indusagi 0.13.10 → 0.13.11

package/dist/agent.js

@@ -4694,6 +4694,33 @@ var MODELS = {
       contextWindow: 1e6,
       maxTokens: 128e3
     },
+    "gpt-6-astra": {
+      id: "gpt-6-astra",
+      name: "GPT-6 Astra",
+      api: "openai-responses",
+      provider: "openai",
+      baseUrl: "https://api.openai.com/v1",
+      reasoning: true,
+      reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
+      input: ["text", "image"],
+      cost: {
+        input: 10,
+        output: 50,
+        cacheRead: 1,
+        cacheWrite: 12.5
+      },
+      longContextPricing: {
+        threshold: 272e3,
+        multipliers: {
+          input: 2,
+          output: 1.5,
+          cacheRead: 2,
+          cacheWrite: 2
+        }
+      },
+      contextWindow: 105e4,
+      maxTokens: 128e3
+    },
     "gpt-5.6-sol": {
       id: "gpt-5.6-sol",
       name: "GPT-5.6 Sol",
@@ -5156,6 +5183,28 @@ var MODELS = {
       },
       contextWindow: 105e4,
       maxTokens: 128e3
+    },
+    "gpt-6-astra": {
+      id: "gpt-6-astra",
+      name: "GPT-6 Astra",
+      api: "openai-codex-responses",
+      provider: "openai-codex",
+      baseUrl: "https://chatgpt.com/backend-api",
+      reasoning: true,
+      reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
+      input: ["text", "image"],
+      cost: {
+        input: 10,
+        output: 50,
+        cacheRead: 1,
+        cacheWrite: 12.5
+      },
+      longContextPricing: {
+        threshold: 272e3,
+        multipliers: { input: 2, output: 1.5, cacheRead: 2, cacheWrite: 2 }
+      },
+      contextWindow: 105e4,
+      maxTokens: 128e3
     }
   },
   "opencode": {
@@ -12668,9 +12717,9 @@ function getModels(provider) {
 function calculateCost(model, usage) {
   return modelRegistry.calculateCost(model, usage);
 }
-var XHIGH_MODEL_IDS = /* @__PURE__ */ new Set(["gpt-5.1-codex-max", "gpt-5.2", "gpt-5.2-codex"]);
+var XHIGH_MODEL_IDS = /* @__PURE__ */ new Set(["gpt-5.1-codex-max", "gpt-5.2", "gpt-5.2-codex", "gpt-6-astra"]);
 function supportsXhigh(model) {
-  return XHIGH_MODEL_IDS.has(model.id);
+  return model.reasoningEfforts?.includes("xhigh") ?? XHIGH_MODEL_IDS.has(model.id);
 }
 
 // src/facade/ml/types.ts
@@ -13358,12 +13407,12 @@ function buildBaseOptions(model, options, apiKey) {
   return base;
 }
 function clampReasoning(effort) {
-  if (effort === "xhigh") return "high";
+  if (effort === "xhigh" || effort === "max") return "high";
   return effort;
 }
 function mapThinkingLevel(level, provider = "clamp-xhigh") {
   if (!level) return void 0;
-  if (provider === "supports-xhigh") return level;
+  if (provider === "supports-xhigh") return level === "max" ? "high" : level;
   return clampReasoning(level);
 }
 var RESERVED_ANSWER_TOKENS = 1024;
@@ -13375,8 +13424,10 @@ var REASONING_BUDGET_DEFAULTS = {
   // 2Ki
   medium: BUDGET_UNIT * 8,
