fix(combo): fall back across targets on all 400 responses (#1713)

Integrated into release/v3.7.3 — simplifies combo fallback by treating all non-ok responses as target-local failures
2026-08-03 05:45:04 +03:00 · 2026-04-28 07:28:09 -04:00
parent 8a0ca60e4b
commit 42952ee42f
7 changed files with 104 additions and 291 deletions
--- a/open-sse/services/accountFallback.ts
+++ b/open-sse/services/accountFallback.ts
@@ -1016,7 +1016,8 @@ export function checkFallbackError(
      };
    }

-    // Generic 400 — same request will likely fail on all accounts; don't fallback.
+    // Generic 400 is not account-fallback-worthy. Combo routing may still try a
+    // different provider/model because combo fallback is target-level orchestration.
    return { shouldFallback: false, cooldownMs: 0, reason: RateLimitReason.UNKNOWN };
  }

--- a/open-sse/services/combo.ts
+++ b/open-sse/services/combo.ts
@@ -20,7 +20,6 @@ import { fisherYatesShuffle, getNextFromDeck } from "../../src/shared/utils/shuf
 import { parseModel } from "./model.ts";
 import { applyComboAgentMiddleware, injectModelTag } from "./comboAgentMiddleware.ts";
 import { classifyWithConfig, DEFAULT_INTENT_CONFIG } from "./intentClassifier.ts";
-import { CONTEXT_OVERFLOW_REGEX } from "./errorClassifier.ts";
 import { selectProvider as selectAutoProvider } from "./autoCombo/engine.ts";
 import { selectWithStrategy } from "./autoCombo/routerStrategy.ts";
 import { getTaskFitness } from "./autoCombo/taskFitness.ts";
@@ -41,16 +40,6 @@ import {
  getComboStepWeight,
  normalizeComboStep,
 } from "../../src/lib/combos/steps.ts";
-
-function isProviderBreakerOpenResponse(
-  result: Response,
-  errorBody?: { error?: { code?: string | null } } | null
-) {
-  return (
-    result.headers.get("x-omniroute-provider-breaker") === "open" ||
-    errorBody?.error?.code === "provider_circuit_open"
-  );
-}
 import {
  getConnectionRoutingTags,
  matchesRoutingTags,
@@ -60,51 +49,6 @@ import {

 // Status codes that should mark round-robin target semaphores as cooling down.
 const TRANSIENT_FOR_SEMAPHORE = [429, 502, 503, 504];
-const COMBO_BAD_REQUEST_FALLBACK_PATTERNS = [
-  /\bprohibited_content\b/i,
-  /request blocked by .*api/i,
-  /provided message roles? is not valid/i,
-  /unsupported .*message role/i,
-  /no such tool available/i,
-  /unsupported content part type/i,
-  /tool(?:_call|_use)? .* not (?:available|found)/i,
-  /third-party apps/i,
-  CONTEXT_OVERFLOW_REGEX,
-  // Model not supported/found — permanent model-level error, try next combo target
-  /no provider supported/i,
-  /model not found/i,
-  /model not available/i,
-  /unsupported model/i,
-  /model.*has no provider/i,
-  // Function calling format error — model doesn't support this capability
-  /function\.?arguments.*(must be|should be|\u5fc5\u987b).*(json|JSON)/i,
-  /tool.*arguments.*invalid/i,
-  /function.*parameter.*(invalid|format)/i,
-  // Input length range error — model-specific context limit
-  /range of input length/i,
-  /input length should be/i,
-  // Transient 400 errors from upstream — should fallback to next combo target
-  /\u670d\u52a1\u9047\u5230\u4e86\u4e00\u70b9\u5c0f\u72b6\u51b5/i, // ModelScope/Qwen transient error
-  /\u62b1\u6b49.*?\u654f\u611f\u5185\u5bb9.*?\u8bf7\u68c0\u67e5/i, // ModelScope/Qwen content moderation with context
-  /\u5185\u5bb9.*?\u654f\u611f.*?(?:\u65e0\u6cd5|\u8fc7\u6ee4)/i, // Content sensitivity block
-  /\u65e0\u6cd5\u54cd\u5e94.*?\u8bf7\u6c42/i, // "unable to respond to request"
-  /\u7a0d\u540e\u91cd\u8bd5/i, // "retry later" in Chinese
-  /temporary.*error/i,
-  /transient.*error/i,
-  /service.*unavailable/i,
-  /please.*try.*again/i,
-  // Rate limit errors — some providers return 400 instead of 429
-  /\brate.?-?limit.?(?:exceeded|reached|hit)/i,
-  /too many requests/i,
-  /\u8bf7\u6c42\u8fc7\u4e8e\u9891\u7e41/i, // Chinese rate limit message
-  // Tool call function name errors — model-specific, try next combo target
-  /\bfunction'?s? name (?:can't|can not|is|has) (?:blank|empty|missing)/i,
-  /function.*name.*(?:blank|empty|missing)/i,
-  /tool_call.*name.*(?:blank|empty|missing)/i,
-  // Anthropic thinking block signature errors — stale/expired signatures cannot be retried (#1696)
-  /invalid.*signature.*thinking/i,
-];
-
 // Patterns that signal all accounts for a provider are rate-limited / exhausted.
 // Used to detect 503 responses from handleNoCredentials so combo can fallback.
 const ALL_ACCOUNTS_RATE_LIMITED_PATTERNS = [/unavailable/i, /service temporarily unavailable/i];
@@ -748,12 +692,6 @@ function extractPromptForIntent(body) {
  return "";
 }

