Released 2026-11-16 · Compare with 1.0.0
Features
- Show the model provider picker about a second sooner when
eve inithands off toeve dev, by running onboarding while the development server starts. (#6472) - Make
npx eve@lateststart much faster with npm by no longer declaring@vercel/connectas an optional peer dependency, which made npm download about 450 MB of package metadata it never installs.
eve uses its own bundled copy of@vercel/connect. Projects that import@vercel/connect/evestill need@vercel/connect2.7.0 or later installed, andeve initalready adds it. (#6453) oidc()refuses a token from a different issuer without fetching its own issuer's discovery document and keys, so a route that also accepts another kind of token, such asvercelOidc(), no longer waits on that fetch on a new server instance. (#6508)- Follow-up messages to a session, including Slack replies in an existing thread, skip a Workflow hook lookup when they reach a server instance that recently resolved that session, without needing
EVE_INBOX_HANDLE_SECRET. (#6506) - A Slack mention or DM that starts its thread now starts its session directly instead of first looking up the thread's session, which takes two lookups off a new conversation's first reply. (#6473)
- A Slack mention or DM reaches its session sooner: the session starts while Slack is still confirming the "Thinking..." status, and a mention that starts its thread no longer fetches the thread to look for earlier files.
A customreceivedrenderer for a mention or DM now runs alongside the session start, so it can finish after the turn's output. (#6474) - A followed session stream resumed past its first line opens sooner on a server instance that hasn't served that session yet, because the stream version check no longer looks up the stream's tail first. (#6507)
- Make
eve initfinish a little sooner by setting up the Git repository while dependencies install. (#6470) - Make
eve initinstall a fresh npm project faster by skipping npm's optional peer dependency lookups, which downloaded hundreds of megabytes of package metadata without installing anything. (#6452) - Let people connect a model while
eve initinstalls dependencies, instead of waiting for the install and the development server to finish first. (#6478) - Make the eve package smaller and faster to install by publishing only the type declarations its entry points reach.
The package drops about 1,800 internal declaration files, 38% of its files and 1.3 MB of its download. (#6464) - Make the eve package about 700 KB smaller to download by bundling fewer duplicate copies of the same dependencies.
eve's bundled Node.js dependencies now share one copy of each dependency version they have in common, and the bundled Vercel Sandbox SDK uses eve's ownundicidependency instead of a second copy of the same version. (#6465)
Bug Fixes
- Place a turn's client context after the user-role instructions and skill or tool catalog updates eve announces in that turn, so the next turn keeps the prompt cache.
User-role instructions and catalog updates join the conversation history, but client context does not. Since #5638, the client context sat ahead of them, so the next turn's prompt diverged at that point. (#6498) - Use current provider MCP endpoints for Datadog, Cloudinary, Make, Natural, and Netlify, and align Notion's API version header with its OpenAPI schema. (#6520)
- Apply a forwarded trace content ceiling to instrumentation files and to AI SDK telemetry in a remote agent, so a downstream hop records no more prompt or response content than the upstream hop allowed.
Before this change, each instrumentation file'stracePolicywas evaluated on its own, and an agent without OpenTelemetry used its audience default. Both paths ignored a forwardedrecordInputs: falseorrecordOutputs: falseceiling. (#6515) - Keep messages queued during
eve initqueued wheneve devstarts, and stop leaving a stray version line above theeve devscreen. (#6531) - Include Better Auth's optional OpenTelemetry dependency in new authenticated web chat projects to avoid unresolved-import warnings during deployment builds. (#5695)
- Keep the changelog URL in the package README from overflowing into the npm sidebar. (#6517)
- Typing in the "Project to link" picker during an integration setup, such as
/addineve dev, no longer crashes withTypeError: n.label is not a function.
Integration setup runs in a child process, and the IPC channel dropped the picker's search action. The child now sends the search label as text and resolves a picked search itself, so the "Search for '…'" row works there as well. (#6014) - Remove the line saying the agent has paused until you connect your account or cancel from the Slack sign-in card that holds a turn, so the card shows only the "Connect " heading above the Connect and Cancel buttons. (#6016)
- Use Parallel instead of Exa as the default
web_searchprovider for AI Gateway models.
Agents without anagent/tools/web_search.tsfile now search with Parallel. To keep Exa, exportwebSearch({ provider: "exa" })from that file. (#6530)