- DEFAULT_CHAIN_ADMISSION.md: reviewed and approved, real artifact refs added - DEFAULT_DATA_IDEMPOTENT_RELEASE_GATE.md: reviewed and approved - scripts/setup_default_data.sh: idempotent init with --dry-run/--apply/artifact - scripts/test/test_default_data.sh: 4 test cases all pass - scripts/acceptance/verify_user_key_self_service.sh: Phase 0 skeleton - .gitignore: add generated artifact directories
200 lines
6.3 KiB
Markdown
200 lines
6.3 KiB
Markdown
# Host Protocol Matrix
|
||
|
||
日期:2026-06-04
|
||
状态:vNext.1 当前真相源(基于首轮 live probe)
|
||
适用范围:宿主协议能力判断、model pool 设计输入、默认链路准入前置判断
|
||
|
||
## 1. 目的
|
||
|
||
把 `scripts/acceptance/verify_host_protocol_matrix.sh` 的首轮 live probe 结果沉淀为可读结论,明确:
|
||
|
||
1. 哪些 provider/model 在“当前本机直连协议层”上已验证 `models/chat/responses`
|
||
2. 哪些结论只能说明 upstream protocol capability,不能外推为 host 或 user-key 已闭环
|
||
3. 哪些模型仍存在缺口,不能进入 vNext.1 已验证集合
|
||
|
||
本文件只陈述当前证据,不扩张为未验证能力。
|
||
|
||
## 2. 真相边界
|
||
|
||
当前 artifact 来自:
|
||
|
||
- `/home/long/project/sub2api-cn-relay-manager/artifacts/host-capability/20260604_212413/protocol-matrix-summary.json`
|
||
|
||
该 artifact 的证据层级是:
|
||
|
||
- 已验证:upstream 直连协议层
|
||
- 未验证:remote43 宿主入口 host probe
|
||
- 未验证:最终 user-key 对外调用层
|
||
|
||
因此,本文件中的 `supported-direct` 结论只能解释为:
|
||
|
||
- 当前本机对该 upstream 的 `models/chat/responses` 三端点探测成功
|
||
- 不能直接解释为:生产宿主一定支持、user-key 一定 200、可直接进入默认消费链路
|
||
|
||
## 3. 首轮 live probe 结果
|
||
|
||
证据文件:
|
||
|
||
- summary: `/home/long/project/sub2api-cn-relay-manager/artifacts/host-capability/20260604_212413/protocol-matrix-summary.json`
|
||
- per-target artifacts:
|
||
- `/home/long/project/sub2api-cn-relay-manager/artifacts/host-capability/20260604_212413/targets/01-deepseek-chat-official`
|
||
- `/home/long/project/sub2api-cn-relay-manager/artifacts/host-capability/20260604_212413/targets/02-kimi-a7m`
|
||
- `/home/long/project/sub2api-cn-relay-manager/artifacts/host-capability/20260604_212413/targets/03-minimax-m3-direct`
|
||
- `/home/long/project/sub2api-cn-relay-manager/artifacts/host-capability/20260604_212413/targets/04-openai-zhongzhuan`
|
||
|
||
### 3.1 DeepSeek Official
|
||
|
||
- provider_id: `deepseek-chat-official`
|
||
- base_url: `https://api.deepseek.com/v1`
|
||
- smoke_model: `deepseek-chat`
|
||
- 结果:
|
||
- `models_status = 200`
|
||
- `chat_status = 200`
|
||
- `responses_status = 200`
|
||
- `support_level = supported-direct`
|
||
- `models_has_smoke_model = false`
|
||
|
||
结论:
|
||
|
||
- DeepSeek 官方 upstream 在当前本机直连协议层上通过 `models/chat/responses`
|
||
- 但 `/v1/models` 返回集合中未直接暴露 `smoke_model=deepseek-chat`
|
||
- 因此后续 model pool / route mapping 必须显式区分:
|
||
- advertised model
|
||
- callable model
|
||
- 不能再假设 `/v1/models` 返回名与真实 callable model 恒等
|
||
|
||
### 3.2 Kimi A7M
|
||
|
||
- provider_id: `kimi-a7m`
|
||
- base_url: `https://kimi.a7m.com.cn/v1`
|
||
- smoke_model: `kimi-k2.6`
|
||
- 结果:
|
||
- `models_status = 200`
|
||
- `chat_status = 200`
|
||
