NVIDIA/SkillEvaluator
Multi-tier framework for evaluating AI agent skills and plugins with quality gates, semantic overlap detection, synthetic evaluation dataset generation, and live agent evaluation that measures how skills affect agent behavior.
The score is jscpd's own --health model: markup, data and text left out, tests kept. The numbers below are the trending scan's.
Most complex files
- src/skillevaluator/tier3/harbor/collector.py2758 cx
- src/skillevaluator/tier3/eval_core/checks.py1768 cx
- src/skillevaluator/tier3/harbor/runner.py1723 cx
- src/skillevaluator/evaluation/tier3_report.py1679 cx
- src/skillevaluator/validators/security.py1588 cx
- src/skillevaluator/tier3/eval_core/plugin_signals.py1536 cx
- src/skillevaluator/tier3/eval_core/plugin_security.py1335 cx
- src/skillevaluator/tier3/harbor/adapter.py1293 cx
- tests/validators/test_security.py1258 cx
- src/skillevaluator/reporting/html.py1198 cx
Dead code 2.06% unused · 60 findings
- src/skillevaluator/tier3/eval_core/plugin_security.pyunused-file
- src/skillevaluator/tier3/harbor/local_environment.pyunused-file
- src/skillevaluator/plugin_dependency_audit.pyunused-file
- src/skillevaluator/validators/claude_plugin_validate.pyunused-file
- src/skillevaluator/tier3/eval_core/runtime_evidence.pyunused-file
- src/skillevaluator/tier1/plugin_checks.pyunused-file
- src/skillevaluator/deduplication/plugin/catalog.pyunused-file
- src/skillevaluator/tier3/harbor/native_agents.pyunused-file
- src/skillevaluator/validators/endpoint_resolution.pyunused-file
- src/skillevaluator/deduplication/plugin/catalog_checks.pyunused-file
The code number leaves out tests, fixtures, docs, examples, data files, styles, translations and vendored code.
Largest clones in the code
- 60 lines in 2 placespython462 tokenssrc/skillevaluator/tier3/eval_core/llm_judge.py:957–1016src/skillevaluator/tier3/harbor/templates/eval.py:848–905
957def _compact_behavior_conversation(conversation_text: str, limit: int | None = None) -> str: 958 """Keep both setup context and late outcome evidence in behavior prompts.""" 959 if limit is None: 960 limit = _behavior_check_budget() 961 if len(conversation_text) <= limit: 962 return conversation_text 963 964 marker = "\n...[middle truncated for behavior check]...\n" 965 if limit <= len(marker) + 1: 966 return conversation_text[:limit]848def _compact_behavior_conversation(conversation_text, limit=None): 849 """Keep both setup context and late outcome evidence in behavior prompts.""" 850 if limit is None: 851 limit = _behavior_check_budget() 852 if len(conversation_text) <= limit: 853 return conversation_text 854 marker = "\n...[middle truncated for behavior check]...\n" 855 if limit <= len(marker) + 1: 856 return conversation_text[:limit] 857 final_limit = _behavior_final_response_limit()First 10 of 60 lines, the second copy on the right.
- 32 lines in 2 placespython432 tokenssrc/skillevaluator/tier3/eval_core/llm_judge.py:902–933src/skillevaluator/tier3/harbor/templates/eval.py:3779–3810
902{behaviors} 903 904For each numbered behavior, check only what that behavior requires. 905A format-only behavior does not require the answer's value to be factually correct: a wrong value \ 906can still satisfy its format. If a behavior also requires factual correctness, evaluate that requirement too. 907For a behavior about final-response format, judge the entire final response literally against only \ 908the stated format requirement, independently of factual correctness. 909When a FINAL RESPONSE section is supplied, its contents are the final answer; other section \ 910headings, tool results, history, and artifacts are context, not part of that answer. 911Inspect every line of the final response, including text after an initially valid answer. \First 10 of 32 lines.
- 49 lines in 2 placespython408 tokenssrc/skillevaluator/tier3/eval_core/llm_judge.py:1152–1200src/skillevaluator/tier3/harbor/templates/eval.py:3870–3918
1152def _behavior_payload_error(parsed: Any, expected_count: int, *, allow_partial: bool = False) -> str | None: 1153 """Describe what is wrong with a behavior reply, or return ``None`` when it is usable. 1154 1155 Each expected behavior is judged exactly once, identified by its 1-based 1156 ``step`` number, with a boolean ``passed`` or an explicit null abstention. *allow_partial* accepts a 1157 salvaged prefix that judges each of its steps once. The score is computed 1158 from the verdicts, so a ``score`` field in the reply is ignored. 1159 """ 1160 if not isinstance(parsed, dict): 1161 return "the reply did not contain one JSON object"First 10 of 49 lines.
