AI Agent Hub
Back to plugins
dsh-evolution-console preview

dsh-evolution-console

Workflow Updated 2026.08.19

Run the following command in DeepSeek Harness:

dsh plugin install yu-xin-c/dsh-evolution-console

Paste the following prompt into your AI chat to install this plugin:

Install this plugin in DeepSeek Harness by running dsh plugin install yu-xin-c/dsh-evolution-console, with the source repository at https://github.com/yu-xin-c/dsh-evolution-console

About this plugin

DSH Creator mode already lets you hot-swap Cordis Packages and reshape the current Agent Runtime inside a workspace. What it still lacks are four things a reliable loop needs: a replayable capability baseline, a systematic before-and-after comparison, an evidence-backed version lineage, and a clean line between "the code runs" and "the code is actually better." Evolution Console fills that gap as a deterministic release gate.

A full loop covers the entire path from observation to recovery: the current Package source (Host and Client) is frozen with a SHA-256 content identity into an immutable candidate; an isolated Headless DSH subprocess runs Champion and Candidate against the same frozen benchmark set, the same Profile, and the same number of runs; a TypeScript gate (not the model scoring itself) decides the outcome with hard rules: score must strictly beat the baseline, no single task may regress, no Runtime or infrastructure errors are tolerated, and only Host-only candidates qualify for auto-promotion. On pass, the candidate is hot-promoted or marked eligible; on fail, it is rejected. The previous Champion can always be rolled back. The plugin also registers five Agent tools, evolution_status, evolution_capture, evolution_evaluate, evolution_promote, and evolution_rollback, so Creator can run the entire loop in one session while high-impact write actions still go through the DSH Approval Policy.

It is built for individual developers or small teams already iterating Cordis Packages in Creator mode who want to replace "try it and see" with "evaluated, gated, evidenced, and rollback-safe." You own the Benchmark and the acceptance rules; the plugin handles versioning, evaluation execution, gate decisions, and lineage memory. Each stays in its own lane.

Screenshots

Use Cases

  • After a Creator-mode Cordis Package change, decide via a deterministic gate whether the new build truly beats the baseline
  • Freeze pre-change capability into a replayable baseline, then run Champion and Candidate in isolated subprocesses
  • Roll back a failed candidate to the last Champion that passed the gate, with full lineage preserved

Best For

  • Solo developers or small teams already iterating Cordis Packages in Creator mode
  • Agent workflow maintainers who want reproducible evaluation rules instead of gut feel
  • Harness users who need version lineage, evidence trails, and rollback to manage Runtime changes