Symbol-aware repo search
repo-search is backed by a lightweight symbol index (internal/indexer).
When the agent searches for a bare identifier — a function, method, type,
class, const or variable name — the result starts with a ranked
definitions: block, followed by the usual full-text matches (the usages):
3 definitions:
internal/agent/searcher.go:31 type RepoSearcher type RepoSearcher struct {
...
42 matches:
internal/agent/llm_runtime.go:222: searcher RepoSearcher
...
Each definition line has the form path:line kind name signature.
Definitions are ranked exact name → case-insensitive exact → prefix →
substring. Queries that are not identifier-shaped (phrases, paths, regexes)
skip the index entirely and behave exactly like before.
Agent manifests are migrated to v7 on load: every agent already allowed
grep is granted repo-search, and the prompt guidance steers symbol
lookups to repo-search first (grep stays the default for regexes, phrases
and non-symbol text).
How the index works
- Backends are pluggable. The built-in backend extracts symbols with
per-language line patterns for Go, Python, and JavaScript/TypeScript
(
.js/.jsx/.ts/.tsx/.mjs/.cjs). Unsupported languages simply fall back to plain full-text search — the index never makes a search slower than grep. - Lazy and cached. Nothing is scanned until the first
repo-searchof a session. The index is persisted at<project>/.spettro/cache/symbols.jsonand reloaded on the next session. - Invalidation. Every lookup re-syncs with the filesystem: files whose
mtime or size changed are re-parsed, deleted files are dropped. The agent's
own
file-write/file-edit/multi-edittools additionally invalidate the touched file directly. - Bounded. Indexing stops at 20k files or 5 seconds, skips files over
1 MiB, and respects
.gitignore(shared matcher ininternal/ignore) plus the usual junk directories (.git,node_modules,vendor,dist,build, virtualenvs). Queries on a ~1k-file repo answer in milliseconds.