[heydryft]

变更说明

Adds Runcrate as a built-in model provider template. Runcrate serves open-weight models (DeepSeek, Qwen, Kimi, GLM, gpt-oss and others) behind an OpenAI-compatible API at https://api.runcrate.ai/v1, so the template uses the default openai provider type, the same way opencode-go and modelscope do.

• 任务类型：feature
• 目标与非目标：goal is one more built-in OpenAI-compatible template that admins can enable and fill with RUNCRATE_API_KEY. Non-goals: no change to defaults, model selection, request handling or reasoning handling.
• substantial / trivial 判断：trivial. It is a data-only template entry plus an avatar fallback, so no decision record was added.

Files:

• backend/yuxi/modules/models/providers/builtin.py: new runcrate entry (base_url, api_key_env, models_endpoint), placed after opencode-go.
• frontend/src/modules/settings/model/modelIcon.js: runcrate uses the neutral default avatar, so the card does not fall back to the OpenAI logo through provider_type.

工程主张与 Owner

• 受影响的工程主张：after app init, a runcrate provider row exists, disabled, and an admin can enable it with a key and pick models from the live model list.
• Owner / commit point / 观察边界：BUILTIN_PROVIDERS owns the template; ensure_builtin_model_providers_in_db seeds it once and never overwrites an edited row; is_enabled stays False because only siliconflow-cn is enabled by default.
• 决策记录：none, data-only template (see above).

验证情况

The template resolves to the OpenAI-compatible path and lists models

• 失败面：wrong provider_type, wrong models endpoint.
• 语义 Owner：_normalize_payload, fetch_remote_models.
• 直接证据 / 命令：a local script (not committed) normalized the template and called fetch_remote_models on it: provider_type openai, 127 models returned, including deepseek-ai/DeepSeek-V4-Flash.
• 负向案例：test_builtin_provider_templates_default_to_openai_provider_type covers every builtin template, including this one.
• 结果：Passed

Chat, tool calls and streaming through load_chat_model

• 失败面：ChatCompletionsAdapter against the real endpoint.
• 语义 Owner：yuxi.modules.models.chat.load_chat_model.
• 直接证据 / 命令：the same script built a ModelInfo for runcrate:deepseek-ai/DeepSeek-V4-Flash with a real key and called load_chat_model: invoke returned ok with finish reason stop; bind_tools with a get_weather tool returned the call get_weather({"city": "Paris"}); stream returned 9 chunks joining to ok.
• 负向案例：不适用
• 结果：Passed

Repo checks

• 直接证据 / 命令：
  • cd backend && uv run pytest test/unit -m "not slow": 3622 passed, 22 failed, 1 skipped. The same 22 tests fail on main without this change on this macOS host (Docker compose slot probe, Cube protocol, office/OCR conversion, renameat2 in workdir copy); the failure lists are identical.
  • uv run pytest test/unit/services/test_model_provider_service.py test/unit/config/test_options.py test/unit/services/test_model_selectors.py -m "not slow": 103 passed.
  • uv run ruff check yuxi, uv run ruff format yuxi --check, uv run ruff check --select I yuxi: clean.
  • python3 scripts/ci/verify_engineering_contracts.py: passed. python3 -m unittest scripts.ci.test_verify_engineering_contracts: 56 tests OK.
  • cd frontend && npx eslint --max-warnings=0 src/modules/settings/model/modelIcon.js: clean. pnpm run test:unit: 1010 passed.
• 结果：Passed

简化 / 删除验收

不涉及

独立语义 Review

A fresh-context reviewer agent read AGENTS.md, backend/AGENTS.md and the full diff. It confirmed the entry shape, the disabled-by-default seeding, and that .env reaches api and worker without a Compose change. It flagged the OpenAI logo fallback (fixed by the modelIcon.js line) and asked for the full unit suite, engineering contract checks and import sort check (results above).

未验证范围与风险

• Not run: the integration and E2E suites, and the change inside the Docker stack (the template was exercised through the service functions directly).
• runcrate is not in the preserve_reasoning set in chat.py, so reasoning text from reasoning models is not carried across turns. Existing tests show a missing reasoning field does not break the tool flow, and the tool call above worked; I left the set unchanged to keep this PR data-only.

事故反馈

不涉及

界面变更

The model provider list shows one more card, "Runcrate", with the neutral default avatar.

关联事项

无

补充说明

No migration: the row is added by the existing builtin seeding on app init, disabled until an admin enables it.

This PR was prepared with AI assistance (omp with Claude): the template entry, the avatar line, the test runs and this description were written with it, and the live calls above were run against the real Runcrate API.