   // 8Ki
-  high: BUDGET_UNIT * 16
+  high: BUDGET_UNIT * 16,
   // 16Ki
+  max: BUDGET_UNIT * 16
+  // clamped to high for generic token-budget providers
 };
 function adjustMaxTokensForThinking(baseMaxTokens, modelMaxTokens, reasoningLevel, customBudgets) {
   const budgets = { ...REASONING_BUDGET_DEFAULTS, ...customBudgets };
@@ -16575,7 +16626,7 @@ var OpenAIResponsesParamBuilder = class {
     if (this.options?.maxTokens) {
       params.max_output_tokens = this.options.maxTokens;
     }
-    if (this.options?.temperature !== void 0) {
+    if (this.model.id !== "gpt-6-astra" && this.options?.temperature !== void 0) {
       params.temperature = this.options.temperature;
     }
     if (this.options?.serviceTier !== void 0) {
@@ -16629,7 +16680,11 @@ var streamSimpleOpenAIResponses = (model, context, options) => {
     throw new Error(`No API key for provider: ${model.provider}`);
   }
   const base = buildBaseOptions(model, options, apiKey);
-  const reasoningEffort = mapThinkingLevel(options?.reasoning, supportsXhigh(model) ? "supports-xhigh" : "clamp-xhigh");
+  const requestedEffort = options?.reasoning;
+  const reasoningEffort = model.reasoningEfforts ? requestedEffort !== void 0 && requestedEffort !== "minimal" && model.reasoningEfforts.includes(requestedEffort) ? requestedEffort : void 0 : mapThinkingLevel(
+    requestedEffort === "max" ? "high" : requestedEffort,
+    supportsXhigh(model) ? "supports-xhigh" : "clamp-xhigh"
+  );
   return streamOpenAIResponses(model, context, {
     ...base,
     reasoningEffort
@@ -16664,8 +16719,8 @@ if (typeof process !== "undefined" && (process.versions?.node || process.version
 }
 var CODEX_URL = "https://chatgpt.com/backend-api/codex/responses";
 var JWT_CLAIM_PATH = "https://api.openai.com/auth";
-var MAX_RETRIES = 3;
-var BASE_DELAY_MS = 1e3;
+var MAX_RETRIES = 4;
+var RETRY_DELAY_MS = 1e4;
 var CODEX_TOOL_CALL_PROVIDERS = /* @__PURE__ */ new Set(["openai", "openai-codex", "opencode"]);
 var CODEX_RESPONSE_STATUSES = /* @__PURE__ */ new Set([
   "completed",
@@ -16679,21 +16734,29 @@ var TRANSIENT_ERROR_PATTERNS = [
   /rate.?limit/i,
   /overloaded/i,
   /service.?unavailable/i,
+  /server.?error/i,
+  /terminated/i,
   /upstream.?connect/i,
   /connection.?refused/i
 ];
 var RETRYABLE_STATUS_CODES = /* @__PURE__ */ new Set([429, 500, 502, 503, 504]);
 var CodexRetryExecutor = class _CodexRetryExecutor {
+  static terminalErrorMarker = /* @__PURE__ */ Symbol("codexTerminalError");
   static isRetryable(status, errorText) {
     if (RETRYABLE_STATUS_CODES.has(status)) {
       return true;
     }
     return TRANSIENT_ERROR_PATTERNS.some((pattern) => pattern.test(errorText));
   }
-  // Exponential backoff where every retry doubles the delay. A left shift is
-  // used because the exponent is always a small non-negative whole number.