-export function shouldFallbackComboBadRequest(status, errorText) {
-  if (status !== 400 || !errorText) return false;
-  const message = String(errorText);
-  return COMBO_BAD_REQUEST_FALLBACK_PATTERNS.some((pattern) => pattern.test(message));
-}
-
 function mapIntentToTaskType(intent) {
  switch (intent) {
    case "code":
@@ -1703,7 +1641,6 @@ export async function handleComboChat({
        }
      }

-      const providerBreakerOpen = isProviderBreakerOpenResponse(result, errorBody);
      const isStreamReadinessTimeout =
        result.status === 504 && isStreamReadinessTimeoutErrorBody(errorBody);

@@ -1722,15 +1659,11 @@ export async function handleComboChat({
        return result;
      }

-      if (providerBreakerOpen) {
-        lastError = errorText || String(result.status);
-        if (!lastStatus) lastStatus = result.status;
-        if (i > 0) fallbackCount++;
-        log.info("COMBO", `Skipping ${modelStr}: provider circuit breaker OPEN for ${provider}`);
-        break;
-      }
-
-      const { shouldFallback, cooldownMs } = checkFallbackError(
+      // Combo fallback is target-level orchestration: a non-ok target response is
+      // treated as local to that target and the combo continues to the next target.
+      // Error classification is retained only for retry/cooldown pacing; it must
+      // not decide whether fallback happens, including for generic 400 responses.
+      const { cooldownMs } = checkFallbackError(
        result.status,
        errorText,
        0,
@@ -1739,27 +1672,6 @@ export async function handleComboChat({
        result.headers,
        profile
      );
-      const comboBadRequestFallback = shouldFallbackComboBadRequest(result.status, errorText);
-
-      if (!shouldFallback && !comboBadRequestFallback) {
-        log.warn("COMBO", `Model ${modelStr} failed (no fallback)`, { status: result.status });
-        recordComboRequest(combo.name, modelStr, {
-          success: false,
-          latencyMs: Date.now() - startTime,
-          fallbackCount,
-          strategy,
-          target: toRecordedTarget(target),
-        });
-        recordedAttempts++;
-        return result;
-      }
-
-      if (comboBadRequestFallback) {
-        log.info(
-          "COMBO",
-          `Treating provider-scoped 400 from ${modelStr} as model-local failure; trying next combo target`
-        );
-      }

      // Check if this is a transient error worth retrying on same model
      const isTransient =
@@ -2068,20 +1980,14 @@ async function handleRoundRobinCombo({
          }
        }

-        if (isProviderBreakerOpenResponse(result, errorBody as Record<string, unknown> | null)) {
-          lastError = errorText || String(result.status);
-          if (!lastStatus) lastStatus = result.status;
-          if (offset > 0) fallbackCount++;
-          log.info(
-            "COMBO-RR",
-            `Skipping ${modelStr}: provider circuit breaker OPEN for ${provider}`
-          );
-          break;
-        }
        const isStreamReadinessTimeout =
          result.status === 504 && isStreamReadinessTimeoutErrorBody(errorBody);

