v4.4.0

更新内容

新的多接口处理架构 提升体验稳定性 继续尽可能对齐原生效果 建议体验此版本

流式性能优化

• TTFT 大幅降低：全新累积缓冲区架构，MinChars=16、MaxWait=10ms、首片即时刷新，显著减少首字延迟
• SSE 解析器升级：从 bufio.Scanner 迁移到 bufio.Reader，消除边界截断 bug，增量文本通过 toolstream 管道分发
• 心跳保活：长请求空闲时发送空 completion chunk + keep-alive heartbeat，防止客户端超时断流
• Assistant turn 语义重构：将 assistant turn 语义和流累积抽成独立包（assistantturn / completionruntime），统一 Claude/Gemini/OpenAI 三协议对接体验

Gemini Thinking 支持

• Gemini API 完整支持 thinking blocks，支持的模型默认启用
• thinking 内容在流式响应中正确归位，reasoning_content 不再泄漏

DSML 工具调用解析加固

• 新增短横线变体（<dsml-tool-calls>）和下划线变体（<dsml_tool_calls>）支持
• CDATA 内 markdown 围栏误判防护，避免代码块内标签被误解析
• CDATA 行偏移精确追踪，保证紧凑工具调用内容不丢失
• Go 端与 Node 端解析器行为对齐

遗留代码清理

• 移除 legacy compatibility 配置项及对应 UI 组件（58 文件，净删 255 行）
• 移除 legacy history split 配置，整合进 completion runtime 管道
• Reference marker 剥离逻辑集中到 textclean 包，消除 4 个协议处理器的重复硬编码

修复

• Thinking-only 空输出自动重试，使用多轮 follow-up 在同一 DeepSeek session 中重生成
• 修复 continuation replay 文本重叠，trim 冗余 thinking 和 response 流内容
• 新增 OpenAI 上传文件元数据检索接口

文档

• 架构文档补充 artifacts/static/ 目录及 7 个辅助包的说明
• 新增 DS2API 项目价值说明文档

---

What's Changed

Streaming Performance

• TTFT optimization: New accumulation buffer architecture with select loop, MinChars=16, MaxWait=10ms, first-flush-immediate
• SSE parser upgrade: Migrated from bufio.Scanner to bufio.Reader for robust stream parsing
• Keep-alive heartbeat: Empty completion chunks prevent client timeout during idle periods
• Assistant turn refactor: Centralized turn semantics and stream accumulation into assistantturn and completionruntime packages

Gemini Thinking

• Full thinking block support for Gemini API, enabled by default for supported models
• Fixed reasoning_content leak and finish_reason=null edge cases

DSML Parser Hardening

• Hyphenated (<dsml-tool-calls>) and underscore (<dsml_tool_calls>) tag variants
• CDATA markdown fence resilience — structural markers inside code blocks no longer cause false parses
• Precise line-offset tracking for compact tool call content preservation
• Node + Go parser alignment with comprehensive test coverage

Legacy Cleanup

• Removed legacy compatibility config and UI components (58 files, -255 net lines)
• Removed legacy history split config, integrated into completion runtime pipeline
• Centralized reference marker stripping into textclean package

Fixes

• Thinking-only empty output auto-retry with multi-turn regeneration
• Continuation replay overlap prevention
• OpenAI file metadata retrieval endpoint
• Tail content truncation fix when max_tokens exceeded

Docs

• Architecture docs: added artifacts/, static/ directories and 7 auxiliary package descriptions
• New DS2API project value note

---

Full Changelog: v4.3.0...v4.4.0

• 修复了使用Zeabur部署会失败的问题 by @VanceHud in #403
• feat(openai): retrieve uploaded file metadata by @NgoQuocViet2001 in #404
• perf(streaming): optimize TTFT and reduce buffering latency by @wyv202011y in #406
• v4.4.0: streaming perf + Gemini thinking + DSML parser hardening by @CJackHwang in #408

New Contributors

• @VanceHud made their first contribution in #403

Full Changelog: v4.3.0...v4.4.0