-  static backoffDelay(attempt) {
-    return BASE_DELAY_MS * (1 << attempt);
+  static backoffDelay(_attempt) {
+    return RETRY_DELAY_MS;
+  }
+  static markTerminal(error) {
+    error[_CodexRetryExecutor.terminalErrorMarker] = true;
+    return error;
+  }
+  static isTerminal(error) {
+    return Boolean(error[_CodexRetryExecutor.terminalErrorMarker]);
   }
   static sleep(ms, signal) {
     return new Promise((resolve5, reject) => {
@@ -16729,12 +16792,15 @@ var CodexRetryExecutor = class _CodexRetryExecutor {
           statusText: response.statusText
         });
         const info = await CodexErrorInterpreter.parse(fakeResponse);
-        throw new Error(info.friendlyMessage || info.message);
+        throw _CodexRetryExecutor.markTerminal(new Error(info.friendlyMessage || info.message));
       } catch (error) {
         if (error instanceof Error && (error.name === "AbortError" || error.message === "Request was aborted")) {
           throw new Error("Request was aborted");
         }
         lastError = error instanceof Error ? error : new Error(String(error));
+        if (_CodexRetryExecutor.isTerminal(lastError)) {
+          throw lastError;
+        }
         if (attempt < MAX_RETRIES && !lastError.message.includes("usage limit")) {
           await _CodexRetryExecutor.sleep(_CodexRetryExecutor.backoffDelay(attempt), options?.signal);
           continue;
@@ -16802,20 +16868,45 @@ var streamOpenAICodexResponses = (model, context, options) => {
       const requestFactory = new CodexRequestFactory(model, context, options, apiKey);
       const { body, bodyJson, headers } = requestFactory.buildPayload();
       options?.onPayload?.(body);
-      const response = await executeWithCodexRetry(
-        () => fetch(CODEX_URL, {
-          method: "POST",
-          headers,
-          body: bodyJson,
-          signal: options?.signal
-        }),
-        options?.signal
-      );
-      if (!response.body) {
-        throw new Error("No response body");
+      let completedOutput;
+      let completedEvents;
+      for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
+        const response = await executeWithCodexRetry(
+          () => fetch(CODEX_URL, {
+            method: "POST",
+            headers,
+            body: bodyJson,
+            signal: options?.signal
+          }),
+          options?.signal
+        );
+        if (!response.body) {
+          throw new Error("No response body");
+        }
+        const attemptOutput = createAssistantMessageOutput(model);
+        const attemptStream = new AssistantMessageEventStream();
+        try {
+          await processStream(response, attemptOutput, attemptStream, model);
+          completedOutput = attemptOutput;
+          completedEvents = [...attemptStream.getHistory()];
+          break;
+        } catch (error) {
+          if (options?.signal?.aborted) {
+            throw new Error("Request was aborted");
+          }
+          const message = error instanceof Error ? error.message : JSON.stringify(error);
+          if (attempt >= MAX_RETRIES || !CodexRetryExecutor.isRetryable(200, message)) {
+            throw error;
+          }
+          await CodexRetryExecutor.sleep(CodexRetryExecutor.backoffDelay(attempt), options?.signal);
+        }
+      }
+      if (!completedOutput || !completedEvents) {
+        throw new Error("Failed after retries");
       }
+      Object.assign(output, completedOutput);
       events.start();
-      await processStream(response, output, stream, model);
+      for (const event of completedEvents) stream.push(event);
       if (options?.signal?.aborted) {
