fix(failover): narrow service-unavailable to require overload indicator (#32828) (#36646)

Merged via squash.

Prepared head SHA: 46fb430612
Co-authored-by: jnMetaCode <12096460+jnMetaCode@users.noreply.github.com>
Co-authored-by: altaywtf <9790196+altaywtf@users.noreply.github.com>
Reviewed-by: @altaywtf

CHANGELOG.md

@@ -139,6 +139,7 @@ Docs: https://docs.openclaw.ai
 - Mattermost/plugin SDK import policy: replace remaining monolithic `openclaw/plugin-sdk` imports in Mattermost mention-gating paths/tests with scoped subpaths (`openclaw/plugin-sdk/compat` and `openclaw/plugin-sdk/mattermost`) so `pnpm check` passes `lint:plugins:no-monolithic-plugin-sdk-entry-imports` on baseline. (#36480) Thanks @Takhoffman.
 
 - Agents/failover cooldown classification: stop treating generic `cooling down` text as provider `rate_limit` so healthy models no longer show false global cooldown/rate-limit warnings while explicit `model_cooldown` markers still trigger failover. (#32972) thanks @stakeswky.
+- Agents/failover service-unavailable handling: stop treating bare proxy/CDN `service unavailable` errors as provider overload while keeping them retryable via the timeout/failover path, so transient outages no longer show false rate-limit warnings or block fallback. (#36646) thanks @jnMetaCode.
 
 ## 2026.3.2
 

src/agents/pi-embedded-helpers.isbillingerrormessage.test.ts

@@ -540,13 +540,20 @@ describe("classifyFailoverReason", () => {
         "This model is currently experiencing high demand. Please try again later.",
       ),
     ).toBe("rate_limit");
-    expect(classifyFailoverReason("LLM error: service unavailable")).toBe("rate_limit");
+    // "service unavailable" combined with overload/capacity indicator → rate_limit
+    // (exercises the new regex — none of the standalone patterns match here)
+    expect(classifyFailoverReason("service unavailable due to capacity limits")).toBe("rate_limit");
     expect(
       classifyFailoverReason(
         '{"error":{"code":503,"message":"The model is overloaded. Please try later","status":"UNAVAILABLE"}}',
       ),
     ).toBe("rate_limit");
   });
+  it("classifies bare 'service unavailable' as timeout instead of rate_limit (#32828)", () => {
+    // A generic "service unavailable" from a proxy/CDN should stay retryable,
+    // but it should not be treated as provider overload / rate limit.
+    expect(classifyFailoverReason("LLM error: service unavailable")).toBe("timeout");
+  });
   it("classifies permanent auth errors as auth_permanent", () => {
     expect(classifyFailoverReason("invalid_api_key")).toBe("auth_permanent");
     expect(classifyFailoverReason("Your api key has been revoked")).toBe("auth_permanent");

src/agents/pi-embedded-helpers/failover-matches.ts

@@ -15,12 +15,16 @@ const ERROR_PATTERNS = {
   overloaded: [
     /overloaded_error|"type"\s*:\s*"overloaded_error"/i,
     "overloaded",
-    "service unavailable",
+    // Match "service unavailable" only when combined with an explicit overload
+    // indicator — a generic 503 from a proxy/CDN should not be classified as
+    // provider-overload (#32828).
+    /service[_ ]unavailable.*(?:overload|capacity|high[_ ]demand)|(?:overload|capacity|high[_ ]demand).*service[_ ]unavailable/i,
     "high demand",
   ],
   timeout: [
     "timeout",
     "timed out",
+    "service unavailable",
     "deadline exceeded",
     "context deadline exceeded",
     "connection error",

src/agents/pi-embedded-runner.run-embedded-pi-agent.auth-profile-rotation.e2e.test.ts

@@ -658,6 +658,16 @@ describe("runEmbeddedPiAgent auth profile rotation", () => {
     expect(usageStats["openai:p1"]?.cooldownUntil).toBeUndefined();
   });
 
+  it("rotates on bare service unavailable without cooling down the profile", async () => {
+    const { usageStats } = await runAutoPinnedRotationCase({
+      errorMessage: "LLM error: service unavailable",
+      sessionKey: "agent:test:service-unavailable-no-cooldown",
+      runId: "run:service-unavailable-no-cooldown",
+    });
+    expect(typeof usageStats["openai:p2"]?.lastUsed).toBe("number");
+    expect(usageStats["openai:p1"]?.cooldownUntil).toBeUndefined();
+  });
+
   it("does not rotate for compaction timeouts", async () => {
     await withAgentWorkspace(async ({ agentDir, workspaceDir }) => {
       await writeAuthStore(agentDir);