[GH-ISSUE #29] 为什么接入claudecode subagent需要必须指定为flash模型? #10

Open
opened 2026-10-26 09:43:29 +00:00 by bsakaguchi · 4 comments
Owner

Originally created by @EwdAger on GitHub (Apr 30, 2026).
Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/29

看到配置中指定了 export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash。想知道为什么这样配置,并且我发现好像不指定 subagent 也会走到 flash 模型,在 subagent 需要处理复杂任务时如何规避?

Originally created by @EwdAger on GitHub (Apr 30, 2026). Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/29 看到配置中指定了 `export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash`。想知道为什么这样配置,并且我发现好像不指定 subagent 也会走到 flash 模型,在 subagent 需要处理复杂任务时如何规避?
Author
Owner

@DillionApple commented on GitHub (Apr 30, 2026):

您可以修改配置,将 CLAUDE_CODE_SUBAGENT_MODEL 设置为 deepseek-v4-pro

<!-- gh-comment-id:4349148252 --> @DillionApple commented on GitHub (Apr 30, 2026): 您可以修改配置,将 CLAUDE_CODE_SUBAGENT_MODEL 设置为 deepseek-v4-pro
Author
Owner

@Agents365-ai commented on GitHub (May 6, 2026):

使用flash会更好,这是特意设置的,subagent不会干太重的或不需要用到pro,我的猜想

<!-- gh-comment-id:4386922504 --> @Agents365-ai commented on GitHub (May 6, 2026): 使用flash会更好,这是特意设置的,subagent不会干太重的或不需要用到pro,我的猜想
Author
Owner

@xuechaow commented on GitHub (May 10, 2026):

It used to be Haiku which is the rather light weight agent. The cost efficient model is anyway main model with powerful model to produce executable tasks and verifications and let light model to execute. Unless cost is not a concern to you.

<!-- gh-comment-id:4416069160 --> @xuechaow commented on GitHub (May 10, 2026): It used to be Haiku which is the rather light weight agent. The cost efficient model is anyway main model with powerful model to produce executable tasks and verifications and let light model to execute. Unless cost is not a concern to you.
Author
Owner

@tianhao8687 commented on GitHub (Aug 17, 2026):

CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash 是当前官方 Claude Code 接入指南里的默认路由选择,主要是成本与延迟取舍,不是协议上的强制要求。

不设置后仍可能落到 Flash,是因为 DeepSeek 的 Anthropic 模型映射会把 claude-haiku* 和 claude-sonnet* 映射到 Flash,只有 claude-opus* 映射到 Pro。

如果希望所有 subagent 都明确使用 Pro,可以在启动 Claude Code 前设置:

export CLAUDE_CODE_SUBAGENT_MODEL='deepseek-v4-pro[1m]'

Windows PowerShell:

$env:CLAUDE_CODE_SUBAGENT_MODEL='deepseek-v4-pro[1m]'

然后完全重启 Claude Code。若只想让少数复杂 subagent 用 Pro,优先在自定义 subagent 的模型配置中单独指定;全局环境变量会影响所有 subagent,成本和延迟都会上升。还要检查 settings.json 与 shell 环境是否同时存在旧值,避免一个覆盖另一个。

<!-- gh-comment-id:5317548362 --> @tianhao8687 commented on GitHub (Aug 17, 2026): `CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash` 是当前[官方 Claude Code 接入指南](https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/claude_code)里的默认路由选择,主要是成本与延迟取舍,不是协议上的强制要求。 不设置后仍可能落到 Flash,是因为 DeepSeek 的 Anthropic 模型映射会把 `claude-haiku*` 和 `claude-sonnet*` 映射到 Flash,只有 `claude-opus*` 映射到 Pro。 如果希望所有 subagent 都明确使用 Pro,可以在启动 Claude Code 前设置: ```bash export CLAUDE_CODE_SUBAGENT_MODEL='deepseek-v4-pro[1m]' ``` Windows PowerShell: ```powershell $env:CLAUDE_CODE_SUBAGENT_MODEL='deepseek-v4-pro[1m]' ``` 然后完全重启 Claude Code。若只想让少数复杂 subagent 用 Pro,优先在自定义 subagent 的模型配置中单独指定;全局环境变量会影响所有 subagent,成本和延迟都会上升。还要检查 `settings.json` 与 shell 环境是否同时存在旧值,避免一个覆盖另一个。
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/awesome-deepseek-agent#10
No description provided.