NemoClaw: Skill for Claude Code

Classify a CI Failure

When the failed job is part of E2E, first load Audit E2E Assertions. This classifier can locate the failure signature. Its bounded output does not replace the complete log review and itemized source audit required before an E2E repair push or expensive rerun.

Run the classifier from a NemoClaw checkout:

node --no-warnings \
  .agents/skills/nemoclaw-maintainer-classify-ci-failure/scripts/classify-ci-failure.mts \
  --workdir "$PWD" --job-id <job-id>

The repository is fixed to NVIDIA/NemoClaw. Optional flags include --artifact-name. --max-lines accepts integers from 1 through 500 and defaults to 120. --clip-mode accepts head or tail and defaults to tail.

The script uses authenticated gh reads. It performs no GitHub writes. It bounds and redacts log output. When an artifact is selected, it bounds pagination, compressed and expanded sizes, entries, paths, file reads, and reported failures. It rejects malformed ZIPs, links, special files, duplicate or unsafe paths, ambiguous listings, and size mismatches. Temporary files use process-owned directories under a fixed private root. The classifier attempts to remove them directly through the filesystem after success or failure. If removal fails, the classifier exits nonzero. It reports a bounded, redacted diagnostic with a direct removal command. Each command runs beneath a stable detached group leader that stays alive until every group descendant exits. This is an internal trusted-subprocess boundary: the classifier invokes Bash and GNU coreutils by fixed paths. It finds gh only in /usr/bin, /usr/local/bin, or $HOME/.local/bin after it validates the path, ownership, permissions, and file type. It treats artifacts as data and never executes artifact content or caller-selected programs. On SIGHUP, SIGINT, or SIGTERM, the classifier rejects new commands, terminates and drains every owned group, synchronously removes tracked directories, and exits with the conventional cancellation code. If normal cleanup fails, it reports a bounded, redacted error and a removal command that remains valid if TMPDIR changes.

Run the script from a Linux NemoClaw checkout with Node.js 22.19 or later, /proc, Bash, authenticated gh in a trusted location, and GNU coreutils (dd, stat, tail, and wc).

Any nonzero log acquisition result is a classifier failure. After the cleanup attempt, the script exits nonzero with no success JSON and reports a bounded, redacted diagnostic. On GitHub authentication or authorization failure, stop, follow the repository GitHub access hard stop, and ask the user to correct the configured gh access (including SSO or token scope) before rerunning. Treat unclassified as bounded evidence, not proof that no known cause exists.