openvino

Adds a core operator to the OpenVINO toolkit. Use when asked to implement a new operation into OpenVINO.

Add a new operation to the OpenVINO CPU plugin — node registration, JIT/oneDNN executors (AVX2/AVX-512/AMX), and functional tests.

Adds a new operation to OpenVINO Frontend pipelines with translator updates, registration, and tests.

Adds a new OpenVINO fusion transformation (subgraph to one or several operations) and corresponding tests.

Add a new operation to the OpenVINO GPU plugin — OpenCL kernel design, oneDNN-backed paths, sub-group/LWS tuning, and functional tests.

Analyze a HuggingFace model and attempt OpenVINO conversion — probe properties, run strategy matrix, classify failures, and produce a structured routing report.

Review OpenVINO changes for correctness, compatibility, performance, security, testing, and maintainability. Use when reviewing a pull request or proposed diff.

Investigate and fix model conversion issues in OpenVINO Frontends (ONNX, PyTorch) — triage, debugging, accuracy comparison, and pre-submission verification.

Author, edit, compile, validate, and debug any OpenVINO gh-aw agentic workflow (e.g. ci-doctor, ci-doctor-mq, and any future ones) and their shared safe-output jobs under .github/workflows/shared/agentic-workflows. Use when a user wants to create a new agentic workflow; change an existing one's prompt, triggers, tools, permissions, or safe-outputs; add or edit a shared job/step; recompile a .lock.yml; update a workflow's persisted-state JSON schemas; or test a workflow via workflow_dispatch. Do NOT use for regular (non-agentic) GitHub Actions workflows, plugin/runtime code, or diagnosing a specific live CI failure.

Debug OpenVINO CPU/GPU plugin inference, accuracy, performance, memory, and device compilation issues, or inspect graph transformations, using debug capabilities such as tensor dumps, execution graphs, profiling, and transformation tracing. Do NOT use for frontend model conversion failures (unsupported operators, translation errors, or source-model loading), build/configuration failures, or generic test/CI failures.

Debug why an OpenVINO MatcherPass transformation is not firing. Use this skill immediately when a user says a transformation is "not applied", a "pass has no effect", a "matcher never triggers", a pattern "doesn't match", a "callback never fires", "WrapType predicate is too strict", a subgraph "not fused" despite the pass being registered, or they see "END: PATTERN DIDN'T MATCH" in matcher logs. Also trigger when a MatcherPass works on one model but silently skips another, when the user wants to add a reproducer test for a transformation that should fire but doesn't, or when they suspect an opset version mismatch preventing a match. Do NOT trigger for: writing a new MatcherPass from scratch, debugging a pass that fires but produces wrong numerical results, crashes in pass registration, or general questions about what MatcherPass is.

Detect and fix clang-format, clang-tidy, and copyright header violations in an OpenVINO C++ codebase. Use when the user complains about code style or formatting, asks to clean up changes, fix linting, add a copyright header, or when a style check or linting CI job is failing. Do not use for build errors, compilation failures, linker errors, test failures, runtime crashes, accuracy issues, or CMake config problems.

Develop and troubleshoot the OpenVINO GGUF frontend (src/frontends/gguf) and the llama.cpp ggml-openvino backend that uses it. Use when checking whether a .gguf or mmproj model is supported; when a .gguf file is rejected as an unsupported architecture; when enabling a llama.cpp model family (llama, qwen, gemma, phi, MoE, hybrid/recurrent) through arch_registry, DecoderBuilder, a ModelBuilder or an ArchitectureExtension; when adding a vision/audio projector or ProjectorExtension; when conversion fails with "Translation for operation type GGML_OP_* is not implemented" or an op translator/op_case in src/op/ needs work; when making a converted model stateful or GenAI/PagedAttention-ready; or when a GGUF model converts but gives garbage, drifting tokens, a llama.cpp mismatch or a decode-only shape failure; or when running or resuming GGUF frontend, backend and GenAI validation batches. Not for general llama.cpp usage or unrelated build/CI failures.

Author, edit, debug, and review the OpenVINO GitHub Actions CI infrastructure — regular (non-agentic) workflows under .github/workflows, reusable job_*.yml workflows, custom composite actions under .github/actions, CI helper scripts under .github/scripts, Dockerfiles under .github/dockerfiles, and the Smart CI / labeler / components configuration. Use when a user wants to add or change a build/test job or step, create or modify a reusable workflow, write or fix a custom action, adjust runners/containers/caches, wire up Smart CI for a component, pin action versions, fix workflow permissions/security, or debug a failing CI workflow's YAML. Do NOT use for gh-aw agentic workflows (*.md with gh-aw frontmatter — use ov-agentic-workflows), for diagnosing a specific product/test failure's root cause in C++/Python code, or for non-CI GitHub configuration.

Write unit tests for OpenVINO ov::Model graph transformations (passes). Use when the user asks to write, add, or refactor tests for a transformation pass (MatcherPass, ModelPass), or to modernize legacy transformation tests.

Upgrade the PyTorch version used by OpenVINO tests (torch / torchvision / torchaudio) and resolve fallout — missing operator translators, new functionalized `*_copy` aten ops, decomposition changes, FX-only tests failing in TorchScript mode, and accuracy regressions caused by stricter typing. Use when the user asks to "bump torch", "update pytorch to X.Y", "upgrade torch tests", or when pytorch_tests / model_hub pytorch tests fail after a torch version change. Do not use for: enabling a single new PyTorch operator unrelated to a version bump, GenAI / Optimum upgrades, or plugin-level numerical bugs unrelated to the frontend.

Write or review OpenVINO ov::Model graph transformations (passes: MatcherPass, ModelPass, GraphRewrite, BackwardGraphRewrite). Use when the user asks to write, add, or refactor a transformation/fusion/decomposition pass, choose between MatcherPass and ModelPass, write pattern-matching predicates, or bring an existing pass in line with OpenVINO transformation conventions.

Install Python dependencies before running verification or test steps. Choose the correct path depending on whether the agent builds OpenVINO from source or not.

Optionally create a draft PR to the upstream repo from a local source directory. Used by coding agents when a local source path is available in the context file.

E2E gate — verifies that applied patches produce a working, numerically sane end-to-end inference through the OpenVINO plugin. Handles HuggingFace/optimum-intel, native OV conversion (ovc/convert_model), and ONNX. Used by orchestrators as the mandatory gate before any PR is published.