js-genai v2.32.0
- Source
- github.com
- Author
- release-please[bot]
- Date

Context
js-genai is Google's official JavaScript and TypeScript library for Gemini. Version 2.32.0 (release notes) is mostly about the Interactions API, which Google DeepMind presented as its rebuilt interface for agents and which stores each conversation turn or task as an interaction on the server. Its main change concerns very long outputs. A request that is not run in the background can produce only a capped number of output . The release adds a continuation_token, which a capped output returns with a requires_continuation status so that the output can be resumed. Google's long-decode documentation still lists this kind of client-side continuation as coming soon for the Interactions API, and for now points to background execution, where the server itself continues long outputs in 64K-token hops. The library's support may therefore arrive before the server's.
The new automatic_continuation option on interactions.create defaults to on. The commit that added it says the library then chains those hops itself, so a plain create() returns the full output as one result, while setting it to false returns the single server response unchanged. On a stream, events from every hop pass through as they arrive, and the commit tells developers to treat an interaction.completed event whose status is requires_continuation as intermediate, not as the end. According to the same commit, a numeric timeout_ms now bounds the whole call across hops, and max_output_tokens is the budget for the whole output, not for each hop.
Code that reads only the first completed event of a stream should change: under the new default it would stop after the first hop of a long answer. Elsewhere, the release adds video understanding on the Gemini Developer API and a video upscale task parameter. The Python library's matching release carries the same continuation support, along with a fix for stale access tokens on long-running Vertex AI servers that this JavaScript release does not list.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Checking sign-in…
Loading comments…



