Summary

Add native support for Zvec (Alibaba's in-process vector database) as a semantic code search backend. No MCP server or external daemon — runs entirely in-process, just like the existing SQLite-based mnemopi memory engine.

Opt-in by design: Code search is disabled by default. Users enable it via /settings → Tools → Available Tools → Code Search (Zvec), then explicitly build the index with /code-index. The code_search tool does not auto-index — it only searches the existing index and tells users to run /code-index when the index is empty.

Architecture

coding-agent
 ├── /code-index slash command ──→ indexWorkspace()
 │    └── uses loadMnemopi() for lazy module loading
 │
 └── code_search tool (discoverable, gated by tools.codeSearchEnabled)
      ├── workspace scanner (gitignore-aware)
      ├── code chunker (100-line chunks, 10-line overlap)
      ├── embedding via mnemopi's embed()/embedQuery()
      └── ZVecCodeStore (from pi-mnemopi)
           ├── HNSW vector index (cosine)
           ├── BM25 full-text index
           └── hybrid search (RRF fusion)

Changes

mnemopi package (@oh-my-pi/pi-mnemopi)

• zvec-store.ts: ZVecCodeStore class — Zvec vector store wrapper with upsert, vector search (HNSW), FTS search (BM25), hybrid search (RRF), removeFile (via deleteByFilterSync), getFileHashes, optimize
• zvec-config.ts: Config functions (OMP_ZVEC_* env vars for index path, chunk size, overlap, top-K)
• @zvec/zvec added as optional peer dependency

coding-agent package (@oh-my-pi/pi-coding-agent)

• settings-schema.ts: Added tools.codeSearchEnabled setting (default: false) under Tools → Available Tools
• builtin-registry.ts: Added /code-index slash command — explicitly builds or rebuilds the Zvec code search index; uses loadMnemopi() for lazy module loading
• code-search-index.ts: Standalone indexWorkspace() function — shared by both the slash command and the tool; uses loadMnemopi() to keep mnemopi off the startup module graph
• code-search-helpers.ts: Shared helpers (chunking, language detection, binary detection, file filtering)
• code-search.ts: CodeSearchTool — searches the existing index only (no auto-indexing); hybrid vector+FTS search with FTS-only fallback; overfetches when a glob pattern filter is provided
• code-search.md: Tool description prompt (mentions /code-index requirement)
• Tool registered as discoverable (loadMode=discoverable), only loads when needed
• createIf returns null when tools.codeSearchEnabled is false or @zvec/zvec native addon not loadable

Infrastructure

• @zvec/zvec: 0.5.0 added to workspace catalog
• trustedDependencies: ["@zvec/zvec"] added to root package.json (Bun lifecycle script allowlist)

Key design decisions

• Opt-in, not default: tools.codeSearchEnabled defaults to false. Users must explicitly enable it in settings and run /code-index to build the index. The tool does not auto-index on first search.
• Embedding source: Uses mnemopi's existing embed()/embedQuery() pipeline (local fastembed via ONNX or API-based). Does NOT use pi-ai (no embedding client) or Zvec's Python-only embedding functions.
• Fallback strategy: When embeddings are unavailable, falls back to BM25-only FTS search — not a grep duplicate (ranked chunks across workspace vs individual matching lines).
• Incremental indexing: Compares persisted chunk hashes via getFileHashes() to skip unchanged files across sessions. removeFile uses deleteByFilterSync for O(1) cleanup. Files that grow past 1MB or become binary have their stale chunks removed automatically.
• Chunk ID format: Uses ${Bun.hash(filePath + "\\0" + content)}-${lineStart} — includes file path to prevent duplicate-content collisions; Zvec rejects :, /, | in document IDs.
• Lazy module loading: The /code-index slash command and indexWorkspace() use loadMnemopi() from src/mnemopi/state.ts (memoized lazy loader) instead of raw await import(), respecting onnxruntime-node NAPI load-ordering constraints.

Verification

• Smoke test: Real ZVecCreateAndOpen → insert → optimize → vector query → FTS query → hybrid multiQuery → removeFile → close — all pass under Bun
• mnemopi tests: 14 pass, 0 fail
• coding-agent tests: 2 pass, 0 fail (edit-then-reindex correctness — no stale chunks)
• Type checks: Both packages pass bun run check:types
• Biome: Clean on all new/modified files

Review fixes addressed

• Doc-ID collision: Include file path in chunk hash (Bun.hash(filePath + "\\0" + content)) so duplicate files don't overwrite each other
• Stale chunks on skip: Move currentFiles.add() after size/binary checks; call store.removeFile() when a previously-indexed file becomes too large or binary
• Pattern filter reliability: Overfetch (topK*5, min 100) when a glob pattern is provided, then slice to topK after filtering
• Test env safety: Pass local env object to config helpers instead of mutating process.env

Test plan

• cd packages/mnemopi && bun test test/zvec-store.test.ts — 14 tests
• cd packages/coding-agent && bun test src/tools/__tests__/code-search-index.test.ts — 2 tests
• cd packages/mnemopi && bun run check:types
• cd packages/coding-agent && bun run check:types