fix: max output tokens when using large thinking budget (anomalyco#2976)

rekram1-node · web-flow · commit 71a7e8ef3668 · 2025-10-04T23:38:41.000-05:00
diff --git a/packages/opencode/src/provider/transform.ts b/packages/opencode/src/provider/transform.ts
@@ -105,18 +105,29 @@ export namespace ProviderTransform {
     return result
   }
 
-  export function maxOutputTokens(providerID: string, outputLimit: number, options: Record<string, any>): number {
+  export function maxOutputTokens(
+    providerID: string,
+    options: Record<string, any>,
+    modelLimit: number,
+    globalLimit: number,
+  ): number {
+    const modelCap = modelLimit || globalLimit
+    const standardLimit = Math.min(modelCap, globalLimit)
+
     if (providerID === "anthropic") {
-      const thinking = options["thinking"]
-      if (typeof thinking === "object" && thinking !== null) {
-        const type = thinking["type"]
-        const budgetTokens = thinking["budgetTokens"]
-        if (type === "enabled" && typeof budgetTokens === "number" && budgetTokens > 0) {
-          return outputLimit - budgetTokens
+      const thinking = options?.["thinking"]
+      const budgetTokens = typeof thinking?.["budgetTokens"] === "number" ? thinking["budgetTokens"] : 0
+      const enabled = thinking?.["type"] === "enabled"
+      if (enabled && budgetTokens > 0) {
+        // Return text tokens so that text + thinking <= model cap, preferring 32k text when possible.
+        if (budgetTokens + standardLimit <= modelCap) {
+          return standardLimit
         }
+        return modelCap - budgetTokens
       }
     }
-    return outputLimit
+
+    return standardLimit
   }
 
   export function schema(_providerID: string, _modelID: string, schema: JSONSchema.BaseSchema) {
diff --git a/packages/opencode/src/session/prompt.ts b/packages/opencode/src/session/prompt.ts
@@ -159,7 +159,7 @@ export namespace SessionPrompt {
       agent,
       model: input.model,
     }).then((x) => Provider.getModel(x.providerID, x.modelID))
-    const outputLimit = Math.min(model.info.limit.output, OUTPUT_TOKEN_MAX) || OUTPUT_TOKEN_MAX
+
     using abort = lock(input.sessionID)
 
     const system = await resolveSystemPrompt({
@@ -266,7 +266,12 @@ export namespace SessionPrompt {
             : undefined,
         maxRetries: 10,
         activeTools: Object.keys(tools).filter((x) => x !== "invalid"),
-        maxOutputTokens: ProviderTransform.maxOutputTokens(model.providerID, outputLimit, params.options),
+        maxOutputTokens: ProviderTransform.maxOutputTokens(
+          model.providerID,
+          params.options,
+          model.info.limit.output,
+          OUTPUT_TOKEN_MAX,
+        ),
         abortSignal: abort.signal,
         providerOptions: {
           [model.npm === "@ai-sdk/openai" ? "openai" : model.providerID]: params.options,
