. No index required — works directly on the source. Use this when a lexical match is inside a very long function and you only need local context, not the whole chunk.
- `find-usages <symbol> [--file <regex>]` — Uses `rg -w` to find usages of a symbol and returns only the calling function signatures. **If there are too many callers, returns a directory-grouped count instead.** `--file` restricts callers to files whose path matches the regex.
- `extract-interface <filepath>` — Extracts only the imports/includes and exported declarations (`func`, `export`, `#include`, `def`, `class`) for a file. Perfect for understanding file contracts.
- `search-signature <pattern> [--file <regex>]` — Searches chunk signatures by regex. No file I/O — purely against the index. Use to find functions by shape: `search-signature "func.*Handler"`, `search-signature "class.*Service"`. `--file` restricts to files matching the regex.
- `largest-chunks [n] [--file <regex>]` — Returns the n largest chunks by line count (default 10). First thing to run on an unfamiliar codebase to identify where complexity lives. `--file` restricts to files matching the regex.
- `callees <filepath> <line_number>` — Returns function names called within the chunk at this line. Callees defined in this codebase are annotated with their location (`→ file:start-end`). Heuristic — accurate for common call patterns.
- `dependents <filepath>` — Finds source files that import or reference this file. Run before refactoring or deleting a file to know what will break.
- `search-related <query> [--file <regex>]` — **Concept-first search.** Tokenises the query, expands it with related terms from the PPMI thesaurus built during `index`, then returns a compact chunk list ranked by relevance (original-query matches first, synonym matches after), capped at 50. Use this when you know _what_ you're looking for but not _what it's called_ — e.g. `search-related "authentication"` may surface code named `login`, `session`, `token`, or `jwt`. Requires `ppmi.json` (produced by `index`). `--file` restricts results to files matching the regex. Use `read-chunk` to read specific chunks.

## The `--file` Flag

Several commands accept `--file <regex>` to restrict results to files whose **relative path** matches the regex. This is order-agnostic — the flag can appear anywhere in the command. Examples:

```bash
# Only results from test files
search-lexical "handleRequest" --file ".*_test\.go"

# Only Python files
search-signature "def " --file ".*\.py"

# Only files under a specific directory
find-usages "parseConfig" --file "internal/config/.*"

# Largest chunks in the renderer subsystem only
largest-chunks 20 --file ".*/renderer/.*"
```

Use `--file` to dramatically narrow results when a query matches broadly across the codebase but you only care about a specific layer (tests, a subsystem, a language).

## Rules

1. **Always Index First**: If the chunk maps don't exist in `~/.garbell/indexes/` or if you've recently refactored heavily, boldly run `~/.garbell/garbell index`. It's incredibly fast and safe to re-run.
2. **Progressive disclosure**: When a tool returns a directory-grouped summary instead of full results, that is a signal to zoom in — narrow your query, or re-run the command on a specific subdirectory or file.
3. **Search-Lexical over Grep**: Prefer `search-lexical` over raw `grep` or `rg`. It resolves matches to function/class boundaries and returns compact summaries. Follow up with `read-chunk` on the specific entries you care about.
4. **Search-Related for concepts, Search-Lexical for symbols**: Use `search-related` when your query is an abstract concept and you are unsure of the exact vocabulary the codebase uses — it will expand the query automatically. Use `search-lexical` for known symbol names, exact strings, or regex patterns where expansion would add noise.
5. **File-Skeleton to Orient**: When you hit an unknown file or directory, run `file-skeleton` first. Then use `read-chunk` to jump directly into the function you care about.
6. **Find-Usages for Refactoring**: Use `find-usages` when renaming or modifying a core struct. It will tell you exactly which functions in which files call it, without spamming your context with the calls themselves.
7. **Extract-Interface to Grasp Modules**: If you want to know what a module provides without reading its implementation, `extract-interface` is your best friend.
8. **Use `peek` for local context in large functions**: If `search-lexical` returns a match deep inside a large chunk and you only need a few lines of context around the match, use `peek <file> <line>` rather than `read-chunk`. It's cheaper and avoids reading hundreds of lines you don't need.
9. **Use `--file` to scope broad searches**: When a query matches across the whole codebase but you only care about one layer (e.g. only tests, only a specific package), add `--file` rather than issuing multiple narrower queries.
10. **Override threshold**: Set `GARBELL_MAX_LINES=<n>` to raise or lower the output threshold for the current invocation (default: 500).

## Errors

- **Missing Index**: `open ~/.garbell/indexes/... no such file or directory`. Run `~/.garbell/garbell index`.
- **Empty Results**: Be cautious with exact queries in `search-lexical`. If it's empty, try a broader term, or use `search-fuzzy` if you suspect the symbol exists but is spelled differently.
- **`search-related` fails with "ppmi.json not found"**: The PPMI thesaurus is generated during `index`. Run `~/.garbell/garbell index` and retry.
- **`read-chunk` returns a line-limit warning for `-1`**: The file exceeds the safe line limit. Add `--unsafe` only if you truly need the whole file; otherwise use `file-skeleton` + `read-chunk` to read only the parts you need, or use your native file-reading tool.
- **Invalid `--file` pattern**: If the regex is malformed, the command returns an error. Correct the regex and retry.
- **Important**: debrief the user about any errors or usability improvements you find in the tooling after using it. This will make it better.
登录后下载/点赞/收藏 ❤ 6 | ★ 0
garbell

目录结构

SKILL.md

举报内容

提示
