fix(vllm): continue automated setup after storage warnings

[sandl99]

Summary

Previously, verified low Docker or model-cache storage stopped express and non-interactive managed-vLLM setup unless NEMOCLAW_IGNORE_VLLM_DISK_SPACE was set. This extracts the first behavior slice from #6965: verified shortages are advisory for automation while the storage checks mature, and the now-redundant override is removed. Interactive setup still requires explicit confirmation, and non-interactive setup still stops when model-cache capacity cannot be verified.

Changes

• Continue express and non-interactive managed-vLLM setup after a verified Docker or model-cache storage warning.
• Preserve interactive confirmation for verified shortages and the fail-closed non-interactive path for an inconclusive model-cache probe.
• Remove NEMOCLAW_IGNORE_VLLM_DISK_SPACE from the implementation and command reference.
• Update focused preflight coverage and managed-vLLM documentation. Express already enters onboarding in non-interactive mode, so no installer-specific branch is needed.

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:
• Docs updated for user-facing behavior changes
• Docs 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: Independent review found no actionable findings; focused tests cover verified-low image and model capacity, interactive responses, post-pull re-probing, and the inconclusive model-cache fail-closed boundary.
• Non-success, skipped, or missing CI check accepted by maintainer — check name, approval link, and follow-up issue:

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 check:diff passed when hooks were skipped or unavailable
• Targeted behavior tests pass for the current change set, or tests are marked not applicable above — npx vitest run --project cli src/lib/inference/vllm.test.ts src/lib/inference/vllm-storage.test.ts (103 passed); npx vitest run test/inference-options-docs.test.ts test/changelog-docs.test.ts (26 passed)
• Applicable broad gate passed — npm test for broad runtime/test-harness changes; npm run check for repo-wide validation/coverage changes — not applicable to this focused preflight behavior change
• 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) — build passed with 0 errors; Fern reported two unrelated warnings for unavailable redirect authentication and the existing light-mode accent contrast
• Doc pages follow the style guide (doc changes only)
• New doc pages include SPDX header and frontmatter (new pages only)

---

Signed-off-by: San Dang sdang@nvidia.com

Summary by CodeRabbit

• Behavior Changes

  • Managed vLLM setup now treats verified storage shortages as advisory during non-interactive installation, allowing guarded downloads to continue.
  • Interactive setup still requires explicit confirmation for verified insufficient-capacity warnings.
  • Non-interactive setup will stop when model-cache capacity can’t be verified and directs users to retry interactively; Docker image-storage verification remains more permissive.
  • Updated warning messaging for Docker image and Hugging Face model-cache capacity.

• Documentation

  • Refreshed guidance for --yes / NEMOCLAW_YES=1, removed deprecated disk-space override guidance, and added an OpenClaw-only NEMOCLAW_MINIMAL_BOOTSTRAP note.

• Tests

  • Updated vLLM installation tests to match the new non-interactive warning/continue/stop behavior.

[coderabbitai]

📝 Walkthrough

Walkthrough

Managed vLLM storage checks now treat verified capacity shortages as advisory during non-interactive setup, while inconclusive model-cache checks still stop non-interactive setup. The disk-space override was removed, with corresponding test and documentation updates.

Changes

Managed vLLM storage handling

Layer / File(s)	Summary
Storage warning control flow src/lib/inference/vllm-storage.ts, src/lib/inference/vllm.ts	Removes the disk-space override and updates image/model storage handling for interactive, non-interactive, advisory, and inconclusive probe results.
Storage behavior validation src/lib/inference/vllm.test.ts	Updates coverage for advisory continuation, interactive confirmation, inconclusive model-cache checks, cold-image re-probing, Docker-storage warnings, and removed override behavior.
Storage behavior documentation docs/changelog/2026-07-14.mdx, docs/inference/set-up-vllm.mdx, docs/reference/commands.mdx	Documents revised storage-warning behavior and replaces the removed environment-variable reference with NEMOCLAW_MINIMAL_BOOTSTRAP documentation.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Sequence Diagram(s)

sequenceDiagram
  participant installVllm
  participant StorageChecks
  participant Docker
  participant Operator
  installVllm->>StorageChecks: evaluate image and model-cache capacity
  StorageChecks-->>installVllm: return storage status
  alt verified shortage in non-interactive mode
    installVllm->>Docker: continue guarded pull
  else inconclusive model-cache capacity
    installVllm-->>installVllm: stop and request interactive rerun
  else interactive warning
    installVllm->>Operator: request confirmation
    Operator-->>installVllm: confirm continuation
    installVllm->>Docker: continue guarded pull
  end

Loading

Possibly related PRs

• NVIDIA/NemoClaw#6883: Adds related managed vLLM storage-warning behavior for Station express/profile setup.
• NVIDIA/NemoClaw#6965: Updates the related managed vLLM storage preflight, override removal, and installation gating flow.

