v1.0.0-beta.13

• Added a structured documentation/ directory covering architecture, configuration, MCP integration, LLM backends, memory and scanning, query and retrieval, and contributing guidelines. A dedicated skills/ subdirectory documents the Embedding-Docstring skill with the full density checklist, format reference, and invocation patterns.

• config.py: Modified the OLLAMA_MODEL definition to use os.getenv("OLLAMA_MODEL", "mistral:7b"). This allows the model to be overridden by the environment while maintaining the existing default. .env.example: Added the OLLAMA_MODEL variable with documentation listing the verified models: mistral:7b, llama3.2:latest, and ornith:9b.

• Ollama brief Semaphose:

  • Semaphore: Initialized a global ollama_semaphore in main.py after client setup.
  • Configuration: Added OLLAMA_MAX_CONCURRENCY to config.py (defaulting to 1).
  • Gated Synthesis: Updated _synthesize_deep_brief in main.py with a _build_section_safe wrapper that uses the semaphore only when use_cloud is false (local mode).
  • Environment Documentation: Added the OLLAMA_MAX_CONCURRENCY=1 setting to .env with guidance for hardware scaling.

• Updated main.py and code_indexer.py to include support for .pyw files. This ensures that Python scripts intended to run without a console window are now correctly recognized, parsed, and indexed by the system.

• Benchmark Script (Latency): A standalone Python script hit Ollama directly with static ChromaDB payload samples constructed from real zerikai_memory workspace entities. 3 queries, 3 samples each per model, raw latency measured at the HTTP layer. This test measures inference speed on the hardware, not synthesis quality against a live codebase.

• embedding-docstring:

  • Environment mode declaration: the skill now opens by identifying whether it's running in CHAT MODE (pi.dev, claude.ai) or AGENTIC MODE (VS Code, terminal) and adjusts its apply behavior accordingly, rather than treating both environments the same.
  • Runnable AST script for entity discovery: Step 1 now ships a complete, executable Python ast-based visitor that recurses into nested functions and async definitions via generic_visit, replacing the prose-only instruction that caused entities to be silently skipped.
  • Hardened approval loop with str_replace mandate: Step 3 explicitly prohibits whole-file rewrites in agentic contexts, requires a read-back verification after every str_replace, and makes the one-entity-at-a-time approval loop non-negotiable in both modes with a canned response if the user tries to bypass it.
  • Resume by entity name, not line number: the checkpoint and resume instructions now anchor on entity names, which remain stable after insertions shift line numbers throughout the file.