TAO ChangeNet Root Cause Analyzer

The skill limits declared tools to Read and Bash and scopes inputs to experiment results, training code, datasets, and KPI definitions; no malware, credential theft, or covert exfiltration is evident. However, Bash is broad, execution confirmation, least-privilege path controls, sensitive-data handling, rollback, and cleanup are unspecified. Hooks also copy session transcripts and configuration, with insufficient data-flow disclosure and consent controls, so substantial points are deducted.

SKILL.md defines a five-phase workflow, report structure, and several validation checks, while BENCHMARK.md records limited evaluation results. The key path nevertheless depends on unprovided reference files, Docker, NVIDIA Container Toolkit, and an Agent tool not declared in the available tool list. Missing-input, malformed-data, and subagent-failure diagnostics are underspecified, and static review cannot establish reproducibility, so the score remains moderate-low.

The audience, ChangeNet/AOI RCA scenarios, input directories, KPI choices, and trigger phrases are relatively clear. The core work uses local images and directories rather than an overseas service, so no extra mainland-network deduction is warranted. Boundaries for non-fit cases, permission prerequisites, Chinese-language output, and variant data layouts are not sufficiently defined, so points are deducted.

Metadata supplies a name, Apache-2.0 license, author, and version, and the README and BENCHMARK provide some repository and evaluation context. Standard Instructions, Examples, and FAQs are absent; dependency installation, maintenance ownership, update path, changelog, and the referenced reference files are not fully supplied. Static validation also reports folder-depth, author-format, and hooks-structure issues, so points are deducted.

The workflow covers score analysis, image review, configuration analysis, counterfactuals, and remediation, and could theoretically produce a structured RCA_Report.md. The limited benchmark report claims strong task performance. However, completion depends heavily on real experiment directories, images, CSVs, configurations, and usable subagents; the mandatory volume of tables and thumbnails is costly, and no representative final report is independently verifiable. The score therefore remains at the static-review ceiling range.

The skill requires major conclusions to trace to specific images, and hooks provide auditable checks for report content, thumbnails, and defect coverage. evals.json also defines one concrete test task. There is no complete committed test suite, CI execution chain, or actual RCA output in the supplied material, and the benchmark results cannot be independently reproduced from this static review. Points are therefore deducted and confidence remains low.