子代理被迫全用 Sol
Subagents were stuck on Sol
要点
- GPT-5.6 Sol 开子代理时，探索和联网检索也被迫用 Sol，额度掉得很快
- 作者想要 Fable 那种顾问执行分流，当时 CLI 做不到
- 临时在 config.toml 打开 multi_agent_v2 并设置 tool_namespace 为 agents
- 相关讨论在 openai/codex 的 31814 号 issue
- 测基准前先确认子代理真能换更便宜模型，否则数字会被额度 bug 污染
原帖开头
If you're wondering why GPT-5.6 Sol with subagents in the Codex CLI burns through your usage so fast, there seems to be a bug. 👀 It can't spawn smaller subagents for basic exploration or web research and forces Sol for everything. Which is annoying because as with Fable, I'd lo …
