Make Claude Code effort level configurable without breaking subagents

[relus-cy]

Background

First, thanks for building ds-cli. It matches exactly the workflow we needed: keep Codex/Claude Code as the planner and delegate bounded implementation or review work to DeepSeek through a simple ds-cli run path. The ds-agent wrapper is also a good fit for keeping the main Codex session clean while the delegated run finishes in the background.

default_model, pro_model, and ds-cli run --pro already cover the model-selection side well. The missing piece I ran into is making Claude Code's effort level configurable without forcing it onto every internal Claude Code request.

Problem

DeepSeek's Claude Code integration docs recommend using:

CLAUDE_CODE_EFFORT_LEVEL=max

with deepseek-v4-pro[1m].

This works for the top-level Claude Code request. However, if users put that env var directly into backend_template.env or a backend's env, it is inherited by all Claude Code child requests as well.

In a local smoke test:

• ds-agent -> ds-cli run -> Claude Code/DeepSeek worked.
• The delegated run successfully used Read and Bash.
• But Claude Code's internal Agent / subagent tool failed before doing any work when CLAUDE_CODE_EFFORT_LEVEL=max was set:

API Error: 400 thinking options type cannot be disabled when reasoning_effort is set

The subagent result showed subagent_tokens: 0 and tool_uses: 0, so this looks like a request-parameter compatibility issue rather than a tool execution failure.

Proposal

Make effort level a first-class ds-cli option instead of asking users to hardcode CLAUDE_CODE_EFFORT_LEVEL in backend env.

Possible config shape:

default_model: deepseek-v4-flash
pro_model: deepseek-v4-pro[1m]

default_effort: null
pro_effort: max

# Optional, only if Claude Code provides a clean way to separate subagent requests.
subagent_model: deepseek-v4-flash
subagent_effort: null

Possible CLI:

ds-cli run --pro --effort max prompt.txt
ds-cli run --effort none prompt.txt

Suggested resolution order:

CLI --effort > backend effort override > pro_effort/default_effort > unset

When effort is null, none, or unset, ds-cli should not export CLAUDE_CODE_EFFORT_LEVEL.

Implementation Sketch

The current model resolution path already does most of the needed work:

• resolve default_model / pro_model
• pass the resolved model into set_backend_env()
• use {model}, {default_model}, and {pro_model} placeholders

The same pattern could be used for effort:

• add default_effort and pro_effort to config
• add optional per-backend effort
• add optional --effort to ds-cli run
• resolve effort next to model in commands/run.py
• pass resolved_effort into set_backend_env()
• only set CLAUDE_CODE_EFFORT_LEVEL when effort is explicitly configured

If Claude Code does not expose a way to keep internal Agent/subagent requests from inheriting this env var, it would still help to make effort opt-in per run and document that max may break nested Agent/subagent calls on DeepSeek's Anthropic-compatible endpoint.

[dazuiba]

我做了非常大的修改, 并顺带完美解决了您的问题.
请安装新版本, 并查看 https://github.com/dazuiba/handoff#configuration

[dazuiba]

//API Error: 400 thinking options type cannot be disabled when reasoning_effort is set

我读了您的发言, 我同时使用了 api.deepseek.com 和 opencode(我使用的是 https://github.com/iTzFaisal/oc-cc-proxy 做的转换) , 并没有 遇到您说的问题.

您尝试安装 handoff后, 如果依然有问题, 可以使用 --verbose命令, 尝试本地复现一下. 将截屏发给我.