green-dalii/dsh-shift-router

DeepSeek Harness 的双层模型路由器——LLM-Judge 路由、多模型回退链、指数退避故障转移,以及任务级编排(pi-shift-router 的 DSH 适配版)

catalog 简介:Two-tier model router for DeepSeek Harness — LLM-Judge routing, multi-model fallback chains, exponential-backoff failover, and task-level orchestration (DSH adaptation of pi-shift-router)

安装

## Features

- **Instant upgrades, trend-gated downgrades** — one `smart` verdict switches to the strong tier immediately; coming back down requires a sliding-window majority (default 5 turns, ≥60%, low-confidence votes ignored).
- **Cache-aware routing** — when Fast and Smart share a provider, the router raises the downgrade threshold (0.9) and holds off while the prompt cache is warm, so switching to a cheaper model never costs more than staying put.
- **Runtime failover** — 429 / 5xx / quota failures put the model into exponential-backoff cooldown (1m → 4m → 16m → 1h → 6h cap; client-side limits start at 16m) and re-resolve the same tier to the next healthy model — same-turn retry, never cross-tier.
- **Task-level orchestration** — complex tasks (a `smart` verdict) run the Smart tier as a **CTO** that plans, delegates implementation to Fast engineer subagents via the harness's `subagent` tool, reviews each result, and iterates. The hard caps are **enforced by the plugin**, not just prompted: each delegation counts a round, each failed worker result counts an escalation, and once a cap is hit the `subagent` tool is denied outright and the system prompt switches to a "wrap up now" notice.
- **Cost telemetry** — per-tier token/throughput tracking and an optional USD pricing table (`/router stats` shows "what this session would have cost on the Smart model").
- **Zero-config startup** — a no-op until you configure tiers; then routing just works. Configuration is editable live via the GUI settings panel **and** `/router config` commands (persisted, no restart).

## Install

### As a bundle (recommended)

green-dalii/dsh-shift-router 加入你的 DSH 配置(web profile)即可启用。

README

展开完整 README 收起

README 内容为空。