[chingyachung]

Integrate Final V2 UML Activity Model Generator and Human–AI Refinement

Base: develop
Head: feature/thesis-generator-final

Summary

Integrate the complete Final V2 Generator application with the laboratory develop branch. The application supports three-candidate generation, candidate selection, human editing and saving, synchronization into a new revision, iterative refinement, and revision restore. It uses the Semantic–Deterministic Pipeline with ActivityGraph validation and AI4MDE conversion.

Main changes

• Add topology and semantic planning, deterministic sketch construction, repair, compilation, conversion, and the required prompts and dependencies.
• Add candidate and revision persistence, authenticated Generator APIs, and the frontend selection and human-in-the-loop workflow.
• Protect saved edits: Sync waits for local saves; unresolved PATCH failures block Sync and Restore; Refinement checks saved process content against the synchronized revision before generation and again before import. Layout-only changes are ignored. Unsupported process fields are rejected instead of silently discarded.
• Compare generated and saved models through the same AI4MDE representation, so generated Decision label/name defaults do not create false conflicts.
• Add offline backend and frontend regression tests and the Generator production-surface documentation.

Verification and demo evidence

• Django model and metadata tests: 75 passed.
• Offline LLM-module pytest: 401 passed, 4 skipped. Live LLM tests require explicit opt-in and an API key.
• Frontend unit tests: 19 passed; frontend production build passed.
• git diff --check passed before commits. No paid LLM calls or official-candidate regeneration were performed.
• The thesis author successfully completed manual E2E testing of three-candidate generation and selection; Human Edit → Save → Sync; AI Refinement; and Revision Restore.
• Demo evidence follow-up: no recording or screenshot of this final E2E run is currently linked. Provide one if the laboratory requires visual evidence for review.
• Implementation scope and limitations: docs/PRODUCTION_SURFACE.md.

Review focus

1. The Semantic–Deterministic Pipeline, validation, and AI4MDE conversion.
2. Candidate selection and the Save → Sync → Revision → Refinement → Restore lifecycle.
3. Bearer authentication on Generator write routes and frontend token handling.
4. Process-field preservation, revision conflict checks, and PATCH failure handling.
5. The project-wide PostgreSQL image pin in docker-compose.yaml.

Known limitations and architect review

• Sync rejects incomplete or unsupported manual process structures with 422; for example, deleting a node may leave a branch without a valid closure. This preserves the saved model and does not create a revision.
• Refinement returns 409 for unsynchronized process edits or fields that cannot safely survive conversion. The saved model is preserved.
• Independent clients or browser tabs do not have a shared atomic edit-version lock. The local editor guards its own pending and failed saves.
• Project and System records inherit the laboratory's shared-access model; the Generator does not add per-user ownership. Other laboratory routes still accept cookie-authenticated writes without CSRF checks. The laboratory architect should decide the project authorization and shared CSRF policy before merge.
• The laboratory's existing tracked environment configuration and any credential exposure in develop history need separate security review and rotation decisions. This branch does not add personal credentials or E2E-only Docker files.
• docker-compose.yaml pins postgres:15-alpine instead of the floating postgres tag to maintain compatibility with the existing development database volume and avoid an implicit major-version change. The laboratory must confirm its deployment version and actual volume compatibility before merge; this PR does not reset or upgrade a volume.

Thesis provenance

This PR submits the Final V2 application from feature/final-v2-integration (56e82629). The official thesis experiment's 40 cases and 120 candidates were generated with the frozen feature/thesis-ui-isolation implementation (3227af4b), as recorded in its run manifest. This PR does not claim that Final V2 generated those frozen candidates and does not include frozen evaluation outputs.

[sonarqubecloud]

Quality Gate failed

Failed conditions
C Security Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE

[chingyachung]

Generator Demo

This video demonstrates the main features of the Final V2 Generator, including:

• Generating three candidate UML Activity Models from a textual process description.
• Selecting a candidate model.
• Manually editing and saving the model.
• Synchronizing manual changes with AI.
• Refining the model using natural language instructions.
• Viewing and restoring previous revisions.

The demo shows how users can interact with the generated models and refine them through the Human–AI Interaction workflow.

demo video link: https://drive.google.com/file/d/1nqbaXcSXsHbaK0DeZdVPvdP8Mww23uFc/view?usp=sharing

[chingyachung]

Closing this PR because the branch has been renamed to feature/chingya-2026-ai-assisted-activity-diagrams, following Irina's feedback. Since the repository is currently being refactored, I will open a new PR when the team is ready.