Repository navigation
v2.12.0: 125 mainframe programs to L4, CICS coverage, two studies
Latestv2.12.0: 125 mainframe programs to L4, CICS coverage, two studies
PyPI: the PyPI package for this release follows in v2.12.1. Until then,
pip install gitgalaxyinstalls v2.11.0.
In short
- GitGalaxy's COBOL-to-Java translator now covers 125 of the 129 programs in six public mainframe codebases. For each of them, the translated Java was run against the original COBOL and matched, with every branch exercised except those shown to be unreachable. At v2.11.0 the translator covered 17 test programs.
- Two new studies, both made with the GitGalaxy scanner: DOOM's DNA and Every Model Has a Tell.
- Python 3.11 or newer is now required (was 3.9). See Upgrading below.
This release covers five weeks and about 2,870 commits since v2.11.0 (2026-09-30), more than usual because no release was cut in October. This note covers the main themes; the full list is in git log v2.11.0..v2.12.0.
Upgrading
- Python 3.11 or newer is required. v2.11.0 supported 3.9 and 3.10; installing v2.12.0 on either now fails at
pip install. - New
gitgalaxy assess ESTATEcommand. An offline, read-only scanner that reports how ready a mainframe codebase is for porting (#4726). - New
translatorextra. The COBOL-to-Java translator's parser dependency is now optional:pip install "gitgalaxy[translator]". Plain scans don't need it. - The
fullextra now includes scikit-learn, which the ML threat layer needs at load time (#2986).
These items come from the package metadata. Removed CLI flags and changed output formats were not checked for this release.
The translator, by codebase
Each program is placed on an evidence ladder:
- L2: the Java was run against the COBOL and produced the same results.
- L3: as L2, and every paragraph of the program was exercised.
- L4: as L2, and every branch was exercised, except branches reviewed and shown to be unreachable.
- L5: as L4, and deliberately planted bugs (mutants) in the translation are each caught by a test or explained.
The precise definitions are in docs/language_status/definition_of_done.md.
| Codebase | Programs | L4 or above | L5 |
|---|---|---|---|
| AWS CardDemo | 44 | 42 | 1 |
| CICS Async API credit-card example | 10 | 10 | 10 |
| CICS Banking Sample Application (CBSA) | 31 | 30 | 0 |
| CICS GenApp | 31 | 31 | 0 |
| DBB Mortgage Application | 6 | 5 | 0 |
| zECS | 7 | 7 | 1 |
| Total | 129 | 125 (96.9%) | 12 |
Four programs are below L4:
- BANKDATA (CBSA) is at L3.
- CBEXPORT and CBIMPORT (CardDemo) contain source errors that IBM's compiler would reject (#5866).
- EPSCSMRD (DBB) is deliberately outside the translator's scope, because it depends on Language Environment condition handling (#5106).
In addition, all 85 programs of the CICS crucible are at L4. The crucible is a test codebase written to break the translator, and at v2.11.0 only 17 of its programs had been translated.
Mutation testing has measured 47 programs so far, and its tests catch 12,350 of the 18,585 planted bugs judged (66.5% before excluding the planted bugs that change nothing).
