refactor(harness): unify context token accounting and charge provider default output schemas

agentscope-core/src/main/java/io/agentscope/core/ReActAgent.java

@@ -5028,6 +5028,16 @@ public boolean supportsNativeStructuredOutput() {
             public int getContextWindowSize() {
                 return activeModel.get().getContextWindowSize();
             }
+
+            @Override
+            public GenerateOptions getDefaultOptions() {
+                return activeModel.get().getDefaultOptions();
+            }
+
+            @Override
+            public GenerateOptions resolveOptions(GenerateOptions requestOptions) {
+                return activeModel.get().resolveOptions(requestOptions);
+            }
         };
     }
 

agentscope-core/src/main/java/io/agentscope/core/model/Model.java

@@ -101,4 +101,30 @@ default boolean supportsToolChoiceSpecific() {
     default int getContextWindowSize() {
         return 0;
     }
+
+    /**
+     * Returns the generation options this model applies when a request does not override them, or
+     * {@code null} if unknown. Used by {@link #resolveOptions} as the provider-agnostic baseline.
+     *
+     * @return default generation options, or {@code null} when the model has none
+     */
+    default GenerateOptions getDefaultOptions() {
+        return null;
+    }
+
+    /**
+     * Resolves the generation options a request will effectively be sent with, applying the
+     * model's defaults the same way the provider does before issuing the call. The harness-layer
+     * budget check consults this so provider-specific mappings (for example a native structured
+     * output field) cannot make the effective request diverge from what was accounted.
+     *
+     * @param requestOptions request-level options, possibly {@code null}
+     * @return effective options for the request
+     */
+    default GenerateOptions resolveOptions(GenerateOptions requestOptions) {
+        GenerateOptions defaults = getDefaultOptions();
+        return defaults == null
+                ? requestOptions
+                : GenerateOptions.mergeOptions(requestOptions, defaults);
+    }
 }

agentscope-core/src/main/java/io/agentscope/core/session/SessionModels.java

@@ -59,6 +59,14 @@ public int getContextWindowSize() {
             return delegate.getContextWindowSize();
         }
 
+        public GenerateOptions getDefaultOptions() {
+            return delegate.getDefaultOptions();
+        }
+
+        public GenerateOptions resolveOptions(GenerateOptions requestOptions) {
+            return delegate.resolveOptions(requestOptions);
+        }
+
         public Flux<ChatResponse> stream(
                 List<Msg> messages, List<ToolSchema> tools, GenerateOptions options) {
             return Flux.defer(

agentscope-extensions/agentscope-extensions-model/agentscope-extensions-model-anthropic/src/main/java/io/agentscope/extensions/model/anthropic/AnthropicChatModel.java

@@ -387,6 +387,11 @@ public String getModelName() {
         return modelName;
     }
 
+    @Override
+    public GenerateOptions getDefaultOptions() {
+        return defaultOptions;
+    }
+
     /**
      * Creates a builder for constructing AnthropicChatModel instances.
      *

agentscope-extensions/agentscope-extensions-model/agentscope-extensions-model-dashscope/src/main/java/io/agentscope/extensions/model/dashscope/DashScopeChatModel.java

@@ -434,6 +434,11 @@ public String getModelName() {
         return modelName;
     }
 
+    @Override
+    public GenerateOptions getDefaultOptions() {
+        return defaultOptions;
+    }
+
     @Override
     public boolean supportsNativeStructuredOutput() {
         if (Boolean.TRUE.equals(enableThinking)) {

agentscope-extensions/agentscope-extensions-model/agentscope-extensions-model-gemini/src/main/java/io/agentscope/extensions/model/gemini/GeminiChatModel.java

@@ -394,6 +394,11 @@ public String getModelName() {
         return modelName;
     }
 
+    @Override
+    public GenerateOptions getDefaultOptions() {
+        return defaultOptions;
+    }
+
     /**
      * Close the Gemini client.
      */

agentscope-extensions/agentscope-extensions-model/agentscope-extensions-model-ollama/src/main/java/io/agentscope/extensions/model/ollama/OllamaChatModel.java

@@ -15,7 +15,10 @@
  */
 package io.agentscope.extensions.model.ollama;
 
+import com.fasterxml.jackson.core.type.TypeReference;
 import io.agentscope.core.formatter.Formatter;
+import io.agentscope.core.formatter.JsonSchema;
+import io.agentscope.core.formatter.ResponseFormat;
 import io.agentscope.core.message.Msg;
 import io.agentscope.core.model.ChatModelBase;
 import io.agentscope.core.model.ChatResponse;
@@ -29,6 +32,8 @@
 import io.agentscope.core.model.transport.HttpTransportFactory;
 import io.agentscope.core.model.transport.OkHttpTransport;
 import io.agentscope.core.model.transport.ProxyConfig;
+import io.agentscope.core.util.JsonException;
+import io.agentscope.core.util.JsonUtils;
 import io.agentscope.extensions.model.ollama.dto.OllamaMessage;
 import io.agentscope.extensions.model.ollama.dto.OllamaRequest;
 import io.agentscope.extensions.model.ollama.dto.OllamaResponse;
@@ -37,6 +42,7 @@
 import io.agentscope.extensions.model.ollama.options.OllamaOptions;
 import java.time.Instant;
 import java.util.List;
+import java.util.Map;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 import reactor.core.publisher.Flux;
@@ -135,6 +141,64 @@ public String getModelName() {
         return this.modelName;
     }
 
