feat: production Stripe webhook async architecture with Redis, Jules worker, and optional AI providers

[Copilot]

Replaces the monolithic Stripe webhook handler with a layered async architecture: the HTTP layer validates, deduplicates, and enqueues in <10ms; all business logic runs in a separate Jules worker process consuming from Redis.

HTTP layer (api/index.py → thin adapter)

• Removed 220-line fat handler; replaced with ~50-line adapter importing from services/
• webhook(): verify sig → Redis idempotency check → mark_event_received → enqueue_event → 200
• Static slug map for WebhookVerificationError responses (fixes CodeQL py/stack-trace-exposure)
• Dropped in-memory processed_events: set / queued_events: list (not safe across Vercel invocations)

Services layer (api/services/)

• redis_client — singleton with _MemoryRedis fallback for dev (no Redis required locally)
• redis_lock — distributed lock via SET NX EX with token-checked release (prevents releasing a different worker's lock on TTL expiry)
• stripe_service — verify_webhook() decoupled from Flask; raises typed WebhookVerificationError
• event_store — Redis hash state machine: RECEIVED → PROCESSING → PROCESSED / FAILED; includes retry counter and TOCTOU guard (re-check after lock acquisition)
• queue_service — LPUSH/BRPOP main queue + sorted-set delayed requeue + DLQ

Jules worker (api/orchestrators/)

• decision_engine — pure Python, no I/O; decide_event_flow(event, context) → Decision
• execution_engine — dispatches to handler; calls AI enrichment via _run_ai_optional() (swallows all errors — AI failure never blocks business logic)
• jules_worker — blocking loop: flush delayed → dequeue → lock → decide → execute → mark done or exponential-backoff retry (2s/4s/8s) → DLQ at MAX_RETRIES

# Run worker externally (not on Vercel)
PYTHONPATH=api python -m orchestrators.jules_worker

AI providers (api/ai_providers/)

• AIProvider ABC + GoogleAIStudioProvider (optional google-generativeai SDK) + ManusAIProvider (plain HTTP)
• Per-provider CircuitBreaker (3-failure threshold, 60s recovery window)
• run_ai_task(task_type, payload, provider=None) — provider selected by AI_PROVIDER env var; never called from the HTTP layer

Models & config

• Decision dataclass: action, reason, priority, requires_ai, provider, retryable, metadata
• EventAction enum covers: mark_order_paid, mark_payment_failed, activate_subscription, handle_refund, request_manual_review, invoke_ai_enrichment, ignore_event
• api/config.py — single source of truth for all env vars with validate() at startup

Infrastructure

• api/requirements.txt — added redis>=5.0.0
• .env.example — documents all vars: Stripe, Redis, AI providers, worker tuning

What's mocked / needs real implementation

• handlers/stripe_handlers.py — TODO stubs for DB writes, emails, fulfillment
• ManusAIProvider._build_request_body / _parse_response — update once Manus API schema is confirmed
• _MemoryRedis — single-process dev only; set REDIS_URL for any multi-process deployment

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
tryonyou-app-1776375445	Error		Jun 2, 2026 2:33pm
tryonyou-app-work	Error		Jun 2, 2026 2:33pm
workspace	Error		Jun 2, 2026 2:33pm

[LVT-ENG]

He'd

[LVT-ENG]

Yes