fix(inference): drop the retired DeepSeek V4 Pro from the featured menu

[laitingsheng]

Summary

The onboard wizard listed deepseek-ai/deepseek-v4-pro in the NVIDIA Endpoints featured model menu after NVIDIA retired the model, so a user who picked it completed onboarding and then failed on the first inference call. The wizard builds that menu from NVIDIA's public featured feed, which still advertises the retired entry, so this change adds the model ID to the featured-catalog retirement deny-list. The menu now shows only the live entries.

Related Issue

Fixes #9611

Changes

• Add deepseek-ai/deepseek-v4-pro to RETIRED_NVIDIA_FEATURED_MODEL_IDS in src/lib/inference/nvidia-featured-models.ts. This is a data entry in the existing deny-list, not a new abstraction or fallback path; the deny-list and its policy comment already exist for models whose catalogs outlive their routes.
• Add one regression test in src/lib/inference/nvidia-featured-models.test.ts that drives getNvidiaFeaturedModelPromptOptions with the current live feed payload and asserts the rendered menu keeps the four live entries and drops the retired one.

Deliberately unchanged, with reasons:

• isDeepSeekV4ProModel in src/lib/inference/openai-probe-models.ts and the matching blueprint payload rule stay. Commit 778811d set this precedent when it retired Kimi K2.6: the retirement is scoped to NVIDIA Endpoints, and those branches still serve custom and compatible routes that host the same model ID. The manual Other... entry path also stays open, and it is already covered because validateNvidiaEndpointModel checks the authenticated /v1/models catalog, which no longer lists the model.
• CLOUD_MODEL_OPTIONS needs no edit. It never contained this model, and src/lib/inference/config.test.ts already asserts its absence.
• The it.each case in config.test.ts titled retires %s only from the NVIDIA Endpoints picker was not extended, because it also asserts the model remains in HERMES_PROVIDER_MODEL_OPTIONS, and this model was never in that list.

Known side effect, consistent with existing behavior: src/lib/onboard/setup-nim-flow.ts assigns openRouterFeaturedModels = nvidiaFeaturedModels, so the OpenRouter picker shares this deny-list and also stops offering the ID. moonshotai/kimi-k2.6 is already denied under the same shared session, so this PR introduces no new coupling.

Runtime evidence

Collected against https://integrate.api.nvidia.com/v1 on 2026-08-19 with a valid key:

• GET /v1/models: HTTP 200, 102 models, deepseek-ai/deepseek-v4-pro absent. This matches the count in the issue report.
• POST /v1/chat/completions with that model: HTTP 410, body {"type":"about:blank","title":"Gone","status":410,"detail":"The model 'deepseek-ai/deepseek-v4-pro' has reached its end of life on 2026-08-07T09:00:00Z and is no longer available."}. NVIDIA reports the exact retirement timestamp, which confirms the 2026-08-07 date in the issue.
• GET https://assets.ngc.nvidia.com/products/api-catalog/featured-models.json: HTTP 200, 5 entries, still including deepseek-ai/deepseek-v4-pro. The regression test uses this exact payload.

Two corrections to the issue report, both minor: the reported menu is produced by the live feed rather than a bundled catalog JSON, and the same shared session means the OpenRouter picker was affected as well, which the report does not mention.

Type of Change

• Code change (feature, bug fix, or refactor)
• Code change with doc updates
• Doc only (prose changes, no code sample modifications)
• Doc only (includes code sample changes)

Quality Gates

• Tests added or updated for changed behavior
• Existing tests cover changed behavior — justification:
• Tests not applicable — justification:
• Sensitive paths changed (security, policy, credentials, preflight, onboarding, inference, runner, sandbox, or messaging)
• Sensitive-path review completed or maintainer-approved waiver recorded — reviewer/approval link/justification: sensitive-path review remains on the normal approval path
• Non-success, skipped, or missing CI check accepted by maintainer — check name, approval link, and follow-up issue:

DGX Station Hardware Evidence

• Tested on DGX Station
• Tested commit: not applicable, scripts/prepare-dgx-station-host.sh is unchanged
• Station profile/scenario: not applicable
• Result: not applicable
• Supporting evidence: not applicable

Verification

• PR description includes a Signed-off-by: line and every commit appears as Verified in GitHub
• Normal pre-commit, commit-msg, and pre-push hooks passed, or npm run validate:pr passed after refreshing origin/main when hooks were skipped or unavailable
• Targeted behavior tests pass for the current change set, or tests are marked not applicable above — command/result or justification: npx vitest run --project cli src/lib/inference src/lib/onboard/nvidia-featured-model-selection.test.ts src/lib/onboard/setup-nim-flow.test.ts passed 2129 tests across 103 files; npx vitest run --project integration test/onboard-selection.test.ts passed 67 tests; npm run typecheck:cli and npm run checks:repository passed. Reverting only the source change fails the new test and leaves the other 17 in that file passing.
• Applicable broad gate passed — npm test for broad runtime/test-harness changes; npm run check for repo-wide validation/coverage changes — command/result: not run; the change is one deny-list entry with no runtime or harness surface
• Quality Gates section completed with required justifications or waivers
• No secrets, API keys, or credentials committed
• npm run docs builds without warnings (doc changes only)
• Doc pages follow the style guide (doc changes only)
• New doc pages include SPDX header and frontmatter (new pages only)

Deferred documentation

docs/inference/model-capability-audit.mdx records retired NVIDIA Endpoints routes and carries a row for Kimi K2.6 from the precedent commit, but has no row for this model. That row is proposed for Docs / Post-Merge Catch-Up. test/model-capability-audit-doc.test.ts asserts the state vocabulary, evidence field names, header row, and navigation entries only, so it does not require a row per retired model and deferral cannot redden CI. No other page needs an edit: docs/inference/choose-model.mdx already states that NVIDIA Endpoints excludes retired choices and corrects catalog lag, and docs/inference/use-nvidia-endpoints.mdx documents the bundled fallback list, which this change does not touch.

