Introduction

In DeepSeek Harness’s goal-mode, a parent agent may delegate multiple background subagents to perform reviews or auxiliary tasks while a goal is still active. The problem is that the parent agent can call update_goal(action=complete) while subagents are still running, and the results are returned later, causing these results to be too late to influence the already completed goal.

dsh-goal-quiescence provides a completion evidence gate for this scenario. It does not replace Harness goal mode, nor is it responsible for scheduling, canceling, or retrying subagents; it constrains the goal completion action itself: it refuses to mark the goal as complete until the observed subagent run has ended and the parent agent has already read the terminal output.

What is it

dsh-goal-quiescence is a community plugin for DeepSeek Harness, positioned as a completion evidence gate for goal-mode subagents.

The maintainer is 1052326311, and the license is MIT. It solves the problem where the parent agent should not directly complete the goal while the subagent evidence is still running, or has not yet explicitly returned to the goal agent.

Core Capabilities

Below is an introduction to what it specifically constrains.

  1. Records subagent lifecycles.

For subagent runs observed after the plugin is loaded and under an active goal, the plugin records the public subagent/start and subagent/end lifecycle pair.

  1. Prevents premature completion.

Before the aforementioned subagent runs have settled, the plugin rejects:

update_goal(action=complete)
  1. Requires the parent agent to acknowledge terminal output.

For each run that has already settled, the goal agent needs to call:

goal_quiescence_ack

This tool returns the terminal assistant output for the corresponding run and returns it as a tool result to the parent agent’s current context.

  1. Queries runs that are still blocking completion.

You can call:

goal_quiescence_status

It returns a bounded list listing runs that are still blocking goal completion.

  1. Clearly defined boundaries.

This plugin does not:

  • Schedule subagents
  • Cancel subagents
  • Retry subagents
  • Replace Harness goal mode

Installation and Enablement

Runtime environment requirements:

Node.js >=22.19.0

The plugin requires the normal Harness services provided by the Web profile:

goal
subagent
tools

Use a locally packaged tarball during installation:

pnpm install
pnpm pack
dsh plugin --profile web add ./dsh-goal-quiescence-<version>.tgz

First execute pnpm install to install dependencies, then execute pnpm pack to generate the local plugin package, and finally use dsh plugin --profile web add to add the plugin to the Web profile. The <version> in the filename must match the packaged artifact.

Typical Usage

The following flow is suitable for use in goal-mode.

  1. Progress the goal normally.

Work normally in goal mode and delegate background reviews as needed.

  1. Handle the case where completion is rejected.

If calling update_goal(action=complete) is rejected, first call:

goal_quiescence_status

Check which runs are still blocking completion.

  1. Wait for runs to end.

Wait for each run in a running state to settle.

  1. Read terminal output.

For each runId returned by goal_quiescence_status, call:

goal_quiescence_ack

Then check the terminal output it returns.

  1. Handle issues and complete the goal.

If the terminal output reports issues that need to be handled, solve the issues first; after handling is complete, call:

update_goal(action=complete)

Scenarios and Considerations

Suitable scenarios:

  • Using subagents to execute background tasks in goal-mode
  • Want to confirm subagents have settled before completing the goal
  • Want to explicitly read back subagent terminal assistant output to the parent agent context
  • Do not want the plugin to automatically schedule, cancel, or retry subagents

Considerations:

  • This is a process-local lifecycle policy.
  • It covers subagent runs started after the plugin is loaded.
  • It covers completion attempts initiated via the model-facing update_goal tool.
  • Direct service calls, process restarts, or child runs that started before the plugin was enabled are not within its observable boundaries.
  • It does not claim to be an atomic core transaction. If end-to-end atomic guarantees are needed, core lifecycle permission is required.
  • The plugin runs with the current dsh process permissions. Before installation, you should check the GitHub repository source code and MIT license to avoid introducing unreviewed community plugin packages.

Conclusion

The value of dsh-goal-quiescence lies not in increasing subagent execution capabilities, but in binding the goal completion action with subagent evidence reading: before completion, confirm the run has ended; when confirming, the terminal output must be brought back to the parent agent context.

Directory page:

https://www.skillhub.cn/plugins/1052326311/dsh-goal-quiescence

GitHub repository:

https://github.com/1052326311/dsh-goal-quiescence