Looks like the tool you used has a lot false positives, I wouldn't trust it too much. Instead we could try to automate the blast method. Run blastx on all transcripts.

• Retain all best hits or all hits with score > some threshold
• Retain all hits with more than one HSP
• Retain all hits where at least one HSP has different frame from the others
• All queries that pass these filters are frame shift candidates

This can be implemented easily using BioPerl or BioPython.

Here is an example in BioPerl that you can adjust for your needs.

Output:

 ./filterBlastFrameShift.pl ~/Downloads/JR132X8F11N-Alignment.txt 
lcl|Query_197237 has 6 hsps with the following frames: (*: e < 1e-10)
1*-3-3-2-23
Query gi|23274247|gb|BC035912.1| has 0 out of 1 hits with frameshifts
lcl|Query_197237 has 5 hsps with the following frames: (*: e < 1e-10)
3*-1-3-32
Query query_1_no_frameshift has 0 out of 1 hits with frameshifts
lcl|Query_197237 has 5 hsps with the following frames: (*: e < 1e-10)
3*-1-3-32
Query query_2_insert_no_frameshift has 0 out of 1 hits with frameshifts
lcl|Query_197237 has 6 hsps with the following frames: (*: e < 1e-10)
1*3*-1-32-1
frame mismatch 0 vs. 2
Query query_2_insert_with_frameshift has 1 out of 1 hits with frameshifts

And the blast example output with fabricated frame shift: