yjh051108/dsh-routing-suite

One repository for the full stack: the runtime injector (restart-free plugin management) plus the reasoning-mode routing presets (task-aware reasoning-mode routing, measured P1-P23) plus the graded task protocol (brainstorm quiz → speced plan → check-in → group close-out → final check, with redteam gate & audit endpoint).

中文 | English

Research & Data (graded)

"Models slack off too — we measured it, and built defenses." — public research slot for the graded plugin:

• 🔬 Model Attention Slack: Observation Notes — real segment-level data of verification-effort decay in long tasks (no-protocol chain: late-stage "1 tool / 0 image checks" vs protocol chain fully on-line) + defense mapping
• 📊 Protocol vs No-Protocol: Measured Comparison — 3-session comparison table + ASCII charts (same-type 3D tasks)
• 🧪 Measurement tool — node scripts/measure.mjs your-session.jsonl: recompute the same metrics on your own sessions (sanitized: numbers only)
• 📁 Raw data — sanitized per-session metric tables + decay profile + audit metrics

Component entry points

Component	Description	Entry
injector	Runtime injector: dev_* tool family (inject / hot-reload / uninject / staging / route self-heal)	injector/README.md · SPEC
preset	Reasoning-mode routing presets (router-standard / router-spec, persona-based routing)	preset/README.md · experiments
graded	Graded mode: session-level two-level task protocol (6 tools + 3 modes + per-item mode + redteam gate + audit endpoint + super panel)	graded/README.md · ARCHITECTURE · THEORY · DATA

Install chain (three steps)

# 1. Clone the suite (single repo: injector/preset content is stored directly, no submodules)
git clone https://github.com/yjh051108/dsh-routing-suite.git
cd dsh-routing-suite

# 2. One-shot install (injector assembly + preset copy + layout self-check + restart prompt)
.\install.ps1

Or manually:

# Step 1: assemble the injector (official assembly; bundles take over after restart)
dsh plugin --profile web add .\injector
# If dsh is not on PATH (npx @deepseek-ai/dsh web): npx '@deepseek-ai/dsh' plugin --profile web add .\injector

# Step 2: install the router presets (one or both; DSH scans one level only,
# so each preset directory must sit FLAT under .agent-presets)
$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

# Step 3: restart DSH → pick Router Standard / Router Spec in a new session

Do NOT copy the preset directory as a whole — the extra nesting hides the presets from DSH discovery.

Components

Path	Repo	Version	Role
injector/	dsh-super-injector	v0.3.3	Runtime injector: dev_* tool family (inject / hot-reload / staging-promote / uninject / route self-heal); git installs build automatically (prepare hook)
preset/	dsh-router-standard	v0.3.0 … mainline v1.19.1/v34	Reasoning-mode routing presets: router-standard (classified persona + full sections) / router-spec (deep-think-first). router-pro is planned but NOT part of v0.3.0
graded/	dsh-graded-mode	v3.2.0	Graded mode: session-level two-level task protocol (brainstorm quiz alignment → North-Star finalization → speced plan → spec-driven injection → check-in → group close-out → final check); 6 tools + 3 modes + per-item mode + redteam gate + audit endpoint; graded/dsh-graded-mode-3.2.0.tgz is a direct dsh plugin --profile web add target

Versions follow each component repo's git tag (links go to the matching Release).

The three components now evolve together with this repository (injector/, preset/ and graded/ are ordinary directories whose contents are committed directly); the standalone upstream repos dsh-super-injector / dsh-router-standard stay for independent releases and may later become mirrors/archives. The preset install directory is preset/router-standard (flat, no extra nesting).

router-standard preset capabilities (P1–P23 measured summary)

• Three behavior bands + weak internal routing: spec (plan-collective) / react (doer) / mixed (trap, avoided) / weak (model self-classifies)
• Model-matched persona: Pro = spec sentence + few-shot (discrimination +5.0); Flash = neutral + classify (+5.7)
• Near-field guidance: fixed guidance after every real user message (cache 92–94% hit), routing 96% + convergence 100% + anti-dilution
• Single-task three anchors (persona-static): recall + converge + anti-runaway — open-task completion 0% → 100%
• plan-mode preserved: only the persona section is replaced; plan boundaries never lose focus
• AI self-optimization tools: dev_router_status / dev_router_mode / dev_mode_subagent

v0.3.0 changes (real-assembly-chain fixes)

• First-turn routing actually works (#13): the first real user message is captured via agent/inbox/claimed BEFORE assembly, so the first request is classified (previously every session's first request was unconditionally weak)
• Near-field guidance moved to agent/pre-step (#34/#36/#55): the guide now rides the SAME request as the user message — reachable on the real chain, and it no longer manufactures a second API call per user message (the 2× cost spike)
• Missing imports (#11), preset.yml YAML quoting (#53), full sections after promotion (#44), install script & docs fixes (#35/#42/#41), injector git-install auto-build (#40)

Docs

• Injector guide (10 rules): injector/README.md
• Routing preset paper & experiments: preset/docs/paper.md + preset/docs/experiments.md (P1–P23)

License

MIT. Credits: xiaobright/modeltest (V4.1b evaluation), xiaobright/dsh-anchored-standard (anchoring mechanism).