🏗️ Architecture: How to validate generated platform output before it reaches ...

Date: 2026-02-28
Time: 21:38:43 UTC
Session: gspace1-156.2
Agent: backend
Confidence: 92%
Status: ✅ Documented

📋 Decision Context

Question:

How to validate generated platform output before it reaches users?

Task: Not specified

🎯 Decision Made

Chosen:

BlueprintValidatorService + ConvergenceLoopService in backend-ai-api. Validator checks all 25 widget types, page structure, config completeness, placeholder text. Loop iterates validate→repair up to 15 times with stall detection (4 consecutive no-progress), using MCPBridgeService to call Claude for repairs.

Options Considered:

• No alternatives documented

💭 Reasoning

ROOT CAUSE: Generated output had ZERO validation. Widgets could have empty configs, pages could be missing required fields, placeholder text like 'TODO' could pass through to users. gmodules1 solved this with a convergence loop (1,831 lines) that catches 90%+ of issues. We ported the pattern but adapted it for our widget-based architecture instead of their code-based one. The validator checks type-specific configs (heading needs text, table needs columns+data, button needs text). The loop tracks persistent errors (3+ iterations → try different strategy) and regressions (new errors → revert). Exit conditions prevent infinite loops. Repair function wired to MCPBridgeService.executeTask() which calls Claude Code CLI with the system-integrate-repair.prompt.md template.

📁 Files Involved

• APIs/backend-ai-api/src/app/workflow/phase-executor/blueprint-validator.service.ts
• APIs/backend-ai-api/src/app/workflow/phase-executor/convergence-loop.service.ts
• APIs/backend-ai-api/src/app/workflow/phase-executor/phase-executor.service.ts

🔗 Related

• Decision ID: dev_1772314723638_a0v77py
• Session: 156
• Workspace: gspace1

Auto-generated by MCP Smart Consolidation