实测：Codex55不降智省Token的配置

本人声明：以下为纯手写，非ai生成。

研究了一下config.toml，总结了一套配置，实测比不配置少用三分之一token，大家都是怎么配置的呢？

核心思路：

model_reasoning_effort = “high”：保留强推理，不牺牲代码质量。

model_verbosity = “low”：减少啰嗦输出，让 Token 花在刀刃上。

关闭记忆相关项：减少任务外的状态消耗和不确定性。

完整配置如下：

model_provider = "codex"

model = "你的模型名称"

approvals_reviewer = "user"

disable_response_storage = true

model_reasoning_effort = "high"

model_verbosity = "low"

network_access = "enabled"

preferred_auth_method = "apikey"

sandbox_mode = "danger-full-access"

windows_wsl_setup_acknowledged = true



[model_providers.codex]

name = "codex"

base_url = "你的API端点名称"

wire_api = "responses"

preferred_auth_method = "apikey"

requires_openai_auth = true



[experimental]

use_freeform_apply_patch = true

use_unified_exec_tool = true



[features]

apply_patch_freeform = true

experimental_windows_sandbox = true

memories = false

parallel = true

plan_tool = true

rmcp_client = true

streamable_shell = false

unified_exec = false

view_image_tool = true



[memories]

generate_memories = false

use_memories = false



[notice]

hide_full_access_warning = true



[profiles.auto]

approval_policy = "never"



[sandbox_workspace_write]

network_access = true