阶跃星辰 Harness v1.0.9

阶跃星辰 Harness v1.0.9

让执行过程更早可见，也让长会话和常用操作更轻快。

• 更早出现具体开场说明：支持的 OpenAI-compatible 推理模型先用当前模型生成简短的目标和第一步，再按原思考档位执行任务。开场请求最多一次，有长度与时间限制；关闭钩子或明确只要结果时维持原流程。
• 准确的思考时长：正文或工具参数生成不再延长已经结束的思考计时。
• 更稳的工具调用：兼容流式工具参数之后重复发送完整快照的端点，同时保留参数完整性校验。
• 更流畅的长会话：优化历史分支遍历、快照复用、统计与索引读取，减少重复扫描和内存占用。
• 更轻快的文件和插件操作：优化附件分段读取、PDF 缓冲、文件搜索和插件目录检查。
• 更可靠的输入队列：恢复未提交的队列内容，减少检查点重复比较。

开场优化会增加一次受限的短请求，服务响应时间仍会波动。有限实测中，Step 订阅 SVG 小任务首句约 6.5 秒，腾讯 WorkBuddy DeepSeek 文件读取任务约 3.3 秒，均完整成功；这不是固定速度保证。

Windows 用户下载 StepFun-Harness-1.0.9-win-x64.exe 安装，已有用户也可在软件内检查更新。底座随本项目统一适配发布。

---

English

Earlier, task-specific progress and smoother everyday operations.

• Earlier opening explanations: supported OpenAI-compatible reasoning models use the selected model for one short, bounded opening request before continuing with the original task reasoning settings. Disabling the hook or requesting only the result preserves the existing flow.
• Accurate thinking duration: subsequent text and tool argument generation no longer extends a completed reasoning block.
• More reliable tool calls: handle repeated, identical complete argument snapshots without weakening argument validation.
• Smoother long conversations: improve history traversal, snapshot reuse, statistics and indexing, reducing repeated reads and memory retention.
• Faster file and plugin operations: improve attachment range reads, PDF buffers, file search and plugin directory checks.
• More reliable input queues: recover unsent entries and avoid redundant checkpoint comparisons.

The opening mechanism adds one bounded short request; latency varies by service. In small live tests, a Step subscription SVG task displayed its opening in about 6.5 seconds and a Tencent WorkBuddy DeepSeek file-reading task in about 3.3 seconds, with both tasks completing successfully. These are observations, not latency guarantees.

Download StepFun-Harness-1.0.9-win-x64.exe on Windows, or check for updates in the app. The underlying runtime is delivered through coordinated Harness updates.