| 1 | import { ProviderDriverKind } from "@t3tools/contracts"; |
| 2 | |
| 3 | import { claudeBackgroundSubagentAfterRootInput } from "./claude_background_subagent_after_root/input.ts"; |
| 4 | import { assertClaudeBackgroundSubagentAfterRootOutput } from "./claude_background_subagent_after_root/output.ts"; |
| 5 | import { claudeBackgroundTaskAfterRootInput } from "./claude_background_task_after_root/input.ts"; |
| 6 | import { assertClaudeBackgroundTaskAfterRootOutput } from "./claude_background_task_after_root/output.ts"; |
| 7 | import { claudeBackgroundSubagentLifecycleInput } from "./claude_background_subagent_lifecycle/input.ts"; |
| 8 | import { assertClaudeBackgroundSubagentLifecycleOutput } from "./claude_background_subagent_lifecycle/output.ts"; |
| 9 | import { claudeBackgroundTaskInterruptInput } from "./claude_background_task_interrupt/input.ts"; |
| 10 | import { claudeBackgroundWakeBeforeQueuedPromptInput } from "./claude_background_wake_before_queued_prompt/input.ts"; |
| 11 | import { assertClaudeBackgroundWakeBeforeQueuedPromptOutput } from "./claude_background_wake_before_queued_prompt/output.ts"; |
| 12 | import { claudeBackgroundWakeBeforeQueuedPromptNoEchoInput } from "./claude_background_wake_before_queued_prompt_no_echo/input.ts"; |
| 13 | import { assertClaudeBackgroundWakeBeforeQueuedPromptNoEchoOutput } from "./claude_background_wake_before_queued_prompt_no_echo/output.ts"; |
| 14 | import { assertClaudeBackgroundTaskInterruptOutput } from "./claude_background_task_interrupt/output.ts"; |
| 15 | import { claudeBackgroundMonitorWakeInput } from "./claude_background_monitor_wake/input.ts"; |
| 16 | import { assertClaudeBackgroundMonitorWakeOutput } from "./claude_background_monitor_wake/output.ts"; |
| 17 | import { claudeBackgroundTaskWakeInput } from "./claude_background_task_wake/input.ts"; |
| 18 | import { assertClaudeBackgroundTaskWakeOutput } from "./claude_background_task_wake/output.ts"; |
| 19 | import { |
| 20 | claudeCompactAfterPeerTurnInput, |
| 21 | claudeCompactAfterPeerTurnNoEchoInput, |
| 22 | } from "./claude_compact_after_peer_turn/input.ts"; |
| 23 | import { assertClaudeCompactAfterPeerTurnOutput } from "./claude_compact_after_peer_turn/output.ts"; |
| 24 | import { assertClaudeCompactAfterPeerTurnNoEchoOutput } from "./claude_compact_after_peer_turn_no_echo/output.ts"; |
| 25 | import { claudeCompactAfterResumeWakeInput } from "./claude_compact_after_resume_wake/input.ts"; |
| 26 | import { assertClaudeCompactAfterResumeWakeOutput } from "./claude_compact_after_resume_wake/output.ts"; |
| 27 | import { claudeIdleResumeInput } from "./claude_idle_resume/input.ts"; |
| 28 | import { assertClaudeIdleResumeOutput } from "./claude_idle_resume/output.ts"; |
| 29 | import { claudeLocalBashTaskInput } from "./claude_local_bash_task/input.ts"; |
| 30 | import { assertClaudeLocalBashTaskOutput } from "./claude_local_bash_task/output.ts"; |
| 31 | import { claudeNestedBackgroundSubagentWakeInput } from "./claude_nested_background_subagent_wake/input.ts"; |
| 32 | import { assertClaudeNestedBackgroundSubagentWakeOutput } from "./claude_nested_background_subagent_wake/output.ts"; |
| 33 | import { claudeNestedSubagentModelInput } from "./claude_nested_subagent_model/input.ts"; |
| 34 | import { assertClaudeNestedSubagentModelOutput } from "./claude_nested_subagent_model/output.ts"; |
| 35 | import { claudeMcpToolPresentationInput } from "./claude_mcp_tool_presentation/input.ts"; |
| 36 | import { assertClaudeMcpToolPresentationOutput } from "./claude_mcp_tool_presentation/output.ts"; |
| 37 | import { claudeResultIsErrorInput } from "./claude_result_is_error/input.ts"; |
| 38 | import { assertClaudeResultIsErrorOutput } from "./claude_result_is_error/output.ts"; |
| 39 | import { grokAutoBlockedCommandInput } from "./grok_auto_blocked_command/input.ts"; |
| 40 | import { assertGrokAutoBlockedCommandOutput } from "./grok_auto_blocked_command/output.ts"; |
| 41 | import { grokBackgroundBashInput } from "./grok_background_bash/input.ts"; |
| 42 | import { assertGrokBackgroundBashOutput } from "./grok_background_bash/output.ts"; |
| 43 | import { grokBackgroundBashFastWakeInput } from "./grok_background_bash_fast_wake/input.ts"; |
| 44 | import { assertGrokBackgroundBashFastWakeOutput } from "./grok_background_bash_fast_wake/output.ts"; |
| 45 | import { grokBackgroundSubagentInput } from "./grok_background_subagent/input.ts"; |
| 46 | import { assertGrokBackgroundSubagentOutput } from "./grok_background_subagent/output.ts"; |
| 47 | import { grokMonitorInput } from "./grok_monitor/input.ts"; |
| 48 | import { assertGrokMonitorOutput } from "./grok_monitor/output.ts"; |
| 49 | import { grokPromptErrorInput } from "./grok_prompt_error/input.ts"; |
| 50 | import { assertGrokPromptErrorOutput } from "./grok_prompt_error/output.ts"; |
| 51 | import { grokSubagentLineageInput } from "./grok_subagent_lineage/input.ts"; |
| 52 | import { assertGrokSubagentLineageOutput } from "./grok_subagent_lineage/output.ts"; |
| 53 | import { assertClaudeMessageSteeringOutput } from "./message_steering/claude_output.ts"; |
| 54 | import { assertMessageSteeringOutput } from "./message_steering/codex_output.ts"; |
| 55 | import { assertCursorMessageSteeringOutput } from "./message_steering/cursor_output.ts"; |
| 56 | import { assertGrokMessageSteeringOutput } from "./message_steering/grok_output.ts"; |
| 57 | import { messageSteeringInput } from "./message_steering/input.ts"; |
| 58 | import { assertPiMessageSteeringOutput } from "./message_steering/pi_output.ts"; |
| 59 | import { MUSE_PERMISSION_DECLINED_FILE, musePermissionInput } from "./muse_permission/input.ts"; |
| 60 | import { assertMusePermissionOutput } from "./muse_permission/output.ts"; |
| 61 | import { museSubagentApprovalInput } from "./muse_subagent_approval/input.ts"; |
| 62 | import { assertMuseSubagentApprovalOutput } from "./muse_subagent_approval/output.ts"; |
| 63 | import { museSubagentInterruptInput } from "./muse_subagent_interrupt/input.ts"; |
| 64 | import { assertMuseSubagentInterruptOutput } from "./muse_subagent_interrupt/output.ts"; |
| 65 | import { museWorkflowInput } from "./muse_workflow/input.ts"; |
| 66 | import { assertMuseWorkflowOutput } from "./muse_workflow/output.ts"; |
| 67 | import { assertMuseSubagentV2Output } from "./subagent_v2/muse_output.ts"; |
| 68 | import { assertMuseMultiTurnOutput } from "./multi_turn/muse_output.ts"; |
| 69 | import { assertMuseTurnInterruptOutput } from "./turn_interrupt/muse_output.ts"; |
| 70 | import { piCompactionInput } from "./pi_compaction/input.ts"; |
| 71 | import { assertPiCompactionOutput } from "./pi_compaction/output.ts"; |
| 72 | import { providerThreadResumeInput } from "./provider_thread_resume/input.ts"; |
| 73 | import { assertPiProviderThreadResumeOutput } from "./provider_thread_resume/pi_output.ts"; |
| 74 | import { assertMultiTurnClaudeOutput } from "./multi_turn/claude_output.ts"; |
| 75 | import { assertMultiTurnOutput } from "./multi_turn/codex_output.ts"; |
| 76 | import { assertPiMultiTurnOutput } from "./multi_turn/pi_output.ts"; |
| 77 | import { multiTurnInput } from "./multi_turn/input.ts"; |
| 78 | import { openCodeChildApprovalInput } from "./opencode_child_approval/input.ts"; |
| 79 | import { assertOpenCodeChildApprovalOutput } from "./opencode_child_approval/output.ts"; |
| 80 | import { openCodeRunningChildApprovalInput } from "./opencode_running_child_approval/input.ts"; |
| 81 | import { assertOpenCodeRunningChildApprovalOutput } from "./opencode_running_child_approval/output.ts"; |
| 82 | import { openCodeSubagentInput } from "./opencode_subagent/input.ts"; |
| 83 | import { openCode2InboxInput } from "./opencode2_inbox/input.ts"; |
| 84 | import { openCode2RevertInput } from "./opencode2_revert/input.ts"; |
| 85 | import { assertOpenCode2RevertOutput } from "./opencode2_revert/output.ts"; |
| 86 | import { assertOpenCode2InboxOutput } from "./opencode2_inbox/output.ts"; |
| 87 | import { openCode2CommandInput } from "./opencode2_command/input.ts"; |
| 88 | import { assertOpenCode2CommandOutput } from "./opencode2_command/output.ts"; |
| 89 | import { openCode2CompactionInput } from "./opencode2_compaction/input.ts"; |
| 90 | import { assertOpenCode2CompactionOutput } from "./opencode2_compaction/output.ts"; |
| 91 | import { openCode2InterruptInput } from "./opencode2_interrupt/input.ts"; |
| 92 | import { assertOpenCode2InterruptOutput } from "./opencode2_interrupt/output.ts"; |
| 93 | import { openCode2PermissionInput } from "./opencode2_permission/input.ts"; |
| 94 | import { assertOpenCode2PermissionOutput } from "./opencode2_permission/output.ts"; |
| 95 | import { openCode2QuestionInput } from "./opencode2_question/input.ts"; |
| 96 | import { assertOpenCode2QuestionOutput } from "./opencode2_question/output.ts"; |
| 97 | import { openCode2BackgroundInput } from "./opencode2_background/input.ts"; |
| 98 | import { assertOpenCode2BackgroundOutput } from "./opencode2_background/output.ts"; |
| 99 | import { openCode2NestedBackgroundInput } from "./opencode2_nested_background/input.ts"; |
| 100 | import { assertOpenCode2NestedBackgroundOutput } from "./opencode2_nested_background/output.ts"; |
| 101 | import { openCode2SubagentInput } from "./opencode2_subagent/input.ts"; |
| 102 | import { assertOpenCode2SubagentOutput } from "./opencode2_subagent/output.ts"; |
