Vercel AI SDK 7

ARCH-07 · Per-request credentials never land in module-level singletons or process.env, where concurrent requests share them

SEC-01 · Provider and service keys never reach a browser bundle or the source: no public-prefixed key variables, no keys in client components, no literal apiKey

SEC-08 · Logs never dump raw request/response bodies, headers, credentials or whole generation results

SEC-09 · DevTools captures (.devtools/) never enter version control

SEC-09 · AI SDK DevTools run only in local development: devToolsMiddleware/DevToolsTelemetry sit behind a development guard

TOOL-05, SEC-01 · Credentials never become model-visible: no secret fields in a tool inputSchema, no secret environment values in prompts or instructions

SEC-03 · A call that gives the model destructive or sensitive local tools sets an explicit toolApproval policy

SEC-04 · Approvals replayed from client-submitted history are protected against forgery (approval signing or trusted server-side state)

SEC-05 · toolApproval is not relied on for provider-executed tools (webSearch, codeExecution, computer use…): the provider runs them

TOOL-02 · Identity, tenancy and permission come from trusted context, not from fields of a tool's model-filled inputSchema

SEC-02, MIG-12 · Request data never becomes privileged instructions: no request values in instructions/system, no system messages from submitted history

SEC-12 · Tools never run commands or code chosen by the model outside a real sandbox (declared in trustedExecutionPaths)

UI-05 · UI messages parsed from a request or storage are validated (validateUIMessages/safeValidateUIMessages) before conversion when tools are in play

SEC-10 · A custom experimental_download keeps the SDK's URL, redirect and credential-origin protections

SEC-11 · MCP clients are closed on completion, failure and cancellation, and keep the default redirect: 'error'

CLI-08, QA-01 · AI SDK code gets no new @ts-ignore, @ts-nocheck or @ts-expect-error: type errors from an SDK mismatch are fixed, not muted

QA-06 · A project override that disables or weakens an ai-sdk-7 rule records its reason, owner and expiry (x-reason, x-owner, x-expires)

ARCH-05 · SDK types are inferred from their definitions (tools, schemas, agents, UI messages), not re-declared or forced with any/double casts

MIG-18 · Tools that read context declare it with contextSchema instead of casting a shared context object

MIG-23 · onChunk narrows chunk.type before reading variant fields: in v7 it also receives lifecycle, finish, error and abort parts

ARCH-08 · Production tools never return random demo values, placeholders or a swallowed failure dressed as success

OUT-02 · Irreversible effects wait for the completed, validated output; partial structured output only feeds previews

OUT-03 · A missing or invalid structured output is handled, never replaced by an empty object or a cast passed off as success

RUN-03 · Streaming failures are handled where the stream is consumed (onError, error parts, iteration), not by a try/catch around streamText

RUN-07 · A generation with tools is not replayed by an outer retry: its tool side effects would run again

TOOL-10 · Pending tool calls are resolved by real outputs, errors or approvals, never by writing output-available states by hand

ARCH-07 · Conversation, approval and cache state is not kept in module-level collections that every request and user shares

RUN-02 · Tool I/O receives the call's abort signal, so cancelling a generation also cancels its tools

RUN-04 · Cleanup and persistence that must happen on cancellation run in onAbort (and onError), not only in onEnd, which an aborted stream never calls

RUN-05, ARCH-06 · Every SDK promise and stream has an owner: awaited, returned, consumed or deliberately supervised, never dropped or hidden behind void

RUN-01 · With requireBudgets, generation calls state or inherit maxOutputTokens and a timeout or abort signal

RUN-01 · Generation deadlines use the SDK timeout or an abort signal, not Promise.race against a timer that leaves the call running

RUN-06 · Retries use the SDK budgets (maxRetries before streaming, streamRetries after it starts), not an outer loop or retry helper around the whole call

ARCH-01 · One growing file does not mix transport, AI orchestration, inline tool execution and persistence

ARCH-02 · A new runtime import in AI SDK code does not close an import cycle (bounded walk of relative imports)

