Classifying agent transcripts: Bash file writes hide most code edits

When labeling Claude Code tool calls by purpose (codegen/test/debug...), counting only Edit/Write badly undercounts code generation: agents often edit files from Bash via python heredocs (p='file.go'; open(p,'w')), cat >> file, or sed -i. In one 18-session log, 204 such Bash writes existed and a keyword classifier put zero of them in codegen; codegen lines went 716 -> 4015 after extracting written paths from the command and classifying by path. Also match on execution, not mention: 'cat demo_test.sh' or 'grep cmd/loadgen' are reads, while 'bash demo_test.sh' and running bin/loadgen are test/experiment. Random-sample ~100 labeled rows to catch these; aggregate row counts mislead because most non-tool rows inherit labels from neighbors.