feat(models): add Atlas Cloud provider preset

[binyangzhu000-sudo]

What

• Add an atlas-cloud provider preset backed by the existing OpenAI Chat Completions transport.
• Configure ATLASCLOUD_API_KEY, https://api.atlascloud.ai/v1, and deepseek-ai/deepseek-v4-pro together without persisting credentials.
• Expose the preset in onboarding guidance and document the generated models.yml contract.

Why

Atlas Cloud is OpenAI-compatible, but configuring its endpoint, credential environment variable, model id, and compatibility flags currently requires error-prone manual YAML.

Testing

• bun test packages/coding-agent/test/provider-onboarding.test.ts packages/coding-agent/test/model-registry.test.ts packages/coding-agent/test/model-onboarding-guidance.test.ts (248 passed)
• bun --cwd=packages/coding-agent run check
• Isolated gjc setup provider --preset atlas-cloud generated a credential-free models.yml.
• Live Atlas Cloud forced tool-call returned HTTP 200 with deepseek-ai/deepseek-v4-pro.

GJC verdict

gajae.pr-review-verdict.v1 needs-human sha256:acf9ba3bbd6b83e44cd62f101b4d0396c287e751acdf3e450e986fdc1fd39249 reviewer:human evidence:local-command

---

• Target branch is dev
• bun check passes (package-scoped coding-agent check passed; full root check was not run)
• Tested locally
• CHANGELOG updated (if user-facing)
• Verdict above matches the exact PR head, not an earlier commit

[Yeachan-Heo]

GJC hostile-review verdict — exact head

Reviewed head: 787d061b5dd9e225d7510dbc825369e273ce89a2
Authoritative dev: 732856b3ccb3fade6e9fbc17908a4fbca5a7682f
Verdict: REQUEST_CHANGES

Blocking findings:

1. Atlas output caps use the wrong documented field. The preset omits compat.maxTokensField: "max_tokens". GJC consequently selects max_completion_tokens, while Atlas's Chat API and exact /v1/models entry advertise max_tokens.
2. Same-ID metadata leaks from NVIDIA into Atlas. The preset writes only deepseek-ai/deepseek-v4-pro; finalizeCustomModel() therefore inherits the bundled NVIDIA same-ID record. Context/output happen to match Atlas, but the effective model inherits NVIDIA reasoning/effort behavior and NVIDIA pricing. Atlas's live catalog does not advertise reasoning for this model and reports approximately $1.68/M input, $3.38/M output, and $0.13/M cache read rather than the inherited $0.435/M, $0.87/M, and $0.003625/M.
3. No exact-head CI exists. This SHA has zero check runs and zero legacy status contexts. Author-reported local tests and the package check are not required-CI evidence.
4. The head is stale and conflicting. It is 9 commits behind current dev; GitHub reports DIRTY, unmergeable, and unrebaseable. Current dev overlaps the environment-variable documentation and generated embedded-doc artifact.

Atlas's provider identity, https://api.atlascloud.ai/v1, Bearer authentication, ATLASCLOUD_API_KEY, and exact model ID are validated by official Atlas documentation and the live catalog. These are ordinary preset/runtime defects, not a provider-existence or owner-confirmation blocker.

Required before re-review: rebase and regenerate; pin max_tokens; represent Atlas-specific effective model metadata rather than inheriting NVIDIA transport economics/capabilities; test the finalized registry model and emitted capped payload; obtain exact-head CI.

PR-3777 head=787d061b5dd9e225d7510dbc825369e273ce89a2 dev=732856b3ccb3fade6e9fbc17908a4fbca5a7682f evidence-sha256=3ba60e486c2f95842e36517c9eec831ab09667e167068a589eb1f297836c529e verdict=REQUEST_CHANGES