Skip to content

Harbor Self-Evolving 0.10.3 · Beta

Make your Agent improve every time.

Continuous evaluation and controlled self-evolution for DeepSeek Harness. Diagnose real Sessions or comparable regressions, change one controlled surface at a time, and let a deterministic Gate return a PROMOTE or REJECT recommendation.

Gate is a recommendation—not deployment. Default Host mode is not a sandbox.

Install and start Understand the product

View source and releases on GitHub

Checkable facts

A complete evaluation surface, not a magic loop

3Collaborating deliverablesDSH Plugin · Skill · Python Adapter
19Strict Agent tools10 approval-gated writes · 9 read-only/in-memory
2Evaluation pathsHistorical diagnosis · Candidate regression
1Deterministic GateKept separate from the Optimizer

Three deliverables, one contract

Native DSH experience, Harbor-grade evidence

Each layer has a narrow responsibility and an explicit trust boundary.

01 / DSH PLUGIN

Work where the Agent already works

◇ Native tools, Workbench, context and reviewed actions

  • 19 strict tools cover setup, diagnosis, evaluation, governance and Gate.
  • The Web UI exposes trials, evidence, artifacts, operations and settings.
  • Mutating Agent tools require one-shot approval; read surfaces stay bounded and redacted.

The Plugin connects DSH interactions to immutable Harbor evaluation identities.

02 / BUNDLED SKILL

Orchestrate the workflow without hiding the boundaries

◇ Identify → diagnose → change once → regress → Gate

  • Uses recent completed Sessions when no Dataset is supplied.
  • Freezes Candidate, Dataset, Stack and Context identities before an expensive Job.
  • Never treats Gate as deployment authority.

The Skill is maintained by this project; it is not DeepSeek endorsement.

Controlled Agent evolution loop from evidence to one change and regression.

03 / PYTHON ADAPTER

Run through Harbor’s evaluation model

◇ Generator · Evaluator · Optimizer · Gate

  • Host execution is the 0.9.7 default; Docker is explicit opt-in.
  • Candidate Context v3 and Historical Context v2 remain distinct protocols.
  • Jobs record stable identities and typed evidence for later comparison.

Identity and evidence become inspectable artifacts—not claims in prose.

DSH Plugin

Two ways in

Start from reality, finish with a controlled decision

  1. Diagnose recent Sessions

    Preview bounded, redacted DSH history; disclose the Judge and data boundary; confirm; run one Trial per selected Session.

  2. Freeze a Candidate

    Snapshot the Candidate and bind Dataset, Evaluation Stack, runtime and Context identities before regression.

  3. Change one surface

    Improve the Agent or Evaluator with an explicit diff, then rerun comparable evidence.

  4. Apply the Gate

    PROMOTE or REJECT is a deterministic recommendation for fixed Gate inputs; external CI/CD still owns deployment.

Design principles

Improvement must be inspectable

01

Identity before score

Candidate

02

Validity before averages

Invalid evidence

03

One controlled change

Attribution remains possible because each iteration changes one reviewed surface.

04

Permission-aware

Mutations

05

Evidence over assertion

Every release claim points back to source

06

Gate is not deployment

Promotion recommendation and production authority remain deliberately separate.

Boundaries first

What Harbor Self-Evolving does—and does not do

Is this CNCF Harbor, the container registry?

No. This project uses Harbor as the name of its evaluation and controlled-evolution runtime for DeepSeek Harness Agents.

Does every completed Job improve the Agent?

No. Job completion is execution status. Valid scores, comparability, policy thresholds and the Gate decide whether promotion is recommended.

Is the Gate deterministic?

The Gate logic is deterministic for fixed inputs. Candidate and Judge model outputs can still be stochastic, so repeated Jobs may differ.

Does PROMOTE deploy anything?

No. Harbor emits evidence and a recommendation. External CI/CD or a human owner keeps deployment and Champion authority.

Is Host mode sandboxed?

No. Default Host mode provides no container isolation, user switching, network policy, or CPU/memory limits; tasks run with the current user’s permissions.

Is this endorsed by DeepSeek?

No. The official Skill is maintained by this project; that does not imply endorsement by DeepSeek.

Normal installation

One command, then restart DSH

npx --yes dsh-harbor-evolution@latest setup --project-root "$PWD"

Run this from your business Agent workspace. Source installation is for contributors only.

Start with evidence you already have.Evaluate recent completed Sessions, or bring an explicit Candidate and Dataset for a comparable regression.

Open the quick start