Back to directory

dsh-harness2

In a curated listMaintenance: Active

asiu-x/dsh-harness2

Runs isolated Propose-Probe-Compose refinement loops, adopts improved workspace harnesses, exposes status/reset controls, and records per-domain lessons.

View on GitHub
$ dsh plugin --profile web add github:Asiu-x/dsh-harness2

14

stars

1

forks

JavaScript

Language

Apache-2.0

License

2026-10-28

Created

2026-11-18

Last push

Apache-2.0 DSH plugin that runs a Propose-Probe-Compose refinement loop on a representative task, adopts the improved harness, and files one transferable lesson per domain in a local library.

DSH integration

Native runtime

Author-claimed

Safety audit

Unaudited

Last verified

2026-11-21

License

Apache-2.0

01What can it help you accomplish?

  • Improve a workspace harness from execution evidence instead of guesswork

    An adopted harness chosen by an anonymous selector, a per-run directory holding probe trajectories, deliverables, proposals and compositions, and a written note of what each edit did

    DSH users who keep hand-tuning prompts and memory files for a recurring task and want the tuning itself to be measured

  • See what the agent is currently running on, and put the seed harness back

    `harness2_status` reporting the adopted harness, its components, the domain library and the latest run; `harness2_reset` archiving the current harness and restoring the seed behind an explicit `confirm: true`

    Anyone who has let an agent edit its own harness and needs an inspectable, reversible state

  • Carry what was learned into the ordinary DSH sessions that do the real work

    An always-on prompt section for every agent in that workspace — the adopted harness text plus the most recent lessons — and one transferable lesson filed under the domain in the local library

    Teams running repeated tasks in one workspace who want each refinement to pay off after the loop ends

02How to install into DeepSeek Harness

Prerequisites

  • DeepSeek Harness with a model provider configured — the README says this release was tested with DSH `0.1.5-rc.1` and Node.js 20+, and that model calls are billed by your DSH provider
  • A representative task plus a small workspace to try it on — the README recommends starting with `k=1` and a small workspace

Installation steps

  1. 01

    From GitHub: `dsh plugin --profile web add github:Asiu-x/dsh-harness2`, then restart DSH (or start `dsh web`).

    $ dsh plugin --profile web add github:Asiu-x/dsh-harness2

  2. 02

    For a local checkout: `git clone https://github.com/Asiu-x/dsh-harness2.git`, then add the absolute path with `dsh plugin --profile web add /absolute/path/to/dsh-harness2`.

    $ git clone https://github.com/Asiu-x/dsh-harness2.git

  3. 03

    Trigger a loop from the chat by asking the agent to use `harness2_refine` on a representative task, with a `domain` and `k: 1`.

Verify the integration

  • Confirm the plugin landed in the composed tree: `dsh --profile web --dump-config | grep -A2 harness2`.

03DSH integration and capability boundaries

DSH integrationNative runtime

Native DSH plugin: `dsh plugin --profile web add github:Asiu-x/dsh-harness2` installs it through the DSH bundle patch it declares, and it then appears as three model tools (`harness2_refine`, `harness2_status`, `harness2_reset`) plus a `/harness2` command

  • Refinement loop (`harness2_refine`)

    A representative task, a domain label and a recursion budget `k` (default 1, capped by `maxK`)→Base probe, two proposals, two candidate probes, a composition, a final probe, an anonymous selection and a reflection — ending in an adopted harness or a deliberate keep-the-base

    spends roughly nine child-agent runs for a `k=1` loop, all billed by your DSH providerwrites `runs/<run>/` under `.dsh-harness2` with trajectories, deliverables, proposals, compositions and isolated workspace copies
  • Harness inspection (`harness2_status`)

    Nothing beyond the workspace the plugin instance is attached to→The adopted harness, its component list, the domain library and the latest run

  • Reset to the seed (`harness2_reset`, `/harness2 reset`)

    An explicit confirmation flag; the slash command needs no flag→The current harness archived and the seed harness restored

    archives the current harness under `.dsh-harness2` and restores the seed into `harness/`
  • Editable harness surface with an always-on budget

    Markdown files under `.dsh-harness2/harness/`: system, guardrails, memory, tool guidance, skills, sub-agents, scripts and middleware rules→A harness whose four always-on files are capped (12,000 characters by default) and whose skills must be named as triggers in `system.md` or `guardrails.md`

    proposals are edited files written into the run directory, not patch descriptionscandidate middleware plugins are read as guidance and never executed
  • Per-domain lesson library

    The reflection written at the end of an adopted run, and the recent entries of the current domain→`INDEX.md` plus one dated folder per adopted harness under `.dsh-harness2/library/<domain>/`, with the full tree, `EXPERIENCE.md` and `experience.json`

    recent lessons are injected into ordinary agents in the workspace, and up to eight library entries are copied into a proposer context

04Who is it for? When not to use it?