Programs at L5 (12)
- CardDemo: CBSTM03A
- CICS Async API: ASYNCPNT, CRDTCHK, CSSTATS2, CSSTATUS, GETADDR, GETNAME, GETPOL, GETSPND, SEQPNT, UPDCSDB
- zECS: ZECSPLT
Programs at L4 (113)
- CardDemo (41): CBACT01C, CBACT02C, CBACT03C, CBACT04C, CBCUS01C, CBPAUP0C, CBSTM03B, CBTRN01C, CBTRN02C, CBTRN03C, COACCT01, COACTUPC, COACTVWC, COADM01C, COBIL00C, COBSWAIT, COBTUPDT, COCRDLIC, COCRDSLC, COCRDUPC, CODATE01, COMEN01C, COPAUA0C, COPAUS0C, COPAUS1C, COPAUS2C, CORPT00C, COSGN00C, COTRN00C, COTRN01C, COTRN02C, COTRTLIC, COTRTUPC, COUSR00C, COUSR01C, COUSR02C, COUSR03C, CSUTLDTC, DBUNLDGS, PAUDBLOD, PAUDBUNL
- CBSA (30): ABNDPROC, ACCTCTRL, BNK1CAC, BNK1CCA, BNK1CCS, BNK1CRA, BNK1DAC, BNK1DCS, BNK1TFN, BNK1UAC, BNKMENU, CRDTAGY1–5, CREACC, CRECUST, CUSTCTRL, DBCRFUN, DELACC, DELCUS, GETCOMPY, GETSCODE, INQACC, INQACCCU, INQCUST, UPDACC, UPDCUST, XFRFUN
- GenApp (31): LGACDB01, LGACDB02, LGACUS01, LGACVS01, LGAPDB01, LGAPOL01, LGAPVS01, LGASTAT1, LGDPDB01, LGDPOL01, LGDPVS01, LGICDB01, LGICUS01, LGICVS01, LGIPDB01, LGIPOL01, LGIPVS01, LGSETUP, LGSTSQ, LGTESTC1, LGTESTP1–4, LGUCDB01, LGUCUS01, LGUCVS01, LGUPDB01, LGUPOL01, LGUPVS01, LGWEBST5
- DBB Mortgage (5): EPSCMORT, EPSCSMRT, EPSMLIST, EPSMPMT, EPSNBRVL
- zECS (6): ECS001, ZECS000, ZECS001, ZECS002, ZECS003, ZECSNC (ZECS002 and ZECSNC are assembler)
Two studies
- DOOM's DNA (text edition). We scanned 43 programs descended from the 1997 DOOM source, written in about twenty languages, by the shape of their code alone, ignoring every name and comment. The shape still shows who copied whom.
- Names are a poor guide: Chocolate Doom keeps 87% of the original's function names, but only 35% of the code inside them is unchanged.
- With every name hidden, a Go translation's functions still land in the right part of the game 60% of the time, against 20% by chance.
- Every Model Has a Tell (text edition). Three AI coding models (Claude Sonnet 5.5, Claude Haiku 5.5 and OpenAI Codex) each built a chess engine and a logic-puzzle solver in C, Java, Python, Rust and COBOL, 60 programs in all.
- Measured by structure alone, a program's closest look-alike came from the same model 83% of the time (43 of 52), against 27% by chance.
- The study's five predictions, written down in advance, were not confirmed after correcting for running several tests.
- All data and code: squid-protocol/agent-habits.
CICS
CICS is the IBM transaction system most of these programs run under. The translator's CICS model now covers:
- 107 of the 257 commands in IBM's CICS reference. Each of the rest is refused by name rather than guessed at.
- 60 of the 68 commands the six codebases use.
- 954 of 1,141 command options.
This cycle added interval control (START, RETRIEVE, CANCEL, RUN TRANSID CHILD, FETCH and DELAY; #4950), RETURN TRANSID IMMEDIATE, and INQUIRE URIMAP browse with WRITE OPERATOR (#7666). It also added RIDFLD record addressing by byte offset, record number and extended offset (RBA, RRN and XRBA) to the estate-crucible scorer (#7680). Twelve CICS resource tables (FCT, PCT, PPT and others) are now read into the resource map. In JCL, a JCLLIB ORDER= library is now treated as a search path for procedures and INCLUDE groups, not as an import (#7693).
Known limitations
- What the results are checked against. The ladder results are measured against GnuCOBOL plus GitGalaxy's own models of CICS, Language Environment and Db2, which are written from IBM's documentation. Nothing is calibrated on real z/OS; see
docs/language_status/oracle_assumptions.md. - L5 counts are unaudited. A sample audit on 2026-11-01 found that 3 of 24 "explained" mutant verdicts were real test gaps. Mutation testing covers 47 of the 129 programs, and none of GenApp yet.
- The six codebases were used to develop the translator. Results on codebases it has never seen may be lower; that measurement is separate work.