fix(ai): retry unavailable Codex upstreams

packages/ai/src/changes.md

@@ -1,5 +1,18 @@
 # AI Source Changes
 
+## 2026-07-25 - Retry transient Codex upstream websocket failures
+
+### What changed and why
+
+- `utils/retry.ts` classifies `upstream_unavailable` provider errors as transient so the existing bounded retry policy
+  retries Codex websocket proxy disconnects such as `ConnectionClosedOK`.
+- The retry classifier and coding-agent event-contract tests pin the exact reported error through the existing retry
+  lifecycle rather than introducing provider-specific retry behavior.
+
+### Expected merge conflict zones
+
+- LOW: `utils/retry.ts` transient transport error patterns.
+
 ## 2026-07-23 - Session-scoped provider resolution via node-only AsyncLocalStorage subpath
 
 ### What changed and why

packages/ai/src/utils/retry.ts

@@ -52,6 +52,7 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
 	"other side closed",
 	"fetch failed",
 	"upstream.?connect",
+	"upstream.?unavailable",
 	"reset before headers",
 	"socket hang up",
 	"socket connection was closed",

packages/ai/test/retry.test.ts

@@ -10,6 +10,8 @@ const nvidiaNIMResourceExhaustedMessage = "ResourceExhausted: Worker local total
 const bunFetchSocketClosedMessage =
 	"The socket connection was closed unexpectedly. For more information, pass `verbose: true` in the second argument to fetch()";
 const openAIResponsesEarlyEofMessage = "OpenAI Responses stream ended before a terminal response event";
+const codexUpstreamUnavailableMessage =
+	"Error: upstream_unavailable: Codex upstream websocket send failed via proxy endpoint unknown: ConnectionClosedOK";
 
 describe("provider retry classification", () => {
 	it("matches explicit provider retry guidance", () => {
@@ -38,6 +40,14 @@ describe("provider retry classification", () => {
 		).toBe(true);
 	});
 
+	it("matches Codex upstream websocket unavailability", () => {
+		expect(
+			isRetryableAssistantError(
+				fauxAssistantMessage("", { stopReason: "error", errorMessage: codexUpstreamUnavailableMessage }),
+			),
+		).toBe(true);
+	});
+
 	it("classifies agent-loop stream idle timeouts as retryable", () => {
 		// Emitted by @earendil-works/pi-agent-core when a provider stream stops
 		// delivering events (e.g. a connection that died after a network change).

packages/coding-agent/test/suite/retry-fallback-engine.test.ts

@@ -6,6 +6,8 @@ import { createHarness, type Harness } from "./harness.ts";
 
 const primary = "faux/faux-1";
 const fallback = "faux/faux-2";
+const codexUpstreamUnavailableMessage =
+	"Error: upstream_unavailable: Codex upstream websocket send failed via proxy endpoint unknown: ConnectionClosedOK";
 
 type EventTranscriptEntry =
 	| { type: "message_start" | "message_end"; role: string }
@@ -544,7 +546,7 @@ describe("retry fallback engine", () => {
 		harness.setResponses([
 			fauxAssistantMessage("", {
 				stopReason: "error",
-				errorMessage: "overloaded_error",
+				errorMessage: codexUpstreamUnavailableMessage,
 			}),
 			fauxAssistantMessage("recovered"),
 		]);
@@ -569,7 +571,7 @@ describe("retry fallback engine", () => {
 				attempt: 1,
 				maxAttempts: 3,
 				delayMs: 1,
-				errorMessage: "overloaded_error",
+				errorMessage: codexUpstreamUnavailableMessage,
 			},
 			{ type: "agent_start" },
 			{ type: "turn_start" },