         throw new Error("Request was aborted");
       }
@@ -16838,7 +16929,8 @@ var streamSimpleOpenAICodexResponses = (model, context, options) => {
     throw new Error(`No API key for provider: ${model.provider}`);
   }
   const base = buildBaseOptions(model, options, apiKey);
-  const reasoningEffort = mapThinkingLevel(options?.reasoning, supportsXhigh(model) ? "supports-xhigh" : "clamp-xhigh");
+  const mappedEffort = options?.reasoning === "max" && model.reasoningEfforts?.includes("max") ? "max" : mapThinkingLevel(options?.reasoning, supportsXhigh(model) ? "supports-xhigh" : "clamp-xhigh");
+  const reasoningEffort = mappedEffort && model.reasoningEfforts && (mappedEffort === "minimal" || !model.reasoningEfforts.includes(mappedEffort)) ? void 0 : mappedEffort;
   return streamOpenAICodexResponses(model, context, {
     ...base,
     reasoningEffort
@@ -16860,7 +16952,10 @@ function buildRequestBody(model, context, options) {
     tool_choice: "auto",
     parallel_tool_calls: true
   };
-  if (options?.temperature !== void 0) {
+  if (options?.maxTokens !== void 0) {
+    body.max_output_tokens = options.maxTokens;
+  }
+  if (options?.temperature !== void 0 && model.id !== "gpt-6-astra") {
     body.temperature = options.temperature;
   }
   if (context.tools) {
@@ -17517,7 +17612,8 @@ var CLAUDE_THINKING_BUDGETS = {
   low: 1500,
   medium: 6e3,
   high: 2e4,
-  xhigh: 2e4
+  xhigh: 2e4,
+  max: 2e4
 };
 function resolveThinkingPayload(model, options) {
   if (!options.reasoning || !model.reasoning) {

package/dist/ai.js

@@ -4818,6 +4818,33 @@ var MODELS = {
       contextWindow: 1e6,
       maxTokens: 128e3
     },
+    "gpt-6-astra": {
+      id: "gpt-6-astra",
+      name: "GPT-6 Astra",
+      api: "openai-responses",
+      provider: "openai",
+      baseUrl: "https://api.openai.com/v1",
+      reasoning: true,
+      reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
+      input: ["text", "image"],
+      cost: {
+        input: 10,
+        output: 50,
+        cacheRead: 1,
+        cacheWrite: 12.5
+      },
+      longContextPricing: {
+        threshold: 272e3,
+        multipliers: {
+          input: 2,
+          output: 1.5,
+          cacheRead: 2,
+          cacheWrite: 2
+        }
+      },
+      contextWindow: 105e4,
+      maxTokens: 128e3
+    },
     "gpt-5.6-sol": {
       id: "gpt-5.6-sol",
       name: "GPT-5.6 Sol",
@@ -5280,6 +5307,28 @@ var MODELS = {
       },
       contextWindow: 105e4,
       maxTokens: 128e3
+    },
+    "gpt-6-astra": {
+      id: "gpt-6-astra",
+      name: "GPT-6 Astra",
+      api: "openai-codex-responses",
+      provider: "openai-codex",
+      baseUrl: "https://chatgpt.com/backend-api",
+      reasoning: true,
+      reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
+      input: ["text", "image"],
+      cost: {
+        input: 10,
+        output: 50,
+        cacheRead: 1,
+        cacheWrite: 12.5
+      },
+      longContextPricing: {
+        threshold: 272e3,
+        multipliers: { input: 2, output: 1.5, cacheRead: 2, cacheWrite: 2 }
+      },
+      contextWindow: 105e4,
+      maxTokens: 128e3
     }
   },
   "opencode": {
@@ -12810,9 +12859,9 @@ function loadCustomModels(models) {
 function calculateCost(model, usage) {
   return modelRegistry.calculateCost(model, usage);
 }
-var XHIGH_MODEL_IDS = /* @__PURE__ */ new Set(["gpt-5.1-codex-max", "gpt-5.2", "gpt-5.2-codex"]);
+var XHIGH_MODEL_IDS = /* @__PURE__ */ new Set(["gpt-5.1-codex-max", "gpt-5.2", "gpt-5.2-codex", "gpt-6-astra"]);
 function supportsXhigh(model) {
-  return XHIGH_MODEL_IDS.has(model.id);
