[Bug] 多个子任务结束后主任务返回 HTTP 400 / HTTP 400 after multi-agent reports

[evangelineak]

中文

问题

当主任务启动多个子任务（subagents）进行并行检查时，子任务完成并把报告返回给主任务后，主任务无法继续生成最终总结，反复出现：

DeepSeek API error (HTTP 400)
code: INVALID_REQUEST
status: 400

复现过程

1. 使用 Windows。
2. 使用 provider: deepseek-official。
3. 使用 model: deepseek-v4-flash。
4. 使用 standard agent preset。
5. 启动多个并行 subagents，读取本地项目文件并返回审查报告。
6. 子任务完成后，主任务继续整理报告。
7. 主任务在下一步生成回复前失败。
8. 系统自动尝试继续时，同样的 HTTP 400 错误又出现两次。

观察到的情况

• 主会话运行时间较长，包含大量工具调用和多个子任务报告。
• 失败前日志显示约 192,000 cached input tokens。
• 日志中的部分中文内容出现乱码。
• 子任务本身大多已经完成工作，主要失败发生在“把多个子任务报告汇总回主任务并生成最终回复”这一步。
• 目前没有上传原始 session 压缩包，因为其中包含本地项目路径和源代码片段。如需要，我可以在脱敏后提供。

期望结果

主任务应该能够：

• 正常汇总已完成的子任务报告；或
• 显示明确的上下文长度/请求大小错误，并提供恢复方式。

请问这是已知的上下文长度、消息拼接、缓存，或多代理报告汇总问题吗？

---

English

Issue

When the main task starts multiple subagents for parallel codebase reviews, the main task fails after the subagents finish and send their reports back. The main task cannot generate the final summary and repeatedly returns:

DeepSeek API error (HTTP 400)
code: INVALID_REQUEST
status: 400

Reproduction steps

1. Windows environment.
2. Provider: deepseek-official.
3. Model: deepseek-v4-flash.
4. Agent preset: standard.
5. Start multiple parallel subagents to inspect local project files and return audit reports.
6. Wait for the subagents to finish.
7. Let the main task continue processing their reports.
8. The main task fails before producing the next response.
9. Automatic continuation attempts produce the same HTTP 400 error twice more.

Observations

• The main session was long and contained many tool calls and reports from multiple subagents.
• The log shows approximately 192,000 cached input tokens immediately before the failure.
• Some Chinese text in the session log appears garbled.
• The subagents mostly completed their work successfully. The failure happened while aggregating their reports back into the main task and generating the final response.
• I have not uploaded the original session archive because it contains local project paths and source-code excerpts. I can provide a sanitized archive if needed.

Expected behavior

The main task should either:

• summarize the completed subagent reports successfully; or
• show a clear context-size/request-size error with recovery guidance.

Could this be a known issue related to context length, message assembly, caching, or multi-agent report aggregation?
dsh-session-sanitized-89d64ca4.zip