v0.4.0: Codebase AST Intelligence, Upward Anchor Discovery, and Tree-Sitter Parser Suite

Added

• Codebase AST Intelligence Tools (Sprint 3):
  • Added kache_explore_symbol(symbol, workspace_id, file_path): Instantaneous (< 50 ns) symbol inspection from SwissTable. Returns signatures, docstrings, line numbers, and file paths in a surgical < 300 token payload with zero file reads.
  • Added kache_index_workspace(path, workspace_id, force): Incremental AST indexer that parses codebases into exact SwissTable keys (sym:<ws>:<symbol>, sym:<ws>:<file>:<symbol>) and vector records (vec:<ws>:<file>:<symbol>:<line>).
  • Added kache_workspace_status(workspace_id, auto_index): Inspects indexing status, total indexed symbols, commit watermark, and cache freshness.
• Multi-Language Tree-Sitter AST Parsing:
  • Built-in AST symbol extraction across 5 major languages: Python (.py), Rust (.rs), Go (.go), TypeScript (.ts, .tsx), and JavaScript (.js, .jsx).
  • Extracts classes, structs, traits, interfaces, methods, functions, docstrings, and type signatures with zero external compiler dependencies.
• Upward Anchor Discovery (find_root_from_path):
  • Added zero-config project root detection from any touched file path, climbing parent directories to locate .git or build manifests (Cargo.toml, pyproject.toml, package.json, go.mod).
  • Completely eliminates slow, guessing directory walks under $HOME.
• Incremental Git Watermarks (kache:meta:<ws>:commit):
  • Sub-millisecond sync verification using commit SHA watermarks in SwissTable. Up-to-date projects skip AST parsing entirely with zero redundant work.
• Multi-Project & Monorepo Support (KACHEDB_WORKSPACE_ROOTS):
  • Configurable root paths via KACHEDB_WORKSPACE_ROOTS supporting multi-repo collections and monorepos, including category directory nesting.
  • Added background startup auto-discovery daemon (_background_startup_index) that scans and indexes configured project roots without blocking MCP client handshake.
• On-Demand Auto-Indexing from File Anchors:
  • kache_explore_symbol automatically resolves workspace roots and lazily indexes codebases on the fly when passed a file_path.

v0.3.0

Added

• Hierarchical Document Chunking & Parent Document Linkage:
  • Added kache_get_parent_document(parent_key) tool for sub-microsecond retrieval of unabridged parent documents.
  • Added auto_chunk=True and parent_key parameters to kache_save_context to automatically chunk large documents (>1,500 chars) with header breadcrumbs linked to their parent KV document.
• Sub-Microsecond Exact Prompt Aliasing:
  • Added prompt_alias parameter to kache_save_context for registering instant SwissTable lookup shortcuts (exact:<alias> or exact:<ws>:<alias>), resolving repeated prompts in < 50ns with 0ms LLM time.
• 64-bit Bitmask Tag Pre-Filtering:
  • Added tags parameter to kache_semantic_search and kache_save_context mapped via kachedb.tags for zero-allocation integer bitwise AND pre-filtering in vector scans.
• Adaptive Fallback & Automatic Resolution:
  • Added exact_first and resolve_parent flags to kache_semantic_search with automatic prompt alias checking, #prompt_anchor semantic indexing, and tag/threshold relaxation fallback.
• Real-Time Telemetry Tracking:
  • Added exact_hits metric and avoided embedding inference time tracking to kache_telemetry.

v0.2.1

Fixed

• Timing Isolation in Semantic Search: Moved vector search stopwatch to start after embedding model inference (embedder.encode()), ensuring reported latency captures true KacheDB SIMD kernel execution time (< 200 µs) rather than embedding runtime.
• Per-Operation Avoided Latency Baselines: Replaced flat 400 ms baseline with calibrated per-operation baselines (semantic_search: 6,200 ms, kv_get: 150 ms), configurable via KACHEDB_BASELINE_SEMANTIC_MS and KACHEDB_BASELINE_KV_MS.
• Avoided Source Sizing: Added avoided_source_chars parameter to record_hit() to calculate savings based on avoided source file sizes rather than cached summary responses.
• Configurable Token Cost: Replaced hardcoded .00/M USD rate with configurable KACHEDB_TOKEN_COST_PER_MILLION setting (defaulting to .00/M for Claude 3.7 Sonnet).

v0.2.0

Added

• Multi-Workspace Memory Isolation:
  • Added workspace_id parameter to kache_save_context, kache_semantic_search, and kache_delete to provide segregated vector index namespaces ({index_name}:{workspace_id}) across different workspaces and projects.
• Native Telemetry Dogfooding:
  • Dogfood native KacheDB atomic keys (kachedb:telemetry:*) for cumulative tracker metrics, tracking hits, misses, tokens saved, and avoided latency without external SQLite or disk-based JSON persistence.
• KacheDB Server & SDK v0.1.1 Parity:
  • Updated dependency to kachedb>=0.1.1 to leverage latest vector indexing and buffer zero-copy stability.
  • Added pythonpath = ["src"] to pytest configuration for reliable local development and test execution.

v0.1.1

Full Changelog: v0.1.1...v0.1.1