Handoff: Platform Generation Pipeline Disconnect

Session 138 | February 23, 2026 | Testing Director

Scope

Complete platform generation pipeline — from AI generation through universe deployment.

• Status: 3 Critical Fixes Applied + 5 Issues for Coding Team
• Test URL: http://localhost:6200/space/chrome-space-1/universe/paks-uni-1/hub
• Backends: backend-api (3000), backend-ai-api (3001)

1. Executive Summary

The platform generation pipeline has a fundamental disconnect between what the AI generates and what appears in the universe hub. Testing confirmed that generation completes successfully (SSE streaming works, modules are generated), but the universe structure remains empty — zero modules persist to the backend.

Root cause analysis identified three bugs in the manifest generation and persistence chain. Fixes have been applied to workflow.service.ts and vibe-universe-generator.component.ts. Both compile clean. The fixes require server restart and full E2E re-testing to validate.

2. Fixes Already Applied (Require Validation)

Fix 1: Reused Modules Lost from Manifest [CRITICAL]

Fix 2: All Entities Assigned to Every Module [HIGH]

Fix 3: Silent Auto-Sync Failure [MEDIUM]

3. Issues for Coding Team

Issue 1: Two Separate Generation Pipelines [CRITICAL — Architectural Decision Required]

There are TWO completely separate generation pipelines that are NOT connected to each other:

• Pipeline A: backend-ai-api/workflow.service.ts — Called by the frontend vibe-universe-generator via SSE at /workflow/start. Uses LLM → splitIntoModules() → code generation.
• Pipeline B: backend-vitto-api/module-generation/ — The 7-phase pipeline (ENHANCE→ANALYZE→DECOMPOSE→PLAN→EXECUTE→INTEGRATE→DEPLOY). Session 137 improvements were applied here.

Impact: Session 137 fixes (route generator, auto-correction, deploy phase null checks) apply to Pipeline B, but the frontend uses Pipeline A. The 7-phase pipeline is unreachable from the UI.

Recommendation: Either: (a) Wire the frontend to call Pipeline B instead of Pipeline A, or (b) Port the session 137 improvements into Pipeline A. This is the single most impactful architectural decision.

Issue 2: Universe Structure Empty After Generation [CRITICAL]

Issue 3: splitIntoModules() Threshold Too Low [MEDIUM]

Issue 4: Naming Bugs in Word Splitting [LOW — Fixed]

Issue 5: Module Catalog Prevents Re-generation [MEDIUM]

4. Data Flow: Generation to Universe Display

Pipeline A (Currently Active)

Frontend: vibe-universe-generator.component.ts
  startGeneration() → WorkflowStateService.startWorkflow()
    → SSE to backend-ai-api/workflow/start

Backend: workflow.service.ts
  LLM generateModule() → splitIntoModules()
    → for each: codeGenerator.generateModule()
    → context.generatedModules.push()
    → generateFinalManifest(context)
    → SSE workflow_complete { manifestJSON }

Frontend receives manifest:
  handleWorkflowComplete() → generatedManifest.set(manifest)
    → autoSyncModulesToCanvas()
    → extractModulesFromNavigation()
    → buildGeneratedUniverseTree()
    → replaceTree() → persistTreeToFlatArchitecture()

Backend persistence:
  PUT /api/spaces/{id}/universes/{id} with { structure: [...] }
    → Universe.structure updated in MongoDB

Hub loading:
  universe-hub.component.ts → universeModuleService.load()
    → POST /structure/hydrate → reads Universe.structure → modules signal

Pipeline B (Session 137, NOT Connected to UI)

POST /vitto-api/module-generation/generate
  → 7-phase pipeline:
    ENHANCE → ANALYZE → DECOMPOSE → PLAN → EXECUTE → INTEGRATE → DEPLOY

⚠️ This pipeline is NOT called by any frontend component. Session 137 improvements (route generator, auto-correction, deploy null-checks) only apply here.