| 103 | import { openCode2ResumeAfterRestartInput } from "./opencode2_resume_after_restart/input.ts"; |
| 104 | import { assertOpenCode2ResumeAfterRestartOutput } from "./opencode2_resume_after_restart/output.ts"; |
| 105 | import { openCode2SimpleInput } from "./opencode2_simple/input.ts"; |
| 106 | import { openCode2SkillInput } from "./opencode2_skill/input.ts"; |
| 107 | import { assertOpenCode2SkillOutput } from "./opencode2_skill/output.ts"; |
| 108 | import { assertOpenCode2SimpleOutput } from "./opencode2_simple/output.ts"; |
| 109 | import { openCode2ToolCallInput } from "./opencode2_tool_call/input.ts"; |
| 110 | import { assertOpenCode2ToolCallOutput } from "./opencode2_tool_call/output.ts"; |
| 111 | import { assertOpenCodeSubagentOutput } from "./opencode_subagent/output.ts"; |
| 112 | import { |
| 113 | assertCodexPlanQuestionsOutput, |
| 114 | assertPlanQuestionsOutput, |
| 115 | } from "./plan_questions/codex_output.ts"; |
| 116 | import { assertOpenCodePlanQuestionsOutput } from "./plan_questions/opencode_output.ts"; |
| 117 | import { planQuestionsInput } from "./plan_questions/input.ts"; |
| 118 | import { assertProposedPlanOutput } from "./proposed_plan/codex_output.ts"; |
| 119 | import { assertProposedPlanCursorOutput } from "./proposed_plan/cursor_output.ts"; |
| 120 | import { proposedPlanInput } from "./proposed_plan/input.ts"; |
| 121 | import { assertQueuedCancelledWhileActiveOutput } from "./queued_cancelled_while_active/codex_output.ts"; |
| 122 | import { queuedCancelledWhileActiveInput } from "./queued_cancelled_while_active/input.ts"; |
| 123 | import { assertQueuedTurnOutput } from "./queued_turn/codex_output.ts"; |
| 124 | import { queuedTurnInput } from "./queued_turn/input.ts"; |
| 125 | import { assertSimpleClaudeOutput } from "./simple/claude_output.ts"; |
| 126 | import { assertSkillInvocationCursorOutput } from "./skill_invocation/cursor_output.ts"; |
| 127 | import { skillInvocationInput } from "./skill_invocation/input.ts"; |
| 128 | import { assertSimpleOutput } from "./simple/codex_output.ts"; |
| 129 | import { assertPiSimpleOutput } from "./simple/pi_output.ts"; |
| 130 | import { simpleInput } from "./simple/input.ts"; |
| 131 | import { assertSubagentOutput } from "./subagent/codex_output.ts"; |
| 132 | import { assertClaudeSubagentOutput } from "./subagent/claude_output.ts"; |
| 133 | import { subagentInput } from "./subagent/input.ts"; |
| 134 | import { assertCursorSubagentOutput } from "./subagent/cursor_output.ts"; |
| 135 | import { assertSubagentContinueOutput } from "./subagent_continue/codex_output.ts"; |
| 136 | import { subagentContinueInput } from "./subagent_continue/input.ts"; |
| 137 | import { assertSubagentV2Output } from "./subagent_v2/codex_output.ts"; |
| 138 | import { subagentV2Input, subagentV2NestedInput } from "./subagent_v2/input.ts"; |
| 139 | import { assertSubagentV2ApprovalOutput } from "./subagent_v2_approval/codex_output.ts"; |
| 140 | import { |
| 141 | SUBAGENT_V2_APPROVAL_POLICY, |
| 142 | subagentV2ApprovalInput, |
| 143 | } from "./subagent_v2_approval/input.ts"; |
| 144 | import { assertSubagentV2NestedOutput } from "./subagent_v2_nested/codex_output.ts"; |
| 145 | import { assertSubagentV2NestedApprovalOutput } from "./subagent_v2_nested_approval/codex_output.ts"; |
| 146 | import { subagentV2NestedApprovalInput } from "./subagent_v2_nested_approval/input.ts"; |
| 147 | import { assertClaudeThreadRollbackOutput } from "./thread_rollback/claude_output.ts"; |
| 148 | import { assertThreadRollbackOutput } from "./thread_rollback/codex_output.ts"; |
| 149 | import { threadRollbackInput } from "./thread_rollback/input.ts"; |
| 150 | import { assertPiThreadRollbackOutput } from "./thread_rollback/pi_output.ts"; |
| 151 | import { assertThreadRollbackAfterRestartOutput } from "./thread_rollback_after_restart/codex_output.ts"; |
| 152 | import { threadRollbackAfterRestartInput } from "./thread_rollback_after_restart/input.ts"; |
| 153 | import { threadRollbackAfterStopInput } from "./thread_rollback_after_stop/input.ts"; |
| 154 | import { assertPiThreadRollbackAfterStopOutput } from "./thread_rollback_after_stop/pi_output.ts"; |
| 155 | import { assertThreadRollbackToStoppedTurnOutput } from "./thread_rollback_to_stopped_turn/codex_output.ts"; |
| 156 | import { threadRollbackToStoppedTurnInput } from "./thread_rollback_to_stopped_turn/input.ts"; |
| 157 | import { assertTodoListOutput } from "./todo_list/codex_output.ts"; |
| 158 | import { assertTodoListCursorOutput } from "./todo_list/cursor_output.ts"; |
| 159 | import { assertTodoListGrokOutput } from "./todo_list/grok_output.ts"; |
| 160 | import { todoListInput } from "./todo_list/input.ts"; |
| 161 | import { assertToolCallDeniedWriteClaudeOutput } from "./tool_call_denied_write/claude_output.ts"; |
| 162 | import { |
| 163 | DENIED_WRITE_POLICY, |
| 164 | TOOL_CALL_DENIED_WRITE_TARGET, |
| 165 | toolCallDeniedWriteInput, |
| 166 | } from "./tool_call_denied_write/input.ts"; |
| 167 | import { assertToolCallReadOnlyClaudeOutput } from "./tool_call_read_only/claude_output.ts"; |
| 168 | import { assertToolCallReadOnlyCursorOutput } from "./tool_call_read_only/cursor_output.ts"; |
| 169 | import { toolCallReadOnlyInput } from "./tool_call_read_only/input.ts"; |
| 170 | import { |
| 171 | assertToolCallReadOnlyOnRequestGrokOutput, |
| 172 | assertToolCallReadOnlyOnRequestOutput, |
| 173 | } from "./tool_call_read_only_on_request/output.ts"; |
| 174 | import { toolCallReadOnlyOnRequestInput } from "./tool_call_read_only_on_request/input.ts"; |
| 175 | import { |
| 176 | stopBackgroundWorkAfterFailedTurnInput, |
| 177 | stopBackgroundWorkAfterReleaseInput, |
| 178 | } from "./stop_background_work_after_failed_turn/input.ts"; |
| 179 | import { assertStopBackgroundWorkAfterFailedTurnOutput } from "./stop_background_work_after_failed_turn/output.ts"; |
| 180 | import { assertToolCallRestrictedGranularClaudeOutput } from "./tool_call_restricted_granular/claude_output.ts"; |
| 181 | import { assertToolCallRestrictedGranularOutput } from "./tool_call_restricted_granular/codex_output.ts"; |
| 182 | import { toolCallRestrictedGranularInput } from "./tool_call_restricted_granular/input.ts"; |
| 183 | import { assertToolCallWorkspaceNeverClaudeOutput } from "./tool_call_workspace_never/claude_output.ts"; |
| 184 | import { assertToolCallWorkspaceNeverOutput } from "./tool_call_workspace_never/codex_output.ts"; |
| 185 | import { toolCallWorkspaceNeverInput } from "./tool_call_workspace_never/input.ts"; |
| 186 | import { assertTurnInterruptClaudeOutput } from "./turn_interrupt/claude_output.ts"; |
| 187 | import { assertTurnInterruptOutput } from "./turn_interrupt/codex_output.ts"; |
| 188 | import { turnInterruptInput } from "./turn_interrupt/input.ts"; |
| 189 | import { assertTurnInterruptMidToolClaudeOutput } from "./turn_interrupt_mid_tool/claude_output.ts"; |
| 190 | import { assertTurnInterruptMidToolCodexOutput } from "./turn_interrupt_mid_tool/codex_output.ts"; |
| 191 | import { assertTurnInterruptMidToolCursorOutput } from "./turn_interrupt_mid_tool/cursor_output.ts"; |
| 192 | import { turnInterruptMidToolInput } from "./turn_interrupt_mid_tool/input.ts"; |
| 193 | import { assertTurnInterruptMidToolPiOutput } from "./turn_interrupt_mid_tool/pi_output.ts"; |
| 194 | import { assertTurnInterruptRestartClaudeOutput } from "./turn_interrupt_restart/claude_output.ts"; |
| 195 | import { turnInterruptRestartInput } from "./turn_interrupt_restart/input.ts"; |
| 196 | import { assertClaudeWebSearchOutput } from "./web_search/claude_output.ts"; |
| 197 | import { assertWebSearchOutput } from "./web_search/codex_output.ts"; |
| 198 | import { webSearchInput } from "./web_search/input.ts"; |
| 199 | import { |
| 200 | ACP_REGISTRY_MODEL_SELECTION, |
| 201 | CLAUDE_MODEL_SELECTION, |
| 202 | CODEX_MODEL_SELECTION, |
| 203 | CURSOR_MODEL_SELECTION, |
| 204 | GROK_MODEL_SELECTION, |
| 205 | MUSE_MODEL_SELECTION, |
| 206 | OPENCODE_MODEL_SELECTION, |
| 207 | OPENCODE2_MODEL_SELECTION, |
| 208 | PI_MODEL_SELECTION, |
| 209 | READ_ONLY_NEVER_POLICY, |
| 210 | READ_ONLY_ON_REQUEST_POLICY, |
| 211 | RESTRICTED_GRANULAR_POLICY, |
| 212 | type OrchestratorReplayFixture, |
| 213 | WORKSPACE_NEVER_POLICY, |
| 214 | } from "./shared.ts"; |
| 215 | |
| 216 | export const ORCHESTRATOR_REPLAY_FIXTURES: ReadonlyArray<OrchestratorReplayFixture> = [ |
| 217 | { |
| 218 | name: "claude_background_subagent_after_root", |
| 219 | buildInput: claudeBackgroundSubagentAfterRootInput, |
| 220 | providers: [ |
| 221 | { |
| 222 | driver: ProviderDriverKind.make("claudeAgent"), |
| 223 | transcriptFile: new URL( |
| 224 | "./claude_background_subagent_after_root/claude_transcript.ndjson", |
| 225 | import.meta.url, |
| 226 | ), |
| 227 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 228 | runContinuationWorker: true, |
| 229 | assertOutput: assertClaudeBackgroundSubagentAfterRootOutput, |
| 230 | }, |
| 231 | ], |
| 232 | }, |
| 233 | { |
| 234 | name: "claude_background_task_after_root", |
| 235 | buildInput: claudeBackgroundTaskAfterRootInput, |
| 236 | providers: [ |
| 237 | { |
| 238 | driver: ProviderDriverKind.make("claudeAgent"), |
| 239 | transcriptFile: new URL( |
| 240 | "./claude_background_task_after_root/claude_transcript.ndjson", |
| 241 | import.meta.url, |
| 242 | ), |
| 243 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 244 | assertOutput: assertClaudeBackgroundTaskAfterRootOutput, |
