[binyangzhu000-sudo]

What

Adds Atlas Cloud as an LLM service. It is an OpenAI-compatible gateway serving DeepSeek, GLM and Kimi models behind one endpoint, so the service follows the Nebius shape: an AiLLM over createOpenAI with the gateway base URL.

• src/llm/services/atlascloud.ts — service + 3 registered models
• src/llm/llmFactory.ts — registry entry
• shared/user/user.model.ts, shared/user/user.schema.ts — atlasCloudKey

Key resolution matches the existing services: currentUser()?.llmConfig.atlasCloudKey falling back to ATLASCLOUD_API_KEY.

Disclosure: I work at Atlas Cloud.

How I verified it

Costs and context windows are from the source, checked by script. Every costPerMilTokens(...) and maxInputTokens was compared field-by-field against the gateway's public catalog (GET https://api.atlascloud.ai/v1/models) — zero mismatches:

model	maxInputTokens	in / out per Mtok
deepseek-ai/DeepSeek-V3.1-Terminus	131,072	$0.30 / $0.95
zai-org/glm-4.7	202,752	$0.52 / $1.85
moonshotai/kimi-k2.6	262,144	$0.95 / $4.00

All three answer. Each was called live against https://api.atlascloud.ai/v1 with the same shape provider() builds. (glm-4.7 returned 200 with empty content on a 5-token cap — it is a reasoning model and spends the budget on reasoning_content — not an error.)

Test plan

• npx tsc --noEmit — 4 errors, identical to the clean tree (verified with git stash -u), 0 mentioning atlascloud
• Catalog cross-check on all 3 models (table above)
• Live call per model
• No lockfile changes

🤖 Generated with Claude Code