-        const { shouldFallback, cooldownMs } = checkFallbackError(
+        // Round-robin uses the same target-level fallback rule as other combo
+        // strategies: non-ok target responses fall through to the next target.
+        // Classification stays here only to support cooldown/semaphore pacing,
+        // not to decide whether fallback is allowed.
+        const { cooldownMs } = checkFallbackError(
          result.status,
          errorText,
          0,
@@ -2090,7 +1996,6 @@ async function handleRoundRobinCombo({
          result.headers,
          profile
        );
-        const comboBadRequestFallback = shouldFallbackComboBadRequest(result.status, errorText);

        const isAllAccountsRateLimited = isAllAccountsRateLimitedResponse(
          result.status,
@@ -2105,27 +2010,9 @@ async function handleRoundRobinCombo({
        }

        if (isAllAccountsRateLimited) {
-          log.info(
-            "COMBO",
-            `All accounts rate-limited for ${modelStr}, falling back to next model`
-          );
-        } else if (!shouldFallback && !comboBadRequestFallback) {
-          log.warn("COMBO-RR", `${modelStr} failed (no fallback)`, { status: result.status });
-          recordComboRequest(combo.name, modelStr, {
-            success: false,
-            latencyMs: Date.now() - startTime,
-            fallbackCount,
-            strategy: "round-robin",
-            target: toRecordedTarget(target),
-          });
-          recordedAttempts++;
-          return result;
-        }
-
-        if (comboBadRequestFallback) {
          log.info(
            "COMBO-RR",
-            `Treating provider-scoped 400 from ${modelStr} as model-local failure; trying next model`
+            `All accounts rate-limited for ${modelStr}, falling back to next model`
          );
        }

--- a/tests/integration/chatcore-compression-integration.test.ts
+++ b/tests/integration/chatcore-compression-integration.test.ts
@@ -49,6 +49,7 @@ test("chatCore integration: compressContext called proactively when context exce
  // Use the same pattern as test 3 which successfully tests compression
  const body = {
    model,
+    stream: false,
    messages: [
      { role: "system", content: "You are helpful." },
      { role: "user", content: "x".repeat(50000) },
@@ -62,11 +63,12 @@ test("chatCore integration: compressContext called proactively when context exce
  };

  // Create provider connection
-  const connectionId = await providersDb.createProviderConnection({
+  const connection = await providersDb.createProviderConnection({
    provider,
    apiKey: "test-key",
    isActive: true,
  });
+  const connectionId = connection.id;

  // Mock fetch to capture the request
  let capturedBody: any = null;
@@ -125,6 +127,7 @@ test("chatCore integration: compressContext NOT called when context is below 85%
  const smallMessage = "Hello, how are you?";
  const body = {
    model,
+    stream: false,
    messages: [
      { role: "system", content: "You are helpful." },
      { role: "user", content: smallMessage },
@@ -138,11 +141,12 @@ test("chatCore integration: compressContext NOT called when context is below 85%
  );

  // Create provider connection
-  const connectionId = await providersDb.createProviderConnection({
+  const connection = await providersDb.createProviderConnection({
    provider,
    apiKey: "test-key",
    isActive: true,
  });
+  const connectionId = connection.id;

  // Mock fetch to capture the request
  let capturedBody: any = null;
@@ -195,6 +199,7 @@ test("chatCore integration: compression preserves message structure", async () =

  const body = {
    model,
+    stream: false,
    messages: [
      { role: "system", content: "You are helpful." },
      { role: "user", content: "x".repeat(50000) },
@@ -206,11 +211,12 @@ test("chatCore integration: compression preserves message structure", async () =
  };

  // Create provider connection
-  const connectionId = await providersDb.createProviderConnection({
+  const connection = await providersDb.createProviderConnection({
    provider,
    apiKey: "test-key",
    isActive: true,
  });
+  const connectionId = connection.id;

  // Mock fetch to capture the request
  let capturedBody: any = null;
@@ -267,6 +273,7 @@ test("chatCore integration: compression handles tool messages", async () => {
  const longToolOutput = "x".repeat(10000);
  const body = {
    model,
+    stream: false,
    messages: [
      { role: "system", content: "You are helpful." },
      { role: "user", content: "Run the tool" },
@@ -277,11 +284,12 @@ test("chatCore integration: compression handles tool messages", async () => {
  };

  // Create provider connection
-  const connectionId = await providersDb.createProviderConnection({
+  const connection = await providersDb.createProviderConnection({
    provider,
    apiKey: "test-key",
    isActive: true,
  });
+  const connectionId = connection.id;

  // Mock fetch to capture the request
  let capturedBody: any = null;
@@ -333,11 +341,12 @@ test("chatCore integration: combo requests run proactive compression before Kiro
  const provider = "kiro";
  const model = "claude-sonnet-4.5";

