feat(ai): replace JSON parsing with structured output via tool_use

[silver-snoopy]

Summary

• Add completeStructured<T>() to AIProvider — uses Claude's tool_choice (forced tool call) and Gemini's responseSchema to guarantee schema-valid JSON responses
• Migrate 3 LLM consumers — report generator, plan tuner, and plan parser all replaced fragile 3-tier JSON parsing (JSON.parse → jsonrepair → brace-matching) with structured output
• Remove jsonrepair dependency — ~250 lines of parsing code eliminated, retry-on-parse-failure loop removed
• Upgrade plan parser — from regex-only to LLM-powered with intelligent exercise classification, regex fallback when AI unavailable
• Security fixes from review — PUT route rawText size cap, prompt injection scanning in plan parser LLM path, Gemini JSON.parse error handling
• Documentation — architecture.md updated with workout-plans service tree, API endpoints, data model tables; ADE Phase 8 checklist added to CLAUDE.md

Files changed (17)

Layer	Files	Change
Shared	ai.ts	+StructuredOutputConfig interface, +completeStructured() on AIProvider
Providers	claude-provider.ts, gemini-provider.ts	Implement completeStructured()
Retry	retry-utils.ts	+completeStructuredWithRetry()
Report	report-generator.ts	Replace 3-tier parser with structured output
Tuner	tuner.ts, tuner-prompt-builder.ts	Replace parser + retry loop with structured output
Parser	plan-parser.ts	LLM structured output + regex fallback
Routes	workout-plans.ts	Pass aiProvider to parser, add PUT size cap
Deps	package.json, package-lock.json	Remove jsonrepair
Docs	architecture.md, CLAUDE.md, plan status	Architecture sync, ADE Phase 8 checklist
Tests	3 test files	Updated mocks from complete → completeStructured

Test plan

• 409 backend tests passing (43 test files)
• 74 frontend tests passing (17 test files)
• npm run lint — 0 errors
• npm run format:check — clean
• tsc build clean (shared + backend)
• Code review: 3 parallel agents (Logic, Conventions, Security) — all HIGH/MEDIUM findings addressed
• Integration test: generate a report with structured output against live AI provider
• Integration test: tune a plan with structured output
• Integration test: parse a free-text plan via LLM path

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
vitals	Ignored		Apr 13, 2026 9:32pm