[cloudgpuapp]

Summary

Adds a cloudgpu text LLM preset to Drama Studio, defined the same way as the zhipu preset (kind: openai, so it goes through the existing OpenAI-compatible /chat/completions path).

CloudGPU (cloudgpu.app) is a model API for DeepSeek, Kimi, GLM, MiniMax, Qwen and more, with an OpenAI-compatible base URL.

Disclosure: I work on CloudGPU.

Related issue

None.

Changes

• studio/server/providers/llm_catalog.py: preset cloudgpu

CLOUDGPU_API_KEY, CLOUDGPU_BASE_URL (default https://cloudgpu.app/v1), CLOUDGPU_MODEL (default deepseek-v4-flash)

• a short models list for the Model settings dropdown (deepseek-v4-flash, deepseek-v4-pro, glm-5.3, kimi-k3, minimax-m3, qwen3.5-397b); any other chat model ID from GET /v1/models can be typed in

• the three env names are picked up by all_secret_envs() automatically, so keys are stored per account and hot-reloaded like the other presets

• studio/server/providers/llm.py: cloudgpu added to the auto-pick order after openrouter, so it only wins when it is the only key set or LLM_PROVIDER=cloudgpu

• studio/.env.example: commented LLM_PROVIDER=cloudgpu block, same shape as suanqiao

Usage:

`LLM_PROVIDER=cloudgpu
CLOUDGPU_API_KEY=sk-...

optional

CLOUDGPU_MODEL=deepseek-v4-flash
CLOUDGPU_BASE_URL=https://cloudgpu.app/v1
`

Testing

• make lint passed (py_compile of the four files in the lint target).

• py_compile of studio/server/providers/llm.py and studio/server/providers/llm_catalog.py passed.

• Public GET https://cloudgpu.app/v1/models (no key) lists every preset model id with "type": "chat" (checked 2026-11-07). Chat completions were not called, and the preset was not exercised in the studio UI.

• make test: 1911 passed, 12 skipped, 3 xfailed, 3 failed. The three failures also fail on main at e4beb3768303cb9630fe4b079d1a810390a70c79 and are outside this diff:

test_planning_skill_mentions_runtime_contract[short-drama]

• test_compose_stage_references_runtime_routing[short-drama]

• test_noncanonical_stage_does_not_raise_keyerror

Checklist

• The change is focused on a single logical concern.

• I ran the relevant tests locally (make test-contracts / make test) where applicable.

• I updated docs/README if behavior or usage changed.

• No unrelated files (build artifacts, local config) are included in the diff.

docs/SELF_HOSTING.md already lists text providers in one sentence. This PR does not edit that sentence. Usage for the new preset is in studio/.env.example.

Same field set as existing presets. Request code is unchanged and still uses the existing compatible chat path. Auto-pick order for existing providers is unchanged (cloudgpu is appended after openrouter).

Open in Web Open in Cursor