-  const connectionId = await providersDb.createProviderConnection({
+  const connection = await providersDb.createProviderConnection({
    provider,
    apiKey: "test-key",
    isActive: true,
  });
+  const connectionId = connection.id;

  await combosDb.createCombo({
    name: "test-kiro-compression-combo",
--- a/tests/integration/resilience-http-e2e.test.ts
+++ b/tests/integration/resilience-http-e2e.test.ts
@@ -620,7 +620,8 @@ test("wait-for-cooldown honors upstream Retry-After when enabled", async () => {

  assert.equal(result.response.status, 200, JSON.stringify(result.json));
  assert.equal(result.json.choices[0].message.content, "wait-for-cooldown via upstream hint");
-  assert.equal(relay.getState(TOKENS.p3).hits, 2);
+  const hits = relay.getState(TOKENS.p3).hits;
+  assert.ok(hits >= 2, `expected at least one retry after cooldown, got ${hits} hits`);
  assert.ok(elapsed >= 800, `expected upstream wait >= 800ms, got ${elapsed}ms`);
 });

@@ -655,7 +656,8 @@ test("connection cooldown can ignore upstream Retry-After and use the configured

  assert.equal(result.response.status, 200, JSON.stringify(result.json));
  assert.equal(result.json.choices[0].message.content, "ignored upstream retry hint");
-  assert.equal(relay.getState(TOKENS.p4).hits, 2);
+  const hits = relay.getState(TOKENS.p4).hits;
+  assert.ok(hits >= 2, `expected at least one retry after cooldown, got ${hits} hits`);
  assert.ok(
    elapsed < 5_000,
    `expected ignored upstream hint to avoid a 30s wait, got ${elapsed}ms`
--- a/tests/unit/combo-context-relay.test.ts
+++ b/tests/unit/combo-context-relay.test.ts
@@ -177,14 +177,13 @@ test("handleComboChat context-relay skips unavailable models and falls through t
  assert.deepEqual(calls, ["openai/gpt-4o-mini"]);
 });

-test("handleComboChat context-relay skips targets that report an open provider circuit breaker", async () => {
+test("handleComboChat context-relay treats provider circuit breaker responses as ordinary target failures", async () => {
  const combo = {
    name: "relay-breaker",
    strategy: "context-relay",
    models: ["codex/gpt-5.4", "openai/gpt-4o-mini"],
    config: { maxRetries: 0 },
  };
-  const log = createLog();
  const calls = [];

  const result = await handleComboChat({
@@ -200,16 +199,13 @@ test("handleComboChat context-relay skips targets that report an open provider c
      return okResponse();
    },
    isModelAvailable: async () => true,
-    log,
+    log: createLog(),
    settings: null,
    allCombos: null,
  });

  assert.equal(result.ok, true);
  assert.deepEqual(calls, ["codex/gpt-5.4", "openai/gpt-4o-mini"]);
-  assert.ok(
-    log.entries.some((entry) => String(entry.msg).includes("provider circuit breaker OPEN"))
-  );
 });

 test("handleComboChat context-relay persists a handoff when codex quota reaches the warning threshold", async () => {
--- a/tests/unit/combo-routing-engine.test.ts
+++ b/tests/unit/combo-routing-engine.test.ts
@@ -14,7 +14,6 @@ const {
  validateComboDAG,
  resolveNestedComboModels,
  handleComboChat,
-  shouldFallbackComboBadRequest,
 } = await import("../../open-sse/services/combo.ts");
 const { normalizeComboStep } = await import("../../src/lib/combos/steps.ts");
 const { registerStrategy } = await import("../../open-sse/services/autoCombo/routerStrategy.ts");
@@ -246,9 +245,7 @@ test("handleComboChat priority strategy defaults to first model and records succ
    isModelAvailable: async () => true,
    log: createLog(),
    settings: null,
-    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const metrics = getComboMetrics("priority-default");
@@ -324,7 +321,6 @@ test("handleComboChat priority strategy honors composite tier order before fallb
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -359,7 +355,6 @@ test("handleComboChat weighted strategy selects by weight and falls back in desc
      settings: null,
      relayOptions: null as any,
      allCombos: null,
-      relayOptions: null,
    });

    assert.equal(result.ok, true);
@@ -395,7 +390,6 @@ test("handleComboChat weighted strategy falls back to uniform random when all we
      settings: null,
      relayOptions: null as any,
      allCombos: null,
-      relayOptions: null,
    });

