Mainframe Migration Toolkit - COBOL/JCL -> Python/SQL
A VS Code extension + language server + CLI toolkit that gives an AI coding agent (GitHub Copilot in agent mode) deterministic navigation of a legacy COBOL/JCL codebase's dependencies (CALL/COPY/EXEC), and a set of scripts that automate the mechanical, low-risk parts of migrating that codebase to Python and SQL.
This project intentionally does not try to auto-translate COBOL business logic into Python - see docs/RESEARCH.md for why that approach is unsafe for financial-grade code, and what large migration programs do instead.
Why "deterministic"?
LLMs are good at reading and writing code, but bad at reliably tracing "what calls what" across thousands of files inside a limited context window - and a hallucinated dependency edge in a mainframe migration can mean a silent production regression. Every tool here is a plain parser or graph traversal: same input always produces the same output, with no model involved. The AI agent's job is to call these tools and reason over their output, not to re-derive the graph itself.
How edge cases are handled
No regex-based parser can understand 100% of real-world COBOL/JCL. When
a tool hits something it can't safely resolve - an unresolved CALL/COPY
target, a dynamic CALL WS-VAR, EXEC CICS/IMS blocks, OCCURS ... DEPENDING ON variable-length tables, an unterminated EXEC SQL block,
a missing PROGRAM-ID, GO TO control flow, an edited/floating-point
PIC clause, a qualified/indicator SQL host variable, a JCL backward
reference to a nonexistent step, etc. - it never guesses. Instead it
emits a structured warning (code, severity, message,
suggested_action, file, line, snippet) defined once in
scripts/common/warnings_model.py /
server/src/warningsModel.ts, and every
tool surfaces it:
- Immediately supplyable missing prerequisites produce
BLOCKand stop the affected file generation before outputs are written. Semantic decisions produce configurable code-localTODO(mainframe-migration): [CODE] explanation / required decisionmarkers while unaffected units continue. - The language server reports the same facts as editor diagnostics on
the offending line, and via the
mainframe/warningscustom LSP request (also folded intomainframe/dependencyGraphandmainframe/impactAnalysisresults). - Known IBM/vendor utility programs (
IEFBR14,SORT,IDCAMS, ...) and system copybooks (SQLCA,DFHCOMMAREA, ...) are allowlisted so real findings aren't buried in expected noise.
See test-fixtures/edge-cases/ for a fixture per edge case, used to regression-test this behavior.
Repository layout
extension/ VS Code extension: starts the language server and
registers 5 vscode.lm Language Model Tools for Copilot
agent mode (see extension/package.json
"languageModelTools").
server/ Deterministic COBOL/JCL language server
(vscode-languageserver, TypeScript). Provides go-to-
definition & hover for CALL/COPY/EXEC PGM=, diagnostics
for unresolved targets, and custom LSP requests
(mainframe/dependencyGraph, mainframe/impactAnalysis,
mainframe/callers, mainframe/resolveCopybook) that back
the extension's LM tools.
scripts/ 17 standalone Python CLI tools for migration planning and
scaffolding (see below). No dependencies beyond the
Python 3 standard library.
scripts/common/ Shared deterministic COBOL/JCL parsing + workspace
indexing code used by every script (and conceptually
mirrored in server/src/*.ts for the LSP).
test-fixtures/ An original, small "bank batch" COBOL/JCL system used to
exercise every tool end-to-end (see
test-fixtures/README.md for its call/copy graph).
docs/RESEARCH.md Research on migration strategies used by large players,
and which tool here implements each one.
Configuration and preflight
Use a workspace-relative mainframe-migration.json to define ordered
primary/fallback COBOL, copybook, and JCL libraries. The schema is
mainframe-migration.schema.json, and
test-fixtures/mainframe-migration.json
is a complete example that intentionally includes edge-cases/ as a
fallback library.
mainframe-migration.json is authoritative for paths, ordered primary and
fallback libraries, source encoding, extensions, external registries, the
generated TODO prefix, and optional named transportProfiles. A transport
profile defines fixed-width text charset, line separator, trim policy, and the
representation of DISPLAY, COMP-3, BINARY/COMP-4, COMP-1, and COMP-2 fields.
Old configs remain valid, but contract/fixture generation requires an explicit
profile or defaultTransportProfile.
Physical copybook layout and transported text layout are separate contracts. For example, a six-byte packed COMP-3 value can become 13 text characters (precision digits, decimal point, and sign) without changing its source offset or physical length. Expanded decimal width is precision plus an explicit decimal point when scale is non-zero plus one sign character when signed; expanded integer width is precision plus one sign character when signed.
Run mandatory phase-zero preflight before generating migration output:
python3 scripts/migration_preflight.py test-fixtures --jcl BANKRUN --format text
python3 scripts/migration_preflight.py test-fixtures --format json
Preflight exits 0 when there are no BLOCK findings and 2 when a
human-suppliable fact is missing or ambiguous. Dynamic calls, transaction
dialects, ODO, REDEFINES, edited PICs, and qualified SQL hosts become
explicit TODO findings and do not stop unrelated analysis. Generate a
starter file with --write-template. Present the initial report grouped by
user_can_supply=true and then false. Resolution order is fixed: current
LSP facts, configured primary libraries, configured fallback libraries,
then verified known externals; exhaustion blocks and requires user input.
