Introduction

The Creator mode in DeepSeek Harness allows creating, running, and hot-updating Cordis Packages within a session, but the original workflow is missing four key steps: capabilities before a change are not frozen into a reproducible baseline; after a change, quality is usually judged by a single manual experience; failed versions, successful versions, and evaluation evidence do not form a long-term lineage; and “the new code runs successfully” is not separated from “the new code is actually better”.

dsh-evolution-console is designed to add a deterministic release gate between Creator and Runtime. It connects pre- and post-change steps into an execution loop, using immutable candidate pinning, isolated evaluation, and code gates to make agent iteration evidence-based.

Core Features

The plugin mainly provides the following capabilities:

  • Pinning Immutable Candidates: Pin the modified Cordis Package into an immutable version node.
  • Baseline/Candidate Controlled Evaluation: Run the baseline and candidate in isolated Headless DSH instances using the same evaluation set.
  • Code Gates: Determine whether a candidate passes based on deterministic rules, preventing the model from self-scoring.
  • Automatic Promotion: Automatically hot-promote after passing gates; mark as eligible when automatic promotion is disabled.
  • Version Lineage: Record all candidates, evaluation reports, and history, with one-click rollback support.
  • Agent Tools: Provide the evolution_* tool series, allowing the Agent to directly execute closed-loop operations in a session.

Installation

The plugin must be installed to both the web and headless Profiles. The Web Profile provides the controller and visual interface; the evaluation subprocess requires the Headless Profile to resolve Evaluator exports.

Install from GitHub

dsh plugin --profile web add github:yu-xin-c/dsh-evolution-console
dsh plugin --profile headless add github:yu-xin-c/dsh-evolution-console

After installation, restart DSH Web:

dsh web

Open a session in an existing workspace, and the Evolution tab will appear at the top.

Install from Local Source

git clone https://github.com/yu-xin-c/dsh-evolution-console.git
cd dsh-evolution-console
pnpm install
pnpm check
dsh plugin --profile web add "$PWD"
dsh plugin --profile headless add "$PWD"

Typical Usage

1. Generate a Candidate in Creator Mode

In the session, use cordis_define to create a Package and use cordis_run to verify that it can be mounted. This is the basis for the Host half.

2. Pin the Version

Go to the Evolution tab of the session:

  1. Select the target Package under “Runtime Package”.
  2. Click “Pin”.
  3. The plugin saves the Package source code, parent version, and SHA-256 identity. If it updates an already running Package, the plugin first pins the current version as the initial Champion, then records the new version as a child node.

3. Run a Before-After Comparison

Select the Candidate and Benchmark, then click “Start Evaluation”. The Champion and Candidate are run in separate new Headless DSH processes, using the same evaluation set and Profile.

4. Review Evidence or Roll Back

The interface displays the version lineage, baseline and candidate scores for each evaluation, pass rate, and gate decision reasons. If a rollback is needed, you can use the interface or the evolution_rollback tool to restore the previous Champion.

Notes

  • Version Limitation: v0.1 does not evaluate or auto-deploy the Client half.
  • Process Overhead: Each Attempt launches a new DSH process; this is reliable but does not target maximum throughput.
  • Gate Rules: Gates are currently deterministic rules and do not include significance testing, cost Pareto frontiers, or manual blind evaluation.
  • Not Static Auditing: Runtime evaluation is not a static source-code security audit. Before publicly running third-party Candidates, the code should still be reviewed.
  • Data Location: Data is stored by default under .dsh/evolution-console/ in the project directory, so the evolution history for the same workspace remains visible after switching sessions.

Summary

dsh-evolution-console is a self-evolution closed-loop tool designed for DeepSeek Harness Creator mode. It is not a generic “self-evolution” dashboard; instead, it creates a true execution loop that connects pre- and post-change work through pinning, isolated evaluation, and deterministic gates. It is suitable for developers who need objective version control, strict regression testing, and traceable iteration workflows.

GitHub Repository