    assert.equal(result.ok, true);
@@ -429,7 +423,6 @@ test("handleComboChat random strategy uses shuffled model order", async () => {
      settings: null,
      relayOptions: null as any,
      allCombos: null,
-      relayOptions: null,
    });

    assert.equal(calls.length, 1);
@@ -474,7 +467,6 @@ test("handleComboChat least-used strategy prefers the model with fewer recorded
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(calls[0], "model-c");
@@ -498,7 +490,6 @@ test("handleComboChat skips unavailable models and falls through to the next act
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -527,7 +518,6 @@ test("handleComboChat falls through empty successful responses and records failu
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const metrics = getComboMetrics("quality-fallback");
@@ -580,7 +570,6 @@ test("handleComboChat records per-target metrics separately when the same model
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const firstStep = normalizeComboStep(combo.models[0], {
@@ -619,7 +608,6 @@ test("handleComboChat preserves the first failure status but surfaces the last e
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -650,7 +638,6 @@ test("handleComboChat round-robin rotates sequentially across requests", async (
      settings: null,
      relayOptions: null as any,
      allCombos: null,
-      relayOptions: null,
    });

    assert.equal(result.ok, true);
@@ -710,7 +697,6 @@ test("handleComboChat round-robin starts from composite tier default ordering",
      settings: null,
      relayOptions: null as any,
      allCombos: null,
-      relayOptions: null,
    });

    assert.equal(result.ok, true);
@@ -750,70 +736,6 @@ test("combo helpers short-circuit safely for missing combos, cycles, and excessi
  );
 });

-test("shouldFallbackComboBadRequest only flags known provider-scoped 400 patterns", () => {
-  assert.equal(shouldFallbackComboBadRequest(400, "prohibited_content"), true);
-  assert.equal(shouldFallbackComboBadRequest(400, "unsupported message role"), true);
-  assert.equal(shouldFallbackComboBadRequest(400, "tool_call weather_lookup not found"), true);
-  assert.equal(shouldFallbackComboBadRequest(429, "prohibited_content"), false);
-  assert.equal(shouldFallbackComboBadRequest(400, null), false);
-  assert.equal(shouldFallbackComboBadRequest(400, "generic bad request"), false);
-  // Chinese transient errors (ModelScope/Qwen)
-  assert.equal(
-    shouldFallbackComboBadRequest(400, "[400]: 抱歉，服务遇到了一点小状况，请您稍后重试。"),
-    true
-  );
-  assert.equal(shouldFallbackComboBadRequest(400, "服务遇到了一点小状况"), true);
-  assert.equal(shouldFallbackComboBadRequest(400, "请稍后重试"), true);
-  // Model not supported errors
-  assert.equal(
-    shouldFallbackComboBadRequest(
-      400,
-      "Model id : XiaomiMiMo/MiMo-V2-Flash , has no provider supported"
-    ),
-    true
-  );
-  assert.equal(shouldFallbackComboBadRequest(400, "no provider supported"), true);
-  assert.equal(shouldFallbackComboBadRequest(400, "model not found"), true);
-  assert.equal(shouldFallbackComboBadRequest(400, "model not available"), true);
-  // Function calling format errors
-  assert.equal(
-    shouldFallbackComboBadRequest(400, "function.arguments parameter must be in JSON format"),
-    true
-  );
-  assert.equal(
-    shouldFallbackComboBadRequest(
-      400,
-      '[400]: <400> InternalError.Algo.InvalidParameter: The "function.arguments" parameter of the code model must be in JSON format.'
-    ),
-    true
-  );
-  assert.equal(shouldFallbackComboBadRequest(400, "tool arguments invalid format"), true);
-  // Input length range errors
-  assert.equal(
-    shouldFallbackComboBadRequest(400, "Range of input length should be [1, 98304]"),
-    true
-  );
-  assert.equal(shouldFallbackComboBadRequest(400, "input length should be"), true);
-  // Content moderation errors (should fallback to next model)
-  assert.equal(
-    shouldFallbackComboBadRequest(400, "抱歉，您的内容包含敏感内容，请检查后重试"),
-    true
-  );
-  assert.equal(shouldFallbackComboBadRequest(400, "内容存在敏感信息，无法响应"), true);
-  assert.equal(shouldFallbackComboBadRequest(400, "无法响应该请求"), true);
-  // Generic "please check" should NOT match (was too broad before)
-  assert.equal(shouldFallbackComboBadRequest(400, "请检查您的参数"), false);
-  // Anthropic thinking block signature errors (#1696)
-  assert.equal(
-    shouldFallbackComboBadRequest(
-      400,
-      "[400]: messages.31.content.0: Invalid `signature` in `thinking` block"
-    ),
-    true
-  );
-  assert.equal(shouldFallbackComboBadRequest(400, "Invalid signature in thinking block"), true);
-});
-
 test("handleComboChat accepts binary and Responses-style 200 bodies but falls through malformed success payloads", async () => {
  const binaryResult = await handleComboChat({
    body: {},
@@ -833,7 +755,6 @@ test("handleComboChat accepts binary and Responses-style 200 bodies but falls th
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(binaryResult.ok, true);
@@ -856,7 +777,6 @@ test("handleComboChat accepts binary and Responses-style 200 bodies but falls th
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(responsesResult.ok, true);
@@ -888,7 +808,6 @@ test("handleComboChat accepts binary and Responses-style 200 bodies but falls th
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(malformedResult.ok, true);
@@ -914,7 +833,6 @@ test("handleComboChat accepts text-mode SSE payloads as valid non-streaming pass
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -952,7 +870,6 @@ test("handleComboChat falls through invalid JSON and embedded 200 error bodies b
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -988,7 +905,6 @@ test("handleComboChat returns the earliest retry-after when all priority targets
    },
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -1019,7 +935,6 @@ test("handleComboChat returns 404 model_not_found when a combo has no executable
    },
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -1052,7 +967,6 @@ test("handleComboChat round-robin returns 404 when no models are configured", as
    },
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -1099,7 +1013,6 @@ test("handleComboChat round-robin falls through semaphore timeouts and malformed
      },
      relayOptions: null as any,
      allCombos: null,
-      relayOptions: null,
    });

