New Skill: rocm-doctor

[volen-silo]

What

Publishes the rocm-doctor skill: diagnoses ROCm / HIP / PyTorch / llama.cpp failures on AMD GPUs (Linux + Windows) against the closed failure-mode catalog now shipped natively in the rocm CLI (ROCm/rocm-cli#39: rocm examine / rocm diagnose / rocm fix). The skill is a thin driver over those commands — it fixes with consent or routes upstream, and never re-implements the probe/catalog/fixes.

Changes

• New skill skills/rocm-doctor/ (SKILL.md, reference.md, skill-card.md), verified against the shipped rocm-core implementation (exit codes, JSON fields, the 15-mode catalog, the 4 auto-applicable fixes).
• Phase 0 install step — the skill installs the rocm CLI (with user consent) when missing, so it's self-sufficient instead of dead-ending.
• Register in .claude-plugin/marketplace.json; regenerate the derived Cursor/Codex/agents manifests.
• README — mark rocm-doctor as in-repo (was _planned_) and genericize the "what is a skill?" example tree.
• Cleanup — remove the superseded staging/rocm-doctor/ copy (old Python-scripts version) and the now-dead skillspector-allow.yml suppressions that pointed at those deleted scripts.

Notes

• In-repo (Path A) skill — no sources.yml / federation entry needed.
• check.sh passes: 0 errors across 7 skills, manifests in sync.

[danielholanda]

@volen-silo Can you please also add a quick walkthrough to this skill? This will make reviewing this much easier.
https://github.com/amd/skills/tree/main/walkthroughs

A good/simple reference is https://github.com/amd/skills/blob/main/walkthroughs/local-ai-use.md

[danielholanda]

@sreeram-11 to provide an initial review

[sreeram-11]

@volen-silo, thank you for adding the rocm-doctor skill!

Tried it out on my Strix Halo laptop (gfx1151, Windows). Overall it works as expected and I'd say it's good to ship. I tested on a clean Windows machine, so the Linux-only fixes and the ones needing real host conditions couldn't be exercised. What I could hit worked correctly.

• What I checked:

  • Real error strings (e.g. hipErrorNoBinaryForGpu, missing vcruntime140_1.dll) matched the correct fix with sensible confidence scores.
  • Junk/unrelated input correctly returned no match and routed upstream, no hallucinated fixes.
  • Confidence tiers work: low-confidence host conditions stay below the match threshold and don't get surfaced as "the answer."

• One thing worth noting: When an app is named only in the error text (Lemonade, Ollama), the CLI defaults to routing to rocm-core. The skill's instructions tell the agent to redirect to the right tracker instead, so that routing depends on the skill layer doing its job, not the CLI. Worked correctly, just good to know that's where the logic lives.

• Takeaway: Working as expected.

[sree-ram-11]

@volen-silo, thank you for adding the rocm-doctor skill!

Tried it out on my Strix Halo laptop (gfx1151, Windows). Overall it works as expected and I'd say it's good to ship. I tested on a clean Windows machine, so the Linux-only fixes and the ones needing real host conditions couldn't be exercised. What I could hit worked correctly.

• What I checked:

  • Real error strings (e.g. hipErrorNoBinaryForGpu, missing vcruntime140_1.dll) matched the correct fix with sensible confidence scores.
  • Junk/unrelated input correctly returned no match and routed upstream, no hallucinated fixes.
  • Confidence tiers work: low-confidence host conditions stay below the match threshold and don't get surfaced as "the answer."

• One thing worth noting: When an app is named only in the error text (Lemonade, Ollama), the CLI defaults to routing to rocm-core. The skill's instructions tell the agent to redirect to the right tracker instead, so that routing depends on the skill layer doing its job, not the CLI. Worked correctly, just good to know that's where the logic lives.

• Takeaway: Working as expected.

[danielholanda]

Please remove all changes to marketplace files. We will add those later to those "showcase" catalogs subject to marketing approval.

[danielholanda]

I really like the behavioral tests here, but it looks like some of those are a bit flaky. Can you take a look at the failing tests?

https://github.com/amd/skills/actions/runs/31060854272/job/92488416742?pr=87

[danielholanda]

Adding @sdevinenamd @meghsat here as a reviewer

[sdevinenamd]

Thank you for creating this skill @volen-silo! It looks solid!

I tested this skill on a Windows machine without ROCm installed. It successfully checked for the ROCm CLI, installed it, and guided me through HIPInfo installation errors as well as ROCm binary installation issues.

I also tried asking about Lemonade-related ROCm errors, and instead of blindly attempting to fix them, it correctly pointed me to the Lemonade documentation and suggested a description for the issue. Additionally, I tested several competitor-related and WSL prompts, and all of them worked as expected.

Suggestions/ questions:

1. I tried asking an unrelated question: "Write Python code for the Fibonacci series and run it on my local Python setup." The skill provided a response and produced the result. I'm not sure whether this is expected behavior or if we should explicitly prevent the skill from answering unrelated questions.
2. Consider including an evaluation script for WSL2. It is currently only mentioned in the skill.md, and adding an eval script could help reinforce that WSL2-related topics are out of scope.
3. I noticed that the skill.md currently pulls the nightly build. I assume this will be changed to use the stable release later?