chore(amdsmi): use the amd-smi review agents in Copilot code review

.github/CODEOWNERS

@@ -2,6 +2,7 @@
 /.azuredevops/                                                        @ROCm/external-ci
 /.github/                                                             @ROCm/rocm-systems-reviewers
 /.github/workflows/amdsmi-*.yml                                       @ROCm/smi-reviewers
+/.github/instructions/amdsmi.instructions.md                          @ROCm/smi-reviewers
 /.github/workflows/hipfile-*.yml                                      @ROCm/hipfile-reviewers
 /.github/workflows/rocjitsu-*.yml                                     @ROCm/rocjitsu-core-team
 /.github/workflows/*rccl*                                             @ROCm/rccl-ci

.github/instructions/amdsmi.instructions.md

@@ -0,0 +1,130 @@
+---
+applyTo: "projects/amdsmi/**"
+excludeAgent: "cloud-agent"
+---
+
+<!--
+Condensed from the amd-smi review agents (projects/amdsmi/.github/agents/,
+mainly amdsmi-review-maintainer.agent.md) and
+projects/amdsmi/.claude/rules/output-conventions.md. Copilot code review only
+reads instruction files from the repository root, so keep this file in sync
+with those sources.
+-->
+
+# amd-smi review rules
+
+Apply these rules, in addition to your normal review, to changes under
+`projects/amdsmi/`. Check correctness against the code first, then the
+conventions.
+
+## 1. Claims match the code
+
+- Check every claim in a CHANGELOG entry, doc, comment, help text and the PR
+  description against the diff and the current code. Flag a fix, feature, test
+  or behavior that the code does not contain, and docs or comments that
+  contradict the code.
+- A fix states its root cause and mechanism. Removing an assert, relaxing a
+  check, adding a test skip or dropping a CLI command is not a fix unless the
+  cause is understood.
+- A fix for specific hardware shows before/after output from that hardware in
+  the PR description. "CI will cover it" is not validation.
+
+## 2. Status codes and N/A values
+
+- Never return `AMDSMI_STATUS_SUCCESS` with partial, default or sentinel data.
+  Flag ignored callee statuses, aggregates that silently drop a failed member,
+  unhandled enum values and ignored parameters.
+- Each errno maps to the status that means it. `ErrnoToRsmiStatus` maps
+  `EPERM`, `ENOENT` and `ENOTSUP` to `NOT_SUPPORTED`, so a failed read can pass
+  for "not supported".
+- A new field reports N/A as all-ones for its width (`UINT32_MAX`,
+  `UINT64_MAX`), never 0. An existing field keeps the N/A its header documents.
+  The same field shows the same N/A in C, Python and the CLI.
+- Skip-and-continue paths log what was skipped and why (index, BDF, status).
+
+## 3. Public API and ABI
+
+- Never change the meaning, units or sign of an existing field, enum value or
+  status in `include/amd_smi/amdsmi.h`.
+- Public structs keep their size and layout: a new field goes at the end and
+  takes its bytes from `reserved`; a new enum value goes before the
+  `*_LAST`/`__MAX` sentinel. Removing or renaming a public function, field,
+  enum value or CLI option needs a deprecation first.
+- A public `amdsmi_*` function calls `AMDSMI_CHECK_INIT()`, rejects null
+  required pointers and out-of-range arguments with `AMDSMI_STATUS_INVAL`
+  before any support probe, and catches exceptions from throwing calls
+  (`std::stoi`, `.at()`, `std::filesystem`) so none escapes the C API. No
+  `assert()` for validation.
+- Hardware and sysfs logic lives in the library, not the CLI. The library never
+  spawns a process (`popen`, `system`).
+- Test hooks and internal helpers stay out of the exported `amdsmi_*` symbols
+  and the installed headers.
+- A changed API updates every layer it reaches: header docs, implementation,
+  Python interface, CLI, Rust and Go bindings, docs.
+  `py-interface/amdsmi_wrapper.py` and the Rust wrapper are generated
+  (`tools/update_wrapper.sh`, `tools/update_rust_wrapper.sh`); flag hand edits.
+
+## 4. Concurrency and lifetime
+
+- For every global or static container, the lock held at each access must cover
+  every device that shares it. A per-device mutex guarding a process-wide map
+  is a data race.
+- No check-then-act across a dropped lock. Lazy singletons use a function-local
+  static or `std::call_once`.
+- Every retry loop and wait is bounded.
+
+## 5. Scope
+
+- Fix the stated bug on its own code path. Flag edits to paths the bug never
+  reaches, drive-by cleanups and changes to other projects.
+- Look for the same defect in siblings (the twin getter, other metric fields,
+  the CPU or NIC variant) and ask for them to be fixed or listed as follow-ups.
+- No hardcoded ASIC names, device IDs, counts, or firmware and metrics
+  versions: use what the driver reports, and keep older drivers working.
+
+## 6. Tests
+
+- Every new guard, branch or early return that changes a public API result has
+  a test that fails without the fix, at the public entry point, asserting the
+  exact status or exit code.
+- Flag assertions that cannot fail (expected values computed by the code under
+  test, unsigned `>= 0`, `ASSERT_*` inside a `void` lambda) and accept lists
+  that take any status. Python tests use `expect_status()` with exact statuses.
+- Skips use `GTEST_SKIP()` or `skipTest()` with a reason, gated on a
+  capability, never on a status any ASIC can return.
+- Unit tests fake every sysfs read. Tests restore any device setting,
+  environment variable or gtest flag they change.
+
+## 7. C++ and Python
+
+- C++: RAII and standard containers instead of raw `new`, C strings and manual
+  `close`; explicit signed/unsigned and narrowing conversions; overflow-checked
+  size arithmetic; trim and range-check sysfs text before use; initialize every
+  local and out-struct; `static_assert` on `sizeof` for structs copied between
+  `amdsmi_*` and `rsmi_*` types; `LOG_*` macros, never `std::cout`, in the
+  library; reuse `amd_smi_utils` and `rocm_smi_utils` helpers; `amd_smi` may
+  include `rocm_smi`, not the reverse.
+- Python: shipped code (`py-interface/`, `amdsmi_cli/`) runs on Python 3.6. CLI
+  changes work in default, `--json`, `--csv` and `--file` output. Each library
+  call in the CLI has its own `try`/`except`, so one failure leaves only that
+  value N/A; only the top level exits.
+- No dead code, debug prints, commented-out code or TODOs.
+
+## 8. Comments, docs and CHANGELOG.md
+
+- Comments explain why, not what, in one or two lines.
+- No ticket IDs (`ROCM-12345`, `SWDEV-123456`, `AILITOOLS-123`) or internal
+  labels ("regression guard", sprint names) in code, comments, docs or
+  CHANGELOG.md. Tickets belong only in the PR's Issue Tracking section.
+- CHANGELOG.md entries are user-facing: a short bold summary plus at most one
+  bullet on what changed for users. Flag internal function, class, member and
+  variable names, enum constants, signal or abort details and file internals.
+  Public API and CLI names are fine. Changes users can't observe (tests, CI)
+  get no entry.
+- `amdsmi.h` docs state units, required permissions, every returned status and
+  the N/A value.
+
+## Not worth a comment
+
+- Formatting and spelling: clang-format, ruff-format, codespell and the gtest
+  suite-name check run in pre-commit and CI.