What is this?
dsh-routing-suite is a three-part kit for the DSH agent runtime: a runtime injector, a task-aware reasoning-mode router preset, and a graded task protocol — all shipped in one repo.
dsh-routing-suite is a three-part kit for the DSH agent runtime: a runtime injector, a task-aware reasoning-mode router preset, and a graded task protocol — all shipped in one repo.Is this one plugin or several separate tools?One repo, three components: the injector (inject/hot-reload/uninstall dev tools), the router-standard/router-spec reasoning presets, and the graded-mode plugin with 6 tools, a red-team gate and audit endpoints.
One repo, three components: the injector (inject/hot-reload/uninstall dev tools), the router-standard/router-spec reasoning presets, and the graded-mode plugin with 6 tools, a red-team gate and audit endpoints.How do I get it installed?Use the one-click three-step chain against DSH >= 0.1.0-rc.6; the graded release adds directly via 'dsh plugin --profile web add'. Don't copy the preset folder by hand — one extra nesting layer and DSH won't discover it.
Use the one-click three-step chain against DSH >= 0.1.0-rc.6; the graded release adds directly via 'dsh plugin --profile web add'. Don't copy the preset folder by hand — one extra nesting layer and DSH won't discover it.What needs to be on my machine first?A DSH runtime (target >= 0.1.0-rc.6) and Node.js >= 22; the injector's git assembly auto-builds through a prepare hook, so no separate build steps are listed.
A DSH runtime (target >= 0.1.0-rc.6) and Node.js >= 22; the injector's git assembly auto-builds through a prepare hook, so no separate build steps are listed.Is DSH itself stable to build on right now?Not fully — DSH is officially a developer preview with declared breaking changes, and router-pro is planned but not shipped with the current v0.3.0 preset.
Not fully — DSH is officially a developer preview with declared breaking changes, and router-pro is planned but not shipped with the current v0.3.0 preset.Does it run on all three desktop OSes?Yes — it's JavaScript on Node.js >= 22, and the repo is tagged for Windows, macOS and Linux.
Yes — it's JavaScript on Node.js >= 22, and the repo is tagged for Windows, macOS and Linux.How does the router change how the model reasons?Each task is classified into a behavior band: spec (plan-collective), react (executor) or mixed (a trap to avoid), with weak letting the model self-classify. Fixed guidance is injected alongside every user message (92-94% cache hits), and personas are tuned per model.
Each task is classified into a behavior band: spec (plan-collective), react (executor) or mixed (a trap to avoid), with weak letting the model self-classify. Fixed guidance is injected alongside every user message (92-94% cache hits), and personas are tuned per model.Why did the authors build the graded protocol?They measured that models slack off on long tasks: without the protocol, late-stage sessions dropped to '1 tool call / 0 image reads', while protocol chains stayed engaged. With the triple anchor (review + convergence + anti-drift), open-task completion went from 0% to 100%.
They measured that models slack off on long tasks: without the protocol, late-stage sessions dropped to '1 tool call / 0 image reads', while protocol chains stayed engaged. With the triple anchor (review + convergence + anti-drift), open-task completion went from 0% to 100%.What does a graded-mode session actually enforce?A session-level two-tier protocol: brainstorm alignment questions, a north-star goal sign-off, a spec'd plan, check-ins along the way, then group wrap-up and final verification — guarded by a red-team gate and audit endpoints.
A session-level two-tier protocol: brainstorm alignment questions, a north-star goal sign-off, a spec'd plan, check-ins along the way, then group wrap-up and final verification — guarded by a red-team gate and audit endpoints.Can I re-run their measurements on my own sessions?Yes — 'node scripts/measure.mjs your-session.jsonl' recomputes the same diligence metrics from any session log and outputs sanitized numbers only, so you can verify the laziness findings yourself.
Yes — 'node scripts/measure.mjs your-session.jsonl' recomputes the same diligence metrics from any session log and outputs sanitized numbers only, so you can verify the laziness findings yourself.What are similar projects?dsh-router-standard, oh-dsh, dsh-anchored-standard, DSH-Plugins-Marketplace, dsh-lowtide, dsh-plugin-bridge, dsh-purge, dsh_desktop
dsh-router-standard, oh-dsh, dsh-anchored-standard, DSH-Plugins-Marketplace, dsh-lowtide, dsh-plugin-bridge, dsh-purge, dsh_desktopWhich topics is it in?model-router, cordis, reasoning, red-team, dsh, dsh-plugin, deepseek-harness, prompt-engineering, deepseek, plugin, research, agent
model-router, cordis, reasoning, red-team, dsh, dsh-plugin, deepseek-harness, prompt-engineering, deepseek, plugin, research, agent