| 245 | }, |
| 246 | ], |
| 247 | }, |
| 248 | { |
| 249 | name: "claude_background_subagent_lifecycle", |
| 250 | buildInput: claudeBackgroundSubagentLifecycleInput, |
| 251 | providers: [ |
| 252 | { |
| 253 | driver: ProviderDriverKind.make("claudeAgent"), |
| 254 | transcriptFile: new URL( |
| 255 | "./claude_background_subagent_lifecycle/claude_transcript.ndjson", |
| 256 | import.meta.url, |
| 257 | ), |
| 258 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 259 | runContinuationWorker: true, |
| 260 | assertOutput: assertClaudeBackgroundSubagentLifecycleOutput, |
| 261 | }, |
| 262 | ], |
| 263 | }, |
| 264 | { |
| 265 | name: "claude_background_task_interrupt", |
| 266 | buildInput: claudeBackgroundTaskInterruptInput, |
| 267 | providers: [ |
| 268 | { |
| 269 | driver: ProviderDriverKind.make("claudeAgent"), |
| 270 | transcriptFile: new URL( |
| 271 | "./claude_background_task_interrupt/claude_transcript.ndjson", |
| 272 | import.meta.url, |
| 273 | ), |
| 274 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 275 | runContinuationWorker: true, |
| 276 | assertOutput: assertClaudeBackgroundTaskInterruptOutput, |
| 277 | }, |
| 278 | ], |
| 279 | }, |
| 280 | { |
| 281 | name: "claude_background_wake_before_queued_prompt", |
| 282 | buildInput: claudeBackgroundWakeBeforeQueuedPromptInput, |
| 283 | providers: [ |
| 284 | { |
| 285 | driver: ProviderDriverKind.make("claudeAgent"), |
| 286 | transcriptFile: new URL( |
| 287 | "./claude_background_wake_before_queued_prompt/claude_transcript.ndjson", |
| 288 | import.meta.url, |
| 289 | ), |
| 290 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 291 | runContinuationWorker: true, |
| 292 | assertOutput: assertClaudeBackgroundWakeBeforeQueuedPromptOutput, |
| 293 | }, |
| 294 | ], |
| 295 | }, |
| 296 | { |
| 297 | name: "claude_background_wake_before_queued_prompt_no_echo", |
| 298 | buildInput: claudeBackgroundWakeBeforeQueuedPromptNoEchoInput, |
| 299 | providers: [ |
| 300 | { |
| 301 | driver: ProviderDriverKind.make("claudeAgent"), |
| 302 | transcriptFile: new URL( |
| 303 | "./claude_background_wake_before_queued_prompt_no_echo/claude_transcript.ndjson", |
| 304 | import.meta.url, |
| 305 | ), |
| 306 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 307 | runContinuationWorker: true, |
| 308 | assertOutput: assertClaudeBackgroundWakeBeforeQueuedPromptNoEchoOutput, |
| 309 | }, |
| 310 | ], |
| 311 | }, |
| 312 | { |
| 313 | name: "claude_compact_after_resume_wake", |
| 314 | buildInput: claudeCompactAfterResumeWakeInput, |
| 315 | providers: [ |
| 316 | { |
| 317 | driver: ProviderDriverKind.make("claudeAgent"), |
| 318 | transcriptFile: new URL( |
| 319 | "./claude_compact_after_resume_wake/claude_transcript.ndjson", |
| 320 | import.meta.url, |
| 321 | ), |
| 322 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 323 | runContinuationWorker: true, |
| 324 | assertOutput: assertClaudeCompactAfterResumeWakeOutput, |
| 325 | }, |
| 326 | ], |
| 327 | }, |
| 328 | { |
| 329 | name: "claude_compact_after_peer_turn", |
| 330 | buildInput: claudeCompactAfterPeerTurnInput, |
| 331 | providers: [ |
| 332 | { |
| 333 | driver: ProviderDriverKind.make("claudeAgent"), |
| 334 | transcriptFile: new URL( |
| 335 | "./claude_compact_after_peer_turn/claude_transcript.ndjson", |
| 336 | import.meta.url, |
| 337 | ), |
| 338 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 339 | runContinuationWorker: true, |
| 340 | assertOutput: assertClaudeCompactAfterPeerTurnOutput, |
| 341 | }, |
| 342 | ], |
| 343 | }, |
| 344 | { |
| 345 | name: "claude_compact_after_peer_turn_no_echo", |
| 346 | buildInput: claudeCompactAfterPeerTurnNoEchoInput, |
| 347 | providers: [ |
| 348 | { |
| 349 | driver: ProviderDriverKind.make("claudeAgent"), |
| 350 | transcriptFile: new URL( |
| 351 | "./claude_compact_after_peer_turn_no_echo/claude_transcript.ndjson", |
| 352 | import.meta.url, |
| 353 | ), |
| 354 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 355 | runContinuationWorker: true, |
| 356 | assertOutput: assertClaudeCompactAfterPeerTurnNoEchoOutput, |
| 357 | }, |
| 358 | ], |
| 359 | }, |
| 360 | { |
| 361 | name: "claude_background_monitor_wake", |
| 362 | buildInput: claudeBackgroundMonitorWakeInput, |
| 363 | providers: [ |
| 364 | { |
| 365 | driver: ProviderDriverKind.make("claudeAgent"), |
| 366 | transcriptFile: new URL( |
| 367 | "./claude_background_monitor_wake/claude_transcript.ndjson", |
| 368 | import.meta.url, |
| 369 | ), |
| 370 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 371 | runContinuationWorker: true, |
| 372 | assertOutput: assertClaudeBackgroundMonitorWakeOutput, |
| 373 | }, |
| 374 | ], |
| 375 | }, |
| 376 | { |
| 377 | name: "claude_background_task_wake", |
| 378 | buildInput: claudeBackgroundTaskWakeInput, |
| 379 | providers: [ |
| 380 | { |
| 381 | driver: ProviderDriverKind.make("claudeAgent"), |
| 382 | transcriptFile: new URL( |
| 383 | "./claude_background_task_wake/claude_transcript.ndjson", |
| 384 | import.meta.url, |
| 385 | ), |
| 386 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 387 | runContinuationWorker: true, |
| 388 | assertOutput: assertClaudeBackgroundTaskWakeOutput, |
| 389 | }, |
| 390 | ], |
| 391 | }, |
| 392 | { |
| 393 | name: "claude_local_bash_task", |
| 394 | buildInput: claudeLocalBashTaskInput, |
| 395 | providers: [ |
| 396 | { |
| 397 | driver: ProviderDriverKind.make("claudeAgent"), |
| 398 | transcriptFile: new URL( |
| 399 | "./claude_local_bash_task/claude_transcript.ndjson", |
| 400 | import.meta.url, |
| 401 | ), |
| 402 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 403 | assertOutput: assertClaudeLocalBashTaskOutput, |
| 404 | }, |
| 405 | ], |
| 406 | }, |
| 407 | { |
| 408 | name: "claude_nested_background_subagent_wake", |
| 409 | buildInput: claudeNestedBackgroundSubagentWakeInput, |
| 410 | providers: [ |
| 411 | { |
| 412 | driver: ProviderDriverKind.make("claudeAgent"), |
| 413 | transcriptFile: new URL( |
| 414 | "./claude_nested_background_subagent_wake/claude_transcript.ndjson", |
| 415 | import.meta.url, |
| 416 | ), |
| 417 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 418 | runContinuationWorker: true, |
| 419 | assertOutput: assertClaudeNestedBackgroundSubagentWakeOutput, |
| 420 | }, |
| 421 | ], |
| 422 | }, |
| 423 | { |
| 424 | name: "claude_nested_subagent_model", |
| 425 | buildInput: claudeNestedSubagentModelInput, |
| 426 | providers: [ |
| 427 | { |
| 428 | driver: ProviderDriverKind.make("claudeAgent"), |
| 429 | transcriptFile: new URL( |
| 430 | "./claude_nested_subagent_model/claude_transcript.ndjson", |
| 431 | import.meta.url, |
| 432 | ), |
| 433 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 434 | assertOutput: assertClaudeNestedSubagentModelOutput, |
| 435 | }, |
| 436 | ], |
| 437 | }, |
| 438 | { |
| 439 | name: "claude_mcp_tool_presentation", |
| 440 | buildInput: claudeMcpToolPresentationInput, |
| 441 | providers: [ |
| 442 | { |
| 443 | driver: ProviderDriverKind.make("claudeAgent"), |
| 444 | transcriptFile: new URL( |
| 445 | "./claude_mcp_tool_presentation/claude_transcript.ndjson", |
| 446 | import.meta.url, |
| 447 | ), |
| 448 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 449 | assertOutput: assertClaudeMcpToolPresentationOutput, |
| 450 | }, |
| 451 | ], |
| 452 | }, |
| 453 | { |
| 454 | name: "claude_idle_resume", |
| 455 | buildInput: claudeIdleResumeInput, |
| 456 | providers: [ |
| 457 | { |
| 458 | driver: ProviderDriverKind.make("claudeAgent"), |
| 459 | transcriptFile: new URL("./claude_idle_resume/claude_transcript.ndjson", import.meta.url), |
| 460 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 461 | assertOutput: assertClaudeIdleResumeOutput, |
| 462 | }, |
| 463 | ], |
| 464 | }, |
| 465 | { |
| 466 | name: "claude_result_is_error", |
| 467 | buildInput: claudeResultIsErrorInput, |
| 468 | providers: [ |
| 469 | { |
| 470 | driver: ProviderDriverKind.make("claudeAgent"), |
| 471 | transcriptFile: new URL( |
| 472 | "./claude_result_is_error/claude_transcript.ndjson", |
| 473 | import.meta.url, |
| 474 | ), |
| 475 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 476 | assertOutput: assertClaudeResultIsErrorOutput, |
| 477 | }, |
| 478 | ], |
| 479 | }, |
| 480 | { |
| 481 | name: "grok_auto_blocked_command", |
| 482 | buildInput: grokAutoBlockedCommandInput, |
| 483 | providers: [ |
| 484 | { |
| 485 | driver: ProviderDriverKind.make("grok"), |
| 486 | transcriptFile: new URL( |
| 487 | "./grok_auto_blocked_command/grok_transcript.ndjson", |
| 488 | import.meta.url, |
| 489 | ), |
| 490 | modelSelection: { ...GROK_MODEL_SELECTION, model: "grok-4.7-build-fast" }, |
| 491 | assertOutput: assertGrokAutoBlockedCommandOutput, |
| 492 | }, |
| 493 | ], |
| 494 | }, |
| 495 | { |
| 496 | name: "grok_background_bash", |
| 497 | buildInput: grokBackgroundBashInput, |
| 498 | providers: [ |
| 499 | { |
| 500 | driver: ProviderDriverKind.make("grok"), |
| 501 | transcriptFile: new URL("./grok_background_bash/grok_transcript.ndjson", import.meta.url), |
| 502 | modelSelection: { ...GROK_MODEL_SELECTION, model: "grok-4.7-build-fast" }, |
| 503 | runContinuationWorker: true, |
| 504 | assertOutput: assertGrokBackgroundBashOutput, |
| 505 | }, |
| 506 | ], |
| 507 | }, |
| 508 | { |
| 509 | name: "grok_background_bash_fast_wake", |
| 510 | buildInput: grokBackgroundBashFastWakeInput, |
