add(smoke): anthropic continuity replay live validation

Author: Tom-Ryder

Tests/AgentRunKitTests/Smoke/AnthropicSmokeTestHelpers.swift

@@ -0,0 +1,46 @@
+@testable import AgentRunKit
+import Foundation
+import Testing
+
+func assertSmokeAnthropicContinuityReplay(client: some LLMClient) async throws {
+    let addTool = try makeSmokeAddTool()
+    let configuration = AgentConfiguration(
+        maxIterations: 5,
+        systemPrompt: """
+        You are a calculator assistant. When asked to add numbers, call the add tool exactly once.
+        After receiving the tool result, immediately call finish with the numeric answer.
+        Do not answer in plain text and do not call add again.
+        """
+    )
+    let agent = Agent<EmptyContext>(client: client, tools: [addTool], configuration: configuration)
+
+    var firstHistory: [ChatMessage]?
+    for try await event in agent.stream(userMessage: "What is 17 + 25?", context: EmptyContext()) {
+        guard case let .finished(_, _, _, history) = event.kind else { continue }
+        firstHistory = history
+    }
+    let history = try #require(firstHistory)
+
+    let replayableAssistant = history.compactMap { message -> AssistantMessage? in
+        guard case let .assistant(assistant) = message else { return nil }
+        guard assistant.continuity?.substrate == .anthropicMessages else { return nil }
+        return assistant
+    }.first
+    let continuity = try #require(replayableAssistant?.continuity)
+
+    #expect(continuity.substrate == .anthropicMessages)
+    guard case let .object(payload) = continuity.payload,
+          case let .array(blocks) = payload["content"]
+    else {
+        Issue.record("Expected Anthropic continuity payload with content array")
+        return
+    }
+    #expect(!blocks.isEmpty)
+
+    let secondResult = try await agent.run(
+        userMessage: "What was the previous sum? Reply with just the number.",
+        history: history,
+        context: EmptyContext()
+    )
+    #expect(try requireContent(secondResult).contains("42"))
+}

Tests/AgentRunKitTests/Smoke/AnthropicSmokeTests.swift

@@ -142,6 +142,10 @@ struct AnthropicSmokeTests {
         try await assertSmokeReasoningGenerate(client: thinkingClient)
     }
 
+    @Test func continuityReplay() async throws {
+        try await assertSmokeAnthropicContinuityReplay(client: client)
+    }
+
     @Test func approvalGate() async throws {
         try await assertSmokeApprovalGate(client: client)
     }

Tests/AgentRunKitTests/Smoke/VertexAnthropicSmokeTests.swift

@@ -17,6 +17,16 @@ private let hasVertexAnthropicConfig =
 ))
 struct VertexAnthropicSmokeTests {
     private func makeClient() throws -> VertexAnthropicClient {
+        try VertexAnthropicClient(
+            projectID: vertexProjectID,
+            location: vertexLocation,
+            model: anthropicModel,
+            authService: GoogleAuthService(),
+            maxTokens: 1024
+        )
+    }
+
+    private func makeBudgetClient() throws -> VertexAnthropicClient {
         try VertexAnthropicClient(
             projectID: vertexProjectID,
             location: vertexLocation,
@@ -29,6 +39,11 @@ struct VertexAnthropicSmokeTests {
 
     @Test
     func budgetHistoryIntegrity() async throws {
-        try await assertSmokeBudgetHistoryIntegrity(client: makeClient())
+        try await assertSmokeBudgetHistoryIntegrity(client: makeBudgetClient())
+    }
+
+    @Test
+    func continuityReplay() async throws {
+        try await assertSmokeAnthropicContinuityReplay(client: makeClient())
     }
 }