[iliyan-velichkov]

Cause

A process setter (setField / setRelationField / clearField) writes one member per step, and every generated step is flowable:async: its own job, its own transaction. A change that is one fact therefore passed through visible half-written states. For example, a job released its slot while still reading RUNNING. Each column also published its own -transitioned event, and an instance that stopped between two steps kept the half-written state.

Change

A step may declare a set: map:

- { name: finish,      kind: serviceTask, args: { set: { slotKey: null, Status: DONE }, next: end } }
- { name: freeMachine, kind: serviceTask, args: { set: { machine.lastReading: null, machine.operator: null }, next: end } }

• Keys are what a single-member setter accepts: a field or to-one relation of the trigger record, or one hop away. All keys must address one record; a mix is refused with a message naming both paths.
• Values follow each key's single-member setter, with the same messages:
  • a literal, "{error}" or null (the clearField erasure) on a string/text field;
  • "{now}" on a timestamp/date field;
  • a seed name or id on a to-one relation, resolved by StatusSymbolResolver.
• Exclusive with setField / clearField / setRelationField / value / delegate / call / javaHandler / notify. An empty map is refused.
• Step kinds: allowed on a serviceTask, and on a userTask for relation keys only (the same rule setRelationField follows there).
• One write: the generated delegate builds one LinkedHashMap in the authored order and calls updateProperties once, so the lifecycle edge, the stampOn stamps and the label judge the whole change, and one -transitioned is published.

Implementation notes:

• SetFieldSupport.Setter now carries an ordered List<Assignment>.
  • Readers of a status write go through Setter.relationValue(...): CheckSupport, NumberingSupport, EdmIntentGenerator.writesStatus, BpmnIntentGenerator.gatesACheck.
  • Readers of raw args use SetFieldSupport.relationWrites(...): lifecycle reachability, StatusStepsSupport.
  • isSetterStep(...) replaces the three-arg checks in the BPMN generator, ServiceTaskHandlerGenerator and the abortOn then: validation.
• GlueGenerator.setterAssignments renders each member's Java value. SetField.java.template gains an #if($assignments) branch; the single-member branches are untouched.
• Byte-identical: a one-key set: projects exactly the single-member descriptor (same delegate), and models without set: produce unchanged descriptors and sources. No existing test was modified.
• Trap fixed: IntentParser's Gson round trip of the YAML tree drops null map values, which would silently delete every erasure. restoreSetMaps puts each step's raw set: map back after the typed mapping.
• Docs: the assistant guide's setter section recommends set: whenever several members change as one fact; intent-dsl-features.md and the engine-intent CLAUDE.md describe the construct.

Verification

• Unit tests, all green, with no existing test changed:
  • engine-intent 1779 tests, including the new SetManyIntentTest (14) and 4 new GlueSettersTest cases;
  • ide-template 309 tests, including a new GlueGeneratorTest case for the rendered values.
• Integration tests, all green:
  • new IntentSetManyIT (generate → publish → compile → run):
    • the trigger-record form clears the error and moves to RUNNING;
    • the one-hop set: { job.slotKey: null, job.Status: DONE } releases the slot, moves the status and stamps finishedAt in one write;
    • a DONE → RUNNING restart that the lifecycle refuses fails the step and leaves the slot unreclaimed.
  • IntentOneHopSetterIT, IntentClearFieldIT, IntentSetFieldNowIT, IntentEngineIT (83 tests).
• mvn -T 1C formatter:validate with the formatter cache wiped; release-profile javadoc build on engine-intent and ide-template.
• Not run: the full IT suite or the PostgreSQL leg; CI covers those.

Found on the way (not fixed here)

A process step named start, of any kind, collides with the BPMN start event's id start, and Flowable refuses the deployment ("multiple occurrences of ID value 'start'"). This predates the change, and the issue's own example uses that name. The IT and the guide's example name the step begin instead. It deserves its own issue: the parser could reserve the name, or the generator could prefix the start event's id.

Fixes #8042

🤖 Generated with Claude Code