SKILL.md's one copy-pasteable Phase 3 command used --remove-duplicates, which argparse rejects; the real flag is --deduplicate. A sweep of every documented invocation against the actual --help output found eleven more flags that do not exist: validate_citations.py --auto-fix, --strict, --output, --check-required-fields, --check-authors, --check-duplicates, --check-syntax; format_bibtex.py --validate, --report, --style; doi_to_bibtex.py --clipboard; and search_google_scholar.py --open-access-only. Several sat in worked examples an agent would run verbatim. citation_validation.md documented an "auto-fix" mode of validate_citations.py that never existed -- the script only reports. Those sections now point at format_bibtex.py, which is what rewrites. Other corrections: - --report wrote a JSON document to report.txt. - three of five declared dependencies were never imported. - Phase 2.5 was marked MANDATORY but needed web access that allowed-tools did not grant; WebSearch and WebFetch are now declared. - no compatibility field, despite needing a Python version, a package, and network access. - the schematics section told a citation skill to produce diagrams by default, and named generate_schematic.py while generate_schematic_ai.py sat beside it unexplained. Reduced to a pointer at scientific-schematics. - venue reference counts are labelled as rules of thumb rather than submission requirements, which is what they are. Documents search_openalex.py and the shared _common.py, and regenerates the workflow diagram, which no longer matched what the skill does. Version 1.8 -> 2.0: format_bibtex.py no longer overwrites its input by default.
1.3 MiB
1536x864px
1.3 MiB
1536x864px