dsh-routing-suite
Manifest valid★ 7kdsh-routing-suite — injector + router-standard kit: install the runtime injector first, then the task-aware reasoning-mode router preset (measured P1-P23).
dsh-routing-suite — Injector × Thinking-Mode Routing × Graded Mode Bundle
A single repo delivers the full three-piece set: runtime injector (no-restart runtime management layer) + thinking-mode routing preset (task-aware reasoning mode, P1-P23 tested) + graded task protocol (brainstorm question → specified plan → check-off regime → group closeout → final verification, with red-team gate and audit endpoint).
Component Entry Points (Quick Jumps)
| Component | Description | Entry |
|---|---|---|
| injector | Runtime injector: full dev_* tool suite (inject / hot-reload / uninstall / sidecar promotion / routing self-heal) | injector/README.md · Docs |
| preset | Thinking-mode routing preset (router-standard / router-spec, routing by model persona) | preset/README.md · Experiment Report |
| graded | Graded mode: session-level two-tier task protocol (6 tools + three modes + sub-class mode granularity + red-team gate + audit endpoint + super panel) | graded/README.md · Architecture · Theory · Test Data |
Research and Data (graded)
"Models slack off too — measured and defended against" — the graded-mode plugin's public research slot:
- 🔬 Model attention slacking: observation notes — real segment-level data showing diligence decay in long-horizon tasks (late-stage "1 tool/0 reads" vs. protocol chain fully online) + counter-mechanism comparison
- 📊 Protocol vs. no protocol: measured comparison — same-shape 3D task, three-session comparison table + ASCII charts
- 🧪 Measurement tool —
node scripts/measure.mjs 你的会话.jsonl: anyone can recompute the above metrics against their own sessions (sanitized: numbers only) - 📁 Measured data — three-session sanitized metrics table + slacking-intensity table + audit criteria
One-Click Install
dsh plugin --profile web add github:yjh051108/dsh-routing-suite
This bundle already ships the three components above (injector/preset/graded are regular directories inside the repo, with content checked in directly); graded artifact:
graded/dsh-external-dsh-graded-mode-0.0.1-rc1.tgz(or a Release attachment).
DSH Target: >=0.1.0-rc.6 <0.2.0 (tracking rc.8 / 0.1.1-rc.2 / 0.1.2-alpha.1)
DSH is currently in developer preview; the official docs explicitly state that breaking changes will occur. See
preset/CHANGELOG.mdin this repo for the version-tracking record.
Install Chain (Three Steps)
# 1. 拉套装(单仓库:injector/preset 内容已直接入库,无需 submodule)
git clone https://github.com/yjh051108/dsh-routing-suite.git
cd dsh-routing-suite
# 2. 一键安装(注入器装配 + 预设复制 + 布局自检 + 提示重启)
.\install.ps1
Or manually:
# 步骤 1:装配注入器(官方装配,重启后由 bundles 接管)
dsh plugin --profile web add .\injector
# dsh 不在 PATH 时:npx '@deepseek-ai/dsh' plugin --profile web add .\injector
# 步骤 2:安装 router 预设(每个预设目录平铺复制到 .agent-presets 下,DSH 只扫一级子目录)
$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\router-standard'
Copy-Item -Recurse .\preset\router-standard $target
$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\router-spec'
Copy-Item -Recurse .\preset\router-spec $target
# 步骤 3:重启 DSH → 新会话选择 Router Standard / Router Spec (experimental)
Note: don't copy the whole
presetdir (adds a nesting level DSH can't detect).
Components
| Path | Repo | Version | Role |
|---|---|---|---|
injector/ | dsh-super-injector | v0.3.3 | Runtime injector: the full dev_* tool suite (inject / hot reload / sidecar promotion / uninstall / route self-healing); github: assembly is auto-built by the prepare hook |
preset/ | dsh-router-standard | v0.3.0 … mainline v1.19.1/v34 | Thinking-mode routing presets: router-standard (categorized persona + full sections) / router-spec (deep-thinking first). router-pro is planned and was not released with v0.3.0 |
graded/ | dsh-graded-mode | v0.0.1-rc1 | Graded mode: session-level two-phase task protocol (brainstorming-choice alignment → north-star finalization → spec planning → spec-based injection → checkpoint system → team wrap-up → final acceptance); 6 tools + three modes + sub-category mode granularity + red-team gate + audit endpoint; graded/dsh-graded-mode-3.2.0.tgz can be directly dsh plugin --profile web add'd |
Version numbers follow each component repo's git tag (the Version column links straight to the corresponding Release).
