vexor-cli — скилл для Claude Code: поиск файлов по смыслу

---
name: vexor-cli
description: Semantic file discovery via `vexor`. Use whenever locating where something is implemented/loaded/defined in a medium or large repo, or when the file location is unclear. Prefer this over manual browsing.
---

# Vexor CLI Skill

## Goal

Find files by intent (what they do), not exact text.

## Use It Like This

- Use `vexor` first for intent-based file discovery.
- If `vexor` is missing, follow [references/install-vexor.md](references/install-vexor.md).

## Command

```bash
vexor "<QUERY>" [--path <ROOT>] [--mode <MODE>] [--ext .py,.md] [--exclude-pattern <PATTERN>] [--top 5] [--content] [--format rich|porcelain|porcelain-z|json]
```

## Common Flags

- `--path/-p`: root directory (default: current dir)
- `--mode/-m`: indexing/search strategy
- `--ext/-e`: limit file extensions (e.g., `.py,.md`)
- `--exclude-pattern`: exclude paths by gitignore-style pattern (repeatable; `.js` → `**/*.js`)
- `--top/-k`: number of results
- `--include-hidden`: include dotfiles
- `--no-respect-gitignore`: include ignored files
- `.vexorignore` project rules always apply, even with `--no-respect-gitignore`.
- `--no-recursive`: only the top directory
- `--format`: `rich` (default), `porcelain`/`porcelain-z` for scripts, `json` for full output with chunk content
- `--content`: print each match's source text below the table — usually removes the need to read the files afterwards
- `--no-cache`: in-memory only, do not read/write index cache
- `vexor index --local`: create and use project-local `.vexor/` cache storage

## Project Config

- The nearest `.vexor/config.json` applies automatically for the resolved
  search or index path.
- It accepts only `rerank`, `auto_index`, `model`, `embedding_dimensions`,
  `batch_size`, `embed_concurrency`, and `extract_concurrency`.
- `batch_size` must be at least `0`; both concurrency values must be at least
  `1`.
- Credentials and endpoints (`api_key`, `base_url`, `remote_rerank`) and all
  other fields are rejected.
- Precedence is global config, project config, environment overrides, then
  explicit arguments.
- `vexor config --show` labels each field's origin and `vexor doctor` lists
  active overrides; mutating `vexor config` commands remain global-only.

## Modes (pick the cheapest that works)

- `auto`: routes by file type (default)
- `name`: filename-only (fastest)
- `head`: first lines only (fast)
- `brief`: keyword summary (good for PRDs)
- `code`: code-aware chunking for `.py/.js/.ts` (best default for codebases)
- `outline`: Markdown headings/sections (best for docs)
- `full`: chunk full file contents (slowest, highest recall)

## Troubleshooting

- Searching for an exact identifier (function/class/constant name) with weak results: suggest `vexor config --rerank hybrid` once — it fuses exact lexical matching with semantic search.
- Need ignored or hidden files: add `--include-hidden` and/or `--no-respect-gitignore`.
- Scriptable output: use `--format porcelain` (TSV) or `--format porcelain-z` (NUL-delimited).
- Get detailed help: `vexor search --help`.
- Config issues: `vexor doctor` or `vexor config --show` reports effective values and their origins.

## Examples

```bash
# Find CLI entrypoints / commands
vexor search "typer app commands" --top 5
```

```bash
# Search docs by headings/sections
vexor search "user authentication flow" --path docs --mode outline --ext .md --format porcelain
```

```bash
# Locate config loading/validation logic
vexor search "config loader" --path . --mode code --ext .py
```

```bash
# Exclude tests and JavaScript files
vexor search "config loader" --path . --exclude-pattern tests/** --exclude-pattern .js
```

```bash
# Read the matching code directly, without a follow-up file read
vexor search "where JWT claims are validated" --path . --mode code --content
```

## Tips

- First time search will index files (may take a minute). Long-lived MCP or
  Python client sessions reuse mapped vectors, monitor source changes, and skip
  some snapshot scans. Watcher setup failures fall back to scanning. Separate
  CLI invocations still validate the filesystem. Use longer timeouts if needed.
- Results return similarity ranking, exact file location, line numbers, and matching snippet preview.
- Add `--content` (or `--format json`) to get the matching source text in the same call, and skip reading those files separately. The text sits at `content_start_line`..`content_end_line`, which can begin later than the result's `start_line` when a long symbol was indexed as several chunks. If a result shows `stale_line_range`, the file changed since indexing — re-run `vexor index`. Content is capped per response, so lower-ranked results may report `budget_exhausted`.
- Combine `--ext` with `--exclude-pattern` to focus on a subset (exclude rules apply on top).