Introduction

In DSH (DeepSeek Harness), sub-agent dispatch usually isn’t just “sending a prompt”. If different sub-tasks require different models, reasoning strengths, or tool scopes, and you also need to review what actually took effect with this dispatch, whether it passed safety checks, and how the cost estimation looks, you need a plugin that can explicitly manage dispatch strategies.

muzyLink/dsh-subagent-profile is the DeepSeek Harness sub-agent dispatch plugin: it selects models, reasoning strengths, and tool scopes for sub-agents based on tasks, saves common combinations as named profiles for reuse, and provides safety checks, cost estimation, and savings analysis, as well as a complete decision ledger.

What is this

dsh-subagent-profile is maintained by muzyLink and is licensed under MIT.

It revolves around a core goal: to ensure every sub-agent dispatch has reusable profiles, checkable safety constraints, readable cost estimates, and traceable decision records. Specifically, it places “which model to use, what reasoning strength, which tools to give, whether to run in the background, how cost is estimated, and how results are recorded” within the same set of plugin capabilities.

Core Features

Below are the capabilities it explicitly provides:

  • Select models, reasoning strengths, and tool scopes for sub-agents based on tasks.
  • Save common combinations as named profiles for reuse at any time.
  • Provide the dispatch tool to dispatch sub-agents.
  • Built-in whitelist, cost, intersection, approval, and budget safety checks and logging.
  • Provide cost estimation and savings comparison.
  • Record decision ledgers for request vs. effective, gates, execution, and settlement.
  • Provide a settings page and a per-session ledger tab.
  • Self-installed orchestrator agent preset.
  • The tool set takes the intersection of the profile and parent, and always removes run_code.
  • Support background dispatch, real-time status, notification center, dispatch suggestions, and sub-agent header badges.

Installation and Enablement

First, install the plugin:

dsh plugin --profile web add dsh-subagent-profile

After installation, restart dsh web and select the self-installed orchestrator preset in the new session preset picker.

Environment requirements are as follows:

  • Node: ^22.19.0 || >=24
  • peerDependencies: @deepseek-ai/dsh-agent, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-subagent, @deepseek-ai/dsh-tools, with version range >=0.1.0-rc.6 <0.2.0, and all marked as optional.

Uninstalling the plugin will remove related data files and the self-installed orchestrator preset directory; reinstalling or restarting will resync presets and generate data files.

Typical Usage

1. Manage profiles in the settings page

Profiles are managed in the settings page. Each profile contains:

  • preset
  • model
  • reasoning effort
  • tool scope
  • optional persona

Two built-in profiles are available: swap-standard and researcher.

Profile files are located at:

~/.dsh/subagent-profiles.json

After editing from the settings page, the configuration takes effect immediately.

2. Dispatch sub-agents using the dispatch tool

Example:

dispatch(
  profile: "researcher",
  prompt: "Survey the DSH plugin ecosystem and compare direct competitors",
  run_in_background: true
)

Here, you specify using the researcher profile as the combination of model, reasoning strength, and tool scope during dispatch, and put the task into the background for execution.

3. View decisions in the per-session ledger tab

Each session has a per-session ledger tab. It records dispatch decisions:

  • request vs. effective
  • gates
  • execution
  • settlement

Used to review the actual effective configuration, execution process, and cost settlement of this sub-agent dispatch.

Safety Checks and Constraints

The default constraints of the plugin are as follows:

  • The tool set takes the intersection of the profile and parent, and always removes run_code.
  • Delegation does not exempt host approval; operations requiring approval are automatically rejected.
  • Non-official presets can be enabled per preset, but are disabled by default; every allowance is audited.

Regarding costs, please note:

  • Cost figures are estimates based on the average per-dispatch price for each model, not fine-grained per-token pricing.
  • Cost estimation and savings analysis are for comparison and should not be directly used as precise billing results.

Background dispatch has dependencies:

  • Requires @deepseek-ai/dsh-jobs and @deepseek-ai/dsh-tool-jobs to be loaded.
  • Otherwise, it will fail with the message: dispatch: Background dispatch unavailable: jobs service missing.

There are also differences in running modes:

  • In Continuable mode, preset swap and reasoningEffort are ignored; the sub-agent inherits the parent session preset and uses the default reasoning strength.
  • In One-shot mode, the final tool set is determined by dsh, and the plugin cannot read the final restricted results.

Applicable Scenarios and Considerations

Suitable for these situations:

  • You are already dispatching sub-agents in DSH but want to use different models and reasoning strengths for different tasks.
  • You want to narrow tool scopes based on tasks while keeping the default constraint that run_code is always removed.
  • You need to view records of request, effective, gates, execution, and settlement for every dispatch.
  • You want to manage common combinations using profiles rather than writing out the full configuration every time.

Please note before installation:

  • The plugin runs with the permissions of the current dsh process; it is recommended to check the source code, dependencies, and license first.
  • The license for this project is MIT.
  • If you rely on background dispatch, confirm that @deepseek-ai/dsh-jobs and @deepseek-ai/dsh-tool-jobs are loaded.
  • If you rely on the final behavior of One-shot or Continuable, note that the plugin cannot read or override final results in certain modes.

Conclusion

The value of dsh-subagent-profile lies in managing model selection, reasoning strength, tool scope, safety checks, cost estimation, and decision ledgers within the same set of profiles and ledgers during sub-agent dispatch.

GitHub Repository

Project Directory Page (from plugin clue, not fully verified)