deepseek-harness-flow

Readme

# DeepSeek Harness Flow

[简体中文](README.zh-CN.md)

DeepSeek Harness Flow is a community plugin that adds a visual workflow canvas and a multi-model evaluation lab to [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness).

![Harness Flow Studio: visual Agent workflow canvas](docs/images/harness-flow-studio.png)

_Studio preview — visual workflow editing remains an additive `Harness Flow` view inside a normal Harness session._

It preserves the original Harness experience. Chat, sessions, settings, tools, and existing workflow records remain unchanged; the plugin only adds a `Harness Flow` view with three focused tabs:

- Studio — edit and run Input, Agent, Map Agent, Condition, Merge, and Output nodes.
- Bench — run preset or imported suites against multiple provider routes and an independent judge.
- Runs — inspect persistent results, non-formal objective results, formal rankings, and redacted exports.

> Community project. Not affiliated with or endorsed by DeepSeek.

## Install

This beta is pinned to DeepSeek Harness commit `47f9438` and its matching RC packages.

```text
dsh plugin --profile web add https://github.com/alison-xx/deepseek-harness-flow/releases/download/v0.1.0-beta.2/deepseek-harness-flow-0.1.0-beta.2.tgz
```

Restart the web profile, open any session, and select `Harness Flow` beside the original `Chat` view. See [the installation guide](docs/INSTALL.md) for local-package installation and troubleshooting.

## Desktop client (Tauri)

The repository also includes a Windows-oriented Tauri desktop launcher. It starts an isolated local Harness profile, installs this plugin bundle into that profile, and opens Harness in a native desktop webview.

![Harness Flow Desktop launcher](docs/images/harness-flow-desktop.png)

```text
pnpm install --frozen-lockfile
pnpm desktop:dev
```

The launcher asks for a local DeepSeek Harness checkout and a Node.js executable, then uses port `4181` by default. `desktop:build` creates an installer once Rust and the Windows build tools are installed. The first desktop beta intentionally uses a local Harness checkout and Node.js runtime rather than silently packaging credentials or a second copy of Harness. See [the desktop guide](docs/DESKTOP.md).

## Benchmark behavior

- Built-in MIT suites: Quick Smoke (6), Comprehensive Agent (18), Coding (8), and Safety (6).
- Default repetitions: 3. Default limits: 2 global jobs and 1 job per provider route.
- DeepSeek, OpenAI, Anthropic, Google, and custom OpenAI-compatible routes.
- Every route has its own Harness credential reference; plaintext keys are sent only to `credentials.set` and are never persisted by this plugin.
- A formal 100-point quality score requires a non-competing independent judge: objective 50%, judge 30%, safety 10%, stability 10%.
- Without a judge, results stay objective-only and no formal rank is assigned.
- Latency, tokens, cache reads, and estimated cost are separate metrics. Cost appears only when a pricing snapshot was entered.
- Imported commands are disabled until a run explicitly authorizes them.

## Development

Requires Node.js 22+ and pnpm.

```text
pnpm install --frozen-lockfile
pnpm verify
pnpm test:browser
pnpm pack:release
```

`pnpm preview` runs a standalone visual test shell at `http://127.0.0.1:4178`. Harness RC dependencies are runtime peers; the repository includes narrow compatibility declarations so a standalone clone can type-check without checking out the upstream monorepo.

## Public API

Typert Remote namespaces:

```text
flow.list/get/put/delete/run/cancel
benchmark.suites/list/import/export
benchmark.runs/start/get/list/cancel/retry/export
```

The primary wire and persistence contracts are `FlowDefinition`, `ModelTarget`, `BenchmarkSuite`, `BenchmarkCase`, `BenchmarkRun`, and `BenchmarkResult`.

## Privacy and security

No telemetry is collected. Test prompts, model answers, reports, and API keys are not uploaded by the plugin. Exporters redact credential-like fields, absolute workspace paths, and sensitive environment-variable names. Please read [SECURITY.md](SECURITY.md) before importing third-party test suites.

## License

MIT. Built-in benchmark prompts are original project content and are released under the same license.