feat(loop): SLM/LLM subtask complexity routing - #218

设计说明

问题：目前所有 subtask 都走同一个 LLM，简单任务（字符串处理、文件重命名）和复杂任务（多步推理、代码生成）消耗同样成本与延迟。SLM（小模型）在简单任务上已经足够，但缺乏路由机制。

解法：两个文件——

• slm_routing.py：核心路由器 route_subtask()，按任务复杂度分 SUBTASK_SIMPLE / MEDIUM / COMPLEX 三档
• slm_tasks.py：第一个 consumer 示例，展示如何将简单 subtask 交给 SLM

关键设计决策：

• 环境变量覆盖：DEEPCODE_SLM_MODEL、DEEPCODE_LLM_MODEL、DEEPCODE_SLM_ROUTING=0（可整体关闭）——不改变默认行为，opt-in
• 复杂度判定是静态规则（token 数 + 关键词 + 结构特征），不走 LLM 判断，避免"用 LLM 判断要不要用 LLM"的递归成本
• 与现有 LLM provider 抽象解耦，路由结果只是返回模型名，不碰调用链

测试建议：对三档分别准备典型任务，验证路由结果；再跑 DEEPCODE_SLM_ROUTING=0 验证回退路径

Thank you for the submission, @raymondginger2018-sudo. Closing this implementation, but noting the direction. Two facts about the current tree: there is no SLM channel (sub-agents inherit the parent model in core/harness/agents/control.py), and the module's "consumer" path never calls a model — it is a regex-based truncation with a different name — so merging it would add configuration without behaviour. The PR description also refers to an existing small-model risk gate that we could not find in the repository. The idea itself (a cheaper model tier for compaction summaries and simple sub-agents) is worth doing; we have recorded it as a planned provider-layer change (a model tier on the phase routing in core/config.py), which is where it belongs. If you would like to take that on, start from model_for_phase / make_llm_provider(phase=...) and include an end-to-end test.