+  return model.reasoningEfforts?.includes("xhigh") ?? XHIGH_MODEL_IDS.has(model.id);
 }
 function modelsAreEqual(a, b) {
   if (!a || !b) return false;
@@ -13551,12 +13600,12 @@ function buildBaseOptions(model, options, apiKey) {
   return base;
 }
 function clampReasoning(effort) {
-  if (effort === "xhigh") return "high";
+  if (effort === "xhigh" || effort === "max") return "high";
   return effort;
 }
 function mapThinkingLevel(level, provider = "clamp-xhigh") {
   if (!level) return void 0;
-  if (provider === "supports-xhigh") return level;
+  if (provider === "supports-xhigh") return level === "max" ? "high" : level;
   return clampReasoning(level);
 }
 var RESERVED_ANSWER_TOKENS = 1024;
@@ -13568,8 +13617,10 @@ var REASONING_BUDGET_DEFAULTS = {
   // 2Ki
   medium: BUDGET_UNIT * 8,
   // 8Ki
-  high: BUDGET_UNIT * 16
+  high: BUDGET_UNIT * 16,
   // 16Ki
+  max: BUDGET_UNIT * 16
+  // clamped to high for generic token-budget providers
 };
 function adjustMaxTokensForThinking(baseMaxTokens, modelMaxTokens, reasoningLevel, customBudgets) {
   const budgets = { ...REASONING_BUDGET_DEFAULTS, ...customBudgets };
@@ -16788,7 +16839,7 @@ var OpenAIResponsesParamBuilder = class {
     if (this.options?.maxTokens) {
       params.max_output_tokens = this.options.maxTokens;
     }
-    if (this.options?.temperature !== void 0) {
+    if (this.model.id !== "gpt-6-astra" && this.options?.temperature !== void 0) {
       params.temperature = this.options.temperature;
     }
     if (this.options?.serviceTier !== void 0) {
@@ -16842,7 +16893,11 @@ var streamSimpleOpenAIResponses = (model, context, options) => {
     throw new Error(`No API key for provider: ${model.provider}`);
   }
   const base = buildBaseOptions(model, options, apiKey);
-  const reasoningEffort = mapThinkingLevel(options?.reasoning, supportsXhigh(model) ? "supports-xhigh" : "clamp-xhigh");
+  const requestedEffort = options?.reasoning;
+  const reasoningEffort = model.reasoningEfforts ? requestedEffort !== void 0 && requestedEffort !== "minimal" && model.reasoningEfforts.includes(requestedEffort) ? requestedEffort : void 0 : mapThinkingLevel(
+    requestedEffort === "max" ? "high" : requestedEffort,
+    supportsXhigh(model) ? "supports-xhigh" : "clamp-xhigh"
+  );
   return streamOpenAIResponses(model, context, {
     ...base,
     reasoningEffort
@@ -16877,8 +16932,8 @@ if (typeof process !== "undefined" && (process.versions?.node || process.version
 }
 var CODEX_URL = "https://chatgpt.com/backend-api/codex/responses";
 var JWT_CLAIM_PATH = "https://api.openai.com/auth";
-var MAX_RETRIES = 3;
-var BASE_DELAY_MS = 1e3;
+var MAX_RETRIES = 4;
+var RETRY_DELAY_MS = 1e4;
 var CODEX_TOOL_CALL_PROVIDERS = /* @__PURE__ */ new Set(["openai", "openai-codex", "opencode"]);
 var CODEX_RESPONSE_STATUSES = /* @__PURE__ */ new Set([
   "completed",
@@ -16892,21 +16947,29 @@ var TRANSIENT_ERROR_PATTERNS = [
   /rate.?limit/i,
   /overloaded/i,
   /service.?unavailable/i,
+  /server.?error/i,
+  /terminated/i,
   /upstream.?connect/i,
   /connection.?refused/i
 ];
 var RETRYABLE_STATUS_CODES = /* @__PURE__ */ new Set([429, 500, 502, 503, 504]);
 var CodexRetryExecutor = class _CodexRetryExecutor {
+  static terminalErrorMarker = /* @__PURE__ */ Symbol("codexTerminalError");
   static isRetryable(status, errorText) {
     if (RETRYABLE_STATUS_CODES.has(status)) {
       return true;
     }
     return TRANSIENT_ERROR_PATTERNS.some((pattern) => pattern.test(errorText));
   }
-  // Exponential backoff where every retry doubles the delay. A left shift is
-  // used because the exponent is always a small non-negative whole number.
