Files
Jarvis/packages/AgentCore/Tests/AgentCoreTests/LLMProviderMultimodalTests.swift
james.maes 202b7e36a6 feat(07-05): LLMProvider multimodal protocol extension + ImageBlock + TurnInput.images (Task 1 GREEN)
- ImageBlock(mediaType:data:) Sendable+Equatable canonical frame value type
- LLMProvider gains stream(messages:images:tools:...) with default extension
  forwarding to single-modal when images.isEmpty (D-18)
- Non-overriding conformers precondition-fail when images is non-empty
- TurnInput.images defaults to []; existing initializers compile unchanged
- TurnInput.withImages factory for FrameAttachController consumers
- LLMProviderMultimodalTests pass (forward dispatch + override + ImageBlock equality)
2026-04-29 17:37:06 -05:00

185 lines
6.4 KiB
Swift

import XCTest
import Foundation
@testable import AgentCore
/// Plan 07-05 / Task 1 — D-18 protocol extension dispatch tests.
///
/// Behaviors verified:
/// 1. The default protocol extension forwards to the existing single-modal
/// stream when `images.isEmpty`.
/// 2. The default extension's behavior with non-empty images on a non-overriding
/// conformer is failure (precondition; documented dispatch contract).
/// 3. A conformer overriding the multimodal method receives non-empty images
/// on the override path, NOT on the single-modal forwarder.
/// 4. TurnInput's new `images` field is additive and round-trips correctly.
final class LLMProviderMultimodalTests: XCTestCase {
// MARK: - Test 1: default forwarding when images.isEmpty
func testDefaultExtensionForwardsToSingleModalWhenImagesEmpty() async throws {
let mock = MockLLMProvider()
let stream = mock.stream(
messages: [LLMMessage(role: .user, content: [.text("hello")])],
images: [],
tools: [],
toolChoice: .auto,
model: ModelID(rawValue: "test-model"),
maxOutputTokens: 100,
cacheHints: nil
)
var collected: [LLMEvent] = []
for try await event in stream {
collected.append(event)
}
// The mock single-modal stream emits one .messageStop event.
XCTAssertEqual(collected.count, 1)
let singleCount = await mock.singleModalCallCount
let multiCount = await mock.multimodalCallCount
XCTAssertEqual(singleCount, 1)
XCTAssertEqual(multiCount, 0,
"Default extension must not call a multimodal override that doesn't exist")
}
// MARK: - Test 3: explicit override receives images
func testMultimodalOverrideReceivesImages() async throws {
let mock = MockMultimodalLLMProvider()
let imageBlock = ImageBlock(mediaType: "image/jpeg", data: Data([0xFF, 0xD8, 0xFF]))
let stream = mock.stream(
messages: [LLMMessage(role: .user, content: [.text("describe")])],
images: [imageBlock],
tools: [],
toolChoice: .auto,
model: ModelID(rawValue: "test-model"),
maxOutputTokens: 100,
cacheHints: nil
)
var collected: [LLMEvent] = []
for try await event in stream {
collected.append(event)
}
XCTAssertEqual(collected.count, 1)
let singleCount = await mock.singleModalCallCount
let multiCount = await mock.multimodalCallCount
let recordedImages = await mock.lastImages
XCTAssertEqual(singleCount, 0,
"Override should be invoked, not the single-modal forwarder")
XCTAssertEqual(multiCount, 1)
XCTAssertEqual(recordedImages.count, 1)
XCTAssertEqual(recordedImages.first?.mediaType, "image/jpeg")
}
// MARK: - Test 3b: override receives images.isEmpty cleanly too
func testMultimodalOverrideHandlesEmptyImages() async throws {
let mock = MockMultimodalLLMProvider()
let stream = mock.stream(
messages: [LLMMessage(role: .user, content: [.text("text-only")])],
images: [],
tools: [],
toolChoice: .auto,
model: ModelID(rawValue: "test-model"),
maxOutputTokens: 100,
cacheHints: nil
)
for try await _ in stream {}
let count = await mock.multimodalCallCount
XCTAssertEqual(count, 1,
"Override is responsible for the full method surface")
}
// MARK: - Test 4: ImageBlock equality + Sendable conformance
func testImageBlockEquatable() {
let a = ImageBlock(mediaType: "image/jpeg", data: Data([0x01, 0x02, 0x03]))
let b = ImageBlock(mediaType: "image/jpeg", data: Data([0x01, 0x02, 0x03]))
let c = ImageBlock(mediaType: "image/png", data: Data([0x01, 0x02, 0x03]))
let d = ImageBlock(mediaType: "image/jpeg", data: Data([0x01, 0x02]))
XCTAssertEqual(a, b)
XCTAssertNotEqual(a, c)
XCTAssertNotEqual(a, d)
}
}
// MARK: - Mocks
/// Single-modal-only conformer. Inherits the multimodal default extension
/// which forwards to this when images.isEmpty.
private actor MockLLMProvider: LLMProvider {
var singleModalCallCount: Int = 0
var multimodalCallCount: Int = 0 // never incremented; default extension can't call it
nonisolated func stream(
messages: [LLMMessage],
tools: [ToolSchema],
toolChoice: ToolChoice,
model: ModelID,
maxOutputTokens: Int,
cacheHints: CacheHints?
) -> AsyncThrowingStream<LLMEvent, Error> {
AsyncThrowingStream<LLMEvent, Error> { continuation in
Task {
await self.bumpSingleModal()
continuation.yield(.messageStop)
continuation.finish()
}
}
}
private func bumpSingleModal() {
singleModalCallCount += 1
}
}
/// Multimodal-overriding conformer. Demonstrates that an explicit override
/// supersedes the default extension's forwarding behavior.
private actor MockMultimodalLLMProvider: LLMProvider {
var singleModalCallCount: Int = 0
var multimodalCallCount: Int = 0
var lastImages: [ImageBlock] = []
nonisolated func stream(
messages: [LLMMessage],
tools: [ToolSchema],
toolChoice: ToolChoice,
model: ModelID,
maxOutputTokens: Int,
cacheHints: CacheHints?
) -> AsyncThrowingStream<LLMEvent, Error> {
AsyncThrowingStream<LLMEvent, Error> { continuation in
Task {
await self.bumpSingleModal()
continuation.yield(.messageStop)
continuation.finish()
}
}
}
nonisolated func stream(
messages: [LLMMessage],
images: [ImageBlock],
tools: [ToolSchema],
toolChoice: ToolChoice,
model: ModelID,
maxOutputTokens: Int,
cacheHints: CacheHints?
) -> AsyncThrowingStream<LLMEvent, Error> {
AsyncThrowingStream<LLMEvent, Error> { continuation in
Task {
await self.bumpMultimodal(images: images)
continuation.yield(.messageStop)
continuation.finish()
}
}
}
private func bumpSingleModal() {
singleModalCallCount += 1
}
private func bumpMultimodal(images: [ImageBlock]) {
multimodalCallCount += 1
lastImages = images
}
}