Legacy analyser

Read-only tools over deterministic analysers for COBOL, copybooks and JCL, Java and SQL DDL (packages/legacy-analyzer, plugin extensions/legacy-analyzer). The analysers provide the facts; the model explains them. Overview: Industrial modernisation.

Tool	What it returns
legacy_index	Parses the sources under path and builds the fact index; counts and diagnostics
legacy_query	Structure: callers, callees, file access, data references, job steps, paragraphs, perform graph, data sets
legacy_rules	Rule candidates: conditions, arithmetic with constants, rounding, 88-level values, numeric constants, literal assignments
legacy_dictionary	Copybook data dictionary: level, PIC, usage, length, offset, positions, OCCURS, REDEFINES, values
legacy_impact	What a change to a data item reaches, with the fact that links each entry
legacy_design_doc	Design document (設計書) from the code, Japanese or English: see design documents and wave plan
legacy_wave_plan	Modernisation wave proposal from dependencies and metrics: see design documents and wave plan
legacy_changes	What changed since the stored baseline, with path:line evidence: see change detection
legacy_baseline_advance	Moves the change-detection baseline forward (or resets it): see change detection
modernise_plan	Entry point for mixed projects: detected languages, inventory per engine, shared tables, coverage, planned assistants
modernise_crud	CRUD of SQL tables (COBOL and PL/I EXEC SQL, Java JPA/@Query/JDBC, JCL utilities) and of files, with evidence
modernise_impact	What a change to a table, column, file, COMMON block, Java class or program reaches across languages and callers
modernise_inventory	Java types, SQL tables, FORTRAN units and COMMON blocks, or PL/I procedures: the list, or one entry in full

Every fact carries file and line. Business meaning is never inferred.

The same analyser runs in engineers’ own editors and IDEs as an MCP server, one project per server: see Legacy analyser in your editor (MCP).

Modernisation assistant (COBOL, Java, SQL DDL, FORTRAN, PL/I, PLC)

modernise_plan detects the languages of a folder (extension, then content of the first bytes), runs every active engine and lists the assistants that have no engine. Active: COBOL/JCL, Java, SQL DDL, FORTRAN, PL/I, PLC IEC 61131-3 (Structured Text sources and PLCopen XML exports; vendor ladder exports are counted, not read). No engine (files are counted, nothing is analysed): batch schedulers, maintenance records, cybersecurity review.

Tables are one node in the knowledge graph across languages, matched by unqualified name, case-insensitive: a COBOL or PL/I EXEC SQL table, a JPA @Table and a CREATE TABLE of the same name are the same table. Programs meet by name the same way: a JCL EXEC PGM=, a COBOL CALL and a FORTRAN or PL/I CALL of one name are one program, so a PL/I batch calling a COBOL routine, or a step running a FORTRAN program, is one path in the graph. The design document gains Java class specifications, FORTRAN program units (with the COMMON block map), PL/I procedure specifications and table definitions when the folder has those sources.

FORTRAN INCLUDE and PL/I %INCLUDE members are looked up among all files of the folder by file or member name, case-insensitive, the including file’s folder first.

Limits:

