NVIDIA · rgsl888prabhu · Jan 14, 2026 · Jan 14, 2026 · Jan 16, 2026 · Jan 16, 2026
diff --git a/.ai/skills/cuopt.yaml b/.ai/skills/cuopt.yaml
@@ -0,0 +1,9 @@
+# cuOpt agent skill manifest (shim)
+#
+# Canonical location:
+# - .github/.ai/skills/cuopt.yaml
+#
+# Many tools scan for `.ai/skills/*` at repo root; this file keeps those tools working.
+manifest_version: 1
+
+canonical: .github/.ai/skills/cuopt.yaml
@@ -0,0 +1,85 @@
+# cuOpt agent skill manifest (machine-readable)
+manifest_version: 1
+
+name: cuOpt
+vendor: NVIDIA
+summary: GPU-accelerated optimization engine for Routing (TSP/VRP/PDP) and Math Optimization (LP/MILP/QP).
+
+capabilities:
+  routing:
+    problems: [TSP, VRP, PDP]
+  mathematical_optimization:
+    problems: [LP, MILP, QP]
+    notes:
+      - QP support exists in the Python API and is currently documented as beta.
+
+compute:
+  accelerator: NVIDIA_GPU
+  requirements:
+    - CUDA 12.0+ (or CUDA 13.0+ depending on package)
+    - NVIDIA driver compatible with CUDA runtime
+    - Compute Capability >= 7.0 (Volta+)
+
+interfaces:
+  - name: python
+    kind: library
+    supports:
+      routing: true
+      lp: true
+      milp: true
+      qp: true
+    packages:
+      - cuopt-cu12
+      - cuopt-cu13
+    docs:
+      - docs/cuopt/source/cuopt-python/quick-start.rst
+      - docs/cuopt/source/cuopt-python/routing/index.rst
+      - docs/cuopt/source/cuopt-python/lp-qp-milp/index.rst
+
+  - name: c_api
+    kind: library
+    library: libcuopt
+    supports:
+      routing: false
+      lp: true
+      milp: true
+      qp: true
+    docs:
+      - docs/cuopt/source/cuopt-c/index.rst
+
+  - name: server_rest
+    kind: service
+    protocol: HTTP
+    supports:
+      routing: true
+      lp: true
+      milp: true
+      qp: false
+    implementation:
+      - python/cuopt_server/cuopt_server/webserver.py
+    openapi:
+      spec_file: docs/cuopt/source/cuopt_spec.yaml
+      served_path: /cuopt.yaml
+      ui_paths:
+        - /cuopt/docs
+        - /cuopt/redoc
+    docs:
+      - docs/cuopt/source/cuopt-server/quick-start.rst
+      - docs/cuopt/source/cuopt-server/server-api/index.rst
+      - docs/cuopt/source/open-api.rst
+
+roles:
+  cuopt_user:
+    description: Use cuOpt to solve routing and math optimization problems without modifying cuOpt internals.
+    rules:
+      - .github/agents/cuopt-user.md
+  cuopt_developer:
+    description: Develop and maintain cuOpt (C++/CUDA/Python/server/docs/CI).
+    rules:
+      - .github/agents/cuopt-developer.md
+      - .github/AGENTS.md
+
+entrypoints:
+  human_router:
+    - .github/AGENTS.md
+    - AGENTS.md
@@ -1,156 +1,15 @@
 # AGENTS.md - AI Coding Agent Guidelines for cuOpt
 
-> This file provides essential context for AI coding assistants (Codex, Cursor, GitHub Copilot, etc.) working with the NVIDIA cuOpt codebase.
+This file is intentionally **minimal**. Choose a role and follow the matching rules:
 
-> **For setup, building, testing, and contribution guidelines, see [CONTRIBUTING.md](../CONTRIBUTING.md).**
+- **Using cuOpt (model + solve problems)**: `.github/agents/cuopt-user.md`
+- **Developing cuOpt (changing this repo)**: `.github/agents/cuopt-developer.md`
 
----
+### Machine-readable skill manifest
 
-## Project Overview
+- **Primary**: `.github/.ai/skills/cuopt.yaml`
+- **Root shim (compat)**: `.ai/skills/cuopt.yaml`
 
-**cuOpt** is NVIDIA's GPU-accelerated optimization engine for:
-- **Mixed Integer Linear Programming (MILP)**
-- **Linear Programming (LP)**
-- **Quadratic Programming (QP)**
-- **Vehicle Routing Problems (VRP)** including TSP and PDP
+### Canonical contribution guide
 