ARCH-04 · No pass-through wrappers around SDK calls: a function that only forwards its arguments to generateText/streamText adds no behaviour

BASE-05 · AI SDK source files do not grow past maxLines (existing debt is judged by its growth only; 0 turns it off)

BASE-05 · Functions of AI SDK code (handlers, tools' execute, callbacks) stay under maxFunctionLines when they are added or grown

CLI-07 · Streams to clients use the SDK helpers (UI message / text stream responses), not hand-written SSE framing or raw provider HTTP calls

TOOL-06 · Multi-step tool use relies on stopWhen/prepareStep or ToolLoopAgent, not a hand-written loop re-calling generation with tool results

TOOL-07, OUT-04 · generateText/streamText with tools state their step budget: their default is one step (ToolLoopAgent's is 20), too few for tools then an answer or output

OUT-01 · Structured results come from output: Output.object/array/choice, not JSON.parse or regexes over generated text

OUT-05 · Tool and output schemas use constructs that convert to JSON Schema (no dates, functions, bigints, maps or transforms)

TOOL-03 · Every tool() states what it does and its limits in a description

UI-02 · useChat is configured through its transport (DefaultChatTransport or a custom ChatTransport), not the removed api/body/headers/maxSteps/initialMessages options

UI-03 · Code reads only what the installed useChat returns: no input, handleInputChange, handleSubmit, append, reload, isLoading or data helpers

UI-04 · UI messages render their parts (text, files, tools, reasoning, data), not the removed content or toolInvocations fields

UI-06 · A UI message stream whose end callback persists the chat receives originalMessages, so message IDs and tool states stay stable

UI-07 · Changing values (tokens, model, options) reach requests per call or through transport functions, not a static body or headers object

UI-08 · Client-side tool results go through addToolOutput with the tool name and call ID, never awaited inside onToolCall

UI-09 · A resumable chat (resume: true) cancels server work through an explicit, authorized endpoint; stop() alone only detaches the client

BASE-01/02/06/08 · At session start, Claude learns the resolved ai and @ai-sdk/* versions, whether these v7 rules apply, and where the version-matched docs and sources are

BASE-02 · Fetching AI SDK docs from the web while the installed package ships its own version-matched docs points Claude to them

BASE-02/06, MIG-01/02/03, UI-04, QA-02 · AI SDK imports name only exports and subpaths the installed package declares (no removed, invented or un-prefixed experimental names)

BASE-07 · AI SDK packages are consumed as ESM: no require() of ai or @ai-sdk/*, no SDK import in a CommonJS file

BASE-07 · Where the project declares a Node.js version (engines, .nvmrc, .node-version, Dockerfile, CI), it meets AI SDK 7's Node.js 22

MIG-03/18 · Generation calls and agents pass no option the installed AI SDK removed (maxSteps, maxTokens, experimental_output, experimental_context…), local spreads included

MIG-01/04/05/08, UI-01, MIG-24 · New imports avoid deprecated AI SDK aliases (stepCountIs, Experimental_Agent, generateObject, Vue Chat, GoogleGenerativeAI*…)

MIG-04/05/22, UI-01 · SDK calls use current option names (instructions, telemetry, include, onEnd, onStepEnd, onStart…), checked per owning API

MIG-06/07/13/15/16 · Generation results are read through current members: stream, partialOutputStream, usage, finalStep, responseMessages and standalone response helpers

MIG-09 · convertToModelMessages is awaited (or returned): its promise is never passed as the messages array

MIG-10/11 · A prepareStep that overrides instructions or messages for some steps restores them for the others: v7 carries overrides forward

MIG-14 · Loops with effects over toolCalls/toolResults/content/files read finalStep or steps, since top-level values now accumulate every step

MIG-17 · Token usage is read from the v7 fields (inputTokenDetails/outputTokenDetails), not the removed cachedInputTokens, reasoningTokens, promptTokens or Anthropic cacheCreationInputTokens

MIG-19 · Approval moves from the deprecated tool-level needsApproval to call or agent toolApproval

