Introduction¶

It is already feasible to decompose a goal into multiple sub-agents in DeepSeek Harness (hereinafter referred to as DSH), but the process is often not transparent enough: sub-sessions run independently, and when issues arise, one can only check the logs; there is a lack of an evidence-based answer to where the difference lies between an approved plan and actual execution; after a failure, when attempting to retry, it is unclear which work can be reused.

dsh-product-subagent-console targets this workflow: a plan that is reviewable first, followed by actual execution; when deviations occur, pinpointing the specific step; and finally preparing a safe recovery. Below, I will introduce its features, installation method, and basic usage.

What is this¶

dsh-product-subagent-console is an independent community plugin maintained by Jokasa7, not an official DSH component. It uses the MIT license and is currently at version 0.9.0.

It adds a draggable Subagents workspace within a DSH session, used to design multi-agent execution, observe real sub-session trees, check what actually happened, and prepare the next recovery actions. The interface language follows the current DSH language settings, and documentation and screenshots are provided in both Chinese and English.

Four Modes¶

Plan: A Reviewable Plan First¶

Convert the goal into editable roles, tasks, dependencies, Providers, tools, checks, and budgets, and perform a preflight check before approval. What you approve is a specific plan revision, not a vague instruction.

Live: View the Real Parent-Child Tree¶

Display the spawned sub-sessions and runs from compatible Providers as a real parent-child tree, allowing you to view tasks, lifecycle states, timing, and sub-session relationships on the same canvas.

Deviation: Locate the First Evidentiary Inconsistency¶

Map approved tasks to actual execution attempts, replay the authoritative event timeline, and find the first inconsistency supported by evidence, rather than guessing based on impressions.

Recovery: Prepare the Safe Next Step¶

Preview which work will be retried and which can be reused, ask questions based on execution facts, export a sanitized Run Capsule, and extract reusable workflow candidates from verified runs.

Installation¶

The plugin is installed locally via GitHub Release, not using an online source. First, download the .tgz and SHA256SUMS.txt from a matching release, then verify the package integrity:

sha256sum --check SHA256SUMS.txt

Then execute the installation according to the profile used. DSH Desktop runs in the Open DSH Terminal:

dsh plugin add ./dsh-product-subagent-console-0.9.0.tgz

For the standard Web profile, use:

dsh plugin --profile web add ./dsh-product-subagent-console-0.9.0.tgz

After installation, restart DSH Desktop or the corresponding Web profile.

Enable Planning and Execution¶

The plugin provides Live observation by default; to execute the task graph directly within a plan, you need to enable the planner tool. The method is to copy an Agent Preset, add the planner tool, save it, and then create a new session using that preset:

- id: agent-planner
  name: dsh-product-subagent-console/plan-tool
  config:
    toolName: design_subagent_plan
    executeToolName: execute_subagent_plan

Preset changes only take effect for new sessions. When the planner tool is not enabled, Live observation for standard compatible delegation remains available.

60-Second Trial¶

After enabling the planner tool, open Subagents → Plan and use the read-only three-agent example provided in the README:

Design a read-only three-agent workflow. Agent A reads the project positioning from README.md; Agent B reads the name and version from package.json; Agent C only uses the results of A and B to return three facts. Do not modify files. Let me review the plan before execution.

The process is: Generate plan → Review roles and boundaries → Preflight → Approve the specified revision → Confirm execution, then switch between the Live, Deviation, and Recovery views to inspect the same run from three perspectives.

The README states that the v0.9 build has been validated through 4 successful read-only workflow runs, 15 observed sub-sessions, 30 test files / 245 tests, and package/installation smoke checks.

Applicable Scenarios and Notes¶

Suitable Scenarios¶

The README lists the following categories:

Parallel code, documentation, or repository reviews where ownership and dependencies need to remain visible;
Phased implementation followed by independent verification or integration;
Diagnosing stalled, incomplete, unexpected branching, or runs lacking evidence;
Turning repeatedly verified workflows into reusable starting points without automatically running them.

Compatibility¶

DeepSeek Harness Web or Desktop profile 0.1.1-rc.2
Node.js ^22.19.0 or >=24.0.0
Plan execution requires a compatible subagent Provider; external coding Agents require the corresponding Provider Bundle

Privacy and Permissions¶

Goals, responsibilities, task briefs, completion criteria, and resource names in the plan are user inputs; do not fill in credentials or private information. See docs/data-and-privacy.md in the repository for details.

It is worth noting: this plugin runs with the permissions of the current dsh process. It is an independent community project. Before installation, it is recommended to check the source code and license (MIT, see THIRD_PARTY_NOTICES.md for third-party notices) to ensure it is acceptable.

Uninstallation¶

dsh plugin remove dsh-product-subagent-console
# or: dsh plugin --profile web remove dsh-product-subagent-console

Restart the corresponding profile after removal. Uninstallation does not automatically delete local Foundry data; see the Data and Privacy documentation for handling. For vulnerabilities, please report via the private channel in SECURITY.md; do not use public Issues.

Summary¶

dsh-product-subagent-console puts the planning, observation, comparison, and recovery of multi-agent execution into the same DSH session, with the core value being that every step of judgment is based on execution facts. You can complete the installation, enablement, and trial through the steps above.

GitHub Repository: https://github.com/Jokasa7/dsh-product-subagent-console
Community Directory Page: https://www.skillhub.cn/plugins/Jokasa7/dsh-product-subagent-console (The community directory is an independent site with no official affiliation with DeepSeek or Hanhwa)

dsh-product-subagent-console: Design, Observe, and Diagnose Multi-Agent Execution within DeepSeek Harness Sessions