dsh-harness2

Harness² (recursive agent harnessing) for DeepSeek Harness.

This is a DSH plugin that lets an agent improve its own harness from execution evidence: it runs a representative task under the current harness, proposes invariant and domain edits, probes them on the same task, composes an improved harness, selects an output anonymously, and stores the adopted harness plus the lesson in a per-domain library.

It ports the method of Google Research's Harness² (paper, Apache-2.0) onto DSH's native agent and prompt surfaces. It does not re-run the paper's benchmarks.

Status: 0.1.0, early release. Tested with DSH 0.1.5-rc.1 and Node.js 20+. See Verification and Limitations.

Install

From GitHub (public plugin package):

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

Restart DSH (or start dsh web) afterwards. The package declares a DSH bundle patch, so it is added to the profile's plugin tree automatically.

Local development:

git clone https://github.com/Asiu-x/dsh-harness2.git
dsh plugin --profile web add /absolute/path/to/dsh-harness2

Check that the plugin is in the composed tree:

dsh --profile web --dump-config | grep -A2 harness2

Quick start

The plugin registers three model tools:

Tool	Purpose
harness2_refine	Run one Harness² improvement loop on a representative task.
harness2_status	Show the adopted harness, its components, the library, and the latest run.
harness2_reset	Archive the current harness and restore the seed. Requires confirm: true.

Ask the agent to use harness2_refine on a task:

Use harness2_refine on this representative task:
"Read data/orders.csv, compute the total of the amount column, and write result.md containing exactly 'Total: <number>' on one line."
domain: data-analysis
k: 1

A k=1 sequential run is roughly nine child-agent runs: one base probe, two proposals, two candidate probes, one composition, one final probe, one selector, and one reflection. Start with k=1 and a small workspace. Model calls are billed by your DSH provider.

The human-facing command:

/harness2            status
/harness2 library    list the current domain's library
/harness2 reset      archive and restore the seed harness

Once a harness is adopted, ordinary DSH agents in that workspace receive its always-on text and recent lessons as a prompt section. A fresh session sees it at startup; the session that ran the refinement receives an adoption context message on its next step.

What it does

task ──► base rollouts ──► PROPOSE (INV + DOM) ──► PROBE ──► COMPOSE ──► final probe
                                                                 │
                                                                 ├─► anonymous SELECT
                                                                 └─► REFLECT ──► library
                                                                             │
                                                                             └─► next task

• Propose — two proposers edit the harness: INV changes how the agent works; DOM adds what a practitioner of the domain knows. Proposals are the edited files, not patch descriptions.
• Probe — each candidate runs the same task in an isolated copy of the workspace. The evidence is the trajectory, the artifact, and whether the run delivered.
• Compose — one composer merges, adopts, or keeps the base, with a specific supporting moment for each decision.
• Select — an anonymous selector compares rendered deliverables, not harnesses. Base is included as a conservative check; a no-op winner is not adopted.
• Reflect — writes what each edit did and at most one transferable lesson into .dsh-harness2/library/<domain>/.

Read docs/method.md for the full method, the eight components, the DSH mapping, and the differences from the reference release.

The editable harness

The harness lives at .dsh-harness2/harness/:

File	Role	Cost
system.md	role and procedure	always on
guardrails.md	standing rules	always on
memory.md	verified workspace facts	always on
tool-guidance.md	how to use tools	always on
skills/<name>/SKILL.md	on-demand workflow	loaded only when invoked
sub_agents/<name>/AGENT.md	delegated role	separate context
scripts/<name>.py	deterministic helper	run on demand
plugins/<name>.md	middleware rule (guidance)	read as a rule

Always-on text is capped (default 12,000 characters). A skill must be named as a trigger in system.md or guardrails.md; referenced scripts must exist. Candidate middleware plugins are not executed by this build.

