Overview¶
DeepSeek Harness (DSH) uses a plugin-based architecture and treats functional extensions as independent contracts. In engineering development scenarios, when an Agent needs to modify code or run release validation, those operations are often non-auditable if there is no governed process. The dsh-engineering-control-plane plugin provides governed engineering task execution by anchoring an implementation, repair, or validation request as a trackable Mission.
What It Is¶
This plugin is the DeepSeek Harness engineering mission orchestration and release gate plugin, maintained by bailong-Hakuryu. It is responsible for converting an Agent’s engineering goals into explicit Missions and controlling their one-way stage progression. By binding Evidence and Quality Gates, it ensures a fully traceable chain from “who made the change” to “why it was approved.”
Core Features¶
The plugin provides the following core capabilities:
* Mission Orchestration and Release Gating: Manages the execution flow of engineering tasks and release admission.
* Status Tracking: Provides task routing, plan execution, validation, and status tracking.
* One-Way Stage Progression: A Mission advances in a fixed order and is irreversible.
* Evidence Binding: Binds summaries such as role outputs and validation results as the basis for review.
* Quality Gate: Determines whether a task passes based on deterministic rules.
* Persistence and Recovery: Uses SQLite to persist data and supports recovery from BLOCKED or REWORK_REQUIRED states.
* Read-Only Integrity Check: Provides a doctor command to validate SQLite data and schema.
* Routing Support: Supports passive routing (natural-language instructions) and active commands (/mission).
Installation and Activation¶
Before installation, ensure that the DeepSeek Harness CLI is installed and that the runtime meets the Node.js version requirement (^22.19.0 || >=24.0.0).
- Download the Release tarball for the corresponding version (currently v0.2.2).
- Install the plugin using the official command.
dsh plugin --profile web add D:\Downloads\dsh-engineering-control-plane-0.2.2.tgz
- Check the final configuration set.
dsh --profile web --dump-config
- Start Harness Web in the target Git repository directory.
dsh web
By default, the plugin binds the current working directory as current-workspace and reads the repository’s package.json to run the test (required), typecheck, and build scripts. During validation, install hooks such as postinstall are not triggered; dependencies must be installed in advance.
Typical Usage¶
The plugin supports two invocation methods:
Passive Invocation (Recommended)
Describe the engineering goal directly; the model routes the request to the Mission tool.
Please implement this feature for the current repository and run full validation after completion.
Please fix this issue; all modifications must pass the Mission release gate.
Active Invocation
Enter the command directly in Harness Web or the CLI.
/mission Run a pre-release validation for the current repository, requiring tests, type checking, and build to all pass
If you explicitly do not need a governed process, you can specify “direct mode, do not create a Mission”.
Usage Guidelines and Notes¶
When using this plugin, note the following facts:
- Source of Validation Commands: Validation commands come from the host configuration and are not improvised by the model.
- Gate Logic: Only a deterministic Quality Gate can produce an
APPROVEDstatus. - Evidence Integrity: Missing, corrupted, truncated, or uncertain evidence causes the mission to fail closed.
- Concurrency Limit: Only one non-terminal Mission is allowed per workspace at a time.
- Dependency Handling: Dependencies are never installed before validation, and hooks such as
postinstallare not triggered. - Scope of Responsibility: This plugin is responsible for engineering governance and is not equivalent to a vulnerability scanner; security assessment is handled by the optional Security Assurance plugin.
Summary¶
dsh-engineering-control-plane introduces Missions, Evidence, and Quality Gates, converting Agent code modification behavior into a controlled and auditable engineering workflow. It is suitable for teams that require strict governance processes. Through one-way stage progression and SQLite persistence, it ensures the integrity and recoverability of the execution process. More details are available in the GitHub repository.