TAO VCN Sample Router
Routes visual-defect gap samples to eligible mining and synthetic-anomaly modules.
The skill limits declared tools to Read/Bash and mainly reads Parquet/CSV files while writing routing artifacts; no malware or credential theft is evident. However, it performs local writes and creates directories, while hooks copy session transcripts and configuration into the output and write /tmp/routing-hook-debug.log, without explicit confirmation, sensitive-data handling, isolation, or rollback, so points are deducted.
The workflow and two isin filters are internally coherent, and hooks check artifacts, schemas, coverage, and empty subsets. However, pandas, a Parquet engine, and CSV schema requirements are not declared; handling is thin for empty CSVs, missing label columns, empty inputs, and unspecified path variables. Static calibration prevents a score above 10 because execution was not performed.
Trigger phrases, input fields, output artifacts, label sets, and unrouted cases are clearly described, with case-insensitive matching. Environment boundaries, Python/pandas/Parquet dependencies, and noncanonical labels are less clear, and the declared standalone compatibility conflicts somewhat with the actual runtime requirements.
The documentation is layered across inputs, method, steps, outputs, report structure, and execution order, and includes version, author, Apache-2.0 licensing, a code recipe, and limitations. However, the benchmark explicitly reports missing recommended Instructions and Examples sections; the author format is incomplete, changelog and maintenance/update ownership are absent, and the hooks directory was flagged as unexpected structure.
The deterministic routing produces two filtered Parquets and a per-label summary, reducing manual work and improving auditability. However, the core path was not executed during this static review, and report generation and packaging depend on external agent behavior, so direct usability receives only a limited score.
A pinned revision, committed evals, benchmark record, and several validation hooks provide some auditability. However, the eval covers recognition and workflow description rather than real Parquet routing, the benchmark has only one task, and there is no independent third-party execution evidence or corroboration of key dependencies.
- Before running, verify gaps_parquet, source_pool_csv, the output path, and a usable Parquet engine; an empty CSV or missing label column may fail the recipe.
- Hooks may copy session transcripts, .claude configuration, and path information into the output directory; review or disable RCA_HOOKS for sensitive projects and verify log permissions.
- Labels absent from both the source pool and the AnomalyGen set are dropped; inspect routing_summary.txt and the report's dropped-label section before continuing the SDA iteration.
What does this skill do, and when should you use it?
This skill provides deterministic routing for an NVIDIA TAO Visual ChangeNet AOI SDA iteration. After gap analysis, it reads a gaps parquet and creates separate subsets for k-NN Mining and AnomalyGen based on case-insensitive label eligibility. It also writes a per-label routing summary and an auditable report, including labels accepted by neither module. It fits teams that need transparent sample handoff, but it does not run the downstream mining or generation workflows.
Reads a gap-analysis parquet containing filepath and label columns while preserving other columns; uppercases labels for comparison; filters a mining subset using labels found in source_pool_csv; filters an AnomalyGen subset using the default PASS, EXCESS_SOLDER, MISSING, and BRIDGE set; writes mining_gaps.parquet, anomalygen_gaps.parquet, routing_summary.txt, and Routing_Report.md; creates an empty, schema-preserving mining parquet when the source pool is missing or unavailable.
- A VCN AOI SDA engineer needs to route weak samples immediately after gap analysis.
- A data team needs to verify which labels are eligible for real-image neighbor mining from the configured source pool.
- An augmentation team needs to isolate labels supported by the AnomalyGen pipeline.
- A pipeline reviewer needs to identify labels rejected by both modules before continuing an SDA iteration.
What are this skill's strengths and limitations?
- Uses explicit, case-insensitive label rules.
- Preserves the input parquet schema and original label values.
- Produces per-label decisions that expose overlap and dropped labels.
- Handles a missing source pool without creating a schema-mismatched output.
- Does not execute k-NN Mining or AnomalyGen itself.
- Requires local parquet and CSV access plus Python and pandas; no test suite or platform validation is documented in the source.
- The default AnomalyGen set covers only four labels.
- A missing source pool disables the Mining subset entirely.
How do you install this skill?
Install the skill from the NVIDIA/skills collection with: npx skills add nvidia/skills --skill tao-route-visual-changenet-samples --yes. The README does not document a separate manual-copy procedure for this skill.
How do you use this skill?
With the skill installed, ask the agent to “route VCN gap samples” or “split AOI gaps for k-NN mining and AnomalyGen.” Provide gaps_parquet, source_pool_csv, and an output directory; anomalygen_supported_labels may optionally override the default set. The skill writes a timestamped routing_results directory beneath the output location. The source does not define a specific argument-passing interface.