fix(workbench): classify runtime dependency outages as retryable (#2021)

internal/cloud/server-workbench-http.test.ts

@@ -7,6 +7,7 @@ import { test } from "bun:test";
 import { createCloudApiServer } from "./server.ts";
 import { createBackendPerformanceStore } from "./backend-performance.ts";
 import { createCodeAgentTraceStore } from "./code-agent-trace-store.ts";
+import { classifyWorkbenchReadModelFailure } from "./server-workbench-http.ts";
 import { createCodeAgentChatResultStore } from "./server-code-agent-http.ts";
 import { createWorkbenchTurnProjection, durableTraceStatus, projectionDiagnostics, traceTerminalEvidence } from "./workbench-turn-projection.ts";
 
@@ -58,6 +59,24 @@ test("workbench projection diagnostics keeps projecting health distinct from cau
   assert.equal(diagnostics.projectionHealth, "projecting");
 });
 
+test("workbench runtime retryable dependency failures classify as 503 (#2020)", () => {
+  const error = new Error("fetch failed");
+  error.name = "WorkbenchRuntimeDependencyError";
+  error.code = "workbench_runtime_unreachable";
+  error.data = {
+    serviceId: "hwlab-workbench-runtime",
+    retryable: true,
+    transient: true,
+    valuesRedacted: true
+  };
+  const classified = classifyWorkbenchReadModelFailure(error);
+  assert.equal(classified.statusCode, 503);
+  assert.equal(classified.code, "workbench_runtime_unreachable");
+  assert.equal(classified.retryable, true);
+  assert.equal(classified.transient, true);
+  assert.equal(classified.serviceId, "hwlab-workbench-runtime");
+});
+
 test("workbench turn projection keeps lastEventAt on the latest event instead of stale admission timing (#1889)", () => {
   const traceId = "trc_turn_timing_latest_event";
   const result = {

internal/cloud/server-workbench-http.ts

@@ -2295,19 +2295,31 @@ function handleWorkbenchReadModelFailure(response, url, options = {}, error) {
     valuesRedacted: true
   });
   if (response.headersSent || response.destroyed) return;
-  sendJson(response, classified.statusCode, workbenchError(classified.code, classified.message, { route, errorName: error?.name ?? "Error" }));
+  sendJson(response, classified.statusCode, workbenchError(classified.code, classified.message, {
+    route,
+    errorName: error?.name ?? "Error",
+    retryable: classified.retryable === true,
+    transient: classified.transient === true,
+    serviceId: classified.serviceId ?? null,
+    runtimeStatus: classified.runtimeStatus ?? null
+  }));
 }
 
-function classifyWorkbenchReadModelFailure(error) {
+export function classifyWorkbenchReadModelFailure(error) {
   const message = String(error?.message ?? error ?? "");
   const code = String(error?.code ?? "").toLowerCase();
   const runtimeStatus = Number(error?.data?.status ?? 0);
   const runtimePayloadError = error?.data?.payload?.error ?? null;
   if (error?.name === "WorkbenchRuntimeDependencyError" || code.startsWith("workbench_runtime_")) {
+    const retryable = error?.data?.retryable === true || error?.data?.transient === true || runtimePayloadError?.retryable === true;
     return {
-      statusCode: Number.isFinite(runtimeStatus) && runtimeStatus >= 500 ? 503 : 500,
+      statusCode: retryable || (Number.isFinite(runtimeStatus) && runtimeStatus >= 500) ? 503 : 500,
       code: String(runtimePayloadError?.code ?? error?.code ?? "workbench_runtime_dependency_failed"),
-      message: String(runtimePayloadError?.message ?? error?.message ?? "Workbench runtime dependency failed.")
+      message: String(runtimePayloadError?.message ?? error?.message ?? "Workbench runtime dependency failed."),
+      retryable,
+      transient: retryable,
+      serviceId: String(error?.data?.serviceId ?? "hwlab-workbench-runtime"),
+      runtimeStatus: Number.isFinite(runtimeStatus) && runtimeStatus > 0 ? runtimeStatus : null
     };
   }
   if (/timeout|connection terminated|terminating connection|econnreset|econnrefused|etimedout/iu.test(message) || /timeout|econnreset|econnrefused|etimedout/u.test(code)) {