[faranm-stack]

Summary

Follow-up fixes for microsoft#212 after #1 merged: update dependencies with high or critical advisories reviewed after the previous update (certifi, pillow, GitPython, aiohttp, filelock, hydra-core), and make python /app/release_smoke.py run unchanged in the deployed image.

Type of change

• New agent (under agents/<name>/)
• Update to an existing agent
• New starter kit (under starter-kits/<name>/)
• Update to an existing starter kit
• Schema change (docs/schemas/**)
• Workflow / script change (.github/**)
• Documentation only
• Other (describe below)

Related issue / tracking

refs microsoft#212; follows #1.

Schema impact

• No schema changes
• Schema changes are backward compatible
• Schema changes are breaking — add the breaking-change label and link the rollout plan below

Validation checklist

• Documentation (README.md, schemas) is updated to match the change.
• No hand-edits to .auto-registry/** (it is generated post-merge).
• No OS / editor artefacts committed (.DS_Store, .env, *.swp, .idea/, .vs/, etc.).
• If model-weight files are added, they are Git-LFS tracked, ≤5 GB each, and in an allowed format. (No model-weight files added.)
• No secrets, customer data, or internal-only URLs are committed.
• All Markdown links resolve.

Reviewer notes (optional)

1. certifi (Dependency Review failure). After #1 merged, Dependency Review still failed on one package: certifi==2026.7.22 is affected by GHSA-6q48-6c7w-9gw2 (critical; compromised emSign intermediate CA keys), published 2026-10-20. Bumped to the first patched release, 2026.10.19. The four packages updated in #1 no longer appear.

2. Release smoke import path (bug introduced in #1). test_download_checkpoints.py imported the downloader from /app/build, which is intentionally not on the runtime PYTHONPATH. The image-build test step added it explicitly, so the build passed, but python /app/release_smoke.py in the deployed image would report one import error and fail the regression gate. The test now locates the downloader itself (repository or /app/build layout), and the image-build step runs with the runtime PYTHONPATH so this cannot regress silently.

Validation on this exact commit:

• Locally: the merged test errors in an image-layout simulation with runtime PYTHONPATH only; the fixed test passes there and in the repository layout.
• Clean image build: 43/43 tests passed with no skips using the runtime PYTHONPATH; pip check passed; certifi-2026.10.19 installed.
• agent.yaml, tool.yaml and all runtime Python modules are unchanged from nvidia-alchemi: hardening fixes and H100 revalidation for #212 #1, so the H100 model and dynamics results from nvidia-alchemi: hardening fixes and H100 revalidation for #212 #1 still apply. A GPU rerun of the release validation on this image is planned.

3. Dependency advisories reviewed after the previous update. A check of all 198 pins against GitHub-reviewed advisories found high or critical advisories affecting these pinned versions. Dependency Review fails at high severity, so these would block microsoft#212 once its advisory data catches up:

Package	Was	Now	Reviewed advisories
pillow	12.3.0	13.0.0	28 high/critical (5 critical)
GitPython	3.1.62	3.2.1	5 high
aiohttp	3.14.3	3.14.4	3 high
filelock	3.32.3	4.2.1	1 high
hydra-core	1.3.7	1.3.8	1 high
setuptools	84.0.0	84.1.0	1 low
yarl	1.24.5	1.25.1	none; required by aiohttp 3.14.4

Validation on commit cee6af3: an offline pip check of every pin's declared requirements finds no conflicts; a clean image build installs all versions above, pip check passes, and all 43 tests pass with no skips; no pinned package has a known advisory. A GPU rerun of the release validation on this image is planned because pillow and filelock change major versions.

New advisories are being published against these packages daily; enabling Dependabot for requirements-build.txt would keep the pins current.