New Skill: amd-skill-finder

[andyluo7]

What changed

• adds amd-skill-finder as a native cross-cutting AMD catalog skill
• searches installed skills and the published AMD catalog before routing to source repositories
• adds a curated registry covering official AMD/ROCm projects and reviewed ecosystem projects
• keeps general GitHub discovery opt-in and distinguishes installable skills from source material
• publishes the new skill through the Claude, Codex, Cursor, and agents manifests

Why

AMD's catalog needs a stable capability detector similar to NVIDIA's skill finder,
while also covering AMD-specific knowledge that currently lives in ROCm and selected
upstream repositories rather than packaged skills.

The finder uses this trust order:

1. installed skills
2. published amd/skills catalog entries
3. official AMD, ROCm, and AMD-AGI repositories
4. reviewed upstream repositories
5. unreviewed GitHub fallback only when explicitly enabled

Repositories remain non-installable source results unless they are published as
validated AMD catalog skills. PyTorch, JAX, and Triton AMD forks are grouped with
their upstream repositories instead of producing duplicate recommendations.

The registry includes Quark, MORI, Hyperloom, PyTorch, JAX, Triton, TorchTitan,
LMCache, Mooncake, TileLang, NIXL, Miles, VERL, VIME, vLLM, SGLang, AITER, ATOM,
and related ROCm sources. Quark and quantization vocabulary are explicit discovery
signals, with a deterministic PTQ routing test. ai-dynamo/dynamo is explicitly
excluded; the separately requested ai-dynamo/nixl source remains included.

Validation

• UV_SYSTEM_CERTS=true bash .github/scripts/check.sh
• python3 -m unittest skills/amd-skill-finder/tests/test_find_skills.py (16 tests)
• uvx --system-certs ruff check skills/amd-skill-finder
• uvx --system-certs ruff format --check ... for finder Python
• git diff --check origin/main...HEAD

Behavioral agent eval definitions are included under skills/amd-skill-finder/evals/.
On this fork PR, ANTHROPIC_API_KEY and the AMD API subscription-key header are empty;
the Claude CLI therefore reports Not logged in before either behavioral assertion runs.
An AMD maintainer must rerun the behavioral gate from a trusted context with repository
secrets. The deterministic unit and repository validation suites pass.

This PR now contains only amd-skill-finder. The Quark skill family is submitted
independently in #122, as requested in the review discussion.

[danielholanda]

@andyluo7 Can you please split this into two PRs? One for amd-skill-finder and another one for quark-*?

[amd-zhaofeng]

This is a useful skill — it turns "look for an existing skill first" into an explicit path. We are on the AMD Quark team and have been reading the finder end to end while working out how Quark's own skills should join the catalog. From that angle there is one thing we would like to raise, only so that the quantization path actually gets walked:

Neither the description nor amd_signals carries any quantization vocabulary. The frontmatter trigger list names vLLM, SGLang, AITER, MORI, TileLang and a dozen other projects, but never Quark, quantization, or any precision format; amd_signals likewise has no quark, so a Quark-only question misses the +8 that score_project gives AMD-official sources. The Quark entry in data/sources.json is already thorough, but it is hard to reach on exactly those queries. Suggest adding Quark and the quantization vocabulary to both the description and amd_signals, plus a deterministic routing test as the policy document itself requires — Quark is currently absent from the 13 projects asserted in test_requested_projects_and_canonical_repositories_are_present.

[andyluo7]

Thanks—addressed in cbbaa02. I added AMD Quark and quantization vocabulary to the skill trigger description; added quark and amd quark as AMD signals while keeping generic precision terms out of the global AMD-signal list to avoid false AMD routing; and added both the amd/Quark canonical-repository assertion and a deterministic Quark PTQ routing test. The finder now passes all 16 unit tests and the repository check.

[amd-zhaofeng]

Thanks—addressed in cbbaa02. I added AMD Quark and quantization vocabulary to the skill trigger description; added quark and amd quark as AMD signals while keeping generic precision terms out of the global AMD-signal list to avoid false AMD routing; and added both the amd/Quark canonical-repository assertion and a deterministic Quark PTQ routing test.

Thanks — I tried this in an agent from our side and it works. On quantization phrasings that don't name toolkit, the agent loads the finder and comes back with amd/Quark, so the registry entry is now actually reachable.

As far as this skill goes, there is nothing further we need to ask for on the Quark side. We'll take up the integration details for the federated Quark skill family in #122 instead.