Enhancement: Extract DWARF spec to queryable JSON

[dguido]

Problem

Claude struggles with specific DWARF questions like "What attributes can exist on a DW_TAG_subprogram DIE?" because:

• The DWARF standard PDFs are too large for context
• Claude does 8-12 web searches, often reading LLVM/libdwarf source as fallback
• Even with web search, answers may be incomplete

Proposed Solution

Create a standalone tool that extracts structured DWARF specification data from PDFs into queryable JSON files. This would enable instant, authoritative answers to structural DWARF questions without web search.

Deliverables

1. JSON Schema Files (data/dwarf{3,4,5}.json)

   • Tag encodings (DW_TAG_* → hex code)
   • Attribute encodings (DW_AT_* → hex code + classes)
   • Tag-attribute mappings (which attributes are valid for each tag, from Appendix A)
   • Operation encodings (DW_OP_* → hex code + operands)

2. Extraction Script (scripts/extract_dwarf_spec.py)

   • Uses pdftotext to extract text
   • Regex parsing of structured tables in Chapter 7 and Appendix A

3. Query Script (scripts/query_dwarf.py)

   • CLI interface for lookups
   • Example: python query_dwarf.py DW_TAG_subprogram returns applicable attributes

Example JSON Schema

{
  "version": "5",
  "tags": {
    "DW_TAG_subprogram": {
      "code": "0x2e",
      "applicable_attributes": ["DW_AT_name", "DW_AT_type", "DW_AT_low_pc", ...]
    }
  },
  "attributes": {
    "DW_AT_name": {
      "code": "0x03",
      "classes": ["string"]
    }
  }
}

Complexity Assessment

Data	Source	Difficulty
Tag codes	Ch 7 table	Easy - clean regex
Attr codes	Ch 7 table	Easy - clean regex
Op codes	Ch 7 table	Easy - clean regex
Tag→Attrs	Appendix A	Medium - multi-line tables

Size Estimate

~1,500-2,000 lines total (scripts + JSON), compared to 20K+ lines of full markdown extraction. This provides programmatic access to the most-queried structural data.

Success Criteria

• python query_dwarf.py DW_TAG_subprogram returns correct attribute list
• JSON files are valid and parseable
• Skill can answer "What attrs can X have?" without web search

---

This enhancement complements the lean skill approach by adding structured, queryable spec data rather than prose.

🤖 Generated with Claude Code