dsh-jev-thinking

Ask TypeSafe's Jev how deep a prompt needs to think. Apply that level to the model request.

English · 中文

What it does

For every message you send, the plugin asks Jev one Choice question: how much internal reasoning does the work in this message require? It writes the answer into the request's reasoningEffort.

Three properties matter.

The levels come from your model. The plugin reads the levels the current route offers, such as off, low and high. It sends exactly those levels as the options. Jev cannot choose an unsupported level, and the plugin never adjusts a level afterwards.

One question per prompt. A turn with ten tool calls still costs one question. Each subagent is its own prompt.

Short replies are judged correctly. The plugin also sends the assistant turn that the reply answers. A reply such as "好" carries no task alone. Without the previous turn, approving a four-step refactor looks like "nothing to reason about".

Install

In a terminal, install into your profile. Replace web with your profile name.

dsh plugin --profile web add dsh-jev-thinking

Alternatively, from inside DSH, ask the agent to install it:

plugin_manager { action: install_bundle, target: "dsh-jev-thinking" }

Set the API key

From DSH. Open the Plugins page, select dsh-jev-thinking, and use the row's Configure control. The page stores the key in the Harness credential store, under the name shown in apiKeyRef (TYPESAFE_API_KEY by default).

By hand. Add the key to ~/.dsh/.credentials.yaml under refs:, or set TYPESAFE_API_KEY for the DSH process.

The plugin re-reads the store for each request, so a rotated key needs no restart.

Settings

Field	Default	Meaning
apiKeyRef	TYPESAFE_API_KEY	Name in the credential store.
endpoint	https://api.typesafe.ai/v1/systemone	System One endpoint.
model	jev-latest	Model that answers the question.
timeoutMs	2500	The plugin falls back after this delay.
level	empty	Level for a failed judgement. The route must offer it, or the plugin ignores it.
verbose	false	Log one line for each decision.

A judgement fails on a timeout, an HTTP error, a missing key, or an answer outside the offered set. The plugin then falls back in this order: the configured level, then the route's own default effort, then no change to the request.

Known limits

• The plugin uses only the last turn. A reference to an older message, such as "do that, as you said", can still be judged wrongly.
• The previous assistant turn goes to TypeSafe. This costs about 750 input tokens per prompt in the common case, and up to about 9600.
• A route that offers xhigh or max adds them to the options. Jev can choose them, so expect the highest price on such routes.
• The plugin does not degrade when credits run out. Each prompt pays one fast failed request before the fallback.
• No setting narrows or excludes levels. Change the code if you need this.

Development

Maintainers only. After changing code, publish with:

npm version patch   # or minor / major
npm publish

Requires an npm account with publish rights for this package.

License

MIT. See LICENSE.