Automated migration repair is capped at 50 total failed attempts.
Configuration-aware CLIs accept --config. For single-file CLIs, the selected file must
exist, match its configured kind, and belong to the selected inventory.
Omitting config retains legacy behavior.
Pinned CardDemo validation
Run the parsed phase-3 scenarios against the exact pinned AWS CardDemo revision with one command (the runner clones/caches and verifies the detached commit):
python3 tests/run_carddemo_scenarios.py
The assertions, observed results, known fixes, and residual TODOs are recorded in
docs/research-carddemo-validation.md.
The 17 tools
| Tool | What it answers |
|---|---|
scripts/migration_preflight.py |
Validates configured source inventory, precedence, static dependencies, and migration blockers before generation |
scripts/dependency_graph.py |
"What calls/copies/executes what across the whole codebase?" (JSON/DOT/text) |
scripts/impact_analysis.py |
"If I change X, what breaks?" (transitive upstream/downstream + risk) |
scripts/dead_code_finder.py |
"Which programs/copybooks are never referenced?" |
scripts/sql_extractor.py |
Pulls embedded EXEC SQL out of COBOL into standalone, bind-parameterized .sql files |
scripts/business_rule_extractor.py |
Mines IF/EVALUATE decision logic (+ preceding comments) into a reviewable rule list |
scripts/copybook_to_dataclass.py |
Copybook PIC layout -> Python @dataclass and/or SQL CREATE TABLE |
scripts/copybook_to_contract.py |
Copybook physical offsets + selected transport profile -> canonical language-neutral JSON record contract with digest, provenance, and structured TODO/BLOCK findings |
scripts/generate_copybook_fixtures.py |
Canonical contract or configured copybook -> deterministic seeded fixed-width input, decoded JSONL, invalid decode cases, and a non-authoritative manifest for ingestion testing only |
scripts/generate_file_readers.py |
FILE-CONTROL/FD/COPY -> standalone domain/model dataclasses and binary-safe fixed-width readers; missing copybooks block writes while semantic layout decisions remain explicit TODO stubs |
scripts/cobol_to_python_skeleton.py |
COBOL paragraph structure -> Python class skeleton (structure preserved, logic left as TODOs) |
scripts/jcl_flow_extractor.py |
JCL step/COND flow -> explicit orchestration description or Python skeleton |
scripts/migration_complexity_report.py |
Ranks every program by migration effort/risk into a wave plan |
scripts/characterization_test_scaffolder.py |
Generates a golden-master pytest harness from a program's LINKAGE SECTION contract |
scripts/generate_program_capsule.py |
Runs selected-JCL preflight, then creates per-program/utility PARTIAL capsules with deterministic DD/I/O evidence, contract copies, explicit semantic TODO IR, placeholder source/tests, and exact paths/commands; any preflight BLOCK writes nothing |
scripts/validate_relational_ir.py |
Pure-stdlib typed validation for relational IR v1; exits 0 only for executable reviewed plans, 2 for BLOCK, and 3 for TODO/non-executable |
scripts/ir_to_pyspark.py |
Deterministically compiles validated executable IR to native PySpark source with run(inputs, outputs, spark) and refuses BLOCK/TODO plans before writing |
Every script is runnable standalone:
python3 scripts/dependency_graph.py test-fixtures --format text
python3 scripts/copybook_to_contract.py test-fixtures/copybooks/TRANREC.cpy --config test-fixtures/mainframe-migration.json --out TRANREC.contract.json
python3 scripts/generate_copybook_fixtures.py TRANREC.contract.json --out-dir synthetic/TRANREC --count 6 --seed 41 --include-invalid
python3 scripts/generate_file_readers.py test-fixtures --out-dir domain
python3 scripts/generate_program_capsule.py test-fixtures --jcl BANKRUN --out-dir migration/BANKRUN/capsules --format text
python3 scripts/validate_relational_ir.py migration/BANKRUN/capsules/mainpgm/ir/plan.scaffold.json --format text
relational-ir.schema.json defines the portable
relational plan format. An AI may author the semantic operations only after the
program semantic brief and target design review are approved. The validator and
PySpark compiler are deterministic; they do not infer missing types or business
rules. A generated program capsule is intentionally PARTIAL, contains an
explicit semantic TODO, and is not a translated or equivalent implementation.
The generated readers open files in binary mode so COMP-3, COMP-4/
BINARY, and DISPLAY fields retain their physical COBOL representation.
Each malformed line yields a structured RecordParseError by default;
valid records continue streaming. Missing copybooks block before package
creation; semantic or unsafe layout decisions generate an empty dataclass
and reader stub with a standardized TODO instead of a guessed layout.
