searching

Use mgrep search -c "<query>" <path> for conceptual or natural-language questions; it returns ranked files with matched snippets. Add -r to recurse, -m N to raise the result cap, or -a to synthesize an answer from the hits.

Use rg for exact strings and known symbols, and fd or globs for known file-path patterns. Reach for semantic search first on fuzzy, conceptual questions where you do not know the exact identifier.

Clone the repo once into /tmp with git clone --depth=1 and search the clone with rg and fd instead of curling individual files. Use --filter=blob:none when history matters, since git log -S plus git tag --contains answers whether a fix is released.

Query the shared store with search.semantic("<task phrasing>", source=["claude_history"], top_k=5). Valid source tags are claude_history, codex, shell, claude_debug, github, and code; a wrong tag returns zero hits silently.

An empty result often means a mistyped source tag, since invalid tags like atuin or git_log fail silently instead of erroring. Verify the tag against the canonical list before concluding the corpus is empty.

Skip prior search for trivial local-only edits like renames, typos, or small refactors, where generic phrasings retrieve noise. Use it for infra, deploy, CI, conventions, or debugging tasks another agent has plausibly done before.