If you would use google you might find some hits. For me the top hit was... my own blog, containing a python solution based on either the NM or MD tag.

Shameless advertisement: plots like that (similar) are available in NanoPlot.

Very old thread, but still few good answers for per read identity. Thought I would mention a bedtools alternative (so not purely bam based, and not so performant, but maybe still interesting)

# Step 1 Create a bed file with the start position, stop position, and edit distance based on the NM tag
bedtools bamtobed -ed -i x.bam > ed.bed &

The format of the bed file looks like this.
chr                                  start     stop      readname                              number_mismatches  read_orientation
chr                                  601       55050     8a6c37a9-44b6-416d-bf88-504e064013b7  2453  -

# Step2: Now calculate read length and then percent mismatches per read (or whatever you like in R/Python/Awk etc)
# Example only without checking the usual bed off-by-one start/stop coordinate nightmare
awk -F"\t" '{print $1"\t"$3-$2"\t"$5"\t"($5/($3-$2)*100)}'  ed.bed | head

# output format
chr   read_len NM     %identity based on NM
chr   18275   1420    7.77
chr   20187   737     3.65

Of course, a purely BAM based solution would be preferable but I haven't run across one yet (at least not in c, c++, rust etc, I don't want to use a python implementation).

Remember to think about removing supplementary alignments etc from the bam before you convert it to bed format.

use the edit distance 'NM' ? Here is a solution using bioalcidaejdk: http://lindenb.github.io/jvarkit/BioAlcidaeJdk.html

$ java -jar dist/bioalcidaejdk.jar -e 'stream().map(R->R.getReadUnmappedFlag()?0:(int)(100.0*(R.getReadLength()-R.getIntegerAttribute("NM"))/(double)R.getReadLength())).collect(Collectors.groupingBy(Function.identity(), Collectors.counting())).forEach((K,V)->{println(K+"\t"+V);});' input.bam