atlas-navlisted
Install: claude install-skill muslewski/token-oracle
# Navigating code and the Atlas
Default to this repo's configured retrieval engine over grep+read-whole-file
— same correctness, far fewer tokens read. If this repo uses the ctx-search
adapter (`adapters/ctx-search/`), that engine is `ctx_search` (the
context-mode MCP tool), kept fresh automatically by a SessionStart hook.
Otherwise use whatever retrieval engine the repo documents, or fall back to
grep — this skill's routing logic applies either way.
**With `ctx_search`: call it with NO `project` param** — it defaults to the
repo (cwd), which is where both buckets are indexed. Results are tagged
`Source: code:…` / `Source: atlas:…`; narrow with `source: code` /
`source: atlas`. Folders listed in `atlas.config.json` →
`retrieval.excludeFromSearch` (default `drafts/`, `visuals/`) are never in
the index — don't expect hits from them.
## Route by intent
| Intent | Tool |
|---|---|
| Known file path | Read directly — no search |
| Exact symbol ("where is X defined / who calls X") | LSP / go-to-definition, if available |
| Broad ("where is X / which note explains Y / find the invariant / which spec") | Configured retrieval engine (`ctx_search` if this repo has the adapter) |
| Engine result empty/wrong, or an exhaustive match list is needed | grep + Read fallback |
## Per-bucket mode
**Atlas hits** (`Source: atlas:…`) — the snippet usually captures the
zone/decision summary inline; no follow-up read needed for orientation.
**Code hits** (`Source: code:…`) — use the engine to