Add GPU Operator DevOps Agent package (prototype, for review & testing) - #450
Open
chenopis wants to merge 3 commits into
Open
Add GPU Operator DevOps Agent package (prototype, for review & testing)#450chenopis wants to merge 3 commits into
chenopis wants to merge 3 commits into
Conversation
A prototype AI DevOps-agent package for the NVIDIA GPU Operator: an AGENTS.md front door + operator skills + references distilled from the public GPU Operator docs and source. It lets an AI coding/ops agent help install, operate, maintain, and troubleshoot the GPU Operator with a probe-first, source-cited, safety-gated operating model. Docs-derived and version-anchored to GPU Operator v26.3.3. This is a review/test prototype, not a validated release. Signed-off-by: Andrew Chen <andrewch@nvidia.com>
Documentation preview |
chenopis
added
feature
A doc-backed uninstall operator skill — Helm release removal, CRD cleanup (operator.cleanupCRD post-delete hook or manual deletion), host driver-module unload (rmmod), and the --no-hooks fallback — grounded in the GPU Operator uninstall documentation. Probe-first preconditions and gated destructive steps. Signed-off-by: Andrew Chen <andrewch@nvidia.com>
chenopis
marked this pull request as ready for review
July 20, 2026 23:53
The gpu-operator-uninstall skill was added to the package but the README 'What's in it' table and 'Skill coverage' feedback list still enumerated six skills and omitted it, and the purpose verb-lists dropped removal. Add gpu-operator-uninstall to the README inventory (now seven SOPs) + feedback list, and add 'uninstall' to the README and AGENTS purpose lines so the docs match the package + AGENTS skill-routing table. Signed-off-by: Andrew Chen <andrewch@nvidia.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This PR adds a prototype GPU Operator DevOps Agent package under a new top-level
gpu-operator-devops-agent/directory — an AI-agent skill package (anAGENTS.mdfront door + operator skills + references) distilled from the public GPU Operator documentation in this repo and the public GPU Operator source. It lets an AI coding/ops agent (Claude, Cursor, etc.) help an operator install / operate / maintain / troubleshoot the GPU Operator with a probe-first, source-cited, safety-gated operating model — instead of a fresh agent guessing from an unstructured search.Opened as a draft for review and hands-on testing — not for merge yet.
What's here
AGENTS.md— the agent operating charter: supported scope, source precedence, bootstrap/command-context policy, safety policy, skill routing, operating invariants, and stop/escalation conditions.skills/— six operator SOPs:gpu-operator-install,-operate,-maintain,-troubleshoot,-evidence-bundle,-improve.references/— the reasoning substrate:models.md(the layered NFD → driver → toolkit → device-plugin → validator → workload success ladder),branch-matrix.md(staged branch decisions),command-cards.md(validated recipes),field-scenario-cards.md+failure-signatures.md(troubleshooting),supported-scope.md,claim-ledger.md, and more.scripts/— a read-only environment inventory plus a few guarded mutation helpers.runtime-memory/— seed files the agent appends to as it learns.README.md— how to test it, plus honesty caveats.Placement
Placed at repo root as
gpu-operator-devops-agent/(outside every Sphinxdocs_root), so the./repo docsPR build ignores it entirely — norepo.tomlchange, no toctree warnings, zero docs-build interaction. If maintainers would rather co-locate it undergpu-operator/, that's a one-linesphinx_exclude_patternsaddition — happy to move it.How to test
See
gpu-operator-devops-agent/README.md. In short: point an agent atgpu-operator-devops-agent/AGENTS.mdas its front door, give it an operator task against a test-owned K3s + A100 sandbox (e.g. "install the GPU Operator on this node," "a GPU workload won't schedule after install," "upgrade and roll back if driver pods crash-loop"), and watch for: probe-first inventory before any mutation, cited doc/source for each load-bearing claim, staged branch selection via the layered model, and safety refusals (blast-radius + rollback) on high-impact steps.Status & caveats
v26.3.3. The package has not itself been rerun end-to-end after authoring (see its Verification Gap Statement inreferences/supported-scope.md). This is a review/test draft, not a validated release.Provenance
Generated from a curated source manifest built by reading the public GPU Operator docs in this repository (
gpu-operator/, snapshot commit77a2daaf234a2cebc8e178f24e18cc4a6150e8b1) and the public GPU Operator source at thev26.3.3release.