dsh-subagent-model-routing

Difficulty-aware model & reasoning-effort routing for DeepSeek Harness subagents.

A Cordis plugin that automatically picks a preset provider / model / reasoning effort for every subagent you spawn, based on how hard the task looks. A small "rename this variable" task gets a cheap, low-effort route; a "refactor the auth module and audit security" task gets the strong model with maximum reasoning.

It applies to every agent preset — standard, code, cordis, minimal, and any custom preset — because it mounts on the host plane, not inside a single preset.

中文文档 →

Features

• Three difficulty tiers — easy / medium / hard, each with its own { provider?, model?, reasoningEffort? } preset.
• Automatic difficulty scoring — heuristic signal words (Chinese & English) plus task length.
• Reasoning-effort aware — off / low / high / max, with graceful fallback when a model doesn't support the requested effort.
• Standalone settings panel — a dedicated Settings → subagent routing page (中文：设置 → 子代理路由), where you edit tiers and thresholds; saves to settings.yaml and hot-applies (no restart).
• Optional diagnostic log — one JSON line per classification, for when you can't see the console.

How it works

When an in-process subagent is created — via subagent, subagent_fork, a workflow worker, or a ralph child — the plugin scores its first task message, lands it in a tier, and swaps that subagent's model route to the tier preset. The {{model}} variable in the persona is kept in sync, and the choice survives continuable-subagent cold restarts.

Out-of-process providers (acp, codex, claude-code, dsh-sdk) never create an agent in this process, so they are skipped automatically.

Install

dsh plugin --profile web add dsh-subagent-model-routing

Then start the web UI:

dsh web

Use

1. Install and start dsh web.
2. (Optional) Open Settings → subagent routing (设置 → 子代理路由) and set each tier's Provider / Model / Reasoning effort. Provider & model are dropdowns sourced from the same model catalog as the chat model picker; leaving a field blank ("inherit") falls back to the parent route.
3. Spawn subagents as usual — routing is automatic.

From the settings panel you can configure each subagent tier's model and reasoning effort:

Configure

All settings are optional and fall back to defaults. The patch config is the baseline layer; whatever you save in the settings panel is layered on top and hot-applies.

| Key | Description | Default | |---|---|---| | tiers.easy / tiers.medium / tiers.hard | Per-tier preset { provider?, model?, reasoningEffort? }. Omitted fields inherit the subagent's parent route; omitting reasoningEffort clears the inherited effort (adapter default). | easy: { reasoningEffort: low }, medium: { reasoningEffort: high }, hard: { reasoningEffort: max } | | hardThreshold | Score ≥ this → hard | 3 | | easyThreshold | Score ≤ this → easy | -1 | | reclassify | first: classify only the first task message (follow-ups reuse the tier); every: re-classify on every message | first | | logFile | Optional path for one-JSON-line-per-event diagnostics | (none) |

Example: switch models per tier

- insert:
    - id: dsh-subagent-model-routing
      name: dsh-subagent-model-routing
      config:
        tiers:
          easy:   { provider: deepseek-official, model: deepseek-v4-flash, reasoningEffort: low }
          medium: { provider: deepseek-official, model: deepseek-v4-flash, reasoningEffort: high }
          hard:   { provider: deepseek-official, model: deepseek-v4-pro,   reasoningEffort: max }
        hardThreshold: 3
        easyThreshold: -1
        reclassify: first

Difficulty scoring (heuristic)

The first task message's text blocks are scored:

• Hard signal words +1 — e.g. refactor / migrate / debug / race condition / security / from scratch / audit / distributed … (and 重构 / 迁移 / 并发 / 安全 / 审计 / 分布式 …).
• Easy signal words −1 — e.g. rename / translate / summarize / typo / comment … (and 重命名 / 翻译 / 总结 / 拼写 …).
• Length — ≤ 120 chars −1; ≥ 800 chars +1; ≥ 2000 chars +2.

Then score ≥ hardThreshold → hard, score ≤ easyThreshold → easy, otherwise medium. classify() is exported so you can unit-test or tune it directly.

Limitations

• In-process subagents only. acp, codex, claude-code, and dsh-sdk subagents are created outside this process and are skipped.
• Heuristic, not guaranteed. Long-tail tasks can be misclassified; tune the thresholds or use reclassify: every.
• Host-half edits need a restart (Node ESM cache); client-half edits need a page refresh.

Development

git clone [email protected]:guangjun-super/dsh-subagent-model-routing.git
cd dsh-subagent-model-routing
npm test          # 16 tests: pure logic + cordis wiring simulation

The wiring tests drive a real cordis Context to simulate scope events. They load it from a local deepseek-harness checkout; if that path doesn't match your machine, set the DSH_CORDIS_CONTEXT env var to a reachable file:// URL, or the wiring block is skipped and only the pure-logic tests (classify / textOf / normalizeConfig) run.

License

MIT