@local/dsh-max-effort
将每个模型的默认推理力度设为其宣传的最高级别，并将未指定力度的请求提升至该级别。
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:kings0527/dsh-max-effort#cd480a820c99380f6ba4857976ecc33ea85ebec9Max Effort 以 @local/dsh-max-effort 发布，当前版本为 1.0.0。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。
A DeepSeek Harness (DSH) Host bundle: every model switch lands on the highest reasoning effort that model advertises, so picking Max never costs a second click. Requests that name no effort are lifted to the same level.
max 就是 Max，否则取最高档），标签也会直接显示，不用再点第二次。github:kings0527/dsh-max-effort；或让 agent 用 plugin_manager 的 install_bundle 安装这个地址。Open the sidebar Plugins page and install the bundle with this spec:
github:kings0527/dsh-max-effort
The full git URL works too: git+https://github.com/kings0527/dsh-max-effort.git.
If the plugin_manager tool is available in a session, the same thing is:
Install the bundle
github:kings0527/dsh-max-effortwithplugin_manager install_bundle.
For a private repository, that machine's git must already be authenticated to GitHub (HTTPS credential or SSH). Public repositories need no extra setup.
git clone https://github.com/kings0527/dsh-max-effort.git
Then install that directory as a bundle (Plugins page or
plugin_manager install_bundle with the absolute path).
Refresh the Web GUI once (Ctrl+R). The model catalog is cached by the open page; the refresh makes it refetch the Host defaults. Afterwards, switching models shows and stores the highest level.
The plugin appears as Max Reasoning Default / 思考强度默认最高.
The composer model seat and the /model popup build a selection from the
clicked model's reasoning.defaultEffort. When that field is absent (the
pi-ai routes such as zai-coding-cn and opencode-go do not set one), the
selection carries no effort, the caption reads “Default”, and picking the
highest level costs a second click every time. DeepSeek advertises high as
its default, so Max takes the same extra click there.
| Seam | Effect |
|---|---|
llm.resolveModelInfo() | Advertises the highest advertised effort as the model's reasoning.defaultEffort, so the model catalog makes every switch carry it and the caption shows it. |
agent/request waterfall | A frozen call configuration without an explicit reasoningEffort is given the highest advertised effort before llm.prepareCall() validates it. Explicit efforts are preserved. |
Models with no reasoning metadata, or whose only level is off, are left
unchanged. For a model whose top level is high rather than max, the highest
advertised level is what gets selected.
ctx.llm is a traceable proxy whose reads wrap every method in a second proxy,
so the override and its guard use Symbol.for('cordis.original') to reach the
raw service. Writing through the proxy also lands on the raw instance, but the
explicit target keeps the guard meaningful. If a future DSH release removes
llm.resolveModelInfo, the plugin row fails at activation instead of silently
reverting to “Default”.
Both switches default to true; override them in the profile patch when
needed.
- id: max-effort
name: '@local/dsh-max-effort'
config:
catalogDefault: true
requestFallback: true
catalogDefault: false keeps the shipped catalog defaults; requestFallback: false stops lifting effort-free requests. Set both to false to park the
plugin without uninstalling it.
node test/highest-effort.mjs
No dependencies; plain Node ESM.
Disable or remove the bundle on the Plugins page. Updating the plugin code requires a DSH restart to drop the previous module generation; a profile change alone reloads the bundle through HMR.
MIT