• Java: declarations are parsed without a compiler. Calls are linked only when the receiver’s declared type is in the sources (same class, supertypes, static imports, fields, parameters, type names, fully qualified names); calls on local variables, expressions and library types are listed as unresolved, overloads are matched by argument count only. Anonymous and local classes, initializer blocks and enum constant bodies are reported, not modelled. SQL is read from string literals, text blocks and String constants; SQL built from variables is marked partial. JPA names without @Table / @Column follow Spring Boot’s default naming (camelCase to UPPER*SNAKE). MyBatis mapper XML (DOCTYPE or <mapper namespace>) is read: each <select> / <insert> / <update> / <delete> gives table access at its own line, bound to the method of the same name of the namespace interface, so callers inherit it; every dynamic branch is read (the tables are the union), <include> is expanded from <sql> fragments of the same file, and a ${...} substitution or a fragment from another file makes the fact partial. Statements whose interface or method is not in the sources are listed as unbound. MyBatis annotations (@Select), resultMap columns and orm.xml are not read. No annotation processing, Lombok; JPQL path joins (JOIN e.dept d) are not mapped to tables. The owning side of @ManyToOne / @OneToOne (targetEntity or the field type) is JavaAnalysis.relations and a JOIN edge owner table -> target table (detail COLS -> TARGET(REFS), via @ManyToOne Type.field; the target’s schema as the edge’s schema). Join columns come from @JoinColumn / @JoinColumns, else JPA’s default FIELD*+ the target’s single@Id column (IDwhen unknown), and are the entity’s column mappings too;mappedByfields have no column.compareJoinKeys(jpa-keys.ts) lists a join with no matching DDL foreign key (owning table in the DDL; the edge detail ends; no foreign key in the DDL) and a foreign key no relationship uses (table mapped by an entity or a join table). @OneToMany+@JoinColumn is a relation on the target’s table to the owner; a join table (@ManyToMany, @OneToManywith@JoinTableor no join column; target fromtargetEntityor the collection’s last type argument) is two relations withjoinTable: true, joinColumnsto the owner andinverseJoinColumnsto the target (JPA default names otherwise; edge detail adds(join table)). Nested annotations carry the named elementthey are the value of.mappedBycollections add nothing. A@SecondaryTable(each one in@SecondaryTables) is an entity table use (via: "@SecondaryTable") and a relation of kind SecondaryTable(emptyfield) from the secondary table to the primary on its pkJoinColumns(default: the primary’s@Idcolumn, implicit); a@Column/@JoinColumnwithtable =maps that column to the secondary table. Inheritance follows the root’s@Inheritance strategy (SINGLE*TABLEwhen not given): aSINGLE_TABLE subclass maps to the root’s table (via: "@Inheritance(SINGLE_TABLE) Root"; its own @Table does not apply) and the root’s table gets the discriminator column (@DiscriminatorColumn, default DTYPE; discriminator: true, empty field); a JOINEDsubclass keeps its own table and is a relation of kindInheritanceto its superclass’s table on its@PrimaryKeyJoinColumns (default: the superclass’s @Idcolumn, implicit), with a discriminator column only when@DiscriminatorColumnis written; every entity of a hierarchy with a discriminator column has discriminator on its entity table use (root, column, and value: its @DiscriminatorValue, else the entity name for a STRINGdiscriminator withimplicit: true, none for CHAR/INTEGERwhose default is provider-specific;duplicate: true when another entity of the hierarchy has the same value, a finding in the design document); aTABLE_PER_CLASSsubclass’s table also maps its superclasses’ columns. A subclass inherits the@Idcolumn for default join column names. A@MappedSuperclasshas no table: its fields,@Id and relationships map on each entity subclass’s own table (typeis the entity),@AttributeOverride(s) on the entity rename those columns, and an entity superclass above a mapped superclass is still the subclass’s parent for inheritance. An embedded field (@Embedded, @EmbeddedId, or a field whose type is @Embeddable) maps no column itself: the embeddable’s fields are the owner table’s columns with field as a path (location.floor), nested embeddables recurse, @AttributeOverride(s) on the field or by path on the entity rename them (the outermost wins), and a to-one relation inside it is the entity’s relation (fieldthe path); collections inside an embeddable are not modelled. @AssociationOverride(s)on the entity (a mapped superclass’s to-one field) or on an embedded or element-collection field (by path; the outermost wins) replace a to-one relation’s join columns with itsjoinColumns; on the entity, its joinTable replaces a mapped superclass’s to-many join table (and the key / order columns go there). An@ElementCollection is an entity table use of its collection table (via: "@CollectionTable", or "@ElementCollection (default name)"for JPA’sENTITY_FIELD) and a relation of kind ElementCollectionfrom that table to the entity’s on itsjoinColumns(default:ENTITY*+ the entity’s@Id column, implicit); its columns are the element’s (@Column, else the field name; an @Embeddableelement’s fields by path), aMap’s @MapKeyColumn(defaultFIELD_KEY; for an entity key, @MapKeyJoinColumn, default FIELD_KEY, also a relation of kind MapKeyto the key entity’s table, as on a to-manyMap’s join table) and an @OrderColumn(defaultFIELD_ORDER); a to-many’s Mapkey and@OrderColumncolumns map the same way on the table holding its rows (the join table, or the target’s for a join-column@OneToMany). The design document’s table definitions list every relation with the foreign key it matches (cited), the finding, or why it is not judged, then the foreign keys no mapping uses.
