complete Studio core and JavaFX boundary migration
Intrepid/Prometeu/Studio/pipeline/pr-master Build started...
JaCoCo Coverage #### Project Overview No changes detected, that affect the code coverage. * Line Coverage: 64.01% (20482/31999) * Branch Coverage: 54.22% (8240/15196) * Lines of Code: 31999 * Cyclomatic Complexity: 13164 #### Quality Gates Summary Output truncated.
Test / Build skipped: 15, passed: 715
Intrepid/Prometeu/Studio/pipeline/head This commit looks good

This commit is contained in:
2026-09-25 11:28:18 +01:00
parent 10f1b3f87c
commit a0a5050d5e
461 changed files with 9319 additions and 7205 deletions
@@ -0,0 +1,103 @@
---
id: LSN-0075
ticket: studio-core-javafx-decoupling
title: Studio Services Must Outlive the Current UI
created: 2026-09-25
tags: [studio, architecture, spring, spring-modulith, javafx, adapters]
---
## Context
Prometeu Studio began as a desktop application in which JavaFX controls, application workflows, provider integrations, background execution, and lifecycle management shared one Gradle module. A static `p.studio.Container` made these dependencies easy to reach, but it also made the visual adapter the practical center of the product. That shape could not support another consumer, such as a future agent adapter, without either driving the UI or duplicating Studio behavior.
The migration needed to preserve the existing Java product and its behavior while making JavaFX replaceable. It also needed to preserve the autonomy of compiler, LSP, runtime, and especially `p.packer`; separating Studio did not authorize absorbing those domains.
## Key Decisions
### Human and non-visual adapters consume the same Studio services
**What:** The Studio is physically divided into `prometeu-studio-api`, `prometeu-studio-core`, `prometeu-studio-ui-javafx`, and `prometeu-app`. JavaFX depends on the neutral API, Core implements application use cases, and the application module composes concrete adapters.
**Why:** The stable boundary is UI versus application services, not JavaFX versus a future web framework. A future agent can depend on the same API without moving application rules out of Core.
**Trade-offs:** More constructor parameters and explicit contracts are required. That cost is intentional: dependencies are visible, testable, and constrained by Gradle instead of hidden behind a global container.
### Spring is a composition mechanism, not a domain dependency
**What:** Spring Boot runs in non-web mode only in `prometeu-app`. Core objects remain constructor-injected POJOs. Spring Modulith verifies the six functional Core modules: project, session, assets, shipper, runtime, and execution.
**Why:** Spring provides deterministic construction and lifecycle without contaminating the public API or Core. Modulith supplies executable documentation and dependency checks while Gradle remains the physical classpath boundary.
**Trade-offs:** The application configuration is deliberately explicit. Broad component scanning and turning JavaFX controls into Spring singletons would recreate hidden ownership and lifecycle problems.
### Provider boundaries follow stability, not uniformity
**What:** Core uses a provider's public API directly when it is stable and neutral. It owns a narrow port when the available integration is static, concrete, lifecycle-oriented, or otherwise unsuitable for Core. Packer's stable API is consumed directly; compiler static access, LSP lifecycle, and runtime process integration are isolated behind consumer-owned gateways implemented in `prometeu-app`.
**Why:** Wrapping every dependency creates redundant abstractions, while importing concrete implementations leaks infrastructure into Core. The useful boundary depends on the provider contract.
**Trade-offs:** The integration matrix must be reviewed case by case. There is no universal rule that every provider receives a gateway.
### Long-running work is a neutral application contract
**What:** Build, pack, and execution expose `OperationHandle<R>` with identity, snapshots, completion, cancellation, updates, and typed results. The application owns the executor implementation; JavaFX only projects operation state and switches to the FX thread.
**Why:** Background work must be equally consumable by JavaFX and future adapters. An executor or JavaFX `Task` is an implementation detail, not an application contract.
**Trade-offs:** Cancellation races and terminal state transitions require explicit tests. In particular, stopping execution must cancel the outer execution handle before its nested preparation handle can complete it successfully.
## Patterns and Algorithms
The dependency pattern is:
```text
JavaFX adapter ---> Studio API <--- future adapters
^
|
Studio Core
^
|
Spring non-web composition
```
External providers remain outside the Core module graph. Stable public provider APIs may be direct Core dependencies; concrete adapters remain in the application module.
Operational failures that a caller can recover from are returned as typed results or diagnostics. For example, metadata writes convert I/O failures into `AssetMetadataWriteResult`. JavaFX renders that result and does not treat an unexpected exception message or stack trace as a functional contract.
The removal sequence for a service locator is also reusable:
1. establish architecture tests and final modules;
2. introduce neutral contracts and explicit factories;
3. move behavior slice by slice into constructor-injected services;
4. reduce static call sites to zero;
5. delete the static container and manual composition together;
6. remove the mixed legacy module;
7. run the full build, packaging, desktop smoke, and shutdown gate as one indivisible acceptance step.
## Pitfalls
- Renaming packages without changing Gradle dependencies does not create a replaceable frontend.
- Replacing `Container` with `ApplicationContext.getBean()` throughout the code is still a service locator. Context lookup belongs only at the explicit application bootstrap boundary.
- Registering every control or dialog as a Spring singleton gives Spring ownership of visual lifecycle and makes multiple windows or future adapters harder.
- A gateway that merely mirrors a stable provider API adds indirection without protection. Conversely, direct use of static or concrete provider entry points leaks infrastructure into Core.
- Moving Packer behavior into `p.studio.core.assets` would reverse domain ownership. Studio coordinates Packer; `p.packer` remains independently consumable.
- Returning raw exception messages to JavaFX turns implementation accidents into an unstable user-facing protocol. Expected failures need neutral results or diagnostics.
- Marking the migration complete after compilation alone misses packaging, restoration, real JavaFX interaction, cancellation, and shutdown leaks.
## Takeaways
- The durable product boundary is the Studio service surface, not the current frontend technology.
- Spring belongs in composition; Spring Modulith belongs in verification and documentation; neither belongs in the neutral API.
- Gradle modules enforce physical isolation, while functional Core modules express ownership and allowed dependencies.
- JavaFX is an equivalent adapter, not the owner of project, asset, build, runtime, or execution rules.
- `Container` removal is a completion criterion, not an optional cleanup step.
- A future agent module can now be added as another API consumer without revisiting this migration or coupling AI concerns to JavaFX.
## References
- Decision: `DEC-0059` — Separate Studio Core from JavaFX with Spring Boot non-web and Spring Modulith.
- Plan: `PLN-0138` — Complete Studio Core and JavaFX Boundary Migration.
- Studio architecture: `docs/specs/studio/9. Studio Core and Adapter Architecture Specification.md`.
- Studio specifications: `docs/specs/studio/README.md`.
- Packer boundary: `docs/specs/packer/1. Domain and Artifact Boundary Specification.md`.
- Packer integration: `docs/specs/packer/5. Diagnostics, Operations, and Studio Integration Specification.md`.