test(orchestration-v2): replay a Claude workflow end to end

The adapter's workflow projection was covered in isolation and the panel
derivation was covered against synthetic agents, but nothing drove real
workflow frames through the orchestrator to the projection the client
receives.

Adds a claude_workflow replay fixture: a local_workflow task with two
phases and two members, asserted on the resulting projection. Pins what
the panel groups by — that members claim declared phases, that their
membership points at the coordinator that actually reached the
projection, and that the coordinator's usage still covers its members
so subtracting theirs cannot go negative.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: shivamhwp

apps/server/src/orchestration-v2/ProviderEventIngestor.test.ts

@@ -303,6 +303,8 @@ layer("ProviderEventIngestorV2", (it) => {
         maxAttempts: 3,
         retryDelayMs: 2_000,
       });
+      assert.isNotNull(errorItem.startedAt);
+      if (errorItem.startedAt === null) return;
       assert.equal(
         DateTime.toEpochMillis(errorItem.startedAt),
         DateTime.toEpochMillis(retryStartedAt),

apps/server/src/orchestration-v2/testkit/fixtures/claude_workflow/claude_transcript.ndjson

@@ -0,0 +1,12 @@
+{"type": "transcript_start", "provider": "claudeAgent", "protocol": "claude-agent-sdk.query", "version": "0.2.111", "scenario": "claude_workflow", "metadata": {"prompts": ["Run a two-phase workflow, then answer exactly: claude workflow fixture complete"], "model": "claude-sonnet-4-6", "nativeSessionId": "3c7f21a0-9d4e-4b21-8f60-000000000201", "queryMode": "streaming", "tools": "claude_code", "permissionMode": "bypassPermissions", "generatedBy": "hand-authored-workflow-observability"}}
+{"type": "expect_outbound", "label": "query.open", "frame": {"type": "query.open", "options": {"model": "claude-sonnet-4-6", "tools": {"type": "preset", "preset": "claude_code"}, "permissionMode": "bypassPermissions", "allowDangerouslySkipPermissions": true, "sessionId": "3c7f21a0-9d4e-4b21-8f60-000000000201"}}}
+{"type": "expect_outbound", "label": "prompt.offer:1", "frame": {"type": "prompt.offer", "message": {"type": "user", "message": {"role": "user", "content": "Run a two-phase workflow, then answer exactly: claude workflow fixture complete"}, "parent_tool_use_id": null}}}
+{"type": "emit_inbound", "label": "system", "frame": {"type": "system", "subtype": "init", "agents": [], "apiKeySource": "none", "claude_code_version": "2.1.183", "cwd": "/tmp/claude-replay-claude_workflow", "tools": [], "mcp_servers": [], "model": "claude-sonnet-4-6", "permissionMode": "bypassPermissions", "slash_commands": [], "output_style": "default", "skills": [], "plugins": [], "fast_mode_state": "off", "uuid": "af3bf4cc-1385-4e19-a5fa-fcbe2a0ca201", "session_id": "3c7f21a0-9d4e-4b21-8f60-000000000201"}}
+{"type": "emit_inbound", "label": "assistant:commentary", "frame": {"type": "assistant", "message": {"model": "claude-sonnet-4-6", "id": "msg_01claudeworkflow1", "type": "message", "role": "assistant", "content": [{"type": "text", "text": "Starting the two-phase workflow."}], "stop_reason": null, "stop_sequence": null, "stop_details": null, "usage": {"input_tokens": 1, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "output_tokens": 1, "service_tier": "standard", "inference_geo": "not_available"}, "context_management": null}, "parent_tool_use_id": null, "session_id": "3c7f21a0-9d4e-4b21-8f60-000000000201", "uuid": "97898e28-ee4e-4fcf-80c3-5c790c502201"}}
+{"type": "emit_inbound", "label": "task_started:local_workflow", "frame": {"type": "system", "subtype": "task_started", "task_id": "wf7ka2mq1", "tool_use_id": "toolu_01ClaudeWorkflowFixture01", "description": "Research and implement", "task_type": "local_workflow", "prompt": "Research the parser, then implement the fix.", "uuid": "4873c95b-1aa9-44a5-9783-dc56dac73201", "session_id": "3c7f21a0-9d4e-4b21-8f60-000000000201"}}