| 511 | providers: [ |
| 512 | { |
| 513 | driver: ProviderDriverKind.make("grok"), |
| 514 | transcriptFile: new URL( |
| 515 | "./grok_background_bash_fast_wake/grok_transcript.ndjson", |
| 516 | import.meta.url, |
| 517 | ), |
| 518 | modelSelection: { ...GROK_MODEL_SELECTION, model: "grok-4.7-build-fast" }, |
| 519 | runContinuationWorker: true, |
| 520 | assertOutput: assertGrokBackgroundBashFastWakeOutput, |
| 521 | }, |
| 522 | ], |
| 523 | }, |
| 524 | { |
| 525 | name: "grok_background_subagent", |
| 526 | buildInput: grokBackgroundSubagentInput, |
| 527 | providers: [ |
| 528 | { |
| 529 | driver: ProviderDriverKind.make("grok"), |
| 530 | transcriptFile: new URL( |
| 531 | "./grok_background_subagent/grok_transcript.ndjson", |
| 532 | import.meta.url, |
| 533 | ), |
| 534 | modelSelection: GROK_MODEL_SELECTION, |
| 535 | runContinuationWorker: true, |
| 536 | assertOutput: assertGrokBackgroundSubagentOutput, |
| 537 | }, |
| 538 | ], |
| 539 | }, |
| 540 | { |
| 541 | name: "grok_monitor", |
| 542 | buildInput: grokMonitorInput, |
| 543 | providers: [ |
| 544 | { |
| 545 | driver: ProviderDriverKind.make("grok"), |
| 546 | transcriptFile: new URL("./grok_monitor/grok_transcript.ndjson", import.meta.url), |
| 547 | modelSelection: GROK_MODEL_SELECTION, |
| 548 | runContinuationWorker: true, |
| 549 | assertOutput: assertGrokMonitorOutput, |
| 550 | }, |
| 551 | ], |
| 552 | }, |
| 553 | { |
| 554 | name: "grok_prompt_error", |
| 555 | buildInput: grokPromptErrorInput, |
| 556 | providers: [ |
| 557 | { |
| 558 | driver: ProviderDriverKind.make("grok"), |
| 559 | transcriptFile: new URL("./grok_prompt_error/grok_transcript.ndjson", import.meta.url), |
| 560 | modelSelection: GROK_MODEL_SELECTION, |
| 561 | assertOutput: assertGrokPromptErrorOutput, |
| 562 | }, |
| 563 | ], |
| 564 | }, |
| 565 | { |
| 566 | name: "grok_subagent_lineage", |
| 567 | buildInput: grokSubagentLineageInput, |
| 568 | providers: [ |
| 569 | { |
| 570 | driver: ProviderDriverKind.make("grok"), |
| 571 | transcriptFile: new URL("./grok_subagent_lineage/grok_transcript.ndjson", import.meta.url), |
| 572 | modelSelection: { |
| 573 | ...GROK_MODEL_SELECTION, |
| 574 | model: "grok-composer-2.5-fast", |
| 575 | }, |
| 576 | assertOutput: assertGrokSubagentLineageOutput, |
| 577 | }, |
| 578 | ], |
| 579 | }, |
| 580 | { |
| 581 | name: "acp_elicitation", |
| 582 | buildInput: planQuestionsInput, |
| 583 | providers: [ |
| 584 | // Grok Build still elicits with the pre-1.0 session/elicitation wire |
| 585 | // method the current spec removed; T3 supports standard ACP only, so |
| 586 | // the scenario covers the registry driver until Grok ships |
| 587 | // elicitation/create. |
| 588 | { |
| 589 | driver: ProviderDriverKind.make("acpRegistry"), |
| 590 | transcriptFile: new URL("./acp_elicitation/registry_transcript.ndjson", import.meta.url), |
| 591 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 592 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 593 | assertOutput: assertPlanQuestionsOutput, |
| 594 | }, |
| 595 | ], |
| 596 | }, |
| 597 | { |
| 598 | name: "simple", |
| 599 | buildInput: simpleInput, |
| 600 | providers: [ |
| 601 | { |
| 602 | driver: ProviderDriverKind.make("codex"), |
| 603 | transcriptFile: new URL("./simple/codex_transcript.ndjson", import.meta.url), |
| 604 | modelSelection: CODEX_MODEL_SELECTION, |
| 605 | assertOutput: assertSimpleOutput, |
| 606 | }, |
| 607 | { |
| 608 | driver: ProviderDriverKind.make("claudeAgent"), |
| 609 | transcriptFile: new URL("./simple/claude_transcript.ndjson", import.meta.url), |
| 610 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 611 | assertOutput: assertSimpleClaudeOutput, |
| 612 | }, |
| 613 | { |
| 614 | driver: ProviderDriverKind.make("cursor"), |
| 615 | transcriptFile: new URL("./simple/cursor_transcript.ndjson", import.meta.url), |
| 616 | modelSelection: CURSOR_MODEL_SELECTION, |
| 617 | assertOutput: assertSimpleOutput, |
| 618 | }, |
| 619 | { |
| 620 | driver: ProviderDriverKind.make("grok"), |
| 621 | transcriptFile: new URL("./simple/grok_transcript.ndjson", import.meta.url), |
| 622 | modelSelection: GROK_MODEL_SELECTION, |
| 623 | assertOutput: assertSimpleOutput, |
| 624 | }, |
| 625 | { |
| 626 | driver: ProviderDriverKind.make("acpRegistry"), |
| 627 | transcriptFile: new URL("./simple/registry_transcript.ndjson", import.meta.url), |
| 628 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 629 | assertOutput: assertSimpleOutput, |
| 630 | }, |
| 631 | { |
| 632 | driver: ProviderDriverKind.make("opencode"), |
| 633 | transcriptFile: new URL("./simple/opencode_transcript.ndjson", import.meta.url), |
| 634 | modelSelection: OPENCODE_MODEL_SELECTION, |
| 635 | assertOutput: assertSimpleOutput, |
| 636 | }, |
| 637 | { |
| 638 | driver: ProviderDriverKind.make("pi"), |
| 639 | transcriptFile: new URL("./simple/pi_transcript.ndjson", import.meta.url), |
| 640 | modelSelection: PI_MODEL_SELECTION, |
| 641 | assertOutput: assertPiSimpleOutput, |
| 642 | }, |
| 643 | { |
| 644 | driver: ProviderDriverKind.make("muse"), |
| 645 | transcriptFile: new URL("./simple/muse_transcript.ndjson", import.meta.url), |
| 646 | modelSelection: MUSE_MODEL_SELECTION, |
| 647 | assertOutput: assertSimpleOutput, |
| 648 | }, |
| 649 | ], |
| 650 | }, |
| 651 | { |
| 652 | name: "skill_invocation", |
| 653 | buildInput: skillInvocationInput, |
| 654 | providers: [ |
| 655 | { |
| 656 | driver: ProviderDriverKind.make("cursor"), |
| 657 | transcriptFile: new URL("./skill_invocation/cursor_transcript.ndjson", import.meta.url), |
| 658 | modelSelection: CURSOR_MODEL_SELECTION, |
| 659 | assertOutput: assertSkillInvocationCursorOutput, |
| 660 | }, |
| 661 | ], |
| 662 | }, |
| 663 | { |
| 664 | name: "tool_call_denied_write", |
| 665 | buildInput: toolCallDeniedWriteInput, |
| 666 | providers: [ |
| 667 | { |
| 668 | driver: ProviderDriverKind.make("claudeAgent"), |
| 669 | transcriptFile: new URL( |
| 670 | "./tool_call_denied_write/claude_transcript.ndjson", |
| 671 | import.meta.url, |
| 672 | ), |
| 673 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 674 | runtimePolicyOverride: DENIED_WRITE_POLICY, |
| 675 | expectedAbsentWorkspacePaths: [TOOL_CALL_DENIED_WRITE_TARGET], |
| 676 | assertOutput: assertToolCallDeniedWriteClaudeOutput, |
| 677 | }, |
| 678 | ], |
| 679 | }, |
| 680 | { |
| 681 | name: "tool_call_read_only", |
| 682 | buildInput: toolCallReadOnlyInput, |
| 683 | providers: [ |
| 684 | { |
| 685 | driver: ProviderDriverKind.make("claudeAgent"), |
| 686 | transcriptFile: new URL("./tool_call_read_only/claude_transcript.ndjson", import.meta.url), |
| 687 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 688 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 689 | assertOutput: assertToolCallReadOnlyClaudeOutput, |
| 690 | }, |
| 691 | { |
| 692 | driver: ProviderDriverKind.make("cursor"), |
| 693 | transcriptFile: new URL("./tool_call_read_only/cursor_transcript.ndjson", import.meta.url), |
| 694 | modelSelection: CURSOR_MODEL_SELECTION, |
| 695 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 696 | assertOutput: assertToolCallReadOnlyCursorOutput, |
| 697 | }, |
| 698 | { |
| 699 | driver: ProviderDriverKind.make("grok"), |
| 700 | transcriptFile: new URL("./tool_call_read_only/grok_transcript.ndjson", import.meta.url), |
| 701 | modelSelection: GROK_MODEL_SELECTION, |
| 702 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 703 | assertOutput: assertToolCallReadOnlyCursorOutput, |
| 704 | }, |
| 705 | { |
| 706 | driver: ProviderDriverKind.make("acpRegistry"), |
| 707 | transcriptFile: new URL( |
| 708 | "./tool_call_read_only/registry_transcript.ndjson", |
| 709 | import.meta.url, |
| 710 | ), |
| 711 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 712 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 713 | assertOutput: assertToolCallReadOnlyCursorOutput, |
| 714 | }, |
| 715 | ], |
| 716 | }, |
| 717 | { |
| 718 | name: "tool_call_read_only_on_request", |
| 719 | buildInput: toolCallReadOnlyOnRequestInput, |
| 720 | providers: [ |
| 721 | { |
| 722 | driver: ProviderDriverKind.make("codex"), |
| 723 | transcriptFile: new URL( |
| 724 | "./tool_call_read_only_on_request/codex_transcript.ndjson", |
| 725 | import.meta.url, |
| 726 | ), |
| 727 | // gpt-6-luna declines the write under a read-only sandbox, so nothing asks for approval. |
| 728 | modelSelection: { ...CODEX_MODEL_SELECTION, model: "gpt-6-sol" }, |
| 729 | runtimePolicyOverride: READ_ONLY_ON_REQUEST_POLICY, |
| 730 | assertOutput: assertToolCallReadOnlyOnRequestOutput, |
| 731 | }, |
| 732 | { |
| 733 | driver: ProviderDriverKind.make("claudeAgent"), |
| 734 | transcriptFile: new URL( |
| 735 | "./tool_call_read_only_on_request/claude_transcript.ndjson", |
| 736 | import.meta.url, |
| 737 | ), |
| 738 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 739 | runtimePolicyOverride: READ_ONLY_ON_REQUEST_POLICY, |
| 740 | assertOutput: assertToolCallReadOnlyOnRequestOutput, |
| 741 | }, |
| 742 | { |
| 743 | driver: ProviderDriverKind.make("grok"), |
| 744 | transcriptFile: new URL( |
| 745 | "./tool_call_read_only_on_request/grok_transcript.ndjson", |
| 746 | import.meta.url, |
| 747 | ), |
| 748 | modelSelection: GROK_MODEL_SELECTION, |
| 749 | runtimePolicyOverride: READ_ONLY_ON_REQUEST_POLICY, |
| 750 | assertOutput: assertToolCallReadOnlyOnRequestGrokOutput, |