Suggested labels: bug-fix, area: inference, provider: vllm

Suggested reviewers: cv

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: automated setup now continues after storage warnings.

✨ 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/vllm-noninteractive-storage-advisory

---

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

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-6985.docs.buildwithfern.com/nemoclaw

[github-code-quality]

Code Coverage Overview

Languages: TypeScript

TypeScript / code-coverage/plugin

The overall coverage remains at 96%, unchanged from the main branch.

TypeScript / code-coverage/cli

The overall coverage in the fix/vllm-noninteract... branch remains at 80%, unchanged from the main branch.

Show a code coverage summary of the most impacted files.

File	main d03e9d7	fix/vllm-noninteract... 6dd9953	+/-
src/lib/inference/vllm.ts	84%	84%	0%
src/lib/inferen...vllm-storage.ts	87%	88%	+1%

---

_{Updated July 16, 2026 01:59 UTC
Code Coverage is in Public Preview. Learn more and provide us with your feedback.}

[github-actions]

PR Review Advisor — Blocking findings reported

Advisor assessment: Blockers require maintainer review
Next action: Review the blockers below.
Findings: 1 blocker · 0 warnings · 0 suggestions

Model lanes

• GPT-5.6 Terra (primary): Completed · high confidence · 1 blocker · 0 warnings · 0 suggestions
• Nemotron 3 Ultra (second opinion): Completed · high confidence · 0 blockers · 0 warnings · 0 suggestions
• Model comparison: normalized findings differ; normalized E2E selections differ; Nemotron reported 1 fewer blocker, the same number of warnings, the same number of suggestions.

Nemotron output stays in workflow artifacts and does not change the assessment above.

E2E guidance

Advisory only. E2E / PR Gate selects and runs jobs independently.

Recommended E2E: inference-routing, network-policy

Blockers

PRA-1 Blocker — Keep verified storage shortages fail-closed in unattended installs

• Location: src/lib/inference/vllm.ts:860
• Category: security
• Problem: `imageStorageAccepted` now returns true for a verified insufficient Docker-storage result whenever `nonInteractive` is set; `modelStorageAccepted` does the same for a verified insufficient model-cache result. The subsequent pull/download can therefore begin after the guard has established that its target filesystem lacks the required capacity, with no explicit operator override.
• Impact: Automated onboarding can start a multi-gigabyte image pull or a registered model download (including the roughly 352-GB Ultra snapshot) on a known-insufficient host filesystem. Exhausting Docker or host-cache storage can impair host and sandbox availability.
• Fix: Restore refusal for verified insufficient capacity in non-interactive mode, or require a narrowly named explicit operator override before continuing. Retain the interactive explicit-confirmation path and the post-pull model-cache re-probe.
• Verification: Read `imageStorageAccepted` and `modelStorageAccepted` in `src/lib/inference/vllm.ts` and run the existing unit cases that mock `availableBytes: 1n` with `nonInteractive: true`; they currently reach Docker pull/model download.
• Test coverage: Cover non-interactive verified-low Docker and model-cache capacity without an explicit override returning `{ ok: false }` and asserting no pull or model-download spawn; separately cover the explicit override continuing.
• Evidence: `src/lib/inference/vllm.ts:856-864` logs a verified Docker-storage warning then returns true in non-interactive mode. `src/lib/inference/vllm.ts:880-897` logs a verified model-cache warning then returns true in non-interactive mode. `src/lib/inference/vllm.test.ts` changed low-storage non-interactive expectations to `{ ok: true }` and asserts pull/download calls.

Workflow run details

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

[sandl99]

PRA-1 is intentional for this scoped PR and does not require a code change.

The maintainer decision is to treat a verified Docker image-storage or model-cache shortage as advisory for express and other non-interactive setup while these estimates mature. The warning, measured capacity, remediation, and diagnostics are still printed before the guarded pull or download. Interactive setup still requires an explicit y or yes, and an inconclusive model-cache probe remains fail-closed in non-interactive setup.

NEMOCLAW_IGNORE_VLLM_DISK_SPACE is deliberately removed because automated setup should not require a hidden escape hatch for this advisory case. The focused tests cover both advisory paths and retain the interactive and inconclusive-probe boundaries. A future hard gate should be considered only after the preflight is stable enough to be enforced without blocking supported automated installs.