---

Signed-off-by: Tinson Lai tinsonl@nvidia.com

Summary by CodeRabbit

• Bug Fixes
  • Removed the retired DeepSeek V4 Pro model from NVIDIA featured-model selections.
  • Preserved other featured models, including Nemotron models with canonicalized identifiers.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 066bcf24-3dba-45fa-b759-4c8d582c2d23

📥 Commits

Reviewing files that changed from the base of the PR and between 2c74cbb and fc3d819.

📒 Files selected for processing (2)

• src/lib/inference/nvidia-featured-models.test.ts
• src/lib/inference/nvidia-featured-models.ts

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

📝 Walkthrough

Walkthrough

The NVIDIA featured-model parser now excludes deepseek-ai/deepseek-v4-pro. A regression test confirms that the four remaining featured models stay selectable.

Changes

NVIDIA featured-model filtering

Layer / File(s)	Summary
Retired model deny-list and regression coverage src/lib/inference/nvidia-featured-models.ts, src/lib/inference/nvidia-featured-models.test.ts	The parser excludes deepseek-ai/deepseek-v4-pro. The test verifies that the remaining four featured models are preserved.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: 🟡 Moderate · up to fc3d8

This change removes an unavailable NVIDIA model from the featured picker and adds a regression test, preventing users from selecting that retired entry. The implementation is localized, but merge readiness remains pending the required sensitive-path review and completion or waiver of the outstanding quality-gate items.

Suggested reviewers: ericksoa, sandl99

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes removing the retired DeepSeek V4 Pro model from the NVIDIA featured menu.
Linked Issues check	✅ Passed	The deny-list update and regression test remove the retired model from NVIDIA featured menus as required by issue #9611.
Out of Scope Changes check	✅ Passed	All changes directly support issue #9611 and remain limited to the deny-list update and its regression test.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/retire-deepseek-v4-pro-featured

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-code-quality]

Code Coverage Overview

Languages: TypeScript

TypeScript / code-coverage/plugin

The overall line coverage in commit f5c86b9 in the fix/retire-deepseek-... branch remains at 96%, unchanged from commit dbf48ba in the main branch.

TypeScript / code-coverage/cli

The overall line coverage in commit f5c86b9 in the fix/retire-deepseek-... branch remains at 82%, unchanged from commit 772ce00 in the main branch.

Show a line coverage summary of the most impacted files.

File	main 772ce00	fix/retire-deepseek-... f5c86b9	+/-
src/lib/actions...rget-staging.ts	95%	75%	-20%
src/lib/actions...rget-runtime.ts	92%	80%	-12%
src/lib/actions...confirmation.ts	88%	81%	-7%
src/lib/actions...ild-pipeline.ts	91%	87%	-4%
src/lib/inferen...board-probes.ts	87%	86%	-1%
src/lib/onboard...-transaction.ts	92%	94%	+2%
src/lib/actions...e-classifier.ts	58%	63%	+5%
src/lib/onboard...on-preflight.ts	85%	91%	+6%
src/lib/shields...ate-mutation.ts	70%	78%	+8%
src/lib/inferen...ntext-window.ts	32%	50%	+18%

---

_{Updated August 19, 2026 17:36 UTC}

[github-actions]

PR Review Advisor — No blocking findings reported

Advisor assessment: No blocking advisor findings reported
Next action: No advisor follow-up needed.
Findings: 0 blockers · 0 warnings · 0 suggestions

Model lanes

• GPT-5.6 Terra (primary): Completed · high confidence · 0 blockers · 0 warnings · 0 suggestions
• Nemotron 3 Ultra (second opinion): Completed · high confidence · 0 blockers · 0 warnings · 0 suggestions
• Model comparison: normalized findings match; normalized terminology decisions differ; normalized E2E selections differ; severity counts match.

1 additional E2E selection from the second opinion

Advisory only. The primary lane did not select these E2E jobs or targets.

• cloud-onboard: The completed second-opinion lane identified E2E coverage that the primary lane omitted.

Second-opinion terminology and E2E selections are advisory. Live E2E does not run automatically for pull requests.

Since last review: 0 prior items resolved · 0 still apply · 0 new items found

1 semantic terminology decision

Terminology decisions are advisory. They affect the assessment only when a separate finding identifies concrete semantic impact.

• established — retired at src/lib/inference/nvidia-featured-models.ts:21: Keep retired for models intentionally excluded after route retirement.

E2E guidance

Advisory only. A maintainer can dispatch the default E2E suite for the commit under review.

Recommended E2E: inference-routing

Manual-only E2E: network-policy
The manual PR workflow does not run these selectors for the commit under review. Run them from reviewed code on main.

Workflow run details

This automated review informs maintainers. Warnings and suggestions do not require a response. A maintainer decides whether to merge.

[prekshivyas]

Reviewed the focused deny-list change and regression coverage; no blocking findings.