OQ-001 Should ComfyUI be replaced with a different image generation backend long term? ComfyUI has a complex workflow JSON API that is brittle to maintain. Alternatives like Automatic1111 or a direct diffusers integration may be simpler. Resolution trigger: after first working fantasy mode demo, evaluate ComfyUI maintenance cost against alternatives.
OQ-002 How should the system behave when a pod fails to start after three provider API retries? Currently the design surfaces a 503 to the client. A dead letter pattern with Telegram admin notification may be more appropriate. Resolution trigger: first time this failure occurs in real usage.
OQ-003 WhatsApp integration is the preferred bot channel long term. Telegram is the build target but the bot service is designed as a protocol adapter so migration does not require changes to any other service. Resolution trigger: when the system is stable and Twilio WhatsApp sandbox evaluation is done.
OQ-004 Phase 2 persistence design is deferred. When Zoltraak Gateway needs to serve data to external apps like Immich or a custom client, a persistence layer will be required. The domain model should not assume OpenWebUI is the permanent store. Resolution trigger: first external integration request.
OQ-005 Should both providers be active simultaneously or is one always primary with the other as fallback? The current design allows configuration-driven selection but does not define automatic failover logic. Resolution trigger: after first provider outage experience.
OQ-006 The gateway returns HTTP 202 with warming-up status for requests that arrive during pod startup. OpenWebUI may not handle 202 responses gracefully on the Ollama-compatible endpoints since it expects the Ollama protocol. This needs testing against real OpenWebUI behaviour. Resolution trigger: first end-to-end OpenWebUI test against a cold pod.
OQ-007 The home server (HP EliteDesk 800 i5 6th gen) runs the local Ollama instance on CPU. Intent classification latency on CPU for small models like phi3:mini needs to be validated. If latency exceeds acceptable bot response time, a pre-loaded persistent model or a rule-based fallback may be needed. Resolution trigger: first real usage test of the Telegram bot.