Configuration

Override the plugin entry in your profile's cordis.patch.yml (or via --patch):

- id: harness2
  config:
    maxK: 3
    concurrency: 2
    timeoutMinutes: 20
    maxWorkspaceMB: 50
    workspaceIsolation: copy
    selection: true
    reflection: true

Key	Default	Meaning
dataDir	.dsh-harness2	Harness, library, and run data.
maxAlwaysOnChars	12000	Cap for the four always-on files.
maxK	3	Maximum recursion budget accepted by the tool.
mode	sequential	Default recursion mode.
concurrency	2	Parallel child agents per phase.
timeoutMinutes	20	Per-child-agent timeout.
toolTimeoutMinutes	120	Cooperative timeout for the whole refine tool.
maxWorkspaceMB	50	Copy cap; larger workspaces need a smaller path or isolation: shared.
workspaceIsolation	copy	copy = one workspace copy per probe; shared = run in place (can interfere).
selection	true	Anonymous selection; if false, adopt the best successful final.
reflection	true	Write the experience record after adoption.
libraryLessons	3	Recent lessons injected into ordinary agents.
libraryEntries	8	Library entries copied into a proposer context.
model / provider / reasoningEffort	session default	Override the child-agent route.

Data layout

.dsh-harness2/
├── harness/                     # the live adopted harness (or seed)
├── state.json                   # adopted, domain, run, inventory
├── library/<domain>/
│   ├── INDEX.md
│   └── harnesses/<date>-<task>/
│       ├── harness/             # full adopted tree
│       ├── EXPERIENCE.md
│       └── experience.json
└── runs/<run>/
    ├── progress.log
    ├── run.json
    ├── rollouts/                # probe trajectories
    ├── deliverables/            # final answers + changed files
    ├── proposals/               # candidate harness + notes + meta
    ├── compositions/            # composed harness + notes + meta
    ├── contexts/                # evidence trees given to internal agents
    └── workspaces/              # isolated probe copies

runs/ and workspaces/ are ignored by the data directory's .gitignore; harness/ and library/ can be committed if you want to version the improvement.

Verification

npm test covers the harness surface and validation rules, the store/library, workspace copy/diffing, and trajectory rendering.

End-to-end, in a disposable DSH headless profile, this release was exercised on:

• a task where the base already succeeded (composer kept the base; selector compared; no fake adoption);
• a task with a hidden workspace rule where a DOM proposal added the verified rule to memory.md, the composer kept it, the selector chose the final, and the harness was adopted;
• a fresh session whose system prompt contained the adopted always-on harness and the library lesson.

The plugin has not been validated against the paper's benchmark numbers, and it is not a benchmark harness.

Limitations

• Each probe runs the real task in a copy of the workspace. Side effects are contained by copy, but the child uses your DSH profile's tools and permissions; review your permission preset before using this on sensitive data.
• Candidate middleware plugins are guidance, not executed. Skills, scripts, subagents, and prompt text are the supported levers.
• shared isolation runs probes in place and can produce collisions; it exists for workspaces too large to copy, not as the default.
• Selection compares deliverables using the model, not a ground-truth grader. It is evidence, not a score.
• Runs are expensive. k=1 is the default; k>1 multiplies model calls.
• The tool can switch the plugin instance to another workspace via its workspace argument, but only one workspace is active per plugin instance.

License and acknowledgements

Apache-2.0. See LICENSE and NOTICE.

The method, component taxonomy, and several prompt rules are adapted from google-research/harness2, Copyright 2026 Google LLC, Apache-2.0. Harness² is a research release and is not an officially supported Google product. This integration is independent and is not affiliated with or endorsed by Google or DeepSeek.

Reference:

@article{xu2026harness2,
  title   = {Harness$^2$: Recursive Agent Harnessing for an Open World},
  author  = {Xu, Ruiyao and Chen, Yanfei and CuiZhu, Zhongying and Dalvi Mishra, Bhavana and Ming, Yifei and Yu, Han and Han, Rujun and Lee, Chen-Yu and Pfister, Tomas},
  journal = {arXiv preprint arXiv:2610.14224},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.14224}
}