feat(loop): SLM/LLM subtask complexity routing

[raymondginger2018-sudo]

Summary

Small language models (SLM) fit local/edge/low-cost niches. This PR adds an explicit subtask-class router: high-frequency, low-complexity subtasks (tool-result cleanup, summarization, classification) should ride the SLM path, while deep reasoning stays on the LLM path.

Components

core/loop/slm_routing.py (new, 152 lines)

Pure decision mechanism:

• route_subtask(task_class, ...) -> RoutingDecision
• Subtask classes: simple / medium / complex
• Env-tunable model overrides: DEEPCODE_SLM_MODEL, DEEPCODE_LLM_MODEL
• DEEPCODE_SLM_ROUTING=0 disables SLM routing (falls back to LLM tier)

core/loop/slm_tasks.py (new, 102 lines)

First consumer: turns the routing decision into a preview-shaping policy for oversized tool results — when cleanup is SLM-grade, the persisted preview is shaped as a clean dense digest; when LLM-grade, raw truncation stays.

Design

• Zero network, zero async — decision-only, deployment-adaptable
• Route by task complexity, not by caller identity
• Clean separation: routing decision vs. execution channel

Files

• core/loop/slm_routing.py
• core/loop/slm_tasks.py

Part of GenAI lesson 19 SLM/LLM cost-tier family.

[raymondginger2018-sudo]

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