| 1 | import { assert } from "@effect/vitest"; |
| 2 | import type { ProviderReplayTranscript } from "@t3tools/contracts"; |
| 3 | |
| 4 | import type { OrchestratorV2ScenarioResult } from "../../OrchestratorScenario.ts"; |
| 5 | import { |
| 6 | assertBaseProjection, |
| 7 | assertRunOrdinals, |
| 8 | assertSemanticProjectionIntegrity, |
| 9 | assertTurnItemTypeSequence, |
| 10 | assertVisibleTurnItemTypeSequence, |
| 11 | assertVisibleUserMessagesExclude, |
| 12 | assertVisibleUserMessagesInclude, |
| 13 | assertUserMessagesInclude, |
| 14 | projectionFor, |
| 15 | THREAD_ROLLBACK_AFTER_PROMPT, |
| 16 | THREAD_ROLLBACK_FIRST_PROMPT, |
| 17 | THREAD_ROLLBACK_SECOND_PROMPT, |
| 18 | } from "../shared.ts"; |
| 19 | |
| 20 | export function assertThreadRollbackOutput( |
| 21 | result: OrchestratorV2ScenarioResult, |
| 22 | transcript: ProviderReplayTranscript, |
| 23 | ) { |
| 24 | assertBaseProjection({ |
| 25 | result, |
| 26 | transcript, |
| 27 | runCount: 3, |
| 28 | runStatuses: ["completed", "rolled_back", "completed"], |
| 29 | }); |
| 30 | |
| 31 | const projection = projectionFor(result, transcript.scenario); |
| 32 | assertSemanticProjectionIntegrity(projection); |
| 33 | assertRunOrdinals(projection, [1, 2, 3]); |
| 34 | assertTurnItemTypeSequence(projection, [ |
| 35 | "user_message", |
| 36 | "assistant_message", |
| 37 | "checkpoint", |
| 38 | "user_message", |
| 39 | "assistant_message", |
| 40 | "checkpoint", |
| 41 | "user_message", |
| 42 | "assistant_message", |
| 43 | "checkpoint", |
| 44 | ]); |
| 45 | assertVisibleTurnItemTypeSequence(projection, [ |
| 46 | "user_message", |
| 47 | "assistant_message", |
| 48 | "checkpoint", |
| 49 | "user_message", |
| 50 | "assistant_message", |
| 51 | "checkpoint", |
| 52 | ]); |
| 53 | assertUserMessagesInclude(projection, [ |
| 54 | THREAD_ROLLBACK_FIRST_PROMPT, |
| 55 | THREAD_ROLLBACK_SECOND_PROMPT, |
| 56 | THREAD_ROLLBACK_AFTER_PROMPT, |
| 57 | ]); |
| 58 | assertVisibleUserMessagesInclude(projection, [ |
| 59 | THREAD_ROLLBACK_FIRST_PROMPT, |
| 60 | THREAD_ROLLBACK_AFTER_PROMPT, |
| 61 | ]); |
| 62 | assertVisibleUserMessagesExclude(projection, [THREAD_ROLLBACK_SECOND_PROMPT]); |
| 63 | // Codex's own context was rewound: the post-rollback recall omits the reverted turn. |
| 64 | const recall = projection.turnItems.findLast((item) => item.type === "assistant_message"); |
| 65 | assert.include(recall?.type === "assistant_message" ? recall.text : "", "first turn complete"); |
| 66 | assert.notInclude(recall?.type === "assistant_message" ? recall.text : "", "second turn"); |
| 67 | assert.isAtLeast(projection.checkpoints.length, 2); |
| 68 | assert.isTrue( |
| 69 | projection.runs.some((run) => run.status === "rolled_back"), |
| 70 | "rollback must be visible in run state", |
| 71 | ); |
| 72 | } |