5. Files Modified in This Session

6. Verification Steps

After restarting all servers:

1. Backend compile check: npx tsc --noEmit --project apis/backend-ai-api/tsconfig.app.json
2. Frontend compile check: npx tsc --noEmit --project libs/libs-structural/vibe-universe/tsconfig.lib.json
3. Navigate to hub: http://localhost:6200/space/chrome-space-1/universe/paks-uni-1/hub
4. Click AI Generator, select Fitness Studio template, run generation
5. Check SSE events: Verify module_generation_complete fires for EACH module (should be >1)
6. Check backend logs: Look for "generateFinalManifest: X total modules (Y generated, Z reused)"
7. After completion: Navigate back to hub. MODULES count should be >0, matching the generated count.
8. Hydrate check: In DevTools console: fetch('/api/spaces/{id}/universes/{id}/structure/hydrate', {method:'POST'}).then(r=>r.json()).then(d=>console.log(d.data.length))

7. Recommended Priority Order

8. S138.2 Continuation — Deep Analysis Results

Agent: S138.2 Continuation Agent | February 23, 2026

8.1 Fix Verification Results

All three fixes were code-verified and TypeScript compilation passes clean for both files.

8.2 Seven Newly Discovered Bugs

Beyond the original 5 issues, deep code analysis revealed 7 additional problems:

8.3 Pipeline Comparison (Complete)

Pipeline A: backend-ai-api/workflow.service.ts

• Size: 1,716 lines (monolithic)
• Frontend: Called via SSE at /ai-api/workflow/start
• Has: LLM generation, splitIntoModules() (threshold <=3 entities), module catalog reuse, code generation, manifest generation, SSE events (ReplaySubject buffer=100)
• Missing: Route generation, auto-correction, validate-fix loop, checkpoint/resume, deploy phase, structured phase orchestration
• Validation: Stub (hardcoded pass)

Pipeline B: backend-vitto-api/module-generation/

• Size: ~3,376 lines across 16 well-structured files
• Frontend: NOT connected (but SSE is fully implemented, ReplaySubject buffer=50)
• Has: 7-phase pipeline (ENHANCE→ANALYZE→DECOMPOSE→PLAN→EXECUTE→INTEGRATE→DEPLOY), auto-correction (12 regex rules), route generation (deterministic), validate-fix loop (2 LLM fix attempts), MongoDB checkpoint/resume, deploy phase, session management, 10 REST endpoints
• Missing: Module catalog reuse, splitIntoModules() heuristic (uses LLM-based DECOMPOSE instead)
• Validation: Structural checks + LLM-based iterative fixing

Feature Matrix

8.4 Architectural Decision: Pipeline Unification

Decision: Wire frontend to Pipeline B. Deprecate Pipeline A's generation orchestration.

Confidence: 0.92 (raw) / 0.53 (calibrated — low historical data for architecture decisions)

Alternatives Considered

Why Option A

1. Pipeline B is objectively superior — structured across 16 files vs 1 monolith, with validation, auto-correction, route generation, and checkpoint/resume
2. Hybrid state already exists — resumeWorkflow() already calls VITTO_API_URL, creating inconsistency
3. Pipeline B's DEPLOY phase already calls backend-ai-api internally for persistence
4. Porting S137 improvements into Pipeline A (Option B) would add ~1000+ lines to an already 1716-line monolithic file
5. Pipeline B uses LLM-based DECOMPOSE (smarter than entity-count heuristic) for module splitting

Migration Plan

What Would Make This Decision Wrong

• If Pipeline B's SSE event format is incompatible with the vibe-universe-generator's expectations at a deep level (more than a simple adapter)
• If backend-vitto-api has reliability/availability issues in production compared to backend-ai-api
• If the module catalog reuse feature (Pipeline A exclusive) is critical for user experience and cannot be cleanly ported to DECOMPOSE

Early Warning Signals

• Frontend changes take >2 days → indicates adapter layer is too complex → reconsider Option B
• Pipeline B SSE events missing data the frontend needs → indicates API surface gap → add missing event types before proceeding

8.5 Updated Priority Order

8.6 MCP Decisions Logged

8.7 Conventions Added

8.8 Lessons Learned

1. Parallel pipelines without coordination = compounding debt. Session 137 improved Pipeline B while the frontend still called Pipeline A. No one connected them.
2. Handoff claims need grep verification. pluralizeName() was marked "FIXED" but never existed. Always verify at code level.
3. Hybrid state is a strong unification signal. startWorkflow() calling Pipeline A but resumeWorkflow() calling Pipeline B proves the architecture is already trending toward B.
4. Monolithic files hide problems. A 1716-line service file accumulated dead code, stub validation, and race conditions that went unnoticed.
5. Persistence gaps are silent. saveConfiguration() dropping reused modules from MongoDB would only surface during future module lookups — a time-bomb.