+    @Override
+    public GenerateOptions getDefaultOptions() {
+        return withSchemaProjection(defaultOptions.toGenerateOptions(), defaultOptions.getFormat());
+    }
+
+    @Override
+    public GenerateOptions resolveOptions(GenerateOptions requestOptions) {
+        OllamaOptions effective =
+                requestOptions == null ? defaultOptions : defaultOptions.merge(requestOptions);
+        return withSchemaProjection(effective.toGenerateOptions(), effective.getFormat());
+    }
+
+    /**
+     * Mirrors the send path's effective {@code format} semantics: the request only clears the
+     * default schema when it maps to a non-null format (json_schema, json_object, or a raw
+     * {@code format} body parameter), exactly as {@code fromGenerateOptions} and
+     * {@code additionalBodyParams} behave before the default merge. The converted options keep
+     * every mapped field; an object-shaped format is additionally projected as the response
+     * format the budget prices.
+     */
+    private static GenerateOptions withSchemaProjection(GenerateOptions converted, Object format) {
+        ResponseFormat schemaFormat = objectSchemaFormat(format);
+        if (schemaFormat == null) {
+            return converted;
+        }
+        return GenerateOptions.mergeOptions(
+                GenerateOptions.builder().responseFormat(schemaFormat).build(), converted);
+    }
+
+    /**
+     * Recognizes a raw {@code format} whose JSON representation is an object, i.e. the schema the
+     * provider actually sends. Values like {@code "json"}, arrays or scalars carry no schema and
+     * are left to the send path untouched.
+     */
+    private static ResponseFormat objectSchemaFormat(Object format) {
+        Map<String, Object> schema = toSchemaMap(format);
+        return schema == null
+                ? null
+                : ResponseFormat.jsonSchema(JsonSchema.builder().schema(schema).build());
+    }
+
+    private static Map<String, Object> toSchemaMap(Object format) {
+        if (format instanceof Map<?, ?> schemaMap) {
+            @SuppressWarnings("unchecked")
+            Map<String, Object> schema = (Map<String, Object>) schemaMap;
+            return schema;
+        }
+        if (format == null) {
+            return null;
+        }
+        try {
+            return JsonUtils.getJsonCodec()
+                    .convertValue(format, new TypeReference<Map<String, Object>>() {});
+        } catch (JsonException e) {
+            return null;
+        }
+    }
+
     @Override
     public boolean supportsToolChoiceSpecific() {
         return true;

agentscope-extensions/agentscope-extensions-model/agentscope-extensions-model-openai-official/src/main/java/io/agentscope/extensions/model/openaiofficial/OpenAIResponsesChatModel.java

@@ -101,6 +101,11 @@ public String getModelName() {
         return configuredOptions != null ? configuredOptions.getModelName() : null;
     }
 
+    @Override
+    public GenerateOptions getDefaultOptions() {
+        return configuredOptions;
+    }
+
     @Override
     public boolean supportsToolChoiceSpecific() {
         return true;

agentscope-extensions/agentscope-extensions-model/agentscope-extensions-model-openai/src/main/java/io/agentscope/extensions/model/openai/OpenAIChatModel.java

@@ -233,6 +233,11 @@ public String getModelName() {
         return configuredOptions != null ? configuredOptions.getModelName() : null;
     }
 
+    @Override
+    public GenerateOptions getDefaultOptions() {
+        return configuredOptions;
+    }
+
     /** Package-private test seam for verifying model-level strict JSON schema configuration. */
     Boolean getStrictJsonSchema() {
         return strictJsonSchema;

agentscope-harness/pom.xml

@@ -52,6 +52,43 @@
             <scope>test</scope>
         </dependency>
 
+        <!-- Price provider-level default response schemas through the real adapter classes. -->
+        <dependency>
+            <groupId>io.agentscope</groupId>
+            <artifactId>agentscope-extensions-model-ollama</artifactId>
+            <scope>test</scope>
+        </dependency>
+
+        <dependency>
+            <groupId>io.agentscope</groupId>
+            <artifactId>agentscope-extensions-model-openai-official</artifactId>
+            <scope>test</scope>
+        </dependency>
+
+        <dependency>
+            <groupId>io.agentscope</groupId>
+            <artifactId>agentscope-extensions-model-openai</artifactId>
+            <scope>test</scope>
+        </dependency>
+
+        <dependency>
+            <groupId>io.agentscope</groupId>
+            <artifactId>agentscope-extensions-model-dashscope</artifactId>
+            <scope>test</scope>
+        </dependency>
+
+        <dependency>
+            <groupId>io.agentscope</groupId>
+            <artifactId>agentscope-extensions-model-anthropic</artifactId>
+            <scope>test</scope>
+        </dependency>
+
+        <dependency>
+            <groupId>io.agentscope</groupId>
+            <artifactId>agentscope-extensions-model-gemini</artifactId>
+            <scope>test</scope>
+        </dependency>
+
         <!-- YAML parsing for subagent config -->
         <dependency>
             <groupId>com.fasterxml.jackson.dataformat</groupId>

agentscope-harness/src/main/java/io/agentscope/harness/agent/context/ContextTokenEstimator.java

