Agent Testing

Test AI SDK v7 agents

Use AI SDK v7's public test surface as the source of truth. Inspect the installed ai/test exports and @ai-sdk/provider V4 types before inventing fixtures; v7 changes can make plausible older examples silently wrong.

Repository test contract

• Reproduce a bug live against the real API before writing the regression test.
• Prefer integration tests that exercise the complete agent flow.
• Run tests with nx run <project>:test; use node --test --no-warnings path/to/file.test.ts only for the narrow reproduction or filtered test.
• Run type checks with nx run <project>:typecheck, never raw tsc.
• Import the package under test by its package specifier, not a relative source path.
• Use node:assert and group related observations in assert.deepStrictEqual when that makes the failure easier to read.

v7-only rules

1. Use V4 mocks (MockLanguageModelV4, MockEmbeddingModelV4, and the other *V4 classes). Do not preserve V3 compatibility.
2. Import test models and helpers from ai/test. Import simulateReadableStream from ai; its ai/test re-export is deprecated.
3. Use the built-ins instead of local substitutes:
   • sequential return values: mockValues(...)
   • deterministic IDs: mockId({ prefix })
   • stream collection: convertReadableStreamToArray(stream)
   • stream construction: convertArrayToReadableStream(values) or simulateReadableStream(...)
   • async iterable construction: convertArrayToAsyncIterable(values)
   • language prompt/settings capture: model.doGenerateCalls and model.doStreamCalls
   • embedding call capture: model.doEmbedCalls
4. MockLanguageModelV4 accepts a function, one result, or an array of results for doGenerate and doStream. Arrays are zero-indexed and return item 0 on the first call. A fixed doStream result contains one ReadableStream and is therefore single-use. Use function form to construct a fresh stream when the same mock can be called more than once, or provide an array containing a distinct stream per expected call.
5. Do not add rawCall; it is not part of LanguageModelV4StreamResult.
6. Do not cast fixtures to any or never to hide protocol errors. Use satisfies LanguageModelV4GenerateResult, LanguageModelV4StreamResult, or LanguageModelV4StreamPart[] so drift fails at compile time.

Default language-model fixture

import type { LanguageModelV4GenerateResult } from '@ai-sdk/provider';
import { generateText } from 'ai';
import { MockLanguageModelV4 } from 'ai/test';
import assert from 'node:assert';
import { test } from 'node:test';

const usage = {
  inputTokens: { total: 3, noCache: 3, cacheRead: 0, cacheWrite: 0 },
  outputTokens: { total: 4, text: 4, reasoning: 0 },
} as const;

const response = {
  content: [{ type: 'text', text: 'A short summary.' }],
  finishReason: { unified: 'stop', raw: 'stop' },
  usage,
  warnings: [],
} satisfies LanguageModelV4GenerateResult;

test('returns the generated summary', async () => {
  const model = new MockLanguageModelV4({ doGenerate: response });
  const result = await generateText({ model, prompt: 'Summarize X' });

  assert.deepStrictEqual(
    { text: result.text, calls: model.doGenerateCalls.length },
    { text: 'A short summary.', calls: 1 },
  );
});

Choose the smallest built-in

Do not wrap language or embedding mocks merely to record options or maintain a call counter; they already capture calls. The other V4 model mocks expose their operation directly without call arrays, so use Node's mock.fn(...) when the test must inspect those calls. Function-form language-model behavior may inspect model.doGenerateCalls.length; mockValues is better when only the return value changes.

Streaming protocol essentials

• Text and reasoning use *-start, zero or more *-delta, then *-end, sharing one id.
• Delta payloads use delta, never textDelta.
• Tool arguments are stringified JSON in the final tool-call.input.
• A stream ends with finish, whose finishReason is { unified, raw } and whose usage has the V4 nested token shape.
• Return { stream } from doStream; add only typed request or response metadata when the test observes it.
• Use initialDelayInMs: null and chunkDelayInMs: null for immediate deterministic delivery; use numeric delays only when timing is the behavior under test.

See references/stream-chunks.md for typed V4 chunks and references/recipes.md for full flows.

Use TypeScript's satisfies LanguageModelV4StreamPart[] for compile-time field validation. When debugging chunks extracted as JSON, use scripts/validate-chunks.mjs for lifecycle and ordering checks that the SDK silently ignores. Keep this script: v7 has stream construction and collection helpers, but no public semantic chunk validator.

Complete public test API catalog

Read references/test-api.md when the task is not a basic language-model test or it consumes SDK streams/messages. It covers:

• MockLanguageModelV4, MockEmbeddingModelV4, MockImageModelV4
• MockSpeechModelV4, MockTranscriptionModelV4, MockRerankingModelV4, MockVideoModelV4
• MockProviderV4
• mockValues, mockId
• convertArrayToAsyncIterable, convertArrayToReadableStream, convertReadableStreamToArray
• simulateReadableStream from ai
• readUIMessageStream, isTextUIPart, toTextStream, and consumeStream from ai, including their incompatible chunk and error boundaries
• timebox from @deepagents/test for status, conversation, and other asynchronous polling

Other references

• references/mock-patterns.md: fixed, sequential, throwing, tool-call, streaming, and provider patterns.
• references/stream-chunks.md: V4 stream part shapes and ordering.
• references/recipes.md: end-to-end retry, tool, structured-output, and stream examples.
• assets/test-template.ts: minimal typed V4 integration-test template.
• scripts/validate-chunks.mjs: V4 JSON chunk lifecycle validator for ordering and matching IDs; use only where static typing cannot help.

When installed types disagree with these files, update the skill to match the installed package rather than adding a compatibility shim.