feat: support Atlas Cloud environment configuration

[binyangzhu000-sudo]

Summary

• use ATLASCLOUD_API_KEY as an opt-in credential when no explicit API_KEY is configured
• select the Atlas Cloud OpenAI-compatible base URL and namespaced DeepSeek V4 model by default
• preserve explicit user, project, and DEEPCODE_* settings precedence
• recognize the namespaced model for thinking-mode and multimodal capability defaults

Validation

• npm run generate && npm run check
• npm test (596 passed, 1 skipped)
• Atlas Cloud model discovery through createOpenAIClient() (123 models; default model present)
• Atlas Cloud chat completion through createOpenAIClient() (non-empty response)
• git diff --check

No README, public documentation, logo, dependency, or lockfile changes.

[binyangzhu000-sudo]

Resolved the merge conflict with current main in follow-up commit 7084a7b9.

The resolution preserves both the Atlas Cloud settings tests and the new upstream context-window tests. The online PR diff remains limited to the original four core files.

Validation:

• Focused settings tests: 35 passed, 0 failed
• Core test suite: 281 passed, 1 skipped, 0 failed
• All workspace typechecks passed
• ESLint, Prettier, conflict-marker, and diff-scope checks passed

The exact-head CI workflow currently shows action_required, so it needs maintainer approval to execute for this fork PR.

[binyangzhu000-sudo]

Synced with main (79 commits). Three conflicts, all where this branch overlapped an upstream refactor:

• model-capabilities.ts — you replaced the NON_MULTIMODAL_MODELS set with supportsMultimodal(model, mode) inferring from -vision. Took the new mechanism and dropped this branch's list rather than reinstating it; I checked first that it has no remaining consumers anywhere in packages/. The Atlas model id stays in DEEPSEEK_V4_MODELS, which still drives the thinking-mode default, and it correctly resolves to non-multimodal under the new rule.
• index.ts — both sides added exports; kept both.
• settings.ts — you hoisted baseURL into a local. I folded the Atlas default into that local rather than patching the return site, so baseURL now mirrors how model already handles useAtlasCloudDefaults:

const baseURL =
  trimString(env.BASE_URL) ||
  (useAtlasCloudDefaults ? ATLASCLOUD_BASE_URL : "") ||
  defaults.baseURL;

Verification — npm run typecheck clean across both workspaces (@vegamo/deepcode-core, deepcode-vscode). Diff against current main is +57/-3.

[binyangzhu000-sudo]

Synced with main (7 commits). Single conflict in packages/core/src/common/model-capabilities.ts: upstream added deepseek-flash to DEEPSEEK_V4_MODELS in the same row this branch adds the Atlas-prefixed deepseek-ai/deepseek-v4-pro. Both entries are kept.

I also checked the auto-merged side, since settings.ts changed on both sides: the ATLASCLOUD_API_KEY fallback in resolveSettingsSources survived intact — it still only supplies the default base URL and model when no BASE_URL / API_KEY is set, so any explicit configuration still wins — and ATLASCLOUD_DEFAULT_MODEL still names a model Atlas serves.

Verification on the merged branch:

• npm run typecheck → clean across workspaces
• npm test → 282/283, 370/374, 59/59

The four failures are cli-args.test.ts and three tool-permission/image cases in packages/core. I ran them in a clean upstream/main worktree and got the identical counts, so they are pre-existing rather than from this branch.

No conflicts remain.