@dsh-external/dsh-turbo-agent
可选择的 DeepSeek Harness TurboAgent 模式，支持为 Flash、Pro 和 Vision 进行隐藏的 Best-of-N 路由。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:qianen6/dsh-turbo-agent#6186588dd3c5fa9a32f81d39447556c546061b32Turbo Agent 以 @dsh-external/dsh-turbo-agent 发布，当前版本为 0.3.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Managed, loopback-only TurboAgent
integration for DeepSeek Harness.
Tested with DeepSeek Harness 0.1.1-rc.2, TurboAgent 0.2.1, and
LLM-as-a-Verifier 0.2.0 on Windows.
The plugin installs a selectable TurboAgent 模式, starts a pinned Python
TurboAgent environment, reuses the Harness
DEEPSEEK_API_KEY credential reference without writing the secret to disk, and
registers a hidden turbo-agent transport through the official DeepSeek
adapter. The native deepseek-official route and its model catalog are not
changed.
DeepSeek-V4-FlashDeepSeek-V4-ProDeepSeek-V4-Flash-Vision-ExpOrdinary presets keep their normal model route. Selecting TurboAgent 模式
automatically maps the session's official Flash, Pro, or Vision choice to a
hidden Best-of-N route; users do not change the provider or select duplicate
Verified models. The hidden provider advertises an empty catalog, so the
model picker continues to show only the three official entries above.
Each request generates three candidates from the model selected in Harness;
models are never mixed into one candidate pool. Vision requests forward the
same prepared images to every candidate and to
deepseek-v4-flash-vision-exp for the fine-grained verifier pass. Text-only
requests use DeepSeek V4 Flash as the verifier. Both paths keep the
two-pivot tournament; only text requests use the majority-vote short circuit,
because identical Vision answers still need to be checked against the image.
0.1.1-rc.2 or newer compatible releaseDEEPSEEK_API_KEY configured in the Harness Models settingsgit clone https://github.com/qianen6/dsh-turbo-agent.git
cd dsh-turbo-agent
.\scripts\Install.ps1 -DshCmd 'C:\path\to\dsh.cmd'
The installer creates a repository-local .venv, installs the pinned Python
dependency, creates or migrates the non-secret configuration at
%USERPROFILE%\.dsh\turbo-agent\turbo-agent.yaml, and adds the plugin to the
selected Harness profile and installs the Agent preset. Existing Flash/Pro
candidate and reasoning settings are preserved when the Vision entry is added.
Restart Harness and choose TurboAgent 模式 from the mode selector.
Open Settings → TurboAgent in Harness to configure:
off, low, high, or maxSaving restarts only the managed proxy. Harness and existing conversations stay online. The values are stored in the non-secret runtime YAML.
Flash, Pro, and Vision are independently configurable from 1 through 16 candidates. The official self-verification tiers, Best-of-3 and Best-of-5, are included. For example:
.\scripts\Set-TurboAgentCandidates.ps1 -Flash 3 -Pro 5 -Vision 3
.\scripts\Set-TurboAgentCandidates.ps1 -ReasoningEffort max
The plugin watches ~/.dsh/turbo-agent/turbo-agent.yaml and automatically
restarts only the managed proxy after an atomic change; Harness stays running.
Candidate cost grows with the selected count and verification comparisons. Start with 3; use 5 for the official higher self-verification tier.
The managed proxy is locked to 127.0.0.1:18888. The upstream CLI defaults to
0.0.0.0; this plugin does not invoke that default. API keys are resolved from
the Harness credential service and passed only in the child-process environment.
TurboAgent traces can contain prompts, responses, and tool arguments. Do not publish the runtime-data directory. See SECURITY.md.
%USERPROFILE%\.dsh\turbo-agent contains the non-secret YAML configuration,
managed PID metadata, TurboAgent traces, and logs.
The visualizer is available locally at http://127.0.0.1:18888/visualizer.
node --check lib/index.js
npm test
$env:PYTHONPATH = Join-Path $PWD 'python'
.\.venv\Scripts\python.exe -m unittest discover -s test -p 'test_*.py' -v
This integration is MIT licensed. TurboAgent is Apache-2.0 licensed; LLM-as-a-Verifier is MIT licensed. They are installed as Python dependencies and are not vendored into this repository.