Implement saver mode (#154)

src/ipc/utils/get_model_client.ts

@@ -1,3 +1,4 @@
+import { LanguageModelV1 } from "ai";
 import { createOpenAI } from "@ai-sdk/openai";
 import { createGoogleGenerativeAI as createGoogle } from "@ai-sdk/google";
 import { createAnthropic } from "@ai-sdk/anthropic";
@@ -8,6 +9,8 @@ import type { LargeLanguageModel, UserSettings } from "../../lib/schemas";
 import { getEnvVar } from "./read_env";
 import log from "electron-log";
 import { getLanguageModelProviders } from "../shared/language_model_helpers";
+import { LanguageModelProvider } from "../ipc_types";
+import { llmErrorStore } from "@/main/llm_error_store";
 
 const AUTO_MODELS = [
   {
@@ -24,11 +27,19 @@ const AUTO_MODELS = [
   },
 ];
 
+export interface ModelClient {
+  model: LanguageModelV1;
+  builtinProviderId?: string;
+}
+
 const logger = log.scope("getModelClient");
 export async function getModelClient(
   model: LargeLanguageModel,
   settings: UserSettings,
-) {
+): Promise<{
+  modelClient: ModelClient;
+  backupModelClients: ModelClient[];
+}> {
   const allProviders = await getLanguageModelProviders();
 
   const dyadApiKey = settings.providerSettings?.auto?.apiKey?.value;
@@ -83,7 +94,44 @@ export async function getModelClient(
       logger.info("Using Dyad Pro API key via Gateway");
       // Do not use free variant (for openrouter).
       const modelName = model.name.split(":free")[0];
-      return provider(`${providerConfig.gatewayPrefix}${modelName}`);
+      const autoModelClient = {
+        model: provider(`${providerConfig.gatewayPrefix}${modelName}`),
+        builtinProviderId: "auto",
+      };
+      const googleSettings = settings.providerSettings?.google;
+
+      // Budget saver mode logic (all must be true):
+      // 1. Pro Saver Mode is enabled
+      // 2. Provider is Google
+      // 3. API Key is set
+      // 4. Has no recent errors
+      if (
+        settings.enableProSaverMode &&
+        providerConfig.id === "google" &&
+        googleSettings &&
+        googleSettings.apiKey?.value &&
+        llmErrorStore.modelHasNoRecentError({
+          model: model.name,
+          provider: providerConfig.id,
+        })
+      ) {
+        return {
+          modelClient: getRegularModelClient(
+            {
+              provider: providerConfig.id,
+              name: model.name,
+            },
+            settings,
+            providerConfig,
+          ).modelClient,
+          backupModelClients: [autoModelClient],
+        };
+      } else {
+        return {
+          modelClient: autoModelClient,
+          backupModelClients: [],
+        };
+      }
     } else {
       logger.warn(
         `Dyad Pro enabled, but provider ${model.provider} does not have a gateway prefix defined. Falling back to direct provider connection.`,
@@ -91,7 +139,14 @@ export async function getModelClient(
       // Fall through to regular provider logic if gateway prefix is missing
     }
   }
+  return getRegularModelClient(model, settings, providerConfig);
+}
 
+function getRegularModelClient(
+  model: LargeLanguageModel,
+  settings: UserSettings,
+  providerConfig: LanguageModelProvider,
+) {
   // Get API key for the specific provider
   const apiKey =
     settings.providerSettings?.[model.provider]?.apiKey?.value ||
@@ -99,30 +154,60 @@ export async function getModelClient(
       ? getEnvVar(providerConfig.envVarName)
       : undefined);
 
+  const providerId = providerConfig.id;
   // Create client based on provider ID or type
-  switch (providerConfig.id) {
+  switch (providerId) {
     case "openai": {
       const provider = createOpenAI({ apiKey });
-      return provider(model.name);
+      return {
+        modelClient: {
+          model: provider(model.name),
+          builtinProviderId: providerId,
+        },
+        backupModelClients: [],
+      };
     }
     case "anthropic": {
       const provider = createAnthropic({ apiKey });
-      return provider(model.name);
+      return {
+        modelClient: {
+          model: provider(model.name),
+          builtinProviderId: providerId,
+        },
+        backupModelClients: [],
+      };
     }
     case "google": {
       const provider = createGoogle({ apiKey });
-      return provider(model.name);
+      return {
+        modelClient: {
+          model: provider(model.name),
+          builtinProviderId: providerId,
+        },
+        backupModelClients: [],
+      };
     }
     case "openrouter": {
       const provider = createOpenRouter({ apiKey });
-      return provider(model.name);
+      return {
+        modelClient: {
+          model: provider(model.name),
+          builtinProviderId: providerId,
+        },
+        backupModelClients: [],
+      };
     }
     case "ollama": {
       // Ollama typically runs locally and doesn't require an API key in the same way
       const provider = createOllama({
         baseURL: providerConfig.apiBaseUrl,
       });
-      return provider(model.name);
+      return {
+        modelClient: {
+          model: provider(model.name),
+        },
+        backupModelClients: [],
+      };
     }
     case "lmstudio": {
       // LM Studio uses OpenAI compatible API
@@ -131,7 +216,12 @@ export async function getModelClient(
         name: "lmstudio",
         baseURL,
       });
-      return provider(model.name);
+      return {
+        modelClient: {
+          model: provider(model.name),
+        },
+        backupModelClients: [],
+      };
     }
     default: {
       // Handle custom providers
@@ -147,7 +237,12 @@ export async function getModelClient(
           baseURL: providerConfig.apiBaseUrl,
           apiKey: apiKey,
         });
-        return provider(model.name);
+        return {
+          modelClient: {
+            model: provider(model.name),
+          },
+          backupModelClients: [],
+        };
       }
       // If it's not a known ID and not type 'custom', it's unsupported
       throw new Error(`Unsupported model provider: ${model.provider}`);