| 751 | }, |
| 752 | { |
| 753 | driver: ProviderDriverKind.make("acpRegistry"), |
| 754 | transcriptFile: new URL( |
| 755 | "./tool_call_read_only_on_request/registry_transcript.ndjson", |
| 756 | import.meta.url, |
| 757 | ), |
| 758 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 759 | runtimePolicyOverride: READ_ONLY_ON_REQUEST_POLICY, |
| 760 | assertOutput: assertToolCallReadOnlyOnRequestOutput, |
| 761 | }, |
| 762 | ], |
| 763 | }, |
| 764 | { |
| 765 | name: "tool_call_workspace_never", |
| 766 | buildInput: toolCallWorkspaceNeverInput, |
| 767 | providers: [ |
| 768 | { |
| 769 | driver: ProviderDriverKind.make("codex"), |
| 770 | transcriptFile: new URL( |
| 771 | "./tool_call_workspace_never/codex_transcript.ndjson", |
| 772 | import.meta.url, |
| 773 | ), |
| 774 | modelSelection: CODEX_MODEL_SELECTION, |
| 775 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 776 | assertOutput: assertToolCallWorkspaceNeverOutput, |
| 777 | }, |
| 778 | { |
| 779 | driver: ProviderDriverKind.make("claudeAgent"), |
| 780 | transcriptFile: new URL( |
| 781 | "./tool_call_workspace_never/claude_transcript.ndjson", |
| 782 | import.meta.url, |
| 783 | ), |
| 784 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 785 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 786 | assertOutput: assertToolCallWorkspaceNeverClaudeOutput, |
| 787 | }, |
| 788 | ], |
| 789 | }, |
| 790 | { |
| 791 | name: "tool_call_restricted_granular", |
| 792 | buildInput: toolCallRestrictedGranularInput, |
| 793 | providers: [ |
| 794 | { |
| 795 | driver: ProviderDriverKind.make("codex"), |
| 796 | transcriptFile: new URL( |
| 797 | "./tool_call_restricted_granular/codex_transcript.ndjson", |
| 798 | import.meta.url, |
| 799 | ), |
| 800 | // gpt-6 models write through the shell; gpt-5.6-terra's apply_patch raises the |
| 801 | // file-change approval this fixture covers. |
| 802 | modelSelection: { ...CODEX_MODEL_SELECTION, model: "gpt-5.6-terra" }, |
| 803 | runtimePolicyOverride: RESTRICTED_GRANULAR_POLICY, |
| 804 | assertOutput: assertToolCallRestrictedGranularOutput, |
| 805 | }, |
| 806 | { |
| 807 | driver: ProviderDriverKind.make("claudeAgent"), |
| 808 | transcriptFile: new URL( |
| 809 | "./tool_call_restricted_granular/claude_transcript.ndjson", |
| 810 | import.meta.url, |
| 811 | ), |
| 812 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 813 | runtimePolicyOverride: RESTRICTED_GRANULAR_POLICY, |
| 814 | assertOutput: assertToolCallRestrictedGranularClaudeOutput, |
| 815 | }, |
| 816 | ], |
| 817 | }, |
| 818 | { |
| 819 | name: "subagent", |
| 820 | buildInput: subagentInput, |
| 821 | providers: [ |
| 822 | { |
| 823 | driver: ProviderDriverKind.make("codex"), |
| 824 | transcriptFile: new URL("./subagent/codex_transcript.ndjson", import.meta.url), |
| 825 | // gpt-5.6-luna still runs multi-agent v1 (collabAgentToolCall); subagent_v2 covers v2. |
| 826 | modelSelection: { ...CODEX_MODEL_SELECTION, model: "gpt-5.6-luna" }, |
| 827 | runtimePolicyOverride: READ_ONLY_ON_REQUEST_POLICY, |
| 828 | assertOutput: assertSubagentOutput, |
| 829 | }, |
| 830 | { |
| 831 | driver: ProviderDriverKind.make("claudeAgent"), |
| 832 | transcriptFile: new URL("./subagent/claude_transcript.ndjson", import.meta.url), |
| 833 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 834 | assertOutput: assertClaudeSubagentOutput, |
| 835 | }, |
| 836 | { |
| 837 | driver: ProviderDriverKind.make("cursor"), |
| 838 | transcriptFile: new URL("./subagent/cursor_transcript.ndjson", import.meta.url), |
| 839 | modelSelection: CURSOR_MODEL_SELECTION, |
| 840 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 841 | assertOutput: assertCursorSubagentOutput, |
| 842 | }, |
| 843 | ], |
| 844 | }, |
| 845 | { |
| 846 | name: "subagent_continue", |
| 847 | buildInput: subagentContinueInput, |
| 848 | providers: [ |
| 849 | { |
| 850 | driver: ProviderDriverKind.make("codex"), |
| 851 | transcriptFile: new URL("./subagent_continue/codex_transcript.ndjson", import.meta.url), |
| 852 | // gpt-5.6-luna still runs multi-agent v1 (collabAgentToolCall); subagent_v2 covers v2. |
| 853 | modelSelection: { ...CODEX_MODEL_SELECTION, model: "gpt-5.6-luna" }, |
| 854 | assertOutput: assertSubagentContinueOutput, |
| 855 | }, |
| 856 | ], |
| 857 | }, |
| 858 | { |
| 859 | name: "subagent_v2", |
| 860 | buildInput: subagentV2Input, |
| 861 | providers: [ |
| 862 | { |
| 863 | driver: ProviderDriverKind.make("codex"), |
| 864 | transcriptFile: new URL("./subagent_v2/codex_transcript.ndjson", import.meta.url), |
| 865 | modelSelection: CODEX_MODEL_SELECTION, |
| 866 | assertOutput: assertSubagentV2Output, |
| 867 | }, |
| 868 | { |
| 869 | driver: ProviderDriverKind.make("muse"), |
| 870 | transcriptFile: new URL("./subagent_v2/muse_transcript.ndjson", import.meta.url), |
| 871 | modelSelection: MUSE_MODEL_SELECTION, |
| 872 | assertOutput: assertMuseSubagentV2Output, |
| 873 | }, |
| 874 | ], |
| 875 | }, |
| 876 | { |
| 877 | name: "subagent_v2_approval", |
| 878 | buildInput: subagentV2ApprovalInput, |
| 879 | providers: [ |
| 880 | { |
| 881 | driver: ProviderDriverKind.make("codex"), |
| 882 | transcriptFile: new URL("./subagent_v2_approval/codex_transcript.ndjson", import.meta.url), |
| 883 | modelSelection: CODEX_MODEL_SELECTION, |
| 884 | runtimePolicyOverride: SUBAGENT_V2_APPROVAL_POLICY, |
| 885 | assertOutput: assertSubagentV2ApprovalOutput, |
| 886 | }, |
| 887 | ], |
| 888 | }, |
| 889 | { |
| 890 | name: "subagent_v2_nested", |
| 891 | buildInput: subagentV2NestedInput, |
| 892 | providers: [ |
| 893 | { |
| 894 | driver: ProviderDriverKind.make("codex"), |
| 895 | transcriptFile: new URL("./subagent_v2_nested/codex_transcript.ndjson", import.meta.url), |
| 896 | modelSelection: CODEX_MODEL_SELECTION, |
| 897 | assertOutput: assertSubagentV2NestedOutput, |
| 898 | }, |
| 899 | ], |
| 900 | }, |
| 901 | { |
| 902 | name: "subagent_v2_nested_approval", |
| 903 | buildInput: subagentV2NestedApprovalInput, |
| 904 | providers: [ |
| 905 | { |
| 906 | driver: ProviderDriverKind.make("codex"), |
| 907 | transcriptFile: new URL( |
| 908 | "./subagent_v2_nested_approval/codex_transcript.ndjson", |
| 909 | import.meta.url, |
| 910 | ), |
| 911 | modelSelection: CODEX_MODEL_SELECTION, |
| 912 | runtimePolicyOverride: SUBAGENT_V2_APPROVAL_POLICY, |
| 913 | assertOutput: assertSubagentV2NestedApprovalOutput, |
| 914 | }, |
| 915 | ], |
| 916 | }, |
| 917 | { |
| 918 | name: "opencode_subagent", |
| 919 | buildInput: openCodeSubagentInput, |
| 920 | providers: [ |
| 921 | { |
| 922 | driver: ProviderDriverKind.make("opencode"), |
| 923 | transcriptFile: new URL("./opencode_subagent/opencode_transcript.ndjson", import.meta.url), |
| 924 | modelSelection: OPENCODE_MODEL_SELECTION, |
| 925 | assertOutput: assertOpenCodeSubagentOutput, |
| 926 | }, |
| 927 | ], |
| 928 | }, |
| 929 | // OpenCode 2 runtime of the same driver, recorded against 2.0.18 over HTTP and SSE. |
| 930 | { |
| 931 | name: "opencode2_simple", |
| 932 | buildInput: openCode2SimpleInput, |
| 933 | providers: [ |
| 934 | { |
| 935 | driver: ProviderDriverKind.make("opencode"), |
| 936 | transcriptFile: new URL("./opencode2_simple/opencode_transcript.ndjson", import.meta.url), |
| 937 | modelSelection: OPENCODE2_MODEL_SELECTION, |
| 938 | assertOutput: assertOpenCode2SimpleOutput, |
| 939 | }, |
| 940 | ], |
| 941 | }, |
| 942 | { |
| 943 | name: "opencode2_tool_call", |
| 944 | buildInput: openCode2ToolCallInput, |
| 945 | providers: [ |
| 946 | { |
| 947 | driver: ProviderDriverKind.make("opencode"), |
| 948 | transcriptFile: new URL( |
| 949 | "./opencode2_tool_call/opencode_transcript.ndjson", |
| 950 | import.meta.url, |
| 951 | ), |
| 952 | modelSelection: { |
| 953 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 954 | model: "opencode/big-pickle", |
| 955 | }, |
| 956 | assertOutput: assertOpenCode2ToolCallOutput, |
| 957 | }, |
| 958 | ], |
| 959 | }, |
| 960 | { |
| 961 | name: "opencode2_interrupt", |
| 962 | buildInput: openCode2InterruptInput, |
| 963 | providers: [ |
| 964 | { |
| 965 | driver: ProviderDriverKind.make("opencode"), |
| 966 | transcriptFile: new URL( |
| 967 | "./opencode2_interrupt/opencode_transcript.ndjson", |
| 968 | import.meta.url, |
| 969 | ), |
| 970 | modelSelection: { |
| 971 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 972 | model: "opencode/big-pickle", |
| 973 | }, |
| 974 | assertOutput: assertOpenCode2InterruptOutput, |
| 975 | }, |
| 976 | ], |
| 977 | }, |
| 978 | { |
| 979 | name: "opencode2_inbox", |
| 980 | buildInput: openCode2InboxInput, |
| 981 | providers: [ |
| 982 | { |
| 983 | driver: ProviderDriverKind.make("opencode"), |
| 984 | transcriptFile: new URL("./opencode2_inbox/opencode_transcript.ndjson", import.meta.url), |
| 985 | modelSelection: { |
| 986 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 987 | model: "opencode/big-pickle", |
| 988 | }, |
| 989 | assertOutput: assertOpenCode2InboxOutput, |
| 990 | }, |
| 991 | ], |
| 992 | }, |
| 993 | { |
| 994 | name: "opencode2_revert", |
| 995 | buildInput: openCode2RevertInput, |
| 996 | providers: [ |
| 997 | { |
| 998 | driver: ProviderDriverKind.make("opencode"), |
| 999 | transcriptFile: new URL("./opencode2_revert/opencode_transcript.ndjson", import.meta.url), |
| 1000 | modelSelection: { |
| 1001 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1002 | model: "opencode/big-pickle", |
| 1003 | }, |