+{"type": "emit_inbound", "label": "task_progress:phase-1", "frame": {"type": "system", "subtype": "task_progress", "task_id": "wf7ka2mq1", "tool_use_id": "toolu_01ClaudeWorkflowFixture01", "description": "Research phase running", "usage": {"input_tokens": 300, "output_tokens": 200, "total_tokens": 500, "tool_uses": 4}, "workflow_progress": [{"type": "workflow_phase", "index": 0, "title": "Research"}, {"type": "workflow_phase", "index": 1, "title": "Implement"}, {"type": "workflow_agent", "index": 0, "label": "Researcher", "state": "active", "phaseIndex": 0, "attempt": 0, "model": "claude-sonnet-4-6", "lastToolName": "Read", "tokens": 200, "toolCalls": 2}], "uuid": "4873c95b-1aa9-44a5-9783-dc56dac73202", "session_id": "3c7f21a0-9d4e-4b21-8f60-000000000201"}}
+{"type": "emit_inbound", "label": "task_progress:phase-2", "frame": {"type": "system", "subtype": "task_progress", "task_id": "wf7ka2mq1", "tool_use_id": "toolu_01ClaudeWorkflowFixture01", "description": "Implement phase running", "usage": {"input_tokens": 700, "output_tokens": 500, "total_tokens": 1200, "tool_uses": 9}, "workflow_progress": [{"type": "workflow_phase", "index": 0, "title": "Research"}, {"type": "workflow_phase", "index": 1, "title": "Implement"}, {"type": "workflow_agent", "index": 0, "label": "Researcher", "state": "done", "phaseIndex": 0, "attempt": 0, "model": "claude-sonnet-4-6", "lastToolName": "Read", "tokens": 400, "toolCalls": 3}, {"type": "workflow_agent", "index": 1, "label": "Implementer", "state": "active", "phase_index": 1, "attempt": 0, "model": "claude-opus-4-1", "last_tool_name": "Edit", "tokens": 500, "tool_calls": 4}], "uuid": "4873c95b-1aa9-44a5-9783-dc56dac73203", "session_id": "3c7f21a0-9d4e-4b21-8f60-000000000201"}}
+{"type": "emit_inbound", "label": "task_notification:local_workflow", "frame": {"type": "system", "subtype": "task_notification", "task_id": "wf7ka2mq1", "tool_use_id": "toolu_01ClaudeWorkflowFixture01", "status": "completed", "output_file": "", "summary": "Workflow finished both phases.", "usage": {"input_tokens": 900, "output_tokens": 600, "total_tokens": 1500, "tool_uses": 11}, "uuid": "4873c95b-1aa9-44a5-9783-dc56dac73204", "session_id": "3c7f21a0-9d4e-4b21-8f60-000000000201"}}
+{"type": "emit_inbound", "label": "assistant:commentary", "frame": {"type": "assistant", "message": {"model": "claude-sonnet-4-6", "id": "msg_01claudeworkflow2", "type": "message", "role": "assistant", "content": [{"type": "text", "text": "claude workflow fixture complete"}], "stop_reason": null, "stop_sequence": null, "stop_details": null, "usage": {"input_tokens": 1, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "output_tokens": 1, "service_tier": "standard", "inference_geo": "not_available"}, "context_management": null}, "parent_tool_use_id": null, "session_id": "3c7f21a0-9d4e-4b21-8f60-000000000201", "uuid": "89be9e65-097b-47e8-849b-be723386b201"}}
+{"type": "emit_inbound", "label": "result", "frame": {"type": "result", "subtype": "success", "is_error": false, "api_error_status": null, "duration_ms": 4000, "duration_api_ms": 3500, "num_turns": 2, "result": "claude workflow fixture complete", "stop_reason": "end_turn", "session_id": "3c7f21a0-9d4e-4b21-8f60-000000000201", "total_cost_usd": 0.001, "usage": {"input_tokens": 2, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "output_tokens": 2, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "inference_geo": "", "iterations": [{"input_tokens": 1, "output_tokens": 1, "cache_read_input_tokens": 0, "cache_creation_input_tokens": 0, "cache_creation": {"ephemeral_5m_input_tokens": 0, "ephemeral_1h_input_tokens": 0}, "type": "message"}], "speed": "standard"}, "modelUsage": {"claude-sonnet-4-6": {"inputTokens": 2, "outputTokens": 2, "cacheReadInputTokens": 0, "cacheCreationInputTokens": 0, "webSearchRequests": 0, "costUSD": 0.001, "contextWindow": 200000, "maxOutputTokens": 32000}}, "permission_denials": [], "terminal_reason": "completed", "fast_mode_state": "off", "uuid": "89be9e65-097b-47e8-849b-be723386b202"}}
+{"type": "runtime_exit", "status": "success"}

