Question 1

Is this actually Jev, or a prompt asking a chat model to guess?

Accepted Answer

It is the paradigm, not the model, unless you point it at the real thing. On the default `openai` channel the adapter asks any OpenAI-compatible endpoint for typed answers and renormalises what comes back — those numbers are the chat model's self-reported estimates, and the README attaches that caveat to every single result. Switch the channel to `typesafe` with a key and it calls the actual Jev API. The honest summary: the structured interface and probability plumbing are real on either channel, the calibration is only real on one.

Question 2

What stops the model from answering 1.0 to everything?

Accepted Answer

A prompt, and the author found that out the hard way. An honesty instruction in the adapter source explicitly tells the model to spread probability when it is uncertain instead of forcing a confident 1.0, because a saturated confidence value destroys the routing signal the whole paradigm depends on — a 1.0 everywhere means the threshold never discriminates. The README says this was validated on real Chinese-language work items. Even so, treat a threshold like 0.85 as something to calibrate on your own data rather than a number to paste in.

Question 3

How is this different from zhangxaochen/dsh-jev?

Accepted Answer

Different scope, and the README puts both in a comparison table. That plugin talks to the official TypeSafe API and includes tool pruning, a loop guard and a safety gate, so it suits someone with a key who wants the guard suite. This one defaults to any OpenAI-compatible endpoint — DeepSeek, Ollama, vLLM, OpenRouter, SiliconFlow, LM Studio, an air-gapped gateway — and carries no guard suite, but it can still reach the real Jev API through its `typesafe` channel. The author describes them as complementary rather than competing.

Question 4

What does installing and configuring it actually involve?

Accepted Answer

DeepSeek Harness, then one plugin command with your profile, pointing at the GitHub repository until the short npm name is published. Set one environment variable for the key and restart dsh; override the base URL and model if you are not on DeepSeek's own API. The fuller configuration is a plugin row in cordis.patch.yml holding channel, key, model, max tokens, timeout and retry count, so you can commit a non-secret config and keep keys in the environment.

Question 5

What exactly does the jev_decide tool return?

Accepted Answer

choice returns the picked option plus a full probability distribution and confidence, up to 255 options; score takes 2 to 10 ordered levels and returns a fractional score with per-level probabilities; boolean takes a statement and returns P(true). One call can carry many questions, and asking one or ten costs about the same.

Question 6

How does it handle messy model output?

Accepted Answer

It renormalises probabilities that do not sum to one, degrades an all-zero answer to a uniform distribution rather than failing, and handles thinking models that leave content empty and answer in reasoning_content. It also retries 429 and 5xx responses with exponential backoff, default three attempts.

Question 7

Which endpoints and config does the openai channel support?

Accepted Answer

Any OpenAI-compatible endpoint: DeepSeek's own API, Ollama, vLLM, OpenRouter, SiliconFlow, LM Studio or an air-gapped internal gateway, set through JEV_BASE_URL and JEV_MODEL with JEV_OPENAI_API_KEY. Shared settings such as maxTokens 4000, timeoutMs 120000 and retries 3 live in cordis.patch.yml, and the package is MIT licensed.

Dsh Jev Adapter

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community