| 1004 | assertOutput: assertOpenCode2RevertOutput, |
| 1005 | }, |
| 1006 | ], |
| 1007 | }, |
| 1008 | { |
| 1009 | name: "opencode2_permission", |
| 1010 | buildInput: openCode2PermissionInput, |
| 1011 | providers: [ |
| 1012 | { |
| 1013 | driver: ProviderDriverKind.make("opencode"), |
| 1014 | transcriptFile: new URL( |
| 1015 | "./opencode2_permission/opencode_transcript.ndjson", |
| 1016 | import.meta.url, |
| 1017 | ), |
| 1018 | modelSelection: { |
| 1019 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1020 | model: "opencode/big-pickle", |
| 1021 | }, |
| 1022 | assertOutput: assertOpenCode2PermissionOutput, |
| 1023 | }, |
| 1024 | ], |
| 1025 | }, |
| 1026 | { |
| 1027 | name: "opencode2_question", |
| 1028 | buildInput: openCode2QuestionInput, |
| 1029 | providers: [ |
| 1030 | { |
| 1031 | driver: ProviderDriverKind.make("opencode"), |
| 1032 | transcriptFile: new URL("./opencode2_question/opencode_transcript.ndjson", import.meta.url), |
| 1033 | modelSelection: { |
| 1034 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1035 | model: "opencode/big-pickle", |
| 1036 | }, |
| 1037 | assertOutput: assertOpenCode2QuestionOutput, |
| 1038 | }, |
| 1039 | ], |
| 1040 | }, |
| 1041 | { |
| 1042 | name: "opencode2_subagent", |
| 1043 | buildInput: openCode2SubagentInput, |
| 1044 | providers: [ |
| 1045 | { |
| 1046 | driver: ProviderDriverKind.make("opencode"), |
| 1047 | transcriptFile: new URL("./opencode2_subagent/opencode_transcript.ndjson", import.meta.url), |
| 1048 | modelSelection: { |
| 1049 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1050 | model: "opencode/big-pickle", |
| 1051 | }, |
| 1052 | assertOutput: assertOpenCode2SubagentOutput, |
| 1053 | }, |
| 1054 | ], |
| 1055 | }, |
| 1056 | { |
| 1057 | name: "opencode2_background", |
| 1058 | buildInput: openCode2BackgroundInput, |
| 1059 | providers: [ |
| 1060 | { |
| 1061 | driver: ProviderDriverKind.make("opencode"), |
| 1062 | transcriptFile: new URL( |
| 1063 | "./opencode2_background/opencode_transcript.ndjson", |
| 1064 | import.meta.url, |
| 1065 | ), |
| 1066 | modelSelection: { |
| 1067 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1068 | model: "opencode/big-pickle", |
| 1069 | }, |
| 1070 | runContinuationWorker: true, |
| 1071 | assertOutput: assertOpenCode2BackgroundOutput, |
| 1072 | }, |
| 1073 | ], |
| 1074 | }, |
| 1075 | { |
| 1076 | name: "opencode2_nested_background", |
| 1077 | buildInput: openCode2NestedBackgroundInput, |
| 1078 | providers: [ |
| 1079 | { |
| 1080 | driver: ProviderDriverKind.make("opencode"), |
| 1081 | transcriptFile: new URL( |
| 1082 | "./opencode2_nested_background/opencode_transcript.ndjson", |
| 1083 | import.meta.url, |
| 1084 | ), |
| 1085 | modelSelection: { |
| 1086 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1087 | model: "openrouter/deepseek/deepseek-v4-flash", |
| 1088 | }, |
| 1089 | assertOutput: assertOpenCode2NestedBackgroundOutput, |
| 1090 | }, |
| 1091 | ], |
| 1092 | }, |
| 1093 | { |
| 1094 | name: "opencode2_compaction", |
| 1095 | buildInput: openCode2CompactionInput, |
| 1096 | providers: [ |
| 1097 | { |
| 1098 | driver: ProviderDriverKind.make("opencode"), |
| 1099 | transcriptFile: new URL( |
| 1100 | "./opencode2_compaction/opencode_transcript.ndjson", |
| 1101 | import.meta.url, |
| 1102 | ), |
| 1103 | modelSelection: { |
| 1104 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1105 | model: "opencode/big-pickle", |
| 1106 | }, |
| 1107 | assertOutput: assertOpenCode2CompactionOutput, |
| 1108 | }, |
| 1109 | ], |
| 1110 | }, |
| 1111 | { |
| 1112 | name: "opencode2_resume_after_restart", |
| 1113 | buildInput: openCode2ResumeAfterRestartInput, |
| 1114 | providers: [ |
| 1115 | { |
| 1116 | driver: ProviderDriverKind.make("opencode"), |
| 1117 | transcriptFile: new URL( |
| 1118 | "./opencode2_resume_after_restart/opencode_transcript.ndjson", |
| 1119 | import.meta.url, |
| 1120 | ), |
| 1121 | modelSelection: { |
| 1122 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1123 | model: "opencode/big-pickle", |
| 1124 | }, |
| 1125 | assertOutput: assertOpenCode2ResumeAfterRestartOutput, |
| 1126 | }, |
| 1127 | ], |
| 1128 | }, |
| 1129 | // Recorded live against 2.0.18 on OpenRouter with the workspace's own command and skill. |
| 1130 | ...( |
| 1131 | [ |
| 1132 | ["opencode2_command", openCode2CommandInput, assertOpenCode2CommandOutput], |
| 1133 | ["opencode2_skill", openCode2SkillInput, assertOpenCode2SkillOutput], |
| 1134 | ] as const |
| 1135 | ).map(([name, buildInput, assertOutput]) => ({ |
| 1136 | name, |
| 1137 | buildInput, |
| 1138 | providers: [ |
| 1139 | { |
| 1140 | driver: ProviderDriverKind.make("opencode"), |
| 1141 | transcriptFile: new URL(`./${name}/opencode_transcript.ndjson`, import.meta.url), |
| 1142 | modelSelection: { |
| 1143 | instanceId: OPENCODE2_MODEL_SELECTION.instanceId, |
| 1144 | model: "openrouter/deepseek/deepseek-v4-flash", |
| 1145 | }, |
| 1146 | assertOutput, |
| 1147 | }, |
| 1148 | ], |
| 1149 | })), |
| 1150 | { |
| 1151 | name: "opencode_child_approval", |
| 1152 | buildInput: openCodeChildApprovalInput, |
| 1153 | providers: [ |
| 1154 | { |
| 1155 | driver: ProviderDriverKind.make("opencode"), |
| 1156 | transcriptFile: new URL( |
| 1157 | "./opencode_child_approval/opencode_transcript.ndjson", |
| 1158 | import.meta.url, |
| 1159 | ), |
| 1160 | modelSelection: OPENCODE_MODEL_SELECTION, |
| 1161 | assertOutput: assertOpenCodeChildApprovalOutput, |
| 1162 | }, |
| 1163 | ], |
| 1164 | }, |
| 1165 | { |
| 1166 | name: "opencode_running_child_approval", |
| 1167 | buildInput: openCodeRunningChildApprovalInput, |
| 1168 | providers: [ |
| 1169 | { |
| 1170 | driver: ProviderDriverKind.make("opencode"), |
| 1171 | transcriptFile: new URL( |
| 1172 | "./opencode_running_child_approval/opencode_transcript.ndjson", |
| 1173 | import.meta.url, |
| 1174 | ), |
| 1175 | modelSelection: OPENCODE_MODEL_SELECTION, |
| 1176 | assertOutput: assertOpenCodeRunningChildApprovalOutput, |
| 1177 | }, |
| 1178 | ], |
| 1179 | }, |
| 1180 | { |
| 1181 | name: "multi_turn", |
| 1182 | buildInput: multiTurnInput, |
| 1183 | providers: [ |
| 1184 | { |
| 1185 | driver: ProviderDriverKind.make("codex"), |
| 1186 | transcriptFile: new URL("./multi_turn/codex_transcript.ndjson", import.meta.url), |
| 1187 | modelSelection: CODEX_MODEL_SELECTION, |
| 1188 | assertOutput: assertMultiTurnOutput, |
| 1189 | }, |
| 1190 | { |
| 1191 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1192 | transcriptFile: new URL("./multi_turn/claude_transcript.ndjson", import.meta.url), |
| 1193 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1194 | assertOutput: assertMultiTurnClaudeOutput, |
| 1195 | }, |
| 1196 | { |
| 1197 | driver: ProviderDriverKind.make("cursor"), |
| 1198 | transcriptFile: new URL("./multi_turn/cursor_transcript.ndjson", import.meta.url), |
| 1199 | modelSelection: CURSOR_MODEL_SELECTION, |
| 1200 | assertOutput: assertMultiTurnOutput, |
| 1201 | }, |
| 1202 | { |
| 1203 | driver: ProviderDriverKind.make("grok"), |
| 1204 | transcriptFile: new URL("./multi_turn/grok_transcript.ndjson", import.meta.url), |
| 1205 | modelSelection: GROK_MODEL_SELECTION, |
| 1206 | assertOutput: assertMultiTurnOutput, |
| 1207 | }, |
| 1208 | { |
| 1209 | driver: ProviderDriverKind.make("acpRegistry"), |
| 1210 | transcriptFile: new URL("./multi_turn/registry_transcript.ndjson", import.meta.url), |
| 1211 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 1212 | assertOutput: assertMultiTurnOutput, |
| 1213 | }, |
| 1214 | { |
| 1215 | driver: ProviderDriverKind.make("pi"), |
| 1216 | transcriptFile: new URL("./multi_turn/pi_transcript.ndjson", import.meta.url), |
| 1217 | modelSelection: PI_MODEL_SELECTION, |
| 1218 | assertOutput: assertPiMultiTurnOutput, |
| 1219 | }, |
| 1220 | { |
| 1221 | driver: ProviderDriverKind.make("muse"), |
| 1222 | transcriptFile: new URL("./multi_turn/muse_transcript.ndjson", import.meta.url), |
| 1223 | modelSelection: MUSE_MODEL_SELECTION, |
| 1224 | assertOutput: assertMuseMultiTurnOutput, |
| 1225 | }, |
| 1226 | ], |
| 1227 | }, |
| 1228 | { |
| 1229 | name: "muse_permission", |
| 1230 | buildInput: musePermissionInput, |
| 1231 | providers: [ |
| 1232 | { |
| 1233 | driver: ProviderDriverKind.make("muse"), |
| 1234 | transcriptFile: new URL("./muse_permission/muse_transcript.ndjson", import.meta.url), |
| 1235 | modelSelection: MUSE_MODEL_SELECTION, |
| 1236 | expectedAbsentWorkspacePaths: [MUSE_PERMISSION_DECLINED_FILE], |
| 1237 | assertOutput: assertMusePermissionOutput, |
| 1238 | }, |
| 1239 | ], |
| 1240 | }, |
| 1241 | { |
| 1242 | name: "muse_subagent_approval", |
| 1243 | buildInput: museSubagentApprovalInput, |
| 1244 | providers: [ |
| 1245 | { |
| 1246 | driver: ProviderDriverKind.make("muse"), |
| 1247 | transcriptFile: new URL("./muse_subagent_approval/muse_transcript.ndjson", import.meta.url), |
| 1248 | modelSelection: MUSE_MODEL_SELECTION, |
| 1249 | assertOutput: assertMuseSubagentApprovalOutput, |
| 1250 | }, |
| 1251 | ], |
| 1252 | }, |
| 1253 | { |
| 1254 | name: "muse_subagent_interrupt", |
| 1255 | buildInput: museSubagentInterruptInput, |
| 1256 | providers: [ |
| 1257 | { |
| 1258 | driver: ProviderDriverKind.make("muse"), |
| 1259 | transcriptFile: new URL( |
| 1260 | "./muse_subagent_interrupt/muse_transcript.ndjson", |
| 1261 | import.meta.url, |
| 1262 | ), |