apps/server/src/orchestration-v2/testkit/fixtures/claude_workflow/input.ts

@@ -0,0 +1,18 @@
+import type { OrchestratorFixtureInput } from "../shared.ts";
+
+export const CLAUDE_WORKFLOW_PROMPT =
+  "Run a two-phase workflow, then answer exactly: claude workflow fixture complete";
+
+/**
+ * End-to-end cover for the Claude workflow coordinator. The adapter-level
+ * projection of phases and workers is unit-tested, and the panel derivation is
+ * unit-tested against synthetic agents — nothing joined the two. This drives
+ * real workflow frames through the orchestrator so the coordinator, its
+ * phases, and its members are asserted on the projection the client actually
+ * receives, then feeds that projection to the panel derivation.
+ */
+export function claudeWorkflowInput(): OrchestratorFixtureInput {
+  return {
+    steps: [{ type: "message", text: CLAUDE_WORKFLOW_PROMPT }],
+  };
+}

apps/server/src/orchestration-v2/testkit/fixtures/claude_workflow/output.ts

@@ -0,0 +1,91 @@
+import { assert } from "@effect/vitest";
+import type { ProviderReplayTranscript } from "@t3tools/contracts";
+
+import type { OrchestratorV2ScenarioResult } from "../../OrchestratorScenario.ts";
+import {
+  assertBaseProjection,
+  assertSemanticProjectionIntegrity,
+  assertUserMessagesInclude,
+  projectionFor,
+} from "../shared.ts";
+import { CLAUDE_WORKFLOW_PROMPT } from "./input.ts";
+
+export function assertClaudeWorkflowOutput(
+  result: OrchestratorV2ScenarioResult,
+  transcript: ProviderReplayTranscript,
+) {
+  assertBaseProjection({ result, transcript, runCount: 1, runStatuses: ["completed"] });
+
+  const projection = projectionFor(result, transcript.scenario);
+  assertSemanticProjectionIntegrity(projection);
+  assertUserMessagesInclude(projection, [CLAUDE_WORKFLOW_PROMPT]);
+
+  const coordinator = projection.subagents.find((subagent) => subagent.kind === "workflow");
+  assert.isDefined(coordinator, "the workflow must reach the projection as a coordinator");
+  assert.deepEqual(coordinator?.role, { name: "workflow-coordinator", source: "app_default" });
+  assert.deepEqual(coordinator?.workflow?.phases, [
+    { index: 0, title: "Research" },
+    { index: 1, title: "Implement" },
+  ]);
+
+  const members = projection.subagents
+    .filter((subagent) => subagent.kind === "workflow_agent")
+    .toSorted(
+      (left, right) =>
+        (left.workflowMembership?.agentIndex ?? 0) - (right.workflowMembership?.agentIndex ?? 0),
+    );
+  assert.lengthOf(members, 2, "both workflow members must reach the projection");
+  assert.deepEqual(
+    members.map((member) => [
+      member.title,
+      member.status,
+      member.workflowMembership?.phaseIndex,
+      member.model,
+      member.usage?.totalTokens,
+    ]),
+    [
+      ["Researcher", "completed", 0, "claude-sonnet-4-6", 400],
+      ["Implementer", "running", 1, "claude-opus-4-1", 500],
+    ],
+  );
+  // Membership must point at the coordinator that actually reached the
+  // projection, not at an id the adapter invented independently.