• SQL DDL: CREATE TABLE / VIEW / INDEX, ALTER TABLE ADD, COMMENT ON, DB2 LABEL ON; ANSI, DB2, Oracle and PostgreSQL basics. Stored procedures, triggers, sequences, grants, partitioning and storage clauses are listed as not modelled. Tables join across languages by bare name. When two explicit schemas define the same name (HR.EMPLOYEE, PAYROLL.EMPLOYEE) the bare-name table is modelled from the first, the second is a gap saying the name is ambiguous, not a duplicate, and each schema’s definition is also its own graph node table:SCHEMA.NAME (columns column:SCHEMA.NAME.COL, a SCHEMA edge from the bare-name node, listed in the SQL model’s schemaTables); an edge whose schema is one of those schemas is bound to that node. An ALTER or COMMENT naming one of those schemas applies to that schema’s definition (the first schema’s is also the bare-name table); one naming no schema for such a table is a gap and is never applied. Foreign keys (REFERENCES PAYROLL.EMPLOYEE), view sources and CREATE INDEX ... ON that name a schema carry it as the edge’s schema (the SQL model keeps it as references.schema, sourceSchemas and tableSchema), so they bind to that schema’s node too; the second schema’s definition gets its own FK and VIEW edges. Such names are listed in the SQL model’s ambiguous, and their graph table node carries ambiguousSchemas (HR,PAYROLL), so a COBOL, Java or PL/I edge to it reads as “one of these”, never as a certain join. A Java edge whose source names the schema carries it as the edge’s schema: @Table(schema = "payroll") or @Table(name = "payroll.employee") on the entity (and so its repository, derived queries and JPQL), or payroll.employee in JDBC, native @Query, constant or MyBatis SQL when every reference to the table in that statement names the same schema. A bare reference has no schema. An unqualified and a qualified definition of one name are still one table (a duplicate). Quoted identifiers are folded to upper case.
• FORTRAN: fixed form (FORTRAN 66/77, columns 1-72, DEC tab format) and free form (Fortran 90 and later). Units, USE, INCLUDE (one level: an include inside an include is not followed), COMMON, EQUIVALENCE (listed), declarations with REAL*8 / KIND types and the IMPLICIT rule (the default I-N rule when none is written). COMMON layouts are compared across units by type and size (DOUBLE PRECISION equals REAL*8, dimensions given by PARAMETER constants compare by value; other KIND values and default sizes are compiler-specific and compared as written). A name with an argument list is a function reference only when a FUNCTION of that name is in scope and the name is not an array or character variable of the unit; statement functions and generic interfaces are not resolved. I/O units are bound to the FILE= literal of an OPEN on the same unit number (the same unit, else the only one in the project); without OPEN, to the mainframe DD name FTnnF001 of the steps whose program reaches the unit through static calls. Units in variables, FILE= expressions and units opened with several files are listed as unresolved. Not modelled, reported: preprocessor lines (#ifdef: every branch is read), computed and assigned GO TO, arithmetic IF, ENTRY (calls to the entry name are linked to its unit), alternate returns, Cray pointers, EXEC SQL, SUBMODULE. Interface blocks and derived type definitions are skipped; type-bound procedures, BLOCK constructs and coarrays are not analysed.
• PL/I: procedures with OPTIONS(MAIN) and nested internal procedures, END matching over DO / BEGIN / SELECT groups (END with a label closes the inner groups). DECLARE: levels, factored names, data attributes, BASED / DEFINED / STATIC / CONTROLLED, INIT; FILE and ENTRY declarations. CALL is resolved to internal procedures in scope, external procedures of the PL/I sources, then by name to any program of the project; CALL through ENTRY variables and function references are not resolved. Record and stream I/O bind to the DD named by the OPEN TITLE literal, else the file name (names longer than 8 characters are not truncated). EXEC SQL tables per procedure; cursors through DECLARE CURSOR. Sequence numbers in columns 73-80 are detected and ignored. Not modelled, reported: macro statements (%IF, %DO, %DECLARE, %name =: not evaluated, the code of every branch is read), FETCH / RELEASE, secondary ENTRY points, EXEC CICS / EXEC DLI, dynamic SQL, EXEC SQL INCLUDE other than SQLCA / SQLDA. LIKE, DEFAULT and %INCLUDE inside an include member are not expanded.
• COBOL: as above; table access comes from EXEC SQL text. A FORTRAN include stored with a .inc extension is also read by the COBOL index as a copybook.