MIG-20 · Message and tool-result content uses current part shapes: file parts with mediaType, no removed media parts or mimeType

MIG-20 · New content parts use the canonical file shape instead of the deprecated image part and image-*/file-* tool-output variants

MIG-22 · Code that reads request or response bodies asks for them with include (bodies are excluded by default in v7)

TOOL-01 · Function tools declare inputSchema (not the v4 parameters), and tool calls/results are read as input/output, not args/result

TOOL-09 · toModelOutput takes its options object ({ toolCallId, input, output }), and messages are converted with the same tools

ARCH-03 · A model ID and its settings have one owner per use case, not a literal repeated across calls

MIG-21 · A call sets portable reasoning or provider-specific reasoning options, deliberately, not both by accident

RUN-08 · A call sets temperature or topP, not both, and does not treat temperature 0 as determinism

RUN-09, BASE-02 · Model IDs are ones the installed provider package declares (or the project allows), not retired or remembered ones

SEC-06 · OpenTelemetry tracing in v7 goes through @ai-sdk/otel registered once; a telemetry flag or a tracer option alone collects no spans

SEC-07 · Registering telemetry decides what it records: inputs/outputs are captured by default, and allowlisted context fields carry no credential

BASE-01, CLI-04 · Installs keep ai on this profile's major, never give @ai-sdk/* packages ai's major, and add no removed SDK package

BASE-01, CLI-04 · package.json keeps ai on this profile's major and gives @ai-sdk/* packages their own majors

CLI-01 · No package-manager initializer (npm/pnpm/yarn/bun init, npm create) runs in a directory that already has a package.json

CLI-01 · Dependencies are installed with the project's package manager (packageManager field, lockfile)

CLI-01/08 · No second lockfile next to another package manager's lockfile

CLI-02 · Lockfiles are written by the package manager only (no hand edit, no script rewriting them)

CLI-02 · A new package.json without dependencies is initialized with the package manager

CLI-02 · Dependencies change through the package manager (add/remove/update), never by editing package.json by hand

CLI-02 · A new package.json comes from the package-manager initializer (or a generator), and gets its dependencies from the package manager

Records the tool calls that run a package manager (they may rewrite package.json and lockfiles; CLI-02/04, CLI-06)

CLI-02 · A turn that changes package.json dependencies also updates the lockfile next to it (once per stop)

CLI-02 · Registry version lookups are not a step before installing

CLI-03 · Packages are installed by name: no guessed exact version, no @latest, no --save-exact without a reason

CLI-04 · After an install naming ai or an @ai-sdk/* package, the resolved ai is still on this profile's major

CLI-05 · A turn adds only the AI SDK packages its code uses: no speculative UI adapter, provider, MCP, DevTools or telemetry install (once per stop)

CLI-06 · Records official AI SDK codemod runs, and asks for the diff to be reviewed afterwards (codemods do not cover every change)

CLI-06 · Repetitive manual migration of names an official codemod renames points to the codemod (from the third file of the session)

CLI-08 · No configuration that forces peers or pins AI SDK packages behind the package manager's back (.npmrc legacy peers, overrides/resolutions)

CLI-08 · Dependency conflicts are fixed, not forced: no --force/--legacy-peer-deps installs, no lockfile deletion to make them pass

QA-02 · Tests use the installed ai/test mocks (V4 at this baseline) instead of older mock versions or module-mocking the SDK

QA-05 · Automated tests do not call live, paid models: provider models and gateway IDs stay out of tests unless they are explicit integration tests

QA-01 · Records successful type-check, build, lint and test commands (the evidence the end-of-turn check reads)

QA-01 · A turn that changed AI SDK code runs the project's type-check, build or tests after the last change (once per stop)

QA-01 · Counts the AI SDK code Claude changes in a turn (the end-of-turn check compares it with the last successful check)

Records the work tree when a prompt starts: the baseline of the AI SDK end-of-turn checks (CLI-02, CLI-05, QA-01)