v2.1.24

• added npcpy.gen.systemone module with OllamaSystem1Client, OllamaSystem1Config, systemone, load_ollama_system1, availability/version checks, url normalization, model listing, a 64 kib request-size guard, and parsers for choice, noul, score, and predict responses (cd8324c, dc5d8a2)
• integrated the ollama backend into npcpy.gen.decision: added DecisionSystem1.from_ollama, ollama-backed decision_* helpers for choice, score, noul, predict, and route, environment-variable backend selection via NPCPY_SYSTEM1_BACKEND and NPCPY_SYSTEM1_OLLAMA_MODEL, tiered routing with threshold fallback, and clear errors for unsupported training or local-path loading (cd8324c, 2ae2d05)
• exported system one client symbols (OllamaSystem1Client, OllamaSystem1Config, OllamaSystem1Error, systemone, load_ollama_system1, etc.) from top-level npcpy and npcpy.gen, and exposed DecisionQuestion, DecisionRouter, DecisionSystem1, and the decision_* helpers in the npc compiler runtime globals (2ae2d05, dc5d8a2, c120788)
• added flask serve endpoints for system one and decision operations (/api/system1/* and /api/decision/*) with backend, model, and base_url parameters and a shared predictor resolver that selects local-trained or ollama predictors (cd8324c, c120788)
• made get_model_context_window respect the OLLAMA_NUM_CTX environment variable when determining the ollama context window size (dc5d8a2)
• refreshed the litellm cost map and included cost-map models when listing locally available models (c120788)
• cached model context-window lookups in /api/models and get_available_models (c120788)
• removed the npc_object.jinxes_dict fallback in execute_jinx to separate tool-capable and jinx-capable paths (c120788)
• bumped package version to 2.1.24 (c120788)
• added trust_remote_code option to EmbeddingConfig and HilbertConfig, and propagated it through all AutoModel and AutoTokenizer loading call sites for custom transformer architectures (2ae2d05, c120788)
• changed the default embedding training device to auto-select cuda, then apple mps, then cpu, while still honoring explicit device overrides (2ae2d05, c120788)
• fixed embedding tokenizer config handling so remote model loading passes config kwargs, while load_embedding_model keeps only model_path and device for local directories (1c4679e)
• added system one decision tests to ci running against a mock ollama server, added a mock ollama server and pytest fixtures, and included system one coverage in flask server tests (665e61e, c120788)
• added comprehensive system one user guide (docs/guides/system-one.md) and api reference page (docs/api/system1.md) covering npcpy.ft.system1, npcpy.gen.systemone, and npcpy.gen.decision, and linked them from the documentation index and mkdocs navigation (665e61e, 2ae2d05)
• reorganized memory api docs: dropped kg_vis and command_history, added knowledge_store, knowledge_index, kg_population, and memory_processor sections, and removed kg_vis.py from the readme's list of large files (665e61e)
• pinned pymdown-extensions>=12.1 in docs requirements (665e61e)
• added examples/example_system1_ollama.py demonstrating system one decisions via a local ollama model (665e61e)
• added optional legend field to ScoreResult and included it in to_dict() output for score interpretation metadata (2ae2d05)

New Contributors

• @Harbor404 made their first contribution in #304

Full Changelog: v2.1.23...v2.1.24