From claudix
This skill should be used when working with claudix semantic search beyond plain queries — building or rebuilding the index, controlling what gets indexed (indexing gitignored files, excluding paths, `.indexignore`/`.indexinclude` rules), switching embedding providers or models (bundled, LM Studio, Ollama), setting up cross-repo search, tuning or disabling grep interception and related-code surfacing, or recovering from a stale or corrupted index.
How this skill is triggered — by the user, by Claude, or both
Slash command
/claudix:claudixThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
One Rust binary serves the MCP tools, the hooks, and a CLI. Prefer `claudix` from PATH; when it is absent (Windows installs create no symlink), run the node bootstrap that ships with the plugin at `bin/claudix-bootstrap.js` under the plugin root, two directories above this skill's base directory:
One Rust binary serves the MCP tools, the hooks, and a CLI. Prefer claudix from PATH; when it is absent (Windows installs create no symlink), run the node bootstrap that ships with the plugin at bin/claudix-bootstrap.js under the plugin root, two directories above this skill's base directory:
claudix <subcommand> 2>&1
# without claudix on PATH:
node "<plugin-root>/bin/claudix-bootstrap.js" <subcommand> 2>&1
Configuration lives in ~/.claude/claudix.toml (global) overridden by <repo>/.claude/claudix.toml (project). Both are optional; every entry point validates them and errors name the exact key to fix.
Agent path: the reindex MCP tool. CLI path: claudix index (--progress streams per-file progress). Pass force: true / --force to wipe the store before rebuilding. Needed for:
Indexing may take a minute on large repositories; if it errors, run /claudix:doctor to diagnose the embedding provider.
.indexignore and .indexinclude files use gitignore syntax (globs, ! negation, comments). Place them at the repo root or in any subdirectory; patterns are relative to the rule file's own directory, like nested .gitignore files. Precedence per path: .indexinclude beats .gitignore, which beats .indexignore.
.indexignore (test fixtures, vendored code, minified bundles)..indexinclude (internal docs/, generated code). A one-line * inside docs/.indexinclude pulls that whole tree. Files without a code chunker index as plain text.[indexing] respect_gitignore = false indexes everything gitignored instead of selected subtrees.Edited rule files apply on the next index run, so reindex after changing them.
[embedding] provider = "bundled" (default, gte-modernbert-base at 768 dims; bge-small-en-v1.5 at 384 dims also selectable, zero setup either way) or "http" (any keyless OpenAI-/v1/embeddings server: LM Studio http://localhost:1234, Ollama http://localhost:11434, vLLM, a LiteLLM proxy). Set model to the id the server reports; set dimensions to the model's output size. Any model or dimension change requires a force rebuild (previous section). Keyed APIs (Voyage, OpenAI) need a local proxy that injects the key.
Each repo must be indexed on its own first, with the same embedding model (mismatches surface per repo in repo_errors). Extra repos are read-only. Two ways to include them:
[search] cross_repos = ["/abs/path/other-repo"] in the active repo's configrepos arg on search_code or find_duplicates (both add to the active project, which is always scanned)Grouped results prefix each directory with its repo (<repo> :: <dir>).
All under [hooks]:
intercept_grep = false disables the conceptual-Grep interception (anchored regexes, globs, short queries always pass through anyway).auto_reembed_on_edit = false stops re-embedding on Write/Edit; reindex_debounce_secs / reindex_max_wait_secs tune the edit-burst coalescing.surface_related_on_edit (default on) injects semantically related locations after an edit; surface_related_on_read = true (opt-in) does the same for ranged Reads. related_top_k caps hits per edit; the cosine floor is corpus-relative (each full index stores the p30 of its own best-neighbor distribution), and related_min_similarity is the fallback plus a hard minimum.watch = true swaps the debounced hook reindex for a live file watcher (claudix watch).claudix status shows file/chunk counts, the embedding model, staleness — SessionStart already reports these, so reach for it only mid-session. /claudix:doctor diagnoses the binary, index, provider health, config errors. Logs: .claudix/logs/index.log; RUST_LOG=debug on any CLI subcommand for verbose output.
npx claudepluginhub uwuclxdy/claudix --plugin claudixGuides collaborative design exploration before implementation: explores context, asks clarifying questions, proposes approaches, and writes a design doc for user approval.
Creates structured, bite-sized implementation plans from specs or requirements before writing code. Useful for breaking down multi-step tasks into testable steps with file structure and task boundaries.
Resolves in-progress git merge or rebase conflicts by analyzing history, understanding intent, and preserving both changes where possible. Runs automated checks after resolution.