-### Architecture
-
-```
-cuopt/
-├── cpp/                    # Core C++ engine (libcuopt, libmps_parser)
-│   ├── include/cuopt/      # Public C/C++ headers
-│   ├── src/                # Implementation (CUDA kernels, algorithms)
-│   └── tests/              # C++ unit tests (gtest)
-├── python/
-│   ├── cuopt/              # Python bindings and routing API
-│   ├── cuopt_server/       # REST API server
-│   ├── cuopt_self_hosted/  # Self-hosted deployment utilities
-│   └── libcuopt/           # Python wrapper for C library
-├── ci/                     # CI/CD scripts and Docker configurations
-├── conda/                  # Conda recipes and environment files
-├── docs/                   # Documentation source
-├── datasets/               # Test datasets for LP, MIP, routing
-└── notebooks/              # Example Jupyter notebooks
-```
-
-### Supported APIs
-
-| API Type | LP | MILP | QP | Routing |
-|----------|:--:|:----:|:--:|:-------:|
-| C API    | ✓  | ✓    | ✓  | ✗       |
-| C++ API  | ✓  | ✓    | ✓  | ✓       |
-| Python   | ✓  | ✓    | ✓  | ✓       |
-| Server   | ✓  | ✓    | ✗  | ✓       |
-
----
-
-## Coding Style and Conventions
-
-### C++ Naming Conventions
-
-- **Base style**: `snake_case` for all names (except test cases: PascalCase)
-- **Prefixes/Suffixes**:
-  - `d_` → device data variables (e.g., `d_locations_`)
-  - `h_` → host data variables (e.g., `h_data_`)
-  - `_t` → template type parameters (e.g., `i_t`, `value_t`)
-  - `_` → private member variables (e.g., `n_locations_`)
-
-```cpp
-// Example naming pattern
-template <typename i_t>
-class locations_t {
- private:
-  i_t n_locations_{};
-  i_t* d_locations_{};  // device pointer
-  i_t* h_locations_{};  // host pointer
-};
-```
-
-### File Extensions
-
-| Extension | Usage |
-|-----------|-------|
-| `.hpp`    | C++ headers |
-| `.cpp`    | C++ source |
-| `.cu`     | CUDA C++ source (nvcc required) |
-| `.cuh`    | CUDA headers with device code |
-
-### Include Order
-
-1. Local headers
-2. RAPIDS headers
-3. Related libraries
-4. Dependencies
-5. STL
-
-### Python Style
-
-- Follow PEP 8
-- Use type hints where applicable
-- Tests use `pytest` framework
-
-### Formatting
-
-- **C++**: Enforced by `clang-format` (config: `cpp/.clang-format`)
-- **Python**: Enforced via pre-commit hooks
-- See [CONTRIBUTING.md](../CONTRIBUTING.md) for pre-commit setup
-
----
-
-## Error Handling Patterns
-
-### Runtime Assertions
-
-```cpp
-// Use CUOPT_EXPECTS for runtime checks
-CUOPT_EXPECTS(lhs.type() == rhs.type(), "Column type mismatch");
-
-// Use CUOPT_FAIL for unreachable code paths
-CUOPT_FAIL("This code path should not be reached.");
-```
-
-### CUDA Error Checking
-
-```cpp
-// Always wrap CUDA calls
-RAFT_CUDA_TRY(cudaMemcpy(&dst, &src, num_bytes));
-```
-
----
-
-## Memory Management Guidelines
-
-- **Never use raw `new`/`delete`** - Use RMM allocators
-- **Prefer `rmm::device_uvector<T>`** for device memory
-- **All operations should be stream-ordered** - Accept `cuda_stream_view`
-- **Views (`*_view` suffix) are non-owning** - Don't manage their lifetime
-
----
-
-## Key Files Reference
-
-| Purpose | Location |
-|---------|----------|
-| Main build script | `build.sh` |
-| Dependencies | `dependencies.yaml` |
-| C++ formatting | `cpp/.clang-format` |
-| Conda environments | `conda/environments/` |
-| Test data download | `datasets/get_test_data.sh` |
-| CI configuration | `ci/` |
-| Version info | `VERSION` |
-
----
-
-## Common Pitfalls
-
-| Problem | Solution |
-|---------|----------|
-| Cython changes not reflected | Rerun: `./build.sh cuopt` |
-| Missing `nvcc` | Set `$CUDACXX` or add CUDA to `$PATH` |
-| CUDA out of memory | Reduce problem size or use streaming |
-| Slow debug library loading | Device symbols cause delay; use selectively |
-
----
-
-*For detailed setup, build instructions, testing workflows, debugging, and contribution guidelines, see [CONTRIBUTING.md](../CONTRIBUTING.md).*
+- `CONTRIBUTING.md`