Question 1

Can I point my existing chat-completions client at this?

Accepted Answer

No, and that is deliberate. The plugin passes the TypeSafe native protocol through and explicitly does not provide or imitate `/v1/chat/completions` or `/v1/responses` compatibility. Clients always send and receive the System One shape, with noul-to-boolean and usage-field translation handled by the plugin, so your code has to send typed questions and read typed answers.

Question 2

How do I get billing right, especially if usage is missing?

Accepted Answer

You configure a usage expression on the channel; for the official input rate it is `tier("base", u("input_tokens") * 42 / 1000000)`, since output is free, and the Vercel gateway uses its own rate with `0.04`. The plugin reserves zero at submit and settles real tokens from the upstream `usage` before responding. With no expression it falls back to per-request pricing, and it reports no number rather than a fabricated one when usage is absent.

Question 3

Do I still need this repo if new-api already ships a TypeSafe plugin?

Accepted Answer

Probably not for basic use. The README notes that when the new-api main repository already includes a factory plugin of the same name there is no need to install this one, and that this repository exists for independent installation, updates, and submitting new versions upstream. Installation is either a Marketplace source with a sha256 check or a manual upload of `plugin.js` and the icon.

Question 4

What happens when a new Jev version ships?

Accepted Answer

Nothing breaks in the plugin, if you use Model Mapping. You add the short name you want clients to send, such as `jev`, to the channel's model list and map it, for example `{ "jev": "jev-latest" }`. Pointing that at a newer version is a mapping edit. The Vercel channel requires the mapping, since the gateway insists on its own model names such as `typesafe-ai/jev`.

Question 5

How do I install it, and what does new-api handle for me?

Accepted Answer

Two ways. Marketplace source: in the NewAPI console go to Task Plugins, Marketplace Sources, add the name FFatTiger Plugins and the index URL for index.json, then find TypeSafe AI in the list and Install, which verifies the sha256 and recompiles the plugin to re-check it. Or upload plugins/tasks/typesafe/1.0.0/plugin.js manually with icon.svg. Either way new-api owns authentication, channel routing, model mapping, billing and logging, and the client only ever posts /v1/systemone.

Question 6

What does a request and response look like end to end?

Accepted Answer

Send state plus typed questions such as a noul for shouldCloseDialog, a choice with retry, stop and ask for nextAction, and a score on low, medium, high for risk. The response comes back in native format with probabilities, including per-option probabilities and confidence on the choice, a weighted score with its legend on the score, and usage of 312 input and 48 output tokens in the documented example. Call semantics are synchronous throughout: one request, one response, no task_id and no polling.

Question 7

How is the plugin packaged and released?

Accepted Answer

The plugin is a single file, plugins/tasks/typesafe/1.0.0/plugin.js, beside its CHANGELOG.md in the marketplace's release format and icon.svg, with index.json generated by the vendored tools/pluginindex generator rather than hand-edited. Regenerate and check the index from tools/pluginindex with PLUGININDEX_NAME set when new-api sits alongside as ../new-api. Published version directories are immutable, so changes ship under a new version number, and the repository is Apache-2.0 to match the official plugin repository.

New API Plugin TypeSafe

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community