Synthetic fixture output never contains authoritative expected program output.
Its manifest says synthetic=true, authoritative=false, and
purpose=ingestion_validation_only. Raw binary fields are not written into a
fixed_width_text artifact; a source_bytes representation requires an
explicit binary-capable output mode and otherwise blocks generation.
Copilot migration workflow
Use the workspace-scoped mainframe JCL migration skill directly, or select the Mainframe JCL Migrator agent for the phase-gated workflow. Example invocation:
/mainframe-jcl-migration /path/to/mainframe-workspace BANKRUN migration/BANKRUN
The first argument is the workspace root, the second is a JCL job name or path, and the optional third argument is the target directory.
The 5 Language Model Tools (for Copilot agent mode)
Registered in extension/package.json under
contributes.languageModelTools and implemented in
extension/src/extension.ts:
mainframe_getDependencyGraph- whole-graph or scoped CALL/COPY/EXEC graph.mainframe_getCallers- who calls a given program, including JCLEXEC PGM=.mainframe_resolveCopybook- resolve a COPY target to its file + users.mainframe_impactAnalysis- transitive blast-radius + risk rating before any change.mainframe_runMigrationScript- runs any of the 17 scripts above, workspace-sandboxed.
These are backed by the same language server used for interactive go-to-definition/hover, so an agent and a human developer see identical dependency information.
Building
cd server && npm install && npm run compile
cd ../extension && npm install && npm run compile
Then press F5 in VS Code (or use the "Run Extension" launch configuration in .vscode/launch.json) to try it against test-fixtures/.
A standalone smoke test for the language server (no VS Code required) is
at server/scripts/smoke_test.js:
node server/scripts/smoke_test.js test-fixtures
node server/scripts/config_parity_test.js test-fixtures
Installing the release artifacts
Build the wheel, source distribution, and bundled VSIX without publishing:
python tools/build_release.py
python -m pip install dist/mainframe_modernization_toolkit-0.1.0-py3-none-any.whl
mainframe-toolkit --version
The wheel contains the standalone VS Code extension, including the bundled
dist/server.js. Export and verify the exact embedded bytes before installing
the VSIX manually:
mainframe-toolkit vsix export --output mainframe-migration-toolkit-0.1.0.vsix
mainframe-toolkit vsix verify mainframe-migration-toolkit-0.1.0.vsix
code --install-extension mainframe-migration-toolkit-0.1.0.vsix
The extension invokes mainframe-toolkit by default. Set
mainframeMigration.executablePath to the installed executable's absolute
path when it is not on the VS Code extension host's PATH. Run
mainframe-toolkit workspace init <workspace> or the
Mainframe Migration: Initialize Workspace command to add the packaged
configuration, schemas, skill, and agent without overwriting existing files.
Source extensions referenced
This project was built with the two extensions supplied in the task as primary references for real-world COBOL/JCL LSP feature scope:
- eclipse-che4z/che-che4z-lsp-for-cobol - full COBOL LSP (ANTLR-based, JVM). Studied for how it models copybook resolution, dialects, and CALL/COPY dependency nodes; not vendored directly (JVM/Maven toolchain would be heavy for an agent-facing tool - see docs/RESEARCH.md for the tradeoff rationale).
- BroadcomMFD/jcl-language-support -
JCL syntax highlighting/snippets/Zowe integration. Studied for JCL file
detection conventions (
.jcl/.cntl, jobcard sniffing) and DD/DSN navigation patterns. - Their Marketplace listings (
broadcomMFD.jcl-language-support,broadcomMFD.cobol-language-support) were fetched for feature summaries.
All COBOL/JCL fixtures under test-fixtures/ are original, written for
this project (not copied from any repository).
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file mainframe_modernization_toolkit-0.1.0.tar.gz.
File metadata
- Download URL: mainframe_modernization_toolkit-0.1.0.tar.gz
- Upload date:
- Size: 378.5 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via:
twine/7.0.0 CPython/3.13.13
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
5182f49fac2ca92f42f640216b319be37547139634006cdee7677172684ed8a6
|
|
| MD5 |
09c659026dc214e7c6e62e95367f3cc3
|
|
| BLAKE2b-256 |
e6952a2b928835c1fe9fe10ae6610f9f486ba0f4e96b5201368ac32ced152b1f
|
File details
Details for the file mainframe_modernization_toolkit-0.1.0-py3-none-any.whl.
File metadata
- Download URL: mainframe_modernization_toolkit-0.1.0-py3-none-any.whl
- Upload date:
- Size: 399.6 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via:
twine/7.0.0 CPython/3.13.13
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
573d4d449f2b2e5e6f49b98ea7d6a06009a305710018c41f044ffdc7ba8b9a63
|
|
| MD5 |
1502a1fc739edf8f07c8b9feb18b9c10
|
|
| BLAKE2b-256 |
e37c75734c29d96256928b0b1c4614100b6448a2dd6a14fd29e05d9885aafe9c
|