# dsh-adversarial-review-preset

An **"adversarial review" agent preset** for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) (`dsh`).

It flips the agent's stance from *confirming the code is right* to *proving it is wrong*: the artifact under review is presumed defective, every finding must carry a locatable authority, and **nothing is modified until you approve it**.

[简体中文](README.md) · English

---

## What it does

This is a **bundle plugin**: it inserts one `@deepseek-ai/dsh-agent-preset` declaration into your profile, adding a selectable mode named 「对抗式审查」 (Adversarial review) to the agent preset roster.

It changes exactly two things; every other capability is the shipped **standard** preset verbatim (same tool set, same plan/compaction/delegation groups, same subagent and workflow tools):

### 1. A replacement persona that makes fault-finding a hard rule

- **Stance**: the code, tests, document, or design under review is wrong until you have personally verified it correct. Conclusions may not contain "should be fine", "probably", or other unfounded hedging.
- **Evidence discipline**: every conclusion must be locatable — a file path plus line number, or a chapter/clause number or link in the authoritative source. Quotations from the authority are verbatim, never paraphrased. An intuition that cannot be located is listed separately as an unverified hypothesis and never counts as a defect.
- **The reviewed artifact is immutable**: do not edit it during review, and **never edit a test to make it pass**. Only when you can cite an exact clause proving the test itself violates the spec may you propose changing it — and that is labelled "test violation", not "code defect".
- **Report before touching anything**: emit the defect report and wait for approval. After approval, fix one defect at a time, starting with a failing test that reproduces it.
- **Adversarial self-check**: for every finding, actively construct the "this is a false positive" rebuttal and try to prove it; if it holds, withdraw the finding.
- **Delegated independent verification**: hand a subagent only the finding and the reviewed files, so your existing conclusions cannot contaminate it.

### 2. Two bundled review skills

The preset mounts this package's `skills/` directory through `@deepseek-ai/dsh-skill-filesystem`'s `customSkillDirs`. The path is resolved by a `!!js` expression from the installed package itself — no hardcoded absolute path, no dependency on a preset directory.

| Skill | Contents |
|---|---|
| `adversarial-review` | The general adversarial review procedure: core rules, A–D coverage checklists (branch completeness / step fidelity / condition polarity / resource and lifetime), chained-defect cross-verification, test-blind-spot checks, adversarial self-check, seven recurring defect patterns, the defect report format and priority table, and the fix discipline |
| `whatwg-spec-adversarial-review` | A WHATWG-spec-specific line-by-line checklist aimed at spec-dense modules (tokenizer, tree construction, and similar) |

Both skills are written in Chinese, matching the preset's persona.

---

## Install

Requires a `dsh` profile with **coding tools** enabled (plugin management capability).

With the plugin manager:

```
plugin_manager(action: "install_bundle", target: "github:Ink-dark/dsh-adversarial-review-preset")
```

Or just ask the agent:

> Install the plugin `github:Ink-dark/dsh-adversarial-review-preset`

A local clone works too — pass the absolute directory path as `target`.

Installation **affects every session in the profile** and needs Full access or one approval.

### Use it

Start a **new task** and pick 「对抗式审查」 in the agent preset selector.

Existing sessions do not switch: a session and its children keep the plugin revision they started with.

---

## Compatibility

`peerDependencies` constrain the DSH runtime to `>=0.1.7-rc.2 <0.2.0`.

DSH validates these `@deepseek-ai/dsh-*` ranges at install time and **blocks installation and activation** when they are unsatisfied — that is the platform's designed guard against a plugin writing a Config schema that has since changed. To force it on an incompatible runtime, grant an exact-version exemption for that precise plugin/runtime pair via `dsh plugin allow-version` or the plugin manager.

This plugin declares as peers only the packages whose **Config it writes** (the version-sensitive ones), not every package it mounts by name, so it only intervenes when those schemas could genuinely have moved.

---

## Repository layout

```
package.json                 bundle manifest, dsh.bundle.patch, peerDependencies
cordis.patch.yml             the one convention file: inserts the preset-adversarial-review declaration
skills/
  adversarial-review/SKILL.md
  whatwg-spec-adversarial-review/SKILL.md
```

The preset declaration contains no executable JavaScript — the entire plugin composes existing plugins from YAML, which is why it is small and easy to audit.

### Modifying it

A preset declaration is carried by a bundle patch and **cannot be edited in place**: after changing `cordis.patch.yml`, reinstall (or have the plugin manager reload the bundle).

Edits under `skills/` are picked up by the `dsh-skill-filesystem` watcher and usually need no reinstall.

---

## License

MIT — see [LICENSE](LICENSE).