DeepSeek Harness (DSH) task orchestration needs to address the trust issue of “whether a task is truly completed.” Relying solely on LLM conclusions can easily lead to hallucinations and makes auditing difficult. The dsh-outcome-loop plugin provides a mechanical-evidence-based acceptance mechanism designed to solve these problems.

This is a local-first, user-owned, vendor-neutral task outcome ledger and acceptance plugin. It consolidates the goals, constraints, acceptance criteria, execution evidence, user feedback, cost, and final outcome of a DSH session into a re-auditable task record.

Core Capabilities

This plugin primarily provides the following capabilities:

  1. Mechanical-evidence acceptance: Acceptance criteria do not rely on subjective LLM judgment, but are based on mechanical evidence such as tests, builds, exit codes, file states, diagnostics, and Git scopes.
  2. Zero additional model cost by default: By default, no extra LLM calls are made, and no model-visible tools are exposed, avoiding increased token consumption.
  3. Local and offline by default: All outcome data is stored by default in the local DSH storage backend and does not depend on external network services.
  4. User owns and controls data: Users can inspect, delete, or export data, and exports require preview and redaction checks.
  5. State separation: Success, failure, and unknown states are clearly distinguished; unknown states are not silently converted into success.

Installation and Prerequisites

To install the plugin, use the provided installation command:

dsh plugin --profile <name> add ./dsh-outcome-loop-0.1.0-beta.8.tgz

Before installation, confirm that the environment meets the prerequisites. This plugin depends on the storageDomain service, while the official dsh-base bundle does not provide this service (only upper-layer bundles such as @deepseek-ai/dsh-web-app provide it). If using a headless or bare-metal setup, you need to manually install the storageDomain service.

Common Commands

The plugin is interacted with through the /outcome command set and supports creating tasks, defining criteria, verifying status, and exporting data.

  1. Create a task contract
    /outcome new 修复登录页按钮在移动端溢出问题

This command creates a task record and records the goal text.
  1. Add acceptance criteria
    You can manually add text criteria, or add mechanical-evidence criteria through commands:
    /outcome criterion add 移动端 375px 宽度下无横向滚动
    /outcome criterion add-command "pnpm test"
    /outcome criterion add-test
    /outcome criterion add-file dist/bundle.js

You can also add structured test-count criteria:

    /outcome criterion add-test --min-passed 2 --max-failed 1
  1. Run verification
    By default, the plugin does not run active verification and only performs passive observation.
    /outcome verify
  1. User adjudication
    Users adjudicate based on mechanical verification results and their own judgment; this is an independent axis:
    /outcome accept
    /outcome reject
    /outcome revise
    /outcome abandon
  1. Export data
    Export supports a two-stage process: preview and confirmation.
    /outcome export [<contract>]
    /outcome export <contract> --approve <digest> --out <path> [--overwrite]
    /outcome exports [<contract>]

Notes

  1. Data localization: Data is fully retained locally and is not uploaded. The plugin never performs any upload operation, and data delivery is entirely determined by the user.
  2. Passive verification: By default, no active verification is run; only passive observation of existing events is performed. Running active verification is controlled by the policy layer.
  3. Contribution mode: Contribution mode is not installed by default and requires manual configuration to use the related commands (such as /contribute).
  4. Service dependency: Ensure the storageDomain service is available in the environment; otherwise, plugin startup will fail.

Summary

dsh-outcome-loop provides a reliable task outcome tracking and acceptance solution for the DSH plugin ecosystem. By replacing subjective judgment with mechanical evidence, it ensures the credibility of acceptance; by using local storage and a zero-additional-cost design, it lowers the barrier to use. For scenarios that require strict auditing of task completion or control of model cost, this plugin is a practical choice.

  • Directory page: https://www.skillhub.cn/plugins/victorzhong0110/dsh-outcome-loop
  • Source code: https://github.com/victorzhong0110/dsh-outcome-loop