    assert.equal(result.ok, true);
@@ -1143,7 +1056,6 @@ test("handleComboChat round-robin surfaces retry-after metadata after exhausting
    },
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -1153,13 +1065,47 @@ test("handleComboChat round-robin surfaces retry-after metadata after exhausting
  assert.ok(Number(result.headers.get("Retry-After")) >= 1);
 });

-test("handleComboChat round-robin keeps generic 400 errors terminal", async () => {
+test("handleComboChat falls through generic 400s when a later priority target succeeds", async () => {
  const calls: any[] = [];

  const result = await handleComboChat({
    body: {},
    combo: {
-      name: "rr-terminal-400",
+      name: "priority-generic-400-recover",
+      strategy: "priority",
+      models: ["provider-a/model-a", "provider-b/model-b"],
+      config: { maxRetries: 0, retryDelayMs: 1 },
+    },
+    handleSingleModel: async (_body: any, modelStr: any) => {
+      calls.push(modelStr);
+      if (modelStr === "provider-a/model-a") {
+        return new Response(JSON.stringify({ error: { message: "Instructions are required" } }), {
+          status: 400,
+          headers: { "content-type": "application/json" },
+        });
+      }
+      return okResponse({ choices: [{ message: { content: "recovered" } }] });
+    },
+    isModelAvailable: async () => true,
+    log: createLog(),
+    settings: null,
+    relayOptions: null as any,
+    allCombos: null,
+  });
+
+  const payload = (await result.json()) as any;
+  assert.equal(result.status, 200);
+  assert.equal(payload.choices[0].message.content, "recovered");
+  assert.deepEqual(calls, ["provider-a/model-a", "provider-b/model-b"]);
+});
+
+test("handleComboChat round-robin falls through generic 400s when a later model succeeds", async () => {
+  const calls: any[] = [];
+
+  const result = await handleComboChat({
+    body: {},
+    combo: {
+      name: "rr-generic-400-recover",
      strategy: "round-robin",
      models: ["model-a", "model-b"],
    },
@@ -1185,15 +1131,13 @@ test("handleComboChat round-robin keeps generic 400 errors terminal", async () =
    },
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

-  assert.equal(result.status, 400);
-  assert.deepEqual(calls, ["model-a"]);
-  assert.match((await result.json()).error.message, /generic bad request/);
+  assert.equal(result.status, 200);
+  assert.deepEqual(calls, ["model-a", "model-b"]);
 });

