Feature request: end-to-end support for OpenAI Codex Standard/Fast service tiers

[oksure]

Feature request

Please consider adding first-class support for OpenAI Codex service tiers, including Standard and Fast, for the native Codex OAuth provider.

Motivation

Codex exposes a speed setting separate from reasoning effort. Fast mode is model- and account-dependent, increases response speed, and consumes usage at a higher rate. DSH currently exposes provider, model, and reasoning effort, but has no equivalent service-tier field that survives model selection, session resume, or the agent loop.

Scope

An end-to-end implementation would ideally cover:

• model metadata and capability discovery for supported service tiers
• session.selectModel and model-selection UI state
• persistence and replay in session headers
• propagation through agent/request and LLM call configuration
• OpenAI Codex Responses adapter request options
• provider-specific wire mapping, for example user-facing Fast to service_tier: "priority"
• tests that prevent silent fallback from Fast to Standard

Safety and compatibility requirements

• Standard should remain the default.
• Fast should only be selectable when the model catalog advertises support.
• A requested Fast tier should not silently downgrade to Standard.
• The active tier should be part of request identity when cached transports are reused.
• The implementation should avoid spoofing authentication or attestation metadata.
• Usage and response metadata should make the selected tier observable.

Would this be a feature the project would consider? I would be happy to help test the behavior against Codex OAuth accounts once the intended transport and session semantics are established.

[jghwwnq]

也希望支持这个功能，并覆盖 OpenAI API key 和原生 Codex OAuth 两种接入方式。希望在保持当前模型和推理强度的同时，显式选择 Fast / Priority，以缩短等待时间。

补充当前源码检查结果（2026-09-14，dsh 0.1.5-rc.2，commit c291e7961a）：

• PiAiProviderProfile 已有 reasoning、thinkingBudgets 等配置，但没有 serviceTier。
• profileOptions() 未传递 serviceTier；实际请求通过同文件中的 models.streamSimple() 发出。
• 本地安装的 @earendil-works/pi-ai@0.85.1 中，OpenAIResponsesOptions 和 OpenAICodexResponsesOptions 均已声明 serviceTier，各自的原生请求构造也会将其写入 service_tier。接入时还需确保该值能通过 dsh 现有的 streamSimple 调用链真正到达请求体。

预期可以独立配置 Standard / Fast：保持现有默认行为，在模型、账号和端点支持时由用户显式启用 Fast，并在界面提示额外额度消耗或费用。降低 reasoning 是另一项设置，不应作为 Fast 的替代。

建议覆盖 openai-responses 和 openai-codex-responses 的请求体验证：启用时发送 service_tier: "priority"，关闭后恢复原有行为，且不改变所选推理强度。如服务端返回实际采用的 tier，也希望能在请求详情中查看。

参考：OpenAI 官方 Fast mode 说明。以上为源码检查结果，未进行付费请求或速度实测。