| 1263 | modelSelection: MUSE_MODEL_SELECTION, |
| 1264 | assertOutput: assertMuseSubagentInterruptOutput, |
| 1265 | }, |
| 1266 | ], |
| 1267 | }, |
| 1268 | { |
| 1269 | name: "muse_workflow", |
| 1270 | buildInput: museWorkflowInput, |
| 1271 | providers: [ |
| 1272 | { |
| 1273 | driver: ProviderDriverKind.make("muse"), |
| 1274 | transcriptFile: new URL("./muse_workflow/muse_transcript.ndjson", import.meta.url), |
| 1275 | modelSelection: MUSE_MODEL_SELECTION, |
| 1276 | runContinuationWorker: true, |
| 1277 | assertOutput: assertMuseWorkflowOutput, |
| 1278 | }, |
| 1279 | ], |
| 1280 | }, |
| 1281 | { |
| 1282 | name: "pi_compaction", |
| 1283 | buildInput: piCompactionInput, |
| 1284 | providers: [ |
| 1285 | { |
| 1286 | driver: ProviderDriverKind.make("pi"), |
| 1287 | transcriptFile: new URL("./pi_compaction/pi_transcript.ndjson", import.meta.url), |
| 1288 | modelSelection: PI_MODEL_SELECTION, |
| 1289 | assertOutput: assertPiCompactionOutput, |
| 1290 | }, |
| 1291 | ], |
| 1292 | }, |
| 1293 | { |
| 1294 | name: "provider_thread_resume", |
| 1295 | buildInput: providerThreadResumeInput, |
| 1296 | providers: [ |
| 1297 | { |
| 1298 | driver: ProviderDriverKind.make("pi"), |
| 1299 | transcriptFile: new URL("./provider_thread_resume/pi_transcript.ndjson", import.meta.url), |
| 1300 | modelSelection: PI_MODEL_SELECTION, |
| 1301 | assertOutput: assertPiProviderThreadResumeOutput, |
| 1302 | }, |
| 1303 | ], |
| 1304 | }, |
| 1305 | { |
| 1306 | name: "multi_turn_restart", |
| 1307 | buildInput: multiTurnInput, |
| 1308 | providers: [ |
| 1309 | { |
| 1310 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1311 | transcriptFile: new URL("./multi_turn_restart/claude_transcript.ndjson", import.meta.url), |
| 1312 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1313 | assertOutput: assertMultiTurnClaudeOutput, |
| 1314 | }, |
| 1315 | ], |
| 1316 | }, |
| 1317 | { |
| 1318 | name: "queued_cancelled_while_active", |
| 1319 | buildInput: queuedCancelledWhileActiveInput, |
| 1320 | providers: [ |
| 1321 | { |
| 1322 | driver: ProviderDriverKind.make("codex"), |
| 1323 | transcriptFile: new URL("./queued_turn/codex_transcript.ndjson", import.meta.url), |
| 1324 | recordedScenario: "queued_turn", |
| 1325 | transcriptEntriesThroughLabel: "turn/completed", |
| 1326 | modelSelection: CODEX_MODEL_SELECTION, |
| 1327 | assertOutput: assertQueuedCancelledWhileActiveOutput, |
| 1328 | }, |
| 1329 | ], |
| 1330 | }, |
| 1331 | { |
| 1332 | name: "queued_turn", |
| 1333 | buildInput: queuedTurnInput, |
| 1334 | providers: [ |
| 1335 | { |
| 1336 | driver: ProviderDriverKind.make("codex"), |
| 1337 | transcriptFile: new URL("./queued_turn/codex_transcript.ndjson", import.meta.url), |
| 1338 | modelSelection: CODEX_MODEL_SELECTION, |
| 1339 | assertOutput: assertQueuedTurnOutput, |
| 1340 | }, |
| 1341 | { |
| 1342 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1343 | transcriptFile: new URL("./queued_turn/claude_transcript.ndjson", import.meta.url), |
| 1344 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1345 | assertOutput: assertQueuedTurnOutput, |
| 1346 | }, |
| 1347 | { |
| 1348 | driver: ProviderDriverKind.make("cursor"), |
| 1349 | transcriptFile: new URL("./queued_turn/cursor_transcript.ndjson", import.meta.url), |
| 1350 | modelSelection: CURSOR_MODEL_SELECTION, |
| 1351 | assertOutput: assertQueuedTurnOutput, |
| 1352 | }, |
| 1353 | { |
| 1354 | driver: ProviderDriverKind.make("grok"), |
| 1355 | transcriptFile: new URL("./queued_turn/grok_transcript.ndjson", import.meta.url), |
| 1356 | modelSelection: GROK_MODEL_SELECTION, |
| 1357 | assertOutput: assertQueuedTurnOutput, |
| 1358 | }, |
| 1359 | { |
| 1360 | driver: ProviderDriverKind.make("acpRegistry"), |
| 1361 | transcriptFile: new URL("./queued_turn/registry_transcript.ndjson", import.meta.url), |
| 1362 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 1363 | assertOutput: assertQueuedTurnOutput, |
| 1364 | }, |
| 1365 | ], |
| 1366 | }, |
| 1367 | { |
| 1368 | name: "todo_list", |
| 1369 | buildInput: todoListInput, |
| 1370 | providers: [ |
| 1371 | { |
| 1372 | driver: ProviderDriverKind.make("codex"), |
| 1373 | transcriptFile: new URL("./todo_list/codex_transcript.ndjson", import.meta.url), |
| 1374 | modelSelection: CODEX_MODEL_SELECTION, |
| 1375 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 1376 | assertOutput: assertTodoListOutput, |
| 1377 | }, |
| 1378 | { |
| 1379 | driver: ProviderDriverKind.make("cursor"), |
| 1380 | transcriptFile: new URL("./todo_list/cursor_transcript.ndjson", import.meta.url), |
| 1381 | modelSelection: CURSOR_MODEL_SELECTION, |
| 1382 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 1383 | assertOutput: assertTodoListCursorOutput, |
| 1384 | }, |
| 1385 | { |
| 1386 | driver: ProviderDriverKind.make("grok"), |
| 1387 | transcriptFile: new URL("./todo_list/grok_transcript.ndjson", import.meta.url), |
| 1388 | modelSelection: GROK_MODEL_SELECTION, |
| 1389 | assertOutput: assertTodoListGrokOutput, |
| 1390 | }, |
| 1391 | { |
| 1392 | driver: ProviderDriverKind.make("acpRegistry"), |
| 1393 | transcriptFile: new URL("./todo_list/registry_transcript.ndjson", import.meta.url), |
| 1394 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 1395 | assertOutput: assertTodoListGrokOutput, |
| 1396 | }, |
| 1397 | ], |
| 1398 | }, |
| 1399 | { |
| 1400 | name: "web_search", |
| 1401 | buildInput: webSearchInput, |
| 1402 | providers: [ |
| 1403 | { |
| 1404 | driver: ProviderDriverKind.make("codex"), |
| 1405 | transcriptFile: new URL("./web_search/codex_transcript.ndjson", import.meta.url), |
| 1406 | modelSelection: CODEX_MODEL_SELECTION, |
| 1407 | assertOutput: assertWebSearchOutput, |
| 1408 | }, |
| 1409 | { |
| 1410 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1411 | transcriptFile: new URL("./web_search/claude_transcript.ndjson", import.meta.url), |
| 1412 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1413 | assertOutput: assertClaudeWebSearchOutput, |
| 1414 | }, |
| 1415 | ], |
| 1416 | }, |
| 1417 | { |
| 1418 | name: "plan_questions", |
| 1419 | buildInput: planQuestionsInput, |
| 1420 | providers: [ |
| 1421 | { |
| 1422 | driver: ProviderDriverKind.make("codex"), |
| 1423 | transcriptFile: new URL("./plan_questions/codex_transcript.ndjson", import.meta.url), |
| 1424 | modelSelection: CODEX_MODEL_SELECTION, |
| 1425 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 1426 | assertOutput: assertCodexPlanQuestionsOutput, |
| 1427 | }, |
| 1428 | { |
| 1429 | driver: ProviderDriverKind.make("grok"), |
| 1430 | transcriptFile: new URL("./plan_questions/grok_transcript.ndjson", import.meta.url), |
| 1431 | modelSelection: GROK_MODEL_SELECTION, |
| 1432 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 1433 | assertOutput: assertPlanQuestionsOutput, |
| 1434 | }, |
| 1435 | { |
| 1436 | driver: ProviderDriverKind.make("opencode"), |
| 1437 | transcriptFile: new URL("./plan_questions/opencode_transcript.ndjson", import.meta.url), |
| 1438 | modelSelection: OPENCODE_MODEL_SELECTION, |
| 1439 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 1440 | assertOutput: assertOpenCodePlanQuestionsOutput, |
| 1441 | }, |
| 1442 | ], |
| 1443 | }, |
| 1444 | { |
| 1445 | name: "proposed_plan", |
| 1446 | buildInput: proposedPlanInput, |
| 1447 | providers: [ |
| 1448 | { |
| 1449 | driver: ProviderDriverKind.make("codex"), |
| 1450 | transcriptFile: new URL("./proposed_plan/codex_transcript.ndjson", import.meta.url), |
| 1451 | modelSelection: CODEX_MODEL_SELECTION, |
| 1452 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 1453 | assertOutput: assertProposedPlanOutput, |
| 1454 | }, |
| 1455 | { |
| 1456 | driver: ProviderDriverKind.make("cursor"), |
| 1457 | transcriptFile: new URL("./proposed_plan/cursor_transcript.ndjson", import.meta.url), |
| 1458 | modelSelection: CURSOR_MODEL_SELECTION, |
| 1459 | runtimePolicyOverride: READ_ONLY_NEVER_POLICY, |
| 1460 | assertOutput: assertProposedPlanCursorOutput, |
| 1461 | }, |
| 1462 | ], |
| 1463 | }, |
| 1464 | { |
| 1465 | name: "message_steering", |
| 1466 | buildInput: messageSteeringInput, |
| 1467 | providers: [ |
| 1468 | { |
| 1469 | driver: ProviderDriverKind.make("codex"), |
| 1470 | transcriptFile: new URL("./message_steering/codex_transcript.ndjson", import.meta.url), |
| 1471 | modelSelection: CODEX_MODEL_SELECTION, |
| 1472 | assertOutput: assertMessageSteeringOutput, |
| 1473 | }, |
| 1474 | { |
| 1475 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1476 | transcriptFile: new URL("./message_steering/claude_transcript.ndjson", import.meta.url), |
| 1477 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1478 | assertOutput: assertClaudeMessageSteeringOutput, |
| 1479 | }, |
| 1480 | { |
| 1481 | driver: ProviderDriverKind.make("cursor"), |
| 1482 | transcriptFile: new URL("./message_steering/cursor_transcript.ndjson", import.meta.url), |
| 1483 | modelSelection: CURSOR_MODEL_SELECTION, |
| 1484 | assertOutput: assertCursorMessageSteeringOutput, |
| 1485 | }, |
| 1486 | { |
| 1487 | driver: ProviderDriverKind.make("grok"), |
| 1488 | transcriptFile: new URL("./message_steering/grok_transcript.ndjson", import.meta.url), |
| 1489 | modelSelection: GROK_MODEL_SELECTION, |
| 1490 | assertOutput: assertGrokMessageSteeringOutput, |
| 1491 | }, |
| 1492 | { |
| 1493 | driver: ProviderDriverKind.make("acpRegistry"), |
