Repository navigation
Conversation
TokenAccountingService 收敛 estimator、固定开销与最终预算调用点； 新增 MediaTokenEstimator 与 TokenCounterUtil 适配层，ContextTokenEstimator 的近似实现改由其驱动。
TokenAccountingService 新增 countBlocks/messageOverhead，TokenCounterUtil 暴露 calculateBlockToken/calculateMessageOverhead；prune 的 minimumTokens 只累计入选块的 token 成本并计一次消息开销，避免同消息内 excluded 大块 借出配额越过阈值；累计变量改为 long 避免超大输入整数回绕。
TokenSurface 增加 responseFormat 维度、TokenEstimate 增加独立计费桶， 默认 estimator 从 GenerateOptions 读取；fixed estimate 按其扣除一次会话 预算，final estimate 计入一次，trigger/keep 算法不变。补充计费桶单次 计入、大 schema 触发预算拒绝、schema 压缩会话预算三处回归。
新增 responseFormat 计费桶后保持无结构化输出 schema 场景的构造兼容。
core：Model 新增 getDefaultOptions()（默认 null），SessionModels.Recorded 包装与 ReActAgent 故障转移匿名模型透传，DashScope/OpenAI/Anthropic/Gemini 返回各自默认配置；harness：ContextTokenEstimator 按 GenerateOptions.mergeOptions(请求, 默认) 计价生效 schema，请求级优先。 补充默认 schema 等值计价、请求级覆盖与预算拒绝回归。
…ex r4） OpenAIResponsesChatModel 的 getDefaultOptions() 返回 configuredOptions； OllamaChatModel 将 schema format 反向映射为 ResponseFormat.jsonSchema 后 返回。harness 增加 ollama/openai-official test 依赖，新增 ProviderDefaultSchemaBudgetTest 走真实适配器验证默认 schema 等值计价、 请求级覆盖与超预算拒绝。
Model 新增 resolveOptions(requestOptions) 钩子（默认 getDefaultOptions 合并），ReActAgent 故障转移模型与 SessionModels.Recorded 透传；Ollama 覆写以镜像发送路径的 format 语义，修正 text 覆盖漏计与 body param 覆盖误计；estimator 改经钩子解析有效选项。补充真实适配器回归。
withSchemaProjection 保留转换后的全部映射字段，仅将对象形 format 投影为 计价的 response format；对象识别改为按发送侧 JSON 表示（Map 直用，其余经 codec 尝试绑定），非对象值不投影。补充 ObjectNode 计价与发送 JSON 回归。
HarnessContextBuilder.inputLimit 经 resolveOptions 并入模型有效默认 maxTokens/maxCompletionTokens；CompactionConfig 文档对齐边界准备器 （allowance 触发）与 standalone onReasoning 两条路径的预算口径。
四类媒体与嵌套工具媒体的触发估值和最终预算精确一致，媒体桶按块计一次。
anthropic/dashscope/gemini/openai 断言默认 response schema 与显式同价、 请求级覆盖默认，OpenAI 非流式线级验证默认 schema 到达 wire；输出预留与 窗口回退改为手算精确临界值。pom 增加四个 model extension test 依赖。
Codecov Report❌ Patch coverage is 📢 Thoughts on this report? Let us know! |
…t-token-accounting
…nting' into refactor/unify-context-token-accounting
…t-token-accounting
…t-token-accounting
dargoner
force-pushed
the
refactor/unify-context-token-accounting
branch
from
October 18, 2026 13:13
b248f4f to
f1ee22c
Compare
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
AgentScope-Java Version
2.1.0-SNAPSHOT (based on
main@ 5f7f8e6)Description
Background
Context-budget accounting was spread across several call sites with different conventions: the final-request budget ignored provider-level default output schemas (Ollama
format, OpenAI/OpenAI-Responsesresponse_format, and the Anthropic/DashScope/Gemini equivalents) that the adapters actually send on the wire; media charging was duplicated between estimators; and the pruning pass accumulated token counts inint, which can overflow on very large histories.Changes
TokenAccountingService,MediaTokenEstimator): a single place that prices a request (text + tools + media + output reservation + response-format schemas), reused by budget checks, compaction and tests.Model,SessionModels,ReActAgent,HarnessContextBuilder): models exposegetDefaultOptions()/resolveOptions(); the context builder budgets what the provider will actually send (explicit options merged over provider defaults). Default response schemas are now charged for OpenAI, OpenAI official (Responses), DashScope, Anthropic, Gemini and Ollama, withformatsemantics aligned to the provider's effective wire format (including JSON-representation detection for Ollama object schemas).ConversationCompactor,TokenCounterUtil): accumulation widened tolongand per-result accounting aligned with the retained/pruned split.Tests
ContextBudgetRegressionTest(9),ProviderDefaultSchemaBudgetTest(21),TokenAccountingServiceTest(10).HarnessContextBuilderTest8 → 12,ConversationCompactorTestextended.main: 6433 tests, 0 failures / 0 errors (43 environment skips).mvn spotless:checkpasses.Checklist
mvn spotless:applymvn test)