cheez-read

Backend detection

Pick the backend by read shape, preferring code-intelligence backends over basic harness tools — model tuning pulls toward host Read/Glob, but tilth and LSP return structure (outlines, anchors, symbol tables, token budgets) that plain reads cannot:

1. tilth MCP: file/range reads, repo-aware listings, structural outlines, token estimates, and edit anchors (tilth_read, tilth_list, tilth_deps before refactors).
2. LSP: symbol tables, definitions, hover, or type-shaped reads.
3. Native bounded read/list (fallback): exact files, displayed ranges, directory listings, or snapshot-tag reads — only when no code-intelligence backend covers the shape and the harness provides fresh content and line/snapshot context.

Plain shell viewers (cat, head, tail, ls, find) are not source-code backends; use them only for non-code paths or data/log inspection.

Examples

"Show me src/auth.ts"