| 1494 | transcriptFile: new URL("./message_steering/registry_transcript.ndjson", import.meta.url), |
| 1495 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 1496 | assertOutput: assertGrokMessageSteeringOutput, |
| 1497 | }, |
| 1498 | { |
| 1499 | driver: ProviderDriverKind.make("pi"), |
| 1500 | transcriptFile: new URL("./message_steering/pi_transcript.ndjson", import.meta.url), |
| 1501 | modelSelection: PI_MODEL_SELECTION, |
| 1502 | assertOutput: assertPiMessageSteeringOutput, |
| 1503 | }, |
| 1504 | ], |
| 1505 | }, |
| 1506 | { |
| 1507 | name: "turn_interrupt", |
| 1508 | buildInput: turnInterruptInput, |
| 1509 | providers: [ |
| 1510 | { |
| 1511 | driver: ProviderDriverKind.make("codex"), |
| 1512 | transcriptFile: new URL("./turn_interrupt/codex_transcript.ndjson", import.meta.url), |
| 1513 | modelSelection: CODEX_MODEL_SELECTION, |
| 1514 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1515 | assertOutput: assertTurnInterruptOutput, |
| 1516 | }, |
| 1517 | { |
| 1518 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1519 | transcriptFile: new URL("./turn_interrupt/claude_transcript.ndjson", import.meta.url), |
| 1520 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1521 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1522 | assertOutput: assertTurnInterruptClaudeOutput, |
| 1523 | }, |
| 1524 | { |
| 1525 | driver: ProviderDriverKind.make("grok"), |
| 1526 | transcriptFile: new URL("./turn_interrupt/grok_transcript.ndjson", import.meta.url), |
| 1527 | modelSelection: GROK_MODEL_SELECTION, |
| 1528 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1529 | assertOutput: assertTurnInterruptOutput, |
| 1530 | }, |
| 1531 | { |
| 1532 | driver: ProviderDriverKind.make("acpRegistry"), |
| 1533 | transcriptFile: new URL("./turn_interrupt/registry_transcript.ndjson", import.meta.url), |
| 1534 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 1535 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1536 | assertOutput: assertTurnInterruptOutput, |
| 1537 | }, |
| 1538 | { |
| 1539 | driver: ProviderDriverKind.make("opencode"), |
| 1540 | transcriptFile: new URL("./turn_interrupt/opencode_transcript.ndjson", import.meta.url), |
| 1541 | modelSelection: OPENCODE_MODEL_SELECTION, |
| 1542 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1543 | assertOutput: assertTurnInterruptOutput, |
| 1544 | }, |
| 1545 | { |
| 1546 | driver: ProviderDriverKind.make("muse"), |
| 1547 | transcriptFile: new URL("./turn_interrupt/muse_transcript.ndjson", import.meta.url), |
| 1548 | modelSelection: MUSE_MODEL_SELECTION, |
| 1549 | assertOutput: assertMuseTurnInterruptOutput, |
| 1550 | }, |
| 1551 | ], |
| 1552 | }, |
| 1553 | { |
| 1554 | name: "turn_interrupt_mid_tool", |
| 1555 | buildInput: turnInterruptMidToolInput, |
| 1556 | providers: [ |
| 1557 | { |
| 1558 | driver: ProviderDriverKind.make("codex"), |
| 1559 | transcriptFile: new URL( |
| 1560 | "./turn_interrupt_mid_tool/codex_transcript.ndjson", |
| 1561 | import.meta.url, |
| 1562 | ), |
| 1563 | modelSelection: CODEX_MODEL_SELECTION, |
| 1564 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1565 | assertOutput: assertTurnInterruptMidToolCodexOutput, |
| 1566 | }, |
| 1567 | { |
| 1568 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1569 | transcriptFile: new URL( |
| 1570 | "./turn_interrupt_mid_tool/claude_transcript.ndjson", |
| 1571 | import.meta.url, |
| 1572 | ), |
| 1573 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1574 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1575 | assertOutput: assertTurnInterruptMidToolClaudeOutput, |
| 1576 | }, |
| 1577 | { |
| 1578 | driver: ProviderDriverKind.make("cursor"), |
| 1579 | transcriptFile: new URL( |
| 1580 | "./turn_interrupt_mid_tool/cursor_transcript.ndjson", |
| 1581 | import.meta.url, |
| 1582 | ), |
| 1583 | modelSelection: CURSOR_MODEL_SELECTION, |
| 1584 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1585 | assertOutput: assertTurnInterruptMidToolCursorOutput, |
| 1586 | }, |
| 1587 | { |
| 1588 | driver: ProviderDriverKind.make("pi"), |
| 1589 | transcriptFile: new URL("./turn_interrupt_mid_tool/pi_transcript.ndjson", import.meta.url), |
| 1590 | modelSelection: PI_MODEL_SELECTION, |
| 1591 | assertOutput: assertTurnInterruptMidToolPiOutput, |
| 1592 | }, |
| 1593 | ], |
| 1594 | }, |
| 1595 | { |
| 1596 | name: "turn_interrupt_restart", |
| 1597 | buildInput: turnInterruptRestartInput, |
| 1598 | providers: [ |
| 1599 | { |
| 1600 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1601 | transcriptFile: new URL( |
| 1602 | "./turn_interrupt_restart/claude_transcript.ndjson", |
| 1603 | import.meta.url, |
| 1604 | ), |
| 1605 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1606 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1607 | assertOutput: assertTurnInterruptRestartClaudeOutput, |
| 1608 | }, |
| 1609 | ], |
| 1610 | }, |
| 1611 | { |
| 1612 | name: "thread_rollback", |
| 1613 | buildInput: threadRollbackInput, |
| 1614 | providers: [ |
| 1615 | { |
| 1616 | driver: ProviderDriverKind.make("codex"), |
| 1617 | transcriptFile: new URL("./thread_rollback/codex_transcript.ndjson", import.meta.url), |
| 1618 | modelSelection: CODEX_MODEL_SELECTION, |
| 1619 | assertOutput: assertThreadRollbackOutput, |
| 1620 | }, |
| 1621 | { |
| 1622 | driver: ProviderDriverKind.make("claudeAgent"), |
| 1623 | transcriptFile: new URL("./thread_rollback/claude_transcript.ndjson", import.meta.url), |
| 1624 | modelSelection: CLAUDE_MODEL_SELECTION, |
| 1625 | assertOutput: assertClaudeThreadRollbackOutput, |
| 1626 | }, |
| 1627 | { |
| 1628 | driver: ProviderDriverKind.make("pi"), |
| 1629 | transcriptFile: new URL("./thread_rollback/pi_transcript.ndjson", import.meta.url), |
| 1630 | modelSelection: PI_MODEL_SELECTION, |
| 1631 | assertOutput: assertPiThreadRollbackOutput, |
| 1632 | }, |
| 1633 | ], |
| 1634 | }, |
| 1635 | { |
| 1636 | name: "thread_rollback_after_restart", |
| 1637 | buildInput: threadRollbackAfterRestartInput, |
| 1638 | providers: [ |
| 1639 | { |
| 1640 | driver: ProviderDriverKind.make("codex"), |
| 1641 | transcriptFile: new URL( |
| 1642 | "./thread_rollback_after_restart/codex_transcript.ndjson", |
| 1643 | import.meta.url, |
| 1644 | ), |
| 1645 | modelSelection: CODEX_MODEL_SELECTION, |
| 1646 | assertOutput: assertThreadRollbackAfterRestartOutput, |
| 1647 | }, |
| 1648 | ], |
| 1649 | }, |
| 1650 | { |
| 1651 | name: "thread_rollback_after_stop", |
| 1652 | buildInput: threadRollbackAfterStopInput, |
| 1653 | providers: [ |
| 1654 | { |
| 1655 | driver: ProviderDriverKind.make("pi"), |
| 1656 | transcriptFile: new URL( |
| 1657 | "./thread_rollback_after_stop/pi_transcript.ndjson", |
| 1658 | import.meta.url, |
| 1659 | ), |
| 1660 | modelSelection: PI_MODEL_SELECTION, |
| 1661 | assertOutput: assertPiThreadRollbackAfterStopOutput, |
| 1662 | }, |
| 1663 | ], |
| 1664 | }, |
| 1665 | { |
| 1666 | name: "thread_rollback_to_stopped_turn", |
| 1667 | buildInput: threadRollbackToStoppedTurnInput, |
| 1668 | providers: [ |
| 1669 | { |
| 1670 | driver: ProviderDriverKind.make("codex"), |
| 1671 | transcriptFile: new URL( |
| 1672 | "./thread_rollback_to_stopped_turn/codex_transcript.ndjson", |
| 1673 | import.meta.url, |
| 1674 | ), |
| 1675 | modelSelection: CODEX_MODEL_SELECTION, |
| 1676 | runtimePolicyOverride: WORKSPACE_NEVER_POLICY, |
| 1677 | assertOutput: assertThreadRollbackToStoppedTurnOutput, |
| 1678 | }, |
| 1679 | ], |
| 1680 | }, |
| 1681 | { |
| 1682 | name: "stop_background_work_after_failed_turn", |
| 1683 | buildInput: stopBackgroundWorkAfterFailedTurnInput, |
| 1684 | providers: [ |
| 1685 | { |
| 1686 | driver: ProviderDriverKind.make("acpRegistry"), |
| 1687 | transcriptFile: new URL( |
| 1688 | "./stop_background_work_after_failed_turn/registry_transcript.ndjson", |
| 1689 | import.meta.url, |
| 1690 | ), |
| 1691 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 1692 | assertOutput: assertStopBackgroundWorkAfterFailedTurnOutput, |
| 1693 | }, |
| 1694 | ], |
| 1695 | }, |
| 1696 | { |
| 1697 | name: "stop_background_work_after_release", |
| 1698 | buildInput: stopBackgroundWorkAfterReleaseInput, |
| 1699 | providers: [ |
| 1700 | { |
| 1701 | driver: ProviderDriverKind.make("acpRegistry"), |
| 1702 | transcriptFile: new URL( |
| 1703 | "./stop_background_work_after_failed_turn/registry_transcript.ndjson", |
| 1704 | import.meta.url, |
| 1705 | ), |
| 1706 | recordedScenario: "stop_background_work_after_failed_turn", |
| 1707 | modelSelection: ACP_REGISTRY_MODEL_SELECTION, |
| 1708 | assertOutput: assertStopBackgroundWorkAfterFailedTurnOutput, |
| 1709 | }, |
| 1710 | ], |
| 1711 | }, |
| 1712 | ]; |
| 1713 | |
| 1714 | // TODO(claude-v2/context-transfer): add provider-switch handoff and return fixtures when portable |
| 1715 | // context handoff is implemented. Cross-reference docs/orchestration-v2/provider-switching-and-context.md |
| 1716 | // and docs/orchestration-v2/thread-lineage-and-context-transfer.md. The return fixture should |
| 1717 | // prefer a delta handoff into an existing Claude provider thread. |
| 1718 | |
| 1719 | // TODO(claude-v2/context-transfer-fixtures): register provider-switch, merge-back, and cross-provider |
| 1720 | // fork fixtures after each path has a real provider transcript. Cross-reference |
| 1721 | // docs/orchestration-v2/provider-switching-and-context.md and |
| 1722 | // docs/orchestration-v2/thread-lineage-and-context-transfer.md. |