kubernetes-sigs/devops-bench

Latest recorded source activity: October 9, 2026 at 22:27

SkillsMP catalog refreshed: October 29, 2026 at 18:30

skills collected: 8

GitHub stars: 34
GitHub forks: 16

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/kubernetes-sigs/devops-bench.

Repository README

Repository README source: README.md

Read the source README

Expand full README

# devops-bench A standardized benchmarking suite to evaluate how well different agents or models perform specific DevOps tasks. Its goal is to provide an open-source, reproducible way to transparently assess agent performance across various infrastructure platforms and operational environments. Most benchmarks stop at "did the model produce reasonable text?" This one runs the agent against live infrastructure and checks the result. It also lets you quantify the payoff of giving agents more to work with — context, operational rules, and tools like MCP servers and skills — so you can see what those additions are actually worth. See the [project roadmap](roadmap.md) for current initiatives and how to get involved. ## How it works For each task, the harness provisions real infrastructure if the task needs it, runs your agent against it, optionally injects chaos and verifies the resulting cluster state, then scores the run with LLM-as-judge metrics — and tears everything down when it's done. A single run, end to end: 1. **Provision** — OpenTofu stands up a cloud cluster or a local kind cluster (or nothing, for no-infra tasks). 2. **Run the agent** — your chosen agent harness drives the task. 3. **Chaos + verify** — optionally break things, then check the live cluster state. 4. **Score** — LLM-as-judge metrics grade the outcome and the agent's tool use. 5. **Teardown** — everything provisioned is cleaned up. ## What's supported **Agent harnesses** — choose with `BENCH_AGENT_TYPE` or `--agent-type` (default `gemini-cli`, an alias for `gemini`): | Key | What it runs | | :-- | :-- | | `gemini` | The Google Gemini CLI. | | `openclaw` | The Openclaw Agent CLI. | | `antigravity` | The Antigravity CLI. | | `api` | In-process: drives a provider SDK directly through a model-agnostic MCP tool loop. | **Model providers** — choose with `AGENT_PROVIDER` and `AGENT_MODEL`: | Key | Backends | | :-- | :-- | | `gemini` | Google AI Studio API key, or Vertex AI. | | `claude` | Anthropic API, Vertex AI, or Bedrock. | | `ollama` | Local models. | **Infrastructure targets** — the OpenTofu deployer supports these targets (set `INFRA_PROVIDER`, or the task's `provider:` key): | Key | Target | | :-- | :-- | | `gcp` | GKE. | | `kind` | Local kind clusters. | `--no-infra` skips provisioning entirely and runs against a pre-existing cluster or none at all. ## Install You need Python 3.12 or newer. The project uses [`uv`](https://docs.astral.sh/uv/): ```bash uv sync ``` The default `dev` group includes the test/lint toolchain and every optional provider SDK (`anthropic`, `openai`). ## Run your first eval The `opa-remediation` task provisions its own local kind cluster, so a real end-to-end eval needs no cloud account. With `tofu`, Docker, `kind`, `kubectl`, and the `gemini` CLI on `PATH`: ```bash export AGENT_API_KEY=... # used by the agent's model provider and the default judge uv run devops-bench tasks/common/opa-remediation \ --project local-kind --cluster devops-bench-kind ``` The results path is printed at the end of the run. The full walkthrough — prerequisites, judge configuration, exit codes — is in [Getting started](docs/getting-started.md); for cloud runs and parallel matrices, see the [run-evals how-to](docs/how-to/run-evals.md). **Working through a coding agent?** Point it at the repo's skills instead of assembling commands yourself — see [the skills overview](docs/getting-started.md#skills-in-this-repo). ## Adding a benchmark task New tasks live under `tasks/<provider>/<name>/task.yaml`, each pairing a `chaos_spec` (what breaks) with a `verification_spec`/`expected_output` (how it's graded). The full schema, placeholders, worked examples, and authoring best practices are in [docs/how-to/add-a-task.md](docs/how-to/add-a-task.md) — read that before you start. Before submitting, run the `task-review` skill over your task — see [the skills overview](docs/getting-started.md#skills-in-this-repo). ## Documentation We welcome contributions around adding new tasks, models, or agent harnesses. Start with [Getting started](docs/getting-started.md), then browse the how-to guides, component docs, known issues, and agent skills from the [documentation index](docs/README.md). ## Community, discussion, contribution, and support Learn how to engage with the Kubernetes community on the [community page](http://kubernetes.io/community/). You can reach the maintainers of this project at: - [Slack channel](https://kubernetes.slack.com/messages/sig-apps) - [Mailing List](https://groups.google.com/a/kubernetes.io/g/sig-apps) ### Code of conduct Participation in the Kubernetes community is governed by the [Kubernetes Code of Conduct](code-of-conduct.md). ## License Apache 2.0 — see [`LICENSE`](LICENSE).

Loading the repository README…

Skills in this repository

Software Developers Software Quality Assurance Analysts & Testers Network & Computer Systems Administrators

3 occupation categories · 75% classified

skill

occupation

description

updated

diagnose-eval-failure

unclassified

Explain WHY a model scored low on an eval — read the judge's reasons and the agent's trajectory, compare against the task rubric, and produce a rationale (capability gap vs. task/rubric problem). Distinct from infrastructure failures. Invoke when someone asks…

2026-10-09

task-review

unclassified

Use when the user adds or changes a benchmark TASK and asks to review it — e.g. "review this new task", "is this task parallel-safe", "review this task.yaml + stack", "will this task collide under the matrix", "check my new tf/prebuilt stack". Reviews a newly…

2026-09-19

validate-eval

Software Quality Assurance Analysts & Testers

Validate that a NEW eval is correct and runnable by running it and iterating in a loop until green — fixing infra/config/task bugs (not the model's score), with opt-in hands-off and unlimited/self-healing modes; invoke when authoring or vetting a task before…

2026-08-28

docs-sync

Software Developers

Keep the docs current after a code change — map changed code areas to the docs that describe them, update those docs in place, and resolve known-issues rows the code has verifiably fixed. Invoke after editing the pipeline, adding a model/agent/task/metric,…

2026-08-28

run-eval

Software Developers

Run ONE evaluation — a single Task × Model × AgentConfig — end to end, local or on the bastion; invoke when the user wants to run, evaluate, or "kick off" a single eval combo (e.g. "run opa-remediation on gemini-3.1-pro-preview", "evaluate this task once and…

2026-08-28

run-parallel-evals

Software Developers

Run a Task × Model × AgentConfig MATRIX of evals in parallel — local or on the bastion, with opt-in hands-off and unlimited/self-healing modes; invoke when the user wants a matrix, a comparison across models/configs, or to "run these evals in parallel" /…

2026-08-28

devops-bench-review

Software Quality Assurance Analysts & Testers

Use when the user asks for a CODE review of devops-bench changes — e.g. "review this PR", "review my changes", "review the working tree", "is this harness/deployer/metric change sound". Reviews a PR (number/URL) or the current working tree and returns ranked…

2026-08-22

cleanup-orphaned-resources

Network & Computer Systems Administrators

Discover and remove cloud or local resources leaked by aborted or failed eval runs — stale per-run state, leftover kind clusters and their node containers, stuck harness processes, and orphaned GKE clusters, node service accounts, secrets, VPCs and their…

2026-08-21

Showing 8 of 8 collected skills.