+  assert.deepEqual(
+    [...new Set(members.map((member) => member.workflowMembership?.workflowSubagentId))],
+    [coordinator?.id],
+  );
+
+  // The panel groups by these three fields, so pin the values it reads rather
+  // than only the ones the coordinator carries. The derivation itself is
+  // covered in client-runtime; the schema is what joins the two sides, so what
+  // matters here is that the server emits a shape that groups cleanly:
+  // every member claims a declared phase, and no member is left ungrouped.
+  const declaredPhases = new Set((coordinator?.workflow?.phases ?? []).map((phase) => phase.index));
+  assert.deepEqual(
+    members.map((member) => declaredPhases.has(member.workflowMembership?.phaseIndex ?? -1)),
+    [true, true],
+  );
+  assert.lengthOf(
+    projection.subagents.filter(
+      (subagent) => subagent.kind === "workflow_agent" && subagent.workflowMembership === null,
+    ),
+    0,
+    "a member without membership would render outside its workflow",
+  );
+  // The coordinator's usage already covers its members, so the panel subtracts
+  // theirs from it. That only stays non-negative if the provider's totals are
+  // carried through unmodified.
+  const memberTotal = members.reduce(
+    (total, member) => total + (member.usage?.totalTokens ?? 0),
+    0,
+  );
+  assert.isAtLeast(coordinator?.usage?.totalTokens ?? 0, memberTotal);
+
+  const assistantTexts = projection.turnItems.flatMap((item) =>
+    item.type === "assistant_message" ? [item.text] : [],
+  );
+  assert.deepEqual(assistantTexts, [
+    "Starting the two-phase workflow.",
+    "claude workflow fixture complete",
+  ]);
+}

apps/server/src/orchestration-v2/testkit/fixtures/index.ts

@@ -4,6 +4,8 @@ import { claudeIdleResumeInput } from "./claude_idle_resume/input.ts";
 import { assertClaudeIdleResumeOutput } from "./claude_idle_resume/output.ts";
 import { claudeLocalBashTaskInput } from "./claude_local_bash_task/input.ts";
 import { assertClaudeLocalBashTaskOutput } from "./claude_local_bash_task/output.ts";
+import { claudeWorkflowInput } from "./claude_workflow/input.ts";
+import { assertClaudeWorkflowOutput } from "./claude_workflow/output.ts";
 import { claudeResultIsErrorInput } from "./claude_result_is_error/input.ts";
 import { assertClaudeResultIsErrorOutput } from "./claude_result_is_error/output.ts";
 import { grokSubagentLineageInput } from "./grok_subagent_lineage/input.ts";
@@ -99,6 +101,18 @@ export const ORCHESTRATOR_REPLAY_FIXTURES = [
       },
     ],
   },
+  {
+    name: "claude_workflow",
+    buildInput: claudeWorkflowInput,
+    providers: [
+      {
+        driver: ProviderDriverKind.make("claudeAgent"),
+        transcriptFile: new URL("./claude_workflow/claude_transcript.ndjson", import.meta.url),
+        modelSelection: CLAUDE_MODEL_SELECTION,
+        assertOutput: assertClaudeWorkflowOutput,
+      },
+    ],
+  },
   {
     name: "claude_idle_resume",
     buildInput: claudeIdleResumeInput,