guangjun-super/dsh-subagent-model-routing
Difficulty-aware model routing for DeepSeek Harness subagents: easy, medium, and hard tiers each get their own provider, model, and reasoning effort.
Automatically picks a provider, model, and reasoning effort for every subagent based on task difficulty. Three tiers — easy, medium, and hard — each carry their own provider/model/reasoning-effort config. A heuristic classifier scores the task prompt at spawn time (or on every message in reclassify:every mode). Settings hot-apply from a dedicated settings panel; no restart is needed for threshold changes.
Install
dsh plugin --profile web add dsh-subagent-model-routingInstalls as dsh-subagent-model-routing 0.1.1 into the web profile; restart is required. The plugin mounts on the host plane so it applies to every agent preset (standard, code, cordis, minimal, and any custom preset). A standalone settings panel at Settings > subagent routing lets you edit tiers and thresholds without a restart.
Compatibility
Applies to in-process subagents only. ACP, Codex, Claude Code, and dsh-sdk subagents are created outside the process and are skipped. Classification is a heuristic — long-tail tasks can be misclassified; the README suggests tuning thresholds or using the reclassify:every mode. Reasoning-effort values fall back gracefully when the target model does not support the requested level.
Details
- Repo: guangjun-super/dsh-subagent-model-routing
- Category: Agent Capabilities
- Stars: 0
- Version: 0.1.1 (npm)
Alternatives
dingminhua/dsh-subagent-default-model · mirelconstantin/dsh-subagents-model · happyren/dsh-agent-messaging