agents: preserve transient probe slot on model-not-found probes

src/agents/model-fallback.test.ts

@@ -1358,6 +1358,46 @@ describe("runWithModelFallback", () => {
       });
       expect(run).toHaveBeenNthCalledWith(2, "groq", "llama-3.3-70b-versatile");
     });
+
+    it("does not consume transient probe slot when first same-provider probe fails with model_not_found", async () => {
+      const { dir } = await makeAuthStoreWithCooldown("anthropic", "rate_limit");
+      const cfg = makeCfg({
+        agents: {
+          defaults: {
+            model: {
+              primary: "anthropic/claude-opus-4-6",
+              fallbacks: [
+                "anthropic/claude-sonnet-4-5",
+                "anthropic/claude-haiku-3-5",
+                "groq/llama-3.3-70b-versatile",
+              ],
+            },
+          },
+        },
+      });
+
+      const run = vi
+        .fn()
+        .mockRejectedValueOnce(new Error("Model not found: anthropic/claude-sonnet-4-5"))
+        .mockResolvedValueOnce("haiku success");
+
+      const result = await runWithModelFallback({
+        cfg,
+        provider: "anthropic",
+        model: "claude-opus-4-6",
+        run,
+        agentDir: dir,
+      });
+
+      expect(result.result).toBe("haiku success");
+      expect(run).toHaveBeenCalledTimes(2);
+      expect(run).toHaveBeenNthCalledWith(1, "anthropic", "claude-sonnet-4-5", {
+        allowTransientCooldownProbe: true,
+      });
+      expect(run).toHaveBeenNthCalledWith(2, "anthropic", "claude-haiku-3-5", {
+        allowTransientCooldownProbe: true,
+      });
+    });
   });
 });
 

src/agents/model-fallback.ts

@@ -532,6 +532,7 @@ export async function runWithModelFallback<T>(params: {
       params.provider === candidate.provider && params.model === candidate.model;
     let runOptions: ModelFallbackRunOptions | undefined;
     let attemptedDuringCooldown = false;
+    let transientProbeProviderForAttempt: string | null = null;
     if (authStore) {
       const profileIds = resolveAuthProfileOrder({
         cfg: params.cfg,
@@ -622,7 +623,7 @@ export async function runWithModelFallback<T>(params: {
           }
           runOptions = { allowTransientCooldownProbe: true };
           if (isTransientCooldownReason) {
-            cooldownProbeUsedProviders.add(candidate.provider);
+            transientProbeProviderForAttempt = candidate.provider;
           }
         }
         attemptedDuringCooldown = true;
@@ -678,6 +679,18 @@ export async function runWithModelFallback<T>(params: {
     }
     const err = attemptRun.error;
     {
+      if (transientProbeProviderForAttempt) {
+        const probeFailureReason = describeFailoverError(err).reason;
+        const shouldPreserveTransientProbeSlot =
+          probeFailureReason === "model_not_found" ||
+          probeFailureReason === "format" ||
+          probeFailureReason === "auth" ||
+          probeFailureReason === "auth_permanent" ||
+          probeFailureReason === "session_expired";
+        if (!shouldPreserveTransientProbeSlot) {
+          cooldownProbeUsedProviders.add(transientProbeProviderForAttempt);
+        }
+      }
       // Context overflow errors should be handled by the inner runner's
       // compaction/retry logic, not by model fallback.  If one escapes as a
       // throw, rethrow it immediately rather than trying a different model