- `responses_status = 200`
|
||
- `support_level = supported-direct`
|
||
|
||
结论:
|
||
|
||
- Kimi A7M 当前不是“协议天然不支持”的证据状态
|
||
- 这次 live probe 说明:在当前时间点、本机直连协议层上,Kimi A7M 的 `models/chat/responses` 全部可用
|
||
- 因此前面遇到的 Kimi 问题,不能再笼统归因为“协议不兼容”;更可能来自:
|
||
- 宿主出口路径
|
||
- 供应商运行状态波动
|
||
- 接入配置或中转层行为差异
|
||
|
||
### 3.3 MiniMax M3
|
||
|
||
- provider_id: `minimax-m3-direct`
|
||
- base_url: `https://mimimax.cn/v1`
|
||
- smoke_model: `MiniMax-M3`
|
||
- 结果:
|
||
- `models_status = 200`
|
||
- `chat_status = 200`
|
||
- `responses_status = 200`
|
||
- `support_level = supported-direct`
|
||
|
||
结论:
|
||
|
||
- MiniMax M3 在当前本机直连协议层上属于稳定的 `supported-direct`
|
||
- 可作为后续 host probe / user-key probe 的优先候选
|
||
|
||
### 3.4 OpenAI Zhongzhuan / asxs
|
||
|
||
- provider_id: `openai-zhongzhuan`
|
||
- base_url: `https://api.asxs.top/v1`
|
||
- smoke_model: `gpt-5.4`
|
||
- 结果:
|
||
- `models_status = 200`
|
||
- `chat_status = 200`
|
||
- `responses_status = 200`
|
||
- `support_level = supported-direct`
|
||
|
||
结论:
|
||
|
||
- asxs 在当前本机直连协议层上可用
|
||
- 但这不能替代“生产宿主出口可用”结论
|
||
- 结合历史记录,应继续区分:
|
||
- 本机 curl / 当前直连可用
|
||
- remote43 宿主出口可能仍受 Cloudflare 1010 或其他边界影响
|
||
|
||
## 4. 当前未纳入已验证集合的目标
|
||
|
||
### 4.1 GLM / 智谱
|
||
|
||
当前缺口:
|
||
|
||
- 缺少 `ZHIPU_API_KEY`
|
||
- 因此没有本轮 live probe artifact
|
||
|
||
当前结论:
|
||
|
||
- GLM 不能被写入“已验证协议矩阵”
|
||
- 后续若要进入 vNext.1 已验证集合,必须先补 key 并生成新的 live artifact
|
||
|
||
## 5. 对 vNext.1 的直接影响
|
||
|
||
### 5.1 对 model pool 设计的影响
|
||
|
||
已确认:
|
||
|
||
- `supported-direct` / `supported-with-plugin-adapter` / `unsupported-by-host` / `upstream-unhealthy` 四类 support level 是合理的最小分类
|
||
- DeepSeek 的 `models_has_smoke_model=false` 强制要求 model pool 区分:
|
||
- `public_model`
|
||
- `advertised_model`
|
||
- `callable_model`
|
||
|
||
### 5.2 对默认链路准入的影响
|
||
|
||
当前不能直接做的事:
|
||
|
||
- 不能仅凭本文件把这些模型直接宣告可进入默认消费链路
|
||
- 不能仅凭 `models/chat/responses=200` 宣告生产宿主闭环完成
|
||
|
||
仍需补的真实门槛:
|
||
|
||
1. host probe
|
||
2. user-key probe
|
||
3. 真实 user-key `chat/completions=200`
|
||
|
||
### 5.3 对 Kimi 结论口径的影响
|
||
|
||
当前推荐口径:
|
||
|
||
- 不再说“Kimi 协议不支持”
|
||
- 改为说:
|
||
- Kimi upstream 直连协议层当前已通过
|
||
- 宿主/用户面闭环仍待进一步 probe 分层验证
|
||
|
||
## 6. 当前未完成项
|
||
|
||
以下事项仍未闭环:
|
||
|
||
1. remote43 宿主入口 host probe 未形成独立 artifact
|
||
2. user-key probe 未形成独立 artifact
|
||
3. GLM 未探测
|
||
4. 当前矩阵脚本虽已补强,但仍不是 production-grade protocol matrix
|
||
|
||
## 7. 当前可执行结论
|
||
|
||
可确认:
|
||
|
||
- DeepSeek Official / Kimi A7M / MiniMax M3 / asxs 的 upstream 直连协议层,在本轮 live probe 中均为 `supported-direct`
|
||
- DeepSeek 存在 advertised/callable name 差异风险,必须进入 model pool 设计真相源
|
||
- Kimi 的历史问题不能继续被笼统归因为“协议不支持”
|
||
|
||
不可确认:
|
||
|
||
- 生产宿主 host 层是否对这些目标同样 `supported-direct`
|
||
- user-key 层是否同样 200
|
||
- 哪些目标已经满足默认消费链路准入
|
||
|
||
## 8. 后续动作
|
||
|
||
vNext.1 后续实施顺序保持不变:
|
||
|
||
1. 用本文件作为 model pool 设计输入
|
||
2. 继续 pool 到 priority failover 运行面的映射
|
||
3. 再补 host / user-key 在线真实验证
|
||
4. 最终再判断默认链路准入
|