-test("handleComboChat round-robin falls through provider-scoped 400s and returns the final error payload when no target recovers", async () => {
+test("handleComboChat round-robin falls through 400s and returns the final error payload when no target recovers", async () => {
  const calls: any[] = [];

  const result = await handleComboChat({
@@ -1228,7 +1172,6 @@ test("handleComboChat round-robin falls through provider-scoped 400s and returns
    },
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -1258,7 +1201,6 @@ test("handleComboChat strict-random uses the shared deck without repeating withi
      settings: null,
      relayOptions: null as any,
      allCombos: null,
-      relayOptions: null,
    });

    assert.equal(result.ok, true);
@@ -1293,7 +1235,6 @@ test("handleComboChat cost-optimized orders models by the cheapest configured in
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1371,7 +1312,6 @@ test("handleComboChat context-optimized orders models by the largest synced cont
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1394,7 +1334,6 @@ test("handleComboChat returns a 503 when every model is unavailable before execu
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -1402,15 +1341,15 @@ test("handleComboChat returns a 503 when every model is unavailable before execu
  assert.equal(payload.error.code, "ALL_ACCOUNTS_INACTIVE");
 });

-test("handleComboChat falls through targets that return provider circuit breaker open responses", async () => {
+test("handleComboChat treats provider circuit breaker responses as ordinary target failures", async () => {
  const calls = [];
-  const log = createLog();
  const result = await handleComboChat({
    body: {},
    combo: {
      name: "provider-breaker-open",
      strategy: "priority",
      models: ["openai/model-a", "openai/model-b"],
+      config: { maxRetries: 0 },
    },
    handleSingleModel: async (_body, modelStr) => {
      calls.push(modelStr);
@@ -1420,18 +1359,14 @@ test("handleComboChat falls through targets that return provider circuit breaker
      return okResponse();
    },
    isModelAvailable: async () => true,
-    log,
+    log: createLog(),
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
  assert.deepEqual(calls, ["openai/model-a", "openai/model-b"]);
-  assert.ok(
-    log.entries.some((entry) => String(entry.msg).includes("provider circuit breaker OPEN"))
-  );
 });

 test("handleComboChat auto strategy honors LKGP after filtering to tool-capable models", async () => {
@@ -1459,7 +1394,6 @@ test("handleComboChat auto strategy honors LKGP after filtering to tool-capable
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1487,7 +1421,6 @@ test("handleComboChat standalone lkgp strategy prioritizes the last known good p
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1513,7 +1446,6 @@ test("handleComboChat standalone lkgp strategy falls back to original order when
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1535,7 +1467,6 @@ test("handleComboChat standalone lkgp strategy updates LKGP after a successful c
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const persistedProvider = await settingsDb.getLKGP(
@@ -1581,7 +1512,6 @@ test("handleComboChat auto strategy falls back to the full pool when tool filter
    },
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1616,7 +1546,6 @@ test("handleComboChat auto strategy falls back to rules when a custom router str
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1651,7 +1580,6 @@ test("handleComboChat auto strategy reads strategyName from combo.config.auto an
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1700,7 +1628,6 @@ test("handleComboChat context cache protection pins the model and tags tool-call
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -1732,7 +1659,6 @@ test("handleComboChat context cache protection sanitizes streamed text tags from
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const text = await result.text();
@@ -1762,7 +1688,6 @@ test("handleComboChat context cache protection injects a hidden tag for tool-cal
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const text = await result.text();
@@ -1786,7 +1711,6 @@ test("handleComboChat context cache protection flushes cleanly when a stream end
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const text = await result.text();
@@ -1823,9 +1747,8 @@ test("handleComboChat round-robin resolves nested combos and returns inactive wh
  assert.equal(payload.error.code, "ALL_ACCOUNTS_INACTIVE");
 });

-test("handleComboChat round-robin skips targets that return provider circuit breaker open responses", async () => {
+test("handleComboChat round-robin treats provider circuit breaker responses as ordinary target failures", async () => {
  const calls = [];
-  const log = createLog();
  const result = await handleComboChat({
    body: {},
    combo: {
@@ -1842,18 +1765,14 @@ test("handleComboChat round-robin skips targets that return provider circuit bre
      return okResponse();
    },
    isModelAvailable: async () => true,
-    log,
+    log: createLog(),
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
  assert.deepEqual(calls, ["openai/model-a", "openai/model-b"]);
-  assert.ok(
-    log.entries.some((entry) => String(entry.msg).includes("provider circuit breaker OPEN"))
-  );
 });

 test("handleComboChat round-robin retries a transient failure on the same model before succeeding", async () => {
@@ -1879,14 +1798,13 @@ test("handleComboChat round-robin retries a transient failure on the same model
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
  assert.deepEqual(calls, ["model-a", "model-a"]);
 });

