localmelo updates

A. Backend registry and infrastructure refactor

• Pluggable backend registry with local (Ollama, MLC-LLM, vLLM, SGLang) and cloud (OpenAI, Anthropic, Gemini, Nvidia) adapters
• Config supports split chat_backend / embedding_backend (e.g. cloud chat + local embedding)
• Removed legacy support/serving/ and support/models/ modules
• Tests reorganized from flat layout to domain-grouped subdirectories
• Removed legacy raw-string Agent constructor

B. Attempt-based agent loop with structured reflection

• Replaced flat step loop with attempt-based outer loop (MAX_ATTEMPTS=5, STEPS_PER_ATTEMPT=10, capped by MAX_AGENT_STEPS=30)
• Extracted _run_attempt() for clean single-responsibility separation
• Renamed ShortTerm → WorkingMemory with reflection entries that persist across attempts
• Structured ReflectionEntry and ReflectionDecision with active-learning fields
• Utility-based continuation gate: utility = info_gain * feasibility * novelty - cost - repeat_risk
• Stuck detection: repeated tool calls or errors trigger early attempt termination
• Reflection context injected into retrieval, planning, and reflect calls
• Reflections promoted to long-term memory only at terminal task state
• Hardened reflection parsing with strict type coercion and [0,1] float clamping

C. Ollama native provider, usage normalization, and smoke benchmark

• Ollama native /api/chat provider with think: true
• Thinking/answer split for both Ollama native and MLC <think> tags
• Provider-boundary usage normalization with total_tokens backfill
• Data-driven smoke benchmark with per-scenario metrics, normalized tokenizer comparison, multi-backend reports
• --report-only mode for regenerating reports from existing JSON

Status

Component	Status
Backend registry (local + cloud)	Done
Split chat/embedding config	Done
Attempt-based agent loop	Done
Structured reflection + utility gate	Done
Working memory with reflection entries	Done
Stuck detection	Done
Ollama native chat provider	Done
Smoke benchmark framework	Done
637 tests passing (ruff, black, mypy, pytest)	Done
Task decomposition (decompose action)	Not yet
Sleep-mode pipeline	Not yet
Long-memory promotion policies	Not yet
Utility threshold empirical tuning	Not yet

A. 后端注册表与基础设施重构

• 可插拔后端注册表，支持本地 (Ollama, MLC-LLM, vLLM, SGLang) 和云端 (OpenAI, Anthropic, Gemini, Nvidia) 适配器
• Config 支持 chat_backend / embedding_backend 分离配置（如云端 chat + 本地 embedding）
• 移除遗留的 support/serving/ 和 support/models/ 模块
• 测试从扁平结构重组为按子系统分目录
• 移除遗留的 Agent 原始字符串构造路径

B. 基于 Attempt 的 Agent 循环与结构化反思

• 用 attempt-based 双层循环替代原来的单层 for 循环（MAX_ATTEMPTS=5，STEPS_PER_ATTEMPT=10，总上限 MAX_AGENT_STEPS=30）
• 提取 _run_attempt() 方法，职责清晰
• ShortTerm 更名为 WorkingMemory，反思条目跨 attempt 保留
• 结构化 ReflectionEntry 和 ReflectionDecision，包含主动学习字段
• 基于效用值的续行门控：utility = info_gain * feasibility * novelty - cost - repeat_risk
• 卡住检测：相同工具调用或错误重复 3 次触发提前终止
• 反思上下文注入到检索、规划和反思调用中
• 反思条目仅在任务终态时推送到长期记忆
• 严格类型强制转换和 [0,1] 浮点数钳位

C. Ollama 原生 Provider、Usage 规范化与 Smoke 基准框架

• Ollama 原生 /api/chat provider，启用 think: true
• 支持 Ollama 和 MLC <think> 标签的 thinking/answer 拆分
• Provider 边界的 usage 规范化，支持 total_tokens 回填
• 数据驱动的 smoke 基准框架：场景指标、统一 tokenizer 对比、多后端报告
• --report-only 模式从已有 JSON 重新生成报告

状态

组件	状态
后端注册表（本地 + 云端）	完成
Chat/Embedding 分离配置	完成
Attempt-based Agent 循环	完成
结构化反思 + 效用门控	完成
Working Memory + 反思条目	完成
卡住检测	完成
Ollama 原生 Chat Provider	完成
Smoke 基准框架	完成
637 测试通过（ruff, black, mypy, pytest）	完成
任务分解（decompose 动作）	待完成
Sleep 模式管线	待完成
长期记忆推送策略	待完成
效用阈值经验校准	待完成

• Split runtime and infrastructure into melo/ and support/
• Introduced provider contracts to reduce coupling
• Established memory and sleep-mode package boundaries
• Added sleep module as foundation for continuous personalization
• Cleaned up local serving paths
• Expanded regression coverage

• 运行时和基础设施拆分为 melo/ 和 support/
• 引入 provider 契约以降低耦合
• 建立 memory 和 sleep-mode 包边界
• 添加 sleep 模块作为持续个性化的基础
• 清理本地 serving 路径
• 扩展回归测试覆盖