swarmxq-video-pipeline-architect

The 6-stage order intent_classification through finalizing is immutable and must never be resequenced, skipped, or merged. Modify behavior inside a stage function while preserving its contract: model tier, max tokens, key output, and error code.

ComfyUI with LTX-Video is used when SWARMX_VIDEO_RENDER_BACKEND=comfyui and the server is reachable; FFmpeg is the CPU-only fallback in auto mode. An explicit comfyui setting with an unreachable server fails with COMFY_UNAVAILABLE rather than falling back.

FFmpeg transcoding and LLM inference share the same CPU core pool, so running both concurrently degrades each other. The render_assembly stage must evict all loaded models, including the Pilot despite its 5-minute keep-alive, before starting FFmpeg.

resumeJob requires artifacts from all prior artifact-producing stages before re-queueing. Resuming from storyboard_generation without a scripting artifact fails, so resume from an earlier stage that has its dependencies present.

No. MAX_CONCURRENT_JOBS is a protected constant of 1 because CPU inference is serial. Additional requests are queued with a position, and clientRequestId serves as the BullMQ dedup key for idempotent submission.

DeepSeek emits <think> reasoning blocks that break JSON parsing. Every Ollama response must pass through sanitizeReasoningOutput before parsing so think blocks never reach intent JSON, script text, or storyboard frames.