Change detection

legacy_changes {path, since?: "baseline", limit?} compares the current fact index with a stored baseline snapshot of the same folder and reports:

• COBOL programs, copybooks, JCL jobs, SQL DDL tables, their columns and stored routines that were added, removed or changed. Programs compare by structure (paragraphs and their statements without comments, data items, files, COPY members), copybooks by file hash, jobs by their steps and DD statements, tables and columns by their definition, routines by their source text with whitespace normalised. A line shift alone is not a change. A changed program lists the paragraphs added and removed and its statement count; a changed table its columns.
• Edges added or removed: CALL, PERFORM, EXEC SQL (table and operation) and JCL STEP (EXEC PGM=).
• path:line evidence on every entry: the current source for added and changed entries, the baseline’s (marked (baseline)) for removed ones. Changed and added programs carry their impact: callers, job steps that run them and tables they touch.

Counts are complete; the entry lists are bounded by limit (default 40, at most 200) with omitted saying how many were left out, in a fixed order (kind, then added, removed, changed, then name). Java, FORTRAN, PL/I and PLC sources are not compared yet.

The first run finds no baseline, creates it from the current sources and returns status: "baseline_created" with the totals: nothing is reported as a change. Later runs return changed or unchanged and current.digest, the snapshot they compared.

The snapshots live in the analyser cache folder, output/.legacy-index/baseline/ of the project (the state directory outside a project), never in the client’s sources: baseline-<id>.json (what the next comparison is against) and candidate-<id>.json (what the last comparison computed).

The baseline moves only with legacy_baseline_advance {path, snapshot?}, which promotes the last compared snapshot (when snapshot is given it must be that current.digest, else the call is refused because the sources were compared again since). A change that lands between the comparison and the advance is therefore still reported next time. The autopilot’s legacy.baseline_advance action calls it after the night’s work item is filed (Autopilot). To reset, a person calls legacy_baseline_advance {path, rebuild: true} (the baseline becomes the current sources), or deletes output/.legacy-index/baseline/ so the next run creates a new baseline.

Enabled per project, not for the gateway

The plugin is off for the gateway. Projects made from the Industry: legacy modernisation template get it for that project only: the template lists "plugins": ["legacy-analyzer"], the project records it, and the tools are added to that project’s sessions alone (a private plugin registry, loadScopedKoLoPlugins in src/plugins/loader.ts). Projects made from that template before projects carried a plugin list get it from their template.

Only bundled plugins whose manifest says "projectScope": true can be enabled this way, and only tool-only plugins. A config that blocks the plugin (plugins.enabled: false, plugins.deny, an allowlist without it, plugins.entries.legacy-analyzer.enabled: false) wins over the template. A tool profile with an allowlist (tools.profile) filters these tools like any plugin tool; add them with tools.alsoAllow.

To use the analyser outside projects, enable it for the gateway: kurahi config set plugins.entries.legacy-analyzer.enabled true.

Modes and scope

The tools are reads: the project guard classifies each call as a read of its path. They are available in all three project modes (参照 Read, 計画 Plan, 実行 Act), reach only the project folder, and never an excluded folder such as _answer-keys (the indexer also never enters it). The analyser’s only write is its own cache folder, output/.legacy-index/ inside the project: the index, and the latest design document and wave plan. Pass the project folder as an absolute path; relative paths resolve against the agent workspace, the same directory the guard uses.

How the agent is told to use it

The template’s engineering pack (templates/projects/packs/industry-legacy-modernisation/) and the project context make the order explicit: legacy_index on the project folder first, then legacy_query and legacy_rules (and legacy_dictionary, legacy_impact), and every rule claim cites file:line from the analyser output.

Test: pnpm exec vitest run --config vitest.e2e.config.ts src/projects/legacy-template.e2e.test.ts (creates a project from the template against a temporary copy of the demo pack and checks the PAYCALC overtime constants come back with file:line; skipped when the pack is not present).