v1.4.17

• Default search to startpage, cascade to searxng then ddgs
• Activity logging tables (activity_log, autocomplete_suggestions, autocomplete_training)
• API endpoints for activity/autocomplete logging and training data export
• Fix memory scope query to not require all filters

v1.4.16

fix: flush SSE events immediately for real-time streaming, prompt chat before stop in agentic loop

v1.4.15

• Fix command injection in _tool_web_search and _tool_file_search (PR #208, thanks @spidershield-contrib)

v1.4.14

• Add Sememolution population-based KG evolution module (kg_population.py) with GeneticEvolver integration, Poisson-sampled search traversal, per-individual graph state, crossover with graph merging, and LLM-judged response ranking
• Add memory extraction pipeline and sememolution documentation to knowledge-graphs guide
• Add MLX Apple Silicon documentation to fine-tuning guide
• Add KG, memory, sememolution, and fine-tuning examples to README
• Remove dead shell=True subprocess fallback in _tool_web_search (closes #207)
• Fix _tool_file_search to use list-form subprocess without shell=True

v1.4.13

• Pass user generation params (temperature, top_p, top_k, max_tokens) to LLM in both chat and tool_agent modes
• Make serve.py robust for Windows — optional imports for redis/flask_sse/mcp, better error handling, NPCSH_BASE env var
• Fix MCP server engine step rendering — action/args from _raw_steps now template-rendered with tool call arguments
• Add Windows CI tests for serve.py imports, settings round-trip, and server startup

v1.4.12

• device routing in ft modules (sft, rl, usft): device='mlx'|'cpu'|'cuda'
• MLX LoRA training via mlx-lm Python API on Apple Silicon
• HF model name → mlx-community resolution
• backwards compatible, default device='cpu'

v1.4.11

stream_events plumbing from jinx execution through generator chain

v1.4.10

• Inline generator in check_llm_command, no separate function
• create_jinx_stream takes (npc, command) directly, no StreamConfig
• Sub-delegation events via shared_context['sub_events']

v1.4.9

• Generator-based streaming for check_llm_command (stream=True yields events)
• No threads/queues — clean generator protocol
• Chat streams token by token, tools emit tool_start/tool_result events

v1.4.8

• Threaded check_llm_command in create_jinx_stream with keepalive SSE events
• Prevents SSE timeout during long delegation
• Event queue for jinxes to push real-time progress