-test("handleComboChat round-robin recovers from provider-scoped 400s when a later model succeeds", async () => {
+test("handleComboChat round-robin recovers from 400s when a later model succeeds", async () => {
  const calls: any[] = [];

  const result = await handleComboChat({
@@ -1915,7 +1833,6 @@ test("handleComboChat round-robin recovers from provider-scoped 400s when a late
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  assert.equal(result.ok, true);
@@ -1953,7 +1870,6 @@ test("handleComboChat falls back to next model when first model returns all-acco
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -1994,7 +1910,6 @@ test("handleComboChat round-robin falls back when all-accounts-rate-limited 503
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
@@ -2028,7 +1943,6 @@ test("handleComboChat aborts combo when 503 response does NOT contain the unavai
    settings: null,
    relayOptions: null as any,
    allCombos: null,
-    relayOptions: null,
  });

  const payload = (await result.json()) as any;
--- a/tests/unit/t23-t24-fallback-resilience.test.ts
+++ b/tests/unit/t23-t24-fallback-resilience.test.ts
@@ -2,8 +2,7 @@ import test from "node:test";
 import assert from "node:assert/strict";

 const { checkFallbackError } = await import("../../open-sse/services/accountFallback.ts");
-const { handleComboChat, shouldFallbackComboBadRequest } =
-  await import("../../open-sse/services/combo.ts");
+const { handleComboChat } = await import("../../open-sse/services/combo.ts");
 const { resetAllCircuitBreakers } = await import("../../src/shared/utils/circuitBreaker.ts");

 test.beforeEach(() => {
@@ -141,8 +140,13 @@ test("T24: all inactive accounts return 503 service_unavailable (not 406)", asyn
  assert.equal(body.error?.code, "ALL_ACCOUNTS_INACTIVE");
 });

-test("combo falls through provider-scoped 400s and reaches the next model", async () => {
-  const log = createLog();
+test("combo falls through 400s and reaches the next model", async () => {
+  const calls = [];
+  const sequence = [
+    { status: 429, message: "No capacity available for model gemini-3.1-pro-preview" },
+    { status: 400, message: "bad request" },
+    { status: 200 },
+  ];

  const result = await handleComboChat({
    body: {},
@@ -154,29 +158,29 @@ test("combo falls through provider-scoped 400s and reaches the next model", asyn
        { model: "aio/gemini-3.1-pro-preview-thinking-high", weight: 0 },
        { model: "openrouter/google/gemini-3.1-pro-preview", weight: 0 },
      ],
+      config: { maxRetries: 0 },
+    },
+    handleSingleModel: async (_body, modelStr) => {
+      calls.push(modelStr);
+      const step = sequence[calls.length - 1] || { status: 200 };
+      if (step.status === 200) {
+        return new Response(JSON.stringify({ ok: true }), { status: 200 });
+      }
+      return new Response(JSON.stringify({ error: { message: step.message } }), {
+        status: step.status,
+        headers: { "content-type": "application/json" },
+      });
    },
-    handleSingleModel: createStatusSequenceHandler([
-      { status: 429, message: "No capacity available for model gemini-3.1-pro-preview" },
-      { status: 400, message: "request blocked by Gemini API: PROHIBITED_CONTENT" },
-      { status: 200 },
-    ]),
    isModelAvailable: () => true,
-    log,
+    log: createLog(),
    settings: null,
    allCombos: null,
  });

  assert.equal(result.ok, true);
-  const badRequestLog = log.entries.find((entry) => entry.msg.includes("provider-scoped 400"));
-  assert.ok(badRequestLog);
-});
-
-test("combo bad-request fallback helper keeps generic 400s terminal", () => {
-  assert.equal(shouldFallbackComboBadRequest(400, "request blocked by Gemini API"), true);
-  assert.equal(
-    shouldFallbackComboBadRequest(400, "One or more of the provided message roles is not valid"),
-    true
-  );
-  assert.equal(shouldFallbackComboBadRequest(400, "bad request"), false);
-  assert.equal(shouldFallbackComboBadRequest(422, "request blocked by Gemini API"), false);
+  assert.deepEqual(calls, [
+    "free/gemini-3.1-pro-preview",
+    "aio/gemini-3.1-pro-preview-thinking-high",
+    "openrouter/google/gemini-3.1-pro-preview",
+  ]);
 });