[Feature Request] 支持在非官方 DeepSeek 风格 OpenAI 兼容网关上选择思考挡位 / Support reasoning-effort selection on non-official DeepSeek-style OpenAI-compatible gateways

[BrokenReflection]

需求背景 / Motivation

我们在一个非官方的 OpenAI 兼容网关上使用 DSH。该网关的模型遵循 DeepSeek
语义，并接受 reasoning_effort 挡位（off / high / max）。我们希望通过现有
配置面（模型条目上的 reasoningEfforts 声明、agent-default-model 中的
reasoningEffort 选择）在非官方网关上按模型、按会话选择思考挡位。目前这条
链路无法端到端工作——请求还没到达模型就失败了。

We run DSH against a non-official OpenAI-compatible gateway whose models
follow DeepSeek semantics and accept reasoning_effort levels (off / high /
max). We want to select the reasoning effort per model and per session, using
the existing configuration surface (reasoningEfforts on model entries and
reasoningEffort in agent-default-model). On such a gateway this currently
does not work end to end — every request fails before reaching the model.

llm-pi-ai:
  providers:
    internal:
      api: openai-completions
      baseURL: https://example.internal/v1   # DeepSeek-style gateway
      compat:
        thinkingFormat: deepseek
      models:
        - id: model-x
          reasoningEfforts:
            off:
            high: high
            max: max
agent-default-model:
  provider: internal
  model: model-x
  reasoningEffort: high

拦路问题 / Blocking issue

一旦模型声明思考挡位，pi-ai 就会用 OpenAI 的 developer 角色发送 system
prompt。DeepSeek 风格的端点只接受 system / user / assistant / tool
四种角色，因此网关对每个请求都返回 400 messages[0].role: unknown variant 'developer'。手工声明的路由目前无法保持 system 角色。

As soon as a model advertises reasoning efforts, pi-ai sends the system
prompt with the OpenAI developer role. DeepSeek-style endpoints only accept
system / user / assistant / tool, so the gateway rejects every request
with 400 messages[0].role: unknown variant 'developer'. Hand-declared routes
currently have no way to keep the system role.

建议方案 / Proposed fix

在 route 级和 model 级的 compat 配置中增加 supportsDeveloperRole（回退顺序：
目录条目 → baseURL 推断），并透传到 resolveModelCompat——与现有
supportsReasoningEffort 开关完全对称：

Add supportsDeveloperRole to the route- and model-level compat profile
(falling back to the installed catalog entry, then pi-ai's baseURL-derived
guess), carried through resolveModelCompat — symmetric with the existing
supportsReasoningEffort switch:

compat:
  thinkingFormat: deepseek
  supportsDeveloperRole: false   # keep role 'system' on DeepSeek-style endpoints

两个开关配合后，非官方 DeepSeek 风格网关上的思考挡位选择即可完整可用。
已有可运行的实现（约 16 行改动，补丁见附件），并已在生产环境的内网
DeepSeek 风格网关上验证三个挡位全部可用。

Together the two switches make reasoning-effort selection fully functional on
non-official DeepSeek-style gateways. A working implementation exists
(~16-line diff, patch attached) and has been verified in production against an
internal DeepSeek-style gateway with all three effort levels.
llm-pi-ai-supports-developer-role.patch

[BrokenReflection]

话说怎么不开放Issues啊，多少人等着提交呢