src/ipc/utils/stream_utils.ts

@@ -0,0 +1,123 @@
+import { streamText } from "ai";
+import log from "electron-log";
+import { ModelClient } from "./get_model_client";
+import { llmErrorStore } from "@/main/llm_error_store";
+const logger = log.scope("stream_utils");
+
+export interface StreamTextWithBackupParams
+  extends Omit<Parameters<typeof streamText>[0], "model"> {
+  model: ModelClient; // primary client
+  backupModelClients?: ModelClient[]; // ordered fall-backs
+}
+
+export function streamTextWithBackup(params: StreamTextWithBackupParams): {
+  textStream: AsyncIterable<string>;
+} {
+  const {
+    model: primaryModel,
+    backupModelClients = [],
+    onError: callerOnError,
+    abortSignal: callerAbort,
+    ...rest
+  } = params;
+
+  const modelClients: ModelClient[] = [primaryModel, ...backupModelClients];
+
+  async function* combinedGenerator(): AsyncIterable<string> {
+    let lastErr: { error: unknown } | undefined = undefined;
+
+    for (let i = 0; i < modelClients.length; i++) {
+      const currentModelClient = modelClients[i];
+
+      /* Local abort controller for this single attempt  */
+      const attemptAbort = new AbortController();
+      if (callerAbort) {
+        if (callerAbort.aborted) {
+          // Already aborted, trigger immediately
+          attemptAbort.abort();
+        } else {
+          callerAbort.addEventListener("abort", () => attemptAbort.abort(), {
+            once: true,
+          });
+        }
+      }
+
+      let errorFromCurrent: { error: unknown } | undefined = undefined; // set when onError fires
+      const providerId = currentModelClient.builtinProviderId;
+      if (providerId) {
+        llmErrorStore.clearModelError({
+          model: currentModelClient.model.modelId,
+          provider: providerId,
+        });
+      }
+      logger.info(
+        "Streaming text with model",
+        currentModelClient.model.modelId,
+        "provider",
+        currentModelClient.model.provider,
+        "builtinProviderId",
+        currentModelClient.builtinProviderId,
+      );
+      const { textStream } = streamText({
+        ...rest,
+        maxRetries: 0,
+        model: currentModelClient.model,
+        abortSignal: attemptAbort.signal,
+        onError: (error) => {
+          const providerId = currentModelClient.builtinProviderId;
+          if (providerId) {
+            llmErrorStore.recordModelError({
+              model: currentModelClient.model.modelId,
+              provider: providerId,
+            });
+          }
+          logger.error(
+            `Error streaming text with ${providerId} and model ${currentModelClient.model.modelId}: ${error}`,
+            error,
+          );
+          errorFromCurrent = error;
+          attemptAbort.abort(); // kill fetch / SSE
+        },
+      });
+
+      try {
+        for await (const chunk of textStream) {
+          /* If onError fired during streaming, bail out immediately. */
+          if (errorFromCurrent) throw errorFromCurrent;
+          yield chunk;
+        }
+
+        /* Stream ended – check if it actually failed */
+        if (errorFromCurrent) throw errorFromCurrent;
+
+        /* Completed successfully – stop trying more models. */
+        return;
+      } catch (err) {
+        if (typeof err === "object" && err !== null && "error" in err) {
+          lastErr = err as { error: unknown };
+        } else {
+          lastErr = { error: err };
+        }
+        logger.warn(
+          `[streamTextWithBackup] model #${i} failed – ${
+            i < modelClients.length - 1
+              ? "switching to backup"
+              : "no backups left"
+          }`,
+          err,
+        );
+        /* loop continues to next model (if any) */
+      }
+    }
+
+    /* Every model failed */
+    if (!lastErr) {
+      throw new Error("Invariant in StreamTextWithbackup failed!");
+    }
+    callerOnError?.(lastErr);
+    logger.error("All model invocations failed", lastErr);
+    // throw lastErr ?? new Error("All model invocations failed");
+  }
+
+  return { textStream: combinedGenerator() };
+}