-  static backoffDelay(attempt) {
-    return BASE_DELAY_MS * (1 << attempt);
+  static backoffDelay(_attempt) {
+    return RETRY_DELAY_MS;
+  }
+  static markTerminal(error) {
+    error[_CodexRetryExecutor.terminalErrorMarker] = true;
+    return error;
+  }
+  static isTerminal(error) {
+    return Boolean(error[_CodexRetryExecutor.terminalErrorMarker]);
   }
   static sleep(ms, signal) {
     return new Promise((resolve, reject) => {
@@ -16942,12 +17005,15 @@ var CodexRetryExecutor = class _CodexRetryExecutor {
           statusText: response.statusText
         });
         const info = await CodexErrorInterpreter.parse(fakeResponse);
-        throw new Error(info.friendlyMessage || info.message);
+        throw _CodexRetryExecutor.markTerminal(new Error(info.friendlyMessage || info.message));
       } catch (error) {
         if (error instanceof Error && (error.name === "AbortError" || error.message === "Request was aborted")) {
           throw new Error("Request was aborted");
         }
         lastError = error instanceof Error ? error : new Error(String(error));
+        if (_CodexRetryExecutor.isTerminal(lastError)) {
+          throw lastError;
+        }
         if (attempt < MAX_RETRIES && !lastError.message.includes("usage limit")) {
           await _CodexRetryExecutor.sleep(_CodexRetryExecutor.backoffDelay(attempt), options?.signal);
           continue;
@@ -17015,20 +17081,45 @@ var streamOpenAICodexResponses = (model, context, options) => {
       const requestFactory = new CodexRequestFactory(model, context, options, apiKey);
       const { body, bodyJson, headers } = requestFactory.buildPayload();
       options?.onPayload?.(body);
-      const response = await executeWithCodexRetry(
-        () => fetch(CODEX_URL, {
-          method: "POST",
-          headers,
-          body: bodyJson,
-          signal: options?.signal
-        }),
-        options?.signal
-      );
-      if (!response.body) {
-        throw new Error("No response body");
+      let completedOutput;
+      let completedEvents;
+      for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
+        const response = await executeWithCodexRetry(
+          () => fetch(CODEX_URL, {
+            method: "POST",
+            headers,
+            body: bodyJson,
+            signal: options?.signal
+          }),
+          options?.signal
+        );
+        if (!response.body) {
+          throw new Error("No response body");
+        }
+        const attemptOutput = createAssistantMessageOutput(model);
+        const attemptStream = new AssistantMessageEventStream();
+        try {
+          await processStream(response, attemptOutput, attemptStream, model);
+          completedOutput = attemptOutput;
+          completedEvents = [...attemptStream.getHistory()];
+          break;
+        } catch (error) {
+          if (options?.signal?.aborted) {
+            throw new Error("Request was aborted");
+          }
+          const message = error instanceof Error ? error.message : JSON.stringify(error);
+          if (attempt >= MAX_RETRIES || !CodexRetryExecutor.isRetryable(200, message)) {
+            throw error;
+          }
+          await CodexRetryExecutor.sleep(CodexRetryExecutor.backoffDelay(attempt), options?.signal);
+        }
+      }
+      if (!completedOutput || !completedEvents) {
+        throw new Error("Failed after retries");
       }
+      Object.assign(output, completedOutput);
       events.start();
-      await processStream(response, output, stream2, model);
+      for (const event of completedEvents) stream2.push(event);
       if (options?.signal?.aborted) {
         throw new Error("Request was aborted");
       }
@@ -17051,7 +17142,8 @@ var streamSimpleOpenAICodexResponses = (model, context, options) => {
     throw new Error(`No API key for provider: ${model.provider}`);
   }
   const base = buildBaseOptions(model, options, apiKey);