7 more
- 34 lines in 2 placespython388 tokens
- 62 lines in 2 placespython378 tokens
- 42 lines in 2 placespython284 tokens
- 45 lines in 2 placespython271 tokens
- 41 lines in 2 placespython250 tokens
- 36 lines in 2 placespython219 tokens
- 21 lines in 2 placespython182 tokens
Largest clones over all files, tests and data included
- 486 lines in 3 placesjsontests1,590 tokenstests/fixtures/skillspector-2.13.3-settings-hook-env-partial-no-llm.json:49–534tests/fixtures/skillspector-2.13.3-settings-hook-handler-partial-no-llm.json:49–534tests/fixtures/skillspector-2.13.3-settings-hook-no-llm.json:49–534
49"finding_id": "finding-7db83e2f2cd2422986a831a7335e0ca7", 50"category": "Bundled Execution Surface", 51"pattern": "Bundled hooks can execute when matching lifecycle events occur.", 52"severity": "LOW", 53"confidence": 0.95, 54"location": { 55 "file": ".claude/settings.json", 56 "start_line": 1, 57 "end_line": null 58},First 10 of 486 lines.
- 453 lines in 3 placesjsontests1,490 tokenstests/fixtures/skillspector-2.13.4-settings-hook-env-partial-no-llm.json:84–536tests/fixtures/skillspector-2.13.4-settings-hook-handler-partial-no-llm.json:84–536tests/fixtures/skillspector-2.13.4-settings-hook-permissions-no-llm.json:84–536
84 "unknown_event_count": 0, 85 "unknown_handler_count": 0 86}, 87"match_fingerprint": "1869d5f9b40b3830990438a7a9ef1f2487ca6219fc898fd8343f7a03a4575e19", 88"occurrences": [ 89 { 90 "file": ".claude/settings.json", 91 "start_line": 1, 92 "end_line": null 93 }First 10 of 453 lines.
- 475 lines in 2 placesjsontests1,571 tokenstests/fixtures/skillspector-2.13.3-bundled-script-hooks-no-llm.json:120–594tests/fixtures/skillspector-2.13.3-upload-hooks-no-llm.json:120–594
120"finding_id": "finding-636438e202a348db8b018bf18008ac8b", 121"category": "Data Flow", 122"pattern": "Tainted flow: 'target' from os.environ.get (line 29, credential/environment) → pathlib.Path.write_text (file write)", 123"severity": "MEDIUM", 124"confidence": 0.65, 125"location": { 126 "file": "scripts/plugin_paths.py", 127 "start_line": 31, 128 "end_line": 31, 129 "start_column": 4,First 10 of 475 lines.
- 464 lines in 2 placesjsontests1,515 tokens
- 174 lines in 2 placestxtdocs and examples1,498 tokens
By format, code only
| Format | Files | Lines | Clones | Duplicated lines | Duplication |
|---|---|---|---|---|---|
| python | 325 | 230,145 | 194 | 2,784 | 1.21% |
| bash | 1 | 91 | 0 | 0 | 0% |
By format, all files
| Format | Files | Lines | Clones | Duplicated lines | Duplication |
|---|---|---|---|---|---|
| python | 849 | 545,777 | 1,471 | 17,356 | 3.18% |
| json | 95 | 33,044 | 296 | 22,778 | 68.93% |
| yaml | 66 | 5,305 | 37 | 609 | 11.48% |
| django | 8 | 5,894 | 17 | 143 | 2.43% |
| markdown | 65 | 10,634 | 10 | 191 | 1.8% |
| txt | 8 | 2,724 | 4 | 333 | 12.22% |
| bash | 15 | 275 | 1 | 6 | 2.18% |
| markup | 6 | 385 | 0 | 0 | 0% |
| toml | 2 | 331 | 0 | 0 | 0% |
| docker | 1 | 27 | 0 | 0 | 0% |
Trending appearances
| Day | Rank | Stars | Files | Lines | Clones | All files | Code | Health | Commit |
|---|---|---|---|---|---|---|---|---|---|
| Nov 17, 2026 | #9 | 2,906 (+873) | 1,120 | 604,451 | 1,836 | 6.85% | 1.21% | C 58.3 | d2cb138 |
| Nov 16, 2026 | #11 | 1,887 (+122) | 1,059 | 581,438 | 1,817 | 7.09% | 1.24% | C 58.1 | ae7e858 |
Want this for your own project? Install jscpd and run jscpd ., or measure it the way the baselines were measured.