neom_ollama 1.2.0

Ollama integration for Dart — model discovery, pull/delete, health checks, setup automation, OpenAI-compatible chat, hardware profiling, reasoning trace parsing, and plain-text tool-call recovery for [...]

1.2.0

• PlainTextToolCallParser — recovers tool calls embedded as text in model output for local models that don't always populate the structured tool_calls field. Supports five formats out of the box:
  • Bracket (LMStudio / Mistral): [name]\n{...}\n[END_TOOL_REQUEST] or [/name]
  • Hermes / Qwen tag: <tool_call>{"name": "...", "arguments": {...}}</tool_call>
  • Function-call tag: <function_call>{"name": "...", "arguments": {...}}</function_call>
  • Fenced JSON: ```json or ```tool_call blocks containing {"name": "...", "arguments": {...}}
  • Bare JSON: whole-message {"name": "...", "arguments": {...}} (last-resort)
• Tolerates parameters as an alias for arguments and string-escaped JSON arguments (some Llama variants).
• Optional allowedToolNames set for host-side allowlist filtering.
• 256 KB per-block payload cap to bound parsing cost.
• PlainTextToolCallParser.strip(text, blocks) — removes recovered blocks and returns the user-visible content for chat UIs.
• 23 unit tests covering all five formats, edge cases, and the strip helper.

1.1.0

• Hardware profiler — cross-platform RAM/CPU/GPU detection (HardwareProfiler.detect()), with native (FFI), web (navigator + WebGL), and stub backends.
• Model advisor — recommends Ollama models based on detected hardware tier.
• Model optimizer — picks optimal num_gpu / num_thread parameters per host.
• Thinking trace support — OllamaClient.chatWithThinking() exposes the reasoning trace for Qwen3, DeepSeek-R1, QwQ, gpt-oss, and other reasoning models. Honours both the new Ollama message.thinking field and inline <think>…</think> / Thought: patterns.
• ThinkingParser — pure-Dart utility (zero deps) to split reasoning from content; handles closed tags, dangling open tags (streaming), and bare Thought: / Answer: formatting.
• OllamaChatResult — new return type with thinking + content fields.
• Web typing fixes — added explicit <Object?> type args to js_util.getProperty and callMethod so the package compiles cleanly under Dart 3.10+ strict typing.
• Tests added: ollama_client_test.dart, thinking_parser_test.dart.

1.0.0

• Initial release.
• OllamaClient — HTTP client for Ollama REST API (status, list, show, pull, create, delete, chat, stream).
• OllamaSetup — Automated setup: install detection, server start, model provisioning with progress streaming.
• OllamaChatService — OpenAI-compatible /v1/chat/completions client with conversation history.
• OllamaModel — Data class with name, size, family, parameters, quantization.
• OllamaPullProgress — Streaming progress for model downloads.
• Pure Dart — no Flutter dependency, works in CLI, desktop, and mobile.