-  const reasoningEffort = mapThinkingLevel(options?.reasoning, supportsXhigh(model) ? "supports-xhigh" : "clamp-xhigh");
+  const mappedEffort = options?.reasoning === "max" && model.reasoningEfforts?.includes("max") ? "max" : mapThinkingLevel(options?.reasoning, supportsXhigh(model) ? "supports-xhigh" : "clamp-xhigh");
+  const reasoningEffort = mappedEffort && model.reasoningEfforts && (mappedEffort === "minimal" || !model.reasoningEfforts.includes(mappedEffort)) ? void 0 : mappedEffort;
   return streamOpenAICodexResponses(model, context, {
     ...base,
     reasoningEffort
@@ -17073,7 +17165,10 @@ function buildRequestBody(model, context, options) {
     tool_choice: "auto",
     parallel_tool_calls: true
   };
-  if (options?.temperature !== void 0) {
+  if (options?.maxTokens !== void 0) {
+    body.max_output_tokens = options.maxTokens;
+  }
+  if (options?.temperature !== void 0 && model.id !== "gpt-6-astra") {
     body.temperature = options.temperature;
   }
   if (context.tools) {
@@ -17730,7 +17825,8 @@ var CLAUDE_THINKING_BUDGETS = {
   low: 1500,
   medium: 6e3,
   high: 2e4,
-  xhigh: 2e4
+  xhigh: 2e4,
+  max: 2e4
 };
 function resolveThinkingPayload(model, options) {
   if (!options.reasoning || !model.reasoning) {

package/dist/cli.js

@@ -410,6 +410,28 @@ var MODEL_CARDS = [
       cacheReadPerMTok: 0.275
     }
   },
+  {
+    id: "gpt-6-astra",
+    provider: "openai",
+    api: "openai-responses",
+    displayName: "GPT-6 Astra",
+    baseUrl: "https://api.openai.com/v1",
+    contextWindow: 105e4,
+    maxOutputTokens: 128e3,
+    modalities: ["text", "image"],
+    reasoning: true,
+    reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
+    longContextPricing: {
+      threshold: 272e3,
+      multipliers: { input: 2, output: 1.5, cacheRead: 2, cacheWrite: 2 }
+    },
+    cost: {
+      inputPerMTok: 10,
+      outputPerMTok: 50,
+      cacheReadPerMTok: 1,
+      cacheWritePerMTok: 12.5
+    }
+  },
   // ── Google (Gemini) ─────────────────────────────────────────────────────
   {
     id: "gemini-2.5-pro",
@@ -1113,6 +1135,7 @@ function anthropicThinkingBudget(level, maxOutput) {
     low: 0.2,
     medium: 0.4,
     high: 0.6,
+    xhigh: 0.8,
     max: 0.8
   };
   if (level === "off") {
@@ -1917,13 +1940,17 @@ function readNumber2(o, key) {
   const v = o[key];
   return typeof v === "number" && Number.isFinite(v) ? v : void 0;
 }
-function reasoningEffort(level) {
+function reasoningEffort(model, level) {
+  if (model.reasoningEfforts?.includes(level)) {
+    return level;
+  }
   switch (level) {
     case "low":
       return "low";
     case "medium":
       return "medium";
     case "high":
+    case "xhigh":
     case "max":
       return "high";
   }
@@ -1964,14 +1991,17 @@ function buildBody2(model, c, opts) {
   if (ceiling > 0) {
     body.max_output_tokens = ceiling;
   }
-  if (opts.temperature !== void 0) {
+  if (model.id !== "gpt-6-astra" && opts.temperature !== void 0) {
     body.temperature = opts.temperature;
   }
-  if (opts.topP !== void 0) {
+  if (model.id !== "gpt-6-astra" && opts.topP !== void 0) {
     body.top_p = opts.topP;
   }
   if (model.reasoning && opts.thinking !== void 0 && opts.thinking !== "off") {
-    body.reasoning = { effort: reasoningEffort(opts.thinking) };
+    const effort = reasoningEffort(model, opts.thinking);
+    if (effort !== void 0) {
+      body.reasoning = { effort };
+    }
   }
   return body;
 }
@@ -2246,6 +2276,7 @@ function geminiThinkingBudget(level, maxOutput) {
     low: 0.2,
     medium: 0.4,
     high: 0.6,
+    xhigh: 0.8,
     max: 0.8
   };
   const budget = Math.floor(maxOutput * fraction[level]);
@@ -2539,6 +2570,7 @@ function thinkingBudget(level, maxOutput) {
     low: 0.2,
     medium: 0.4,
     high: 0.6,
+    xhigh: 0.8,
     max: 0.8
   };
   return Math.max(0, Math.floor(maxOutput * fraction[level]));