[yoruuuchan]

Summary

Add an experimental Huawei Cloud CodeArts Agent (码道) provider.

• In-process OAuth PKCE adapter; short-lived STS credentials; direct OpenAI-compatible chat/SSE to snap-access.cn-north-4.myhuaweicloud.com.
• No CodeArts IDE/CLI per request.
• Catalog merges agent-center, builtin, and benefit sources; native model_id is kept separate from display model_name.
• Vision models receive image content parts (supports_images from the builtin catalog); text-only models reject images with 400.
• Reasoning levels come from /v1/model/thinking-variants; the selected max level is forwarded as reasoning_effort instead of being dropped to the upstream default.
• Quota meters and daily/new-user credit check-in.
• Registered as provider=codearts, region cn-north-4; console login and model settings updated.

Live acceptance

Single account:

• OAuth login -> ready; 10 models called through Chat Completions.
• Chat Completions, Anthropic Messages, and Responses: non-stream, stream, tools, reasoning, usage.
• Image input verified on all three APIs with Qwen3-VL-235B; image tokens reflected in usage.
• Reasoning options {high, max} exposed for GLM-5.2; console selection max persists as reasoning_effort=max and the chat request succeeds.
• Quota read and check-in verified.

Two accounts:

• Both ready; sticky session on account 1.
• Disable account 1 -> same session fails over to account 2.
• Re-enable account 1 -> session stays on account 2.
• Disable account 2 -> session fails back to account 1.

Validation

• go test ./... green on Linux (golang:1.25 container).
• go vet, gofmt, git diff --check, changelog validation.
• npm run lint and npm run sync with the locked frontend toolchain.
• GitHub CI: go, windows-host, worker, frontend, container, secrets all pass.

Known limitations

• Quota-exhaustion failover and normal load distribution were not live-tested.
• Benefit-success path was not live-tested; both accounts are credit-only trials. Credit fallback is verified.
• Image URLs/dimensions are subject to upstream behavior; data-URL images are verified.
• Unusable-model learning is in-memory with a 30-minute TTL.

Experimental provider; not claiming full production acceptance.

[yoruuuchan]

The previous head failed the frontend check because the embedded assets were generated locally with Vite 8.3.5 while \rontend/package-lock.json\ pins 8.2.2. \�c97b3b\ regenerates \internal/webui/static\ with a clean Linux
pm ci\ (Vite 8.2.2). The new head currently has no workflow run on the fork branch; could a maintainer approve the workflow run? Thanks.

[caigee-cmd]

感谢提交码道适配器。进程内实现、凭据刷新，以及原生 model id 和展示名分开处理，这些都没问题，CI 也是绿的。现有账号不会受影响。

合并前有几处控制面的小缺口，都不影响其它供应商：

1. codearts/ 模型前缀还没接到请求路由上。internal/executor/prepare.go 里的 providerPrefixes 仍停在 kimi/，所以 codearts/<model> 钉不住这个供应商。跨供应商池关闭时，这个 id 会落到 Qoder；池打开时，前缀会被忽略。README 和控制台提示已经写了这个前缀，请把它加进这张表。

2. 调用方的 system prompt 默认仍会被丢掉。控制台给码道传的是 drop_system_prompt: false，但 accounts.DefaultDropSystemPrompt 只对 trae 和 kimi 返回 false。导入或通过 API 创建、又没带这个字段时，system prompt 会被剥掉。请把 codearts 加进这个 switch。

3. 控制台里保存的 reasoning 档位刷新后不会回来。internal/console/models.go 的 modelSettingResponse 只给 trae、workbuddy、kimi 返回 reasoning_effort，码道会落到通用响应里，已保存的档位刷新后就没了。请按 kimi 的方式处理（只返回 reasoning，不带 max-mode）。

另外两处顺手即可：internal/app/architecture_test.go 的 isConcreteProvider 还没列入 codearts，分层检查覆盖不到新适配器；catalog.go 里 toModelInfo 的注释仍写着不声明图片，和现在的转发实现不一致。

额度耗尽切换还没实测，作为实验性供应商可以接受。上面三条补上之后就可以合。

[yoruuuchan]

已按 review 补齐，当前 head 9b4d43c：

1. internal/executor/prepare.go 的 providerPrefixes 加入 codearts/，并补了 ProviderPrefix / ResolveProviderFilter 的测试；codearts/<model> 现在能正常钉住供应商。
2. accounts.DefaultDropSystemPrompt 的 switch 加入 codearts，API/导入创建且未显式传字段时默认保留 caller system prompt；测试同步覆盖。
3. console.modelSettingResponse 把 codearts 按 Kimi 处理，只返回 reasoning_effort，不返回 max_mode；补了 PATCH/GET 响应测试。

顺手项也处理了：

• internal/app/architecture_test.go 的 isConcreteProvider 加入 codearts；
• internal/providers/codearts/catalog.go 的 toModelInfo 注释改为图片/思考档位按目录声明转发。

CI 六项全绿（go / windows-host / worker / frontend / container / secrets）。额度耗尽切换仍如你所说，保持实验性未实测。可以继续看是否还有需要补的。