@@ -15,13 +15,8 @@
  */
 package io.agentscope.harness.agent.context;
 
-import io.agentscope.core.message.ContentBlock;
-import io.agentscope.core.message.TextBlock;
-import io.agentscope.core.message.ThinkingBlock;
-import io.agentscope.core.message.ToolResultBlock;
-import io.agentscope.core.message.ToolUseBlock;
 import io.agentscope.core.middleware.ModelCallInput;
-import io.agentscope.core.util.JsonUtils;
+import io.agentscope.core.model.GenerateOptions;
 
 /** Replace with an adapter-specific tokenizer when available. Estimates are not provider usage. */
 @FunctionalInterface
@@ -36,39 +31,19 @@ record Estimate(long tokens, String method, boolean exact) {
     }
 
     static ContextTokenEstimator approximate() {
+        TokenAccountingService service = TokenAccountingService.defaults();
         return input -> {
-            long count = 0;
-            for (var msg : input.messages()) {
-                count += 12 + text(msg.getName());
-                for (var block : msg.getContent()) count += block(block);
-            }
-            for (var tool : input.tools()) {
-                count += 16 + text(JsonUtils.getJsonCodec().toJson(tool));
-            }
-            return new Estimate(count, "mixed-text-2.5chars-media-2048-v1", false);
+            GenerateOptions options =
+                    input.model() == null
+                            ? input.options()
+                            : input.model().resolveOptions(input.options());
+            var estimate =
+                    service.estimate(
+                            new TokenAccountingService.TokenSurface(
+                                    input.messages(),
+                                    input.tools(),
+                                    options == null ? null : options.getResponseFormat()));
+            return new Estimate(estimate.totalTokens(), estimate.method(), estimate.exact());
         };
     }
-
-    private static long text(String value) {
-        return value == null ? 0 : (long) Math.ceil(value.length() / 2.5);
-    }
-
-    private static long block(ContentBlock block) {
-        if (block instanceof TextBlock text) return text(text.getText());
-        if (block instanceof ThinkingBlock thinking) return text(thinking.getThinking());
-        if (block instanceof ToolUseBlock use) {
-            return 12
-                    + text(use.getId())
-                    + text(use.getName())
-                    + text(JsonUtils.getJsonCodec().toJson(use.getInput()))
-                    + text(use.getContent());
-        }
-        if (block instanceof ToolResultBlock result) {
-            long tokens = 12 + text(result.getId()) + text(result.getName());
-            for (var output : result.getOutput()) tokens += block(output);
-            return tokens;
-        }
-        // Explicitly approximate: media cost depends on provider, resolution and duration.
-        return 2048;
-    }
 }

agentscope-harness/src/main/java/io/agentscope/harness/agent/context/HarnessContextBuilder.java

@@ -27,6 +27,7 @@
 import io.agentscope.core.middleware.ModelRequestPreparer;
 import io.agentscope.core.middleware.ReasoningInput;
 import io.agentscope.core.middleware.TaskContextProjection;
+import io.agentscope.core.model.GenerateOptions;
 import io.agentscope.core.state.AgentState;
 import io.agentscope.core.state.TaskContextState;
 import io.agentscope.core.util.JsonUtils;
@@ -548,11 +549,15 @@ private long inputLimit(ModelCallInput input) {
         long window = input.model().getContextWindowSize();
         long output = policy.reservedOutputTokens();
         if (output == 0) output = window > 0 ? Math.min(4096, Math.max(1, window / 8)) : 4096;
-        if (input.options() != null) {
-            if (input.options().getMaxTokens() != null)
-                output = Math.max(output, input.options().getMaxTokens());
-            if (input.options().getMaxCompletionTokens() != null)
-                output = Math.max(output, input.options().getMaxCompletionTokens());
+        // The provider applies the model's default options to the outgoing request, so the
+        // output reservation must be read from the same effective options: a default
+        // maxTokens that never appears in request options still consumes budget.
+        GenerateOptions effective = input.model().resolveOptions(input.options());
+        if (effective != null) {
+            if (effective.getMaxTokens() != null)
+                output = Math.max(output, effective.getMaxTokens());
+            if (effective.getMaxCompletionTokens() != null)
+                output = Math.max(output, effective.getMaxCompletionTokens());
         }
         long safety =
                 policy.safetyMarginTokens() > 0