Introduction

When engaging in multi-model collaboration within DeepSeek Harness (DSH), the common practice is to manually select a default model or temporarily switch providers within a session. Once a task is decomposed into multiple dependent work packages, aligning costs, quality floors, and execution boundaries (read-only preview vs. writable workspace, or whether to use vendor CLIs) often requires mental bookkeeping. There has been a lack of a local workspace to plan, compare routing schemes, and then execute.

dsh-model-router (npm package name @ljwei-stak/dsh-model-router, repository Alice-Marx/dsh-model-router) is a community-maintained DSH plugin by Alice-Marx, categorized under model inference capabilities in the community directory. It integrates “task planning, routing assignment, dependent work package execution, and official CLI discovery and installation” into a single workbench, emphasizing cost previews and explicit execution permissions. Since version 0.12.0, the routing plugin and GAL have been split into independent plugins; the GALGame capabilities mentioned in the intro require separate installation via @ljwei-stak/dsh-galgame. This article discusses only the model-router.

What is this

In one sentence: A cost-aware model routing workbench for DeepSeek Harness Desktop—performs local task classification and complexity analysis, compares configured routes, checks execution permissions, and then runs dependent work packages via model APIs or supported official vendor CLIs.

The current release version is 0.22.2, licensed under MIT. The plugin entry point is the apply(ctx) function exported from .dsh-plugin/index.mjs within the package. The README declares host compatibility for DeepSeek Harness Desktop 0.2.0-rc.1, 0.2.0-rc.2, and 0.2.1-rc.2 (including Hermes/Pi installation and terminal support); compatibility beyond this range is not declared in the materials.

Core Features

The following summarizes current behavior based on the README capability table, without expanding on interface details not present in the source materials.

Task Planning & Routing

  • Task Planning: Local task classification and complexity analysis; provides single-route recommendations, dependent team work packages, or explicit specification of configured models.
  • Model Configuration: Reads precise provider/model routes from Harness; optional supplements for quality, USD pricing, specialization, CLI model names, and execution/billing preferences.
  • Routing Presets: Economy, Balanced, Quality-first presets; supports quality floors, estimated token costs, and bounded allocation search; keeps pricing as unknown when unit prices are missing.
  • Routing Algorithms (0.22.x): Supports task-relevant quality evidence, joint model and effort selection, historical P90 resource estimation, and dependency-aware search; algorithms recommend or manually specify API/official tools for each decomposed package, preserving dependencies and refreshing cost previews and retry targets. Manual hybrid execution is read-only; tools use their configured default models, and editable official team workflows remain available.

Execution & Official Tools

  • Read-only Execution: Preview and confirm routing runs in the workbench, or invoke within an official Harness session; supported headless CLIs and model directory API calls return logged results.
  • Editable CLI Execution: Workbench sequential CLI teams or single supported CLI within a session; uses approved isolated Git worktrees and conditional patch integration. Selecting a cli-native editable team runs under normal OS user permissions for the vendor CLI; worktrees collect edits but do not restrict OS access; CLIs retain their respective configured models unless explicitly mapped.
  • Official Tools: Detection, installation/updates, launch readiness checks, and login status checks covering twelve tools in the registry; installation tasks feature logging and cancellation.
  • Interactive Terminal: PowerShell/System shell and twelve registered CLI targets (including ZCode TUI on verified Windows Desktop installations), with pinned login commands, live output, and multi-session support.

Cost, History & Boundaries

  • Cost & Subscriptions: Daily/monthly budget checks, economy degradation or pausing, subscription-priority routing, quota cooldowns, and handling of other subscription failures.
  • History & Quality: Per-package result logging, dependency views, failure details, rerun/reassignment, scoring, and optional review by stronger models.
  • Execution Boundaries: Displays the read/write scope for each route and restrictions such as whether the process uses the Harness sandbox (the full expression in the materials’ table was not truncated; refer to the README for details).

The planning phase is completed locally without initiating model calls; the execution phase sends selected tasks/contexts to the chosen API or vendor CLI, which may incur costs.

Installation & Enablement

The installation command must use the npm package spec from the materials; do not construct dsh plugin add github:... based on the repository name. A standalone npm install -g will not automatically register the Desktop plugin.

Install via DSH CLI (web profile):

dsh plugin --profile web add @ljwei-stak/dsh-model-router@0.22.2

Use in Harness Desktop:

Launch Desktop once to initialize the desktop profile, fully exit Desktop, then execute:

dsh plugin --profile desktop add @ljwei-stak/dsh-model-router@0.22.2

Alternatively, paste the package spec above in Harness Desktop → Plugins → Add plugin. A Desktop restart is required after installation.

Configuration: Importing user-written JSON/YAML configuration files is not supported. Configure providers, models, and credentials in the Models page of Harness; configure routing profiles and preferences in the plugin settings. Routing persistent state is located at $DSH_HOME/model-router/state.json (default ~/.dsh/model-router/state.json). Vendor CLIs retain their own logins and configurations; the router does not copy API keys into model profiles or terminal logs.

Typical Usage

The materials point to more complete walkthroughs for reference:

The reproducible mainline is as follows:

  1. Configure provider/model and credentials (and optional pricing, quality, etc.) on the Models page.
  2. Open the plugin workbench, input or import tasks, perform local classification and decomposition, and view recommended routes or manually assign each work package.
  3. Compare cost previews and quality floors under Economy/Balanced/Quality presets, and confirm execution permissions (read-only preview, workspace-write, or CLI native environment).
  4. Execute after confirmation in the workbench, or trigger a routed run in an official session; if vendor CLIs are needed, complete detection, installation, and login checks using the official tool flow within the plugin first.
  5. View per-package results, dependencies, and failure details in the history view, and rerun or reassign as needed.

The routing workflow diagram in the README summarizes the execution flow; the 0.20.0 Windows validation report records evidence of real vendor execution and limitations. The materials note that version 0.22.2 currently has no verified end-to-end vendor runs; workbench screenshots use fixture data and are simulated UI previews rather than live runs.

Scenarios & Notes

Who it is for: Teams or individuals who have configured multi-model routing on DSH Desktop and need to decompose complex tasks into packages and allocate them by cost/quality; scenarios requiring unified discovery and installation of twelve registered official CLIs and performing read-only or editable execution within Harness.

Permissions & Risks: The plugin runs with the permissions of the current DSH process; review the repository source code and MIT license before installation, and assess data export and billing terms independently. The default Harness restricted environment distinguishes between read-only and approved workspace-write; explicitly selecting a cli-native editable team grants permissions equivalent to the current OS user and does not run inside the Harness sandbox. External providers and CLI services are subject to their respective data processing and billing terms.

Ecosystem Note: The DSH community plugin directory (such as the skillhub directory page) is an independent site with no official affiliation with DeepSeek / High-Flyer; the “Model Inference” classification comes from community categorization and is not an official DSH app store category.

Conclusion

dsh-model-router consolidates local planning, cost-aware routing, dependent work package execution, and official CLI operations into a single workbench, suitable for systematically managing multi-model tasks on Harness Desktop rather than switching models ad-hoc in a single session. Current version 0.22.2, GitHub repository approximately 205 stars (subject to the directory/repository page).

  • Directory Page: https://www.skillhub.cn/plugins/Alice-Marx/dsh-model-router
  • GitHub: https://github.com/Alice-Marx/dsh-model-router
  • npm: https://www.npmjs.com/package/@ljwei-stak/dsh-model-router