The three components evolve together with this repository (injector/, preset/, and graded/ are now plain directories in the repo, with content committed directly); the upstream standalone repositories are retained for independent releases and can later be converted to mirrors/archives. The preset installation directory is preset/router-standard (already flattened, no extra nesting).
router-standard Preset Capabilities (P1-P23 Empirical Summary)
- Three behavior bands + weak internal routing: spec (planning-collective) / react (executor) / mixed (trap, avoid) / weak (model self-classifies)
- Persona selection by model: Pro=spec phrasing+few-shot (distinctiveness +5.0); Flash=neutral+classify (+5.7)
- Close-range guidance: fixed guidance injected after each user message (92-94% cache hit rate), routing 96% + convergence 100% + anti-dilution
- Single-task three anchors (static persona): review + convergence + off-topic resistance —— open-task completion rate 0% → 100%
- plan-mode preservation: only the persona section is replaced; plan boundaries don't lose memory
- AI self-optimization tools:
dev_router_status/dev_router_mode/dev_mode_subagent
v0.3.0 Changes (Real Assembly Chain Fixes)
- First-turn routing now really works (#13): the first real user message is captured via
agent/inbox/claimedbefore assembly, so the very first request is already classified by task (previously, the first turn of every session was unconditionally weak) - Close-range guidance moved to
agent/pre-step(#34/#36/#55): guidance is now injected in the same request as the user message, making it reachable on the real chain, and it no longer triggers an extra second API call (previously each round added one more call = 2× cost) - Missing-import fix (#11), preset.yml YAML quoting (#53), full regression after promotion (#44), install-script and documentation fixes (#35/#42/#41), automatic build of injector git assembly (#40)
Docs
- Injector bootstrap (10 iron rules of the spec):
injector/README.md - Routing preset paper and experiments:
preset/docs/paper.md+preset/docs/experiments.md(P1-P23) - Repository structure migration (submodule → direct files): docs/FLATTEN-MIGRATION.md
License
MIT. Acknowledgments: xiaobright/modeltest (V4.1b evaluation), xiaobright/dsh-anchored-standard (anchoring mechanism).
Comments
Loading…
Similar plugins
by CARVIN94
An AI routing gateway that runs as a DSH plugin. Adds a Routing System sidebar panel for suppliers, account pool, fallback combos and API keys, and serves an OpenAI-compatible endpoint at http://local
★ 15
↓ 725/wk
MIT
TypeScript
Oct 28, 2026
dsh plugin --profile web add dsh-router-coreby yjh051108
推荐组件(非必须):DeepSeek Harness 运行时注入器;已随 dsh-routing-suite 单仓库化保留,本仓库继续维护/发布。
★ 167
TypeScript
Oct 16, 2026
dsh plugin --profile web add @dsh-external/dsh-super-injectorby zhangzhangco
Automatic tier-based model routing for DeepSeek Harness (dsh): a virtual `smart` model classifies every request by difficulty (hard / normal / easy) and by vision need, then delegates it to the models you already configured.
★ 0
NOASSERTION
JavaScript
Sep 22, 2026
dsh plugin --profile web add dsh-tier-routerby neo805
DSH plugin: context-aware per-step reasoning-effort routing — classify → escalate → hysteresis → capability clamp; Auto mask in the native selector; default effort fill + wire mapping for custom llm-pi-ai models. AI-authored (kimi-k3), produced by @neo805. No Issues/PRs; forks welcome (MIT).
★ 0
NOASSERTION
JavaScript
Oct 14, 2026
dsh plugin --profile web add dsh-effort-routerby CTWCTW9999
Auto Mode for DSH: a flash-model router that auto-picks the best agent preset (standard/code/minimal/cordis) for each new session and executes immediately, no confirmation.
★ 0
MIT
JavaScript
Aug 18, 2026
dsh plugin --profile web add dsh-agent-preset-routerby green-dalii
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)
★ 4
↓ 237/wk
MIT
TypeScript
Oct 18, 2026
dsh plugin --profile web add dsh-shift-router