DECISIONS

2026-02-15 — Profile-first RunPod build flow

We use profile files as inputs and generate lock files as build/runtime truth.
Profile inputs:
- Runpod_Comfy/profiles/runpod.<id>.yaml
- Runpod_Comfy/profiles/nodes.<id>.yaml
- Runpod_Comfy/profiles/models.<id>.yaml
- Optional Runpod_Comfy/profiles/urls.<id>.txt
Compiler output:
- Runpod_Comfy/config/comfyui.lock
- Runpod_Comfy/config/custom_nodes.lock.json
- Runpod_Comfy/config/models.json
Docker base image is taken from generated comfyui.lock (BASE_IMAGE build arg).
Goal: normal updates (new node repo/model URL/version bumps) should require editing config files only, not Python/Docker logic.
Expected operator UX:
- Edit 3–4 simple profile files with real values.
- Run compiler.
- Build/push image.
- Deploy and smoke-test on RunPod.

Planning/serious execution in OpenClaw web dashboard.
Telegram for quick pings.
Repo files (TODO/WORKLOG/DECISIONS/WORKFLOW) are continuity anchor across devices.

Use a 3-zone layout inside Runpod_Comfy/:
- setup/ (operator-facing edits/tools; keep lightweight)
- image/ (generated outputs + active locks + deploy runbook)
- core/ (runtime internals + technical references)
Keep one recipe folder per image variant under setup/recipes/<id>/.
Compiler supports explicit args (--id, --out, --image-tag, --no-activate).
Default outputs go to image/builds/<recipe-id>/ and can be activated to image/active/.