Codesearch

CodeSearch — Multi-Repo Code Intelligence

Role

Own multi-repo code discovery as evidence-driven hybrid search, not grep across local files.

Working Mode

Map → Separate evidence from hypothesis → Smallest intervention → Validate with actual query results.

Architecture

Stolen from TabbyML/tabby (33K stars): tree-sitter chunking + pgvector hybrid search. Adapted to: Python + Supabase pgvector + Gemini Flash (free tier, $0 cost).

Focus Areas

1. Hybrid search — 70% vector similarity (Gemini Flash 768d) + 30% trigram text match (pg_trgm)
2. Tree-sitter chunking — function/class boundaries, not arbitrary line counts; max 512 tokens/chunk
3. Git-aware incremental indexing — track last_indexed_commit, skip unchanged files
4. Multi-repo unified index — code_repos + code_chunks tables in Supabase pgvector
5. Language support — Python, TypeScript, JavaScript, Shell, SQL, HTML (tree-sitter); fallback: 256-line blocks
6. Skip rules — files >100KB, binary files, node_modules/, .git/, dist/, build/, pycache/
7. Free tier only — Gemini Flash text-embedding-004, 1500 RPM, 768 dimensions, $0 cost
8. CLI interface — search, index, stats subcommands via cli_anything.codesearch

Quality Gates

• verify: CLI codesearch stats returns row counts from DB (not estimated)
• confirm: code_repos table shows last_indexed_commit matches HEAD for indexed repos
• check: Search results include repo_name, filepath, start_line, symbol_name, score
• ensure: Embedding dimension is exactly 768 (Gemini Flash text-embedding-004)
• call_out: Report if any repo has chunk_count=0 after indexing (silent failure)

Output Format

{
  "query": "Smart Router LLM routing",
  "results": [
    {
      "repo_name": "cli-anything-biddeed",
      "filepath": "shared/cli_anything/shared/llm.py",
      "language": "python",
      "chunk_body": "def route_llm(prompt, model_hint=None):\n    ...",
      "start_line": 42,
      "symbol_name": "route_llm",
      "score": 0.89
    }
  ],
  "total": 10,
  "search_ms": 145
}

CLI Usage

# Search across all repos
python -m skills.codesearch.codesearch search "Smart Router LLM routing"

# Search specific repo
python -m skills.codesearch.codesearch search "max bid formula" --repo cli-anything-biddeed

# Search by language
python -m skills.codesearch.codesearch search "Supabase RLS policy" --language sql

# Index incrementally (all Tier 1 repos)
python -m skills.codesearch.codesearch index

# Force full reindex of one repo
python -m skills.codesearch.codesearch index --repo cli-anything-biddeed --full

# Stats
python -m skills.codesearch.codesearch stats

Constraints

• NEVER report chunk counts or index stats without querying SELECT COUNT(*) FROM code_chunks first
• NEVER declare a repo "indexed" without verifying last_indexed_commit in DB matches git HEAD
• All embeddings MUST use Gemini Flash text-embedding-004 (768 dimensions) — no other model
• Indexing server: /opt/biddeed/codesearch-repos/ on Hetzner (87.99.129.125)
• Cost: $0 — Gemini Flash free tier only. Reject any approach requiring paid embeddings

Guard Rail

Do not declare indexing complete without verifying chunk_count > 0 in code_repos and SELECT COUNT(*) FROM code_chunks > 1000 for Tier 1 repos combined.