Skip to content

[DRAFT] Add Claude skills for QEC and solvers#530

Draft
kvmto wants to merge 12 commits into
NVIDIA:mainfrom
kvmto:qx_skills
Draft

[DRAFT] Add Claude skills for QEC and solvers#530
kvmto wants to merge 12 commits into
NVIDIA:mainfrom
kvmto:qx_skills

Conversation

@kvmto
Copy link
Copy Markdown
Collaborator

@kvmto kvmto commented May 5, 2026

Summary

Adds an agent-skills framework for CUDA-QX (modeled on cuda-quantum), with multi-agent mirroring and an evaluation harness.

  • 14 skills under .agents/skills/ — cross-cutting (quickstart, build, benchmarking, profiling, testing, contributing, skills-authoring), QEC family (decode, realtime, ai-decoders, extending), solvers family (algorithms, chemistry, extending)
  • Multi-agent: .agents/skills/ is canonical (Codex reads natively); scripts/sync_agents_skills.sh mirrors to .claude/skills/ and .cursor/skills/ (gitignored)
  • Evals in .agents/evals/: prompts, assertions, three graders (programmatic / executable / LLM-judge), Cohen's κ aggregation, HTML viewer
  • AGENTS.md entry point; shared _shared/scripts/ (preflight, pick_workflow, import_smoke); new top-level helpers (build_one_lib.sh, clean.sh, doctor.sh)

Testing

  • Markdown-only change
  • Checked editor diagnostics for the new files; no issues reported

How to try this PR locally

bash scripts/sync_agents_skills.sh

@vedika-saravanan vedika-saravanan marked this pull request as draft May 5, 2026 19:35
kvmto and others added 5 commits May 12, 2026 22:40
Signed-off-by: kvmto <kmato@nvidia.com>
Signed-off-by: vedika-saravanan <vsaravanan@nvidia.com>
Signed-off-by: kvmto <kmato@nvidia.com>
Signed-off-by: kvmto <kmato@nvidia.com>
CI's PR sanity check (yapf) flagged 8 reformatting changes in the new
skill-eval pipeline. Reformat in place using the repo's
`scripts/run_yapf_format.sh` rules (`based_on_style = google`).

Also gitignore `.claude/evals/workspaces/` so per-iteration eval
artifacts (`responses.json`, `grading.*.json`, generated reports) don't
land in commits.

No functional changes; pipeline smoke-tested end-to-end after reformat
(preflight, import_smoke, pick_workflow, runner, programmatic grader,
aggregate, viewer).

Signed-off-by: kvmto <kmato@nvidia.com>
Signed-off-by: vedika-saravanan <vsaravanan@nvidia.com>
…s, shellcheck pass, fix latent SC2259 in submodule pipeline)

Signed-off-by: vedika-saravanan <vsaravanan@nvidia.com>
Signed-off-by: vedika-saravanan <vsaravanan@nvidia.com>
Signed-off-by: vedika-saravanan <vsaravanan@nvidia.com>
Signed-off-by: vedika-saravanan <vsaravanan@nvidia.com>
Signed-off-by: vedika-saravanan <vsaravanan@nvidia.com>
Rename skills to the cudaq-* namespace, fold test/CI guidance into contributing, add public-safe QEC triage references, and expand the eval harness coverage. Track the generated Claude and Cursor skill mirrors so fresh clones work without requiring users to run the sync script manually.

Signed-off-by: kvmto <kmato@nvidia.com>
@kvmto kvmto requested a review from caldwellshane June 3, 2026 15:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants