Next.js cleans up Node deprecations, eve batches workflow model calls

Next.js spent the day clearing Node.js deprecation warnings out of its own code. The test harness moves off node-fetch now that fetch is built in, motivated by the deprecated url.parse it called, with fetchViaHTTP keeping call sites working through a compatibility type that translates node-fetch-only options: agent goes to http(s) directly, timeout becomes AbortSignal.timeout [1]. The same cleanup lands in testmode, where unstable_getResponseFromNextConfig no longer parses the request URL with the deprecated url.parse, constructing the pathname/query/protocol/host shape from the WHATWG URL API instead and keeping protocol and host null for relative URLs so the existing rewrite base fallback keeps working [2].

On the Turbopack side, get_relative_request_to joins get_relative_path_to for callers that need a relative path handable to the resolver as a module request, returning ./c, ../c or . where the older helper returns c, ../c or .. The fix targets import.meta.glob's base-relative keys, which currently come back without the leading ./: { base: '../outside' } produces one.js where Vite does not [3]. Strict route matching also gets a diagnostic: a layout can have sibling slots whose pages can never form a complete route, and catchall pruning removes broad matchers, so incompatible parallel route slots are now reported [4].

In vercel/ai, OpenAI web-search tools automatically added web_search_call.action.source, and Bedrock Mantle reused that logic but rejects the include value, breaking web search. An opt-out includeWebSearchSources sets an internal provider capability flag [5]. Mistral's reasoning configuration was dropped for supported model IDs and aliases outside a hardcoded allowlist; model IDs for chat, embedding, speech and transcription are refreshed, and streaming reasoning was verified live for mistral-small-latest, mistral-medium-3-5 and zai-glm-5 [6]. HarnessAgent lifecycle operations could close an adapter channel while asynchronous host tools were running, losing results from persisted continuation state and risking duplicate execution after resume; continuation state is now preserved across suspension and session finalization [7].

eve lands the most consequential API change of the day. It used to checkpoint after every model call, adding Workflow orchestration overhead to sequential tool loops; the opt-in experimental.workflow.modelCallsPerStep setting now batches several model calls in one Workflow step, with a default of 1 and stops at authorization, input, coordination, background-task, interrupt, and terminal boundaries [8]. Separately, ctx.agent used to require workflow authors to pass both a replay key and a target name even though the replay key is framework bookkeeping; it now takes ctx.agent(target, input) and derives invocation identity from its durable reply hook, including repeated and parallel calls to the same subagent. The PR states flatly that this is a breaking pre-1.0 API change [9].