Good for

  • DSH users who keep hand-tuning prompts and memory files for a recurring task and want the tuning itself to be measured
  • Anyone who has let an agent edit its own harness and needs an inspectable, reversible state
  • Teams running repeated tasks in one workspace who want each refinement to pay off after the loop ends

Not for

  • A `k=1` loop is about nine child-agent runs and every model call is billed by your DSH provider; `k>1` multiplies the calls, so the README says to start with `k=1` and a small workspace.
  • Each probe runs the real task, and the child agent uses your DSH profile's tools and permissions — the README asks you to review the permission preset before pointing it at sensitive data.
  • Probes copy the workspace, capped at 50 MB by default; `shared` isolation runs them in place and can produce collisions, so it exists only for workspaces too large to copy.

05Compatibility, maintenance and safety notes

  • A `k=1` loop is about nine child-agent runs and every model call is billed by your DSH provider; `k>1` multiplies the calls, so the README says to start with `k=1` and a small workspace.
  • Each probe runs the real task, and the child agent uses your DSH profile's tools and permissions — the README asks you to review the permission preset before pointing it at sensitive data.
  • Probes copy the workspace, capped at 50 MB by default; `shared` isolation runs them in place and can produce collisions, so it exists only for workspaces too large to copy.
  • Candidate middleware plugins are guidance only, selection is a model judgement of deliverables rather than a ground-truth grade, and the plugin has not been validated against the paper's benchmarks — it is not a benchmark harness.
2026-10-282026-11-18v0.1.1

Apache-2.0 · early but active (latest release v0.1.1, 2026-11-18; repo created 2026-10-28, last push 2026-11-18)

06Frequently asked questions

How do I install it in DeepSeek Harness, and how do I know it loaded?

Run `dsh plugin --profile web add github:Asiu-x/dsh-harness2`, then restart DSH or start `dsh web`; the package declares a bundle patch, so the profile picks it up on its own. Check with `dsh --profile web --dump-config | grep -A2 harness2`. For a local checkout, clone the repo and add the absolute path instead.

Where does the adopted harness live, and how do I undo an adoption?

Everything sits in `.dsh-harness2/` next to the workspace: the live harness in `harness/`, state in `state.json`, the per-domain library in `library/<domain>/`, and each run in `runs/<run>/`. `harness2_reset` (or `/harness2 reset`) archives the current harness and puts the seed back; the tool form requires `confirm: true`.

What does one refinement run actually cost?

A `k=1` sequential loop is roughly nine child-agent runs — one base probe, two proposals, two candidate probes, a composition, a final probe, a selector and a reflection — and every model call is billed by your DSH provider. Start with `k=1` and a small workspace; raising `k` multiplies the calls, and `concurrency` decides how many run at once.

Which DSH version and Node.js does it need?

DeepSeek Harness, tested by the author with DSH `0.1.5-rc.1` on Node.js 20+; this is an early `0.1.1` release. It has not been validated against the numbers in the Harness² paper and is not a benchmark harness, so treat selection output as evidence rather than a score.

How does it sit next to other DSH plugins and my own prompt files?

It is a standalone DSH plugin: it registers `harness2_refine`, `harness2_status` and `harness2_reset` plus the `/harness2` command, and it keeps everything it writes under the workspace's own `.dsh-harness2/` directory. Once a harness is adopted, ordinary agents in that workspace get its always-on text and recent lessons as a prompt section — that is the part that changes their behaviour, and it is why you would not keep a hand-written prompt file for the same rules.

08Data and sources

  • Author-claimedgithub.com55ddce2a0693…

    This is a DSH plugin that lets an agent improve its own harness from execution evidence

  • Author-claimedgithub.com55ddce2a0693…

    dsh plugin --profile web add github:Asiu-x/dsh-harness2

  • Author-claimedgithub.com55ddce2a0693…

    The package declares a DSH bundle patch, so it is added to the profile's plugin tree automatically.

This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-21. Found an error? Submit a correction.

🏆

Best DeepSeek Harness Plugins

Twelve plugins worth installing first — picked from the whole catalog, across every category.

DSH Plugins is an independent community directory of DeepSeek Harness plugins. Not affiliated with or endorsed by DeepSeek. Third-party plugins are not security-audited — review the source before installing.

New DeepSeek Harness plugins, weekly. No spam.