GPU arbiter daemon mediating VRAM between Ollama (LLM) and ComfyUI (image-gen) on a single-GPU host. Go-based, OpenAI-compatible chat API + ComfyUI-shaped image endpoints, intelligent routing via state-machine with hysteresis. Caymans-resident.
Find a file
Quartermaster aff5f772a2 fix: extend README + drop accidentally-committed build artifact
README now carries the design-phase framing (vs auto-init one-liner).
Build artifact ($module/harbormaster from `go build ./...`) was
committed in the previous commit by accident; removing + gitignoring
so future builds don't recommit it.
2026-06-19 21:01:50 +02:00
cmd/harbormaster skeleton: placeholder main + go.mod + README stub 2026-06-19 21:01:25 +02:00
.gitignore fix: extend README + drop accidentally-committed build artifact 2026-06-19 21:01:50 +02:00
go.mod skeleton: placeholder main + go.mod + README stub 2026-06-19 21:01:25 +02:00
README.md fix: extend README + drop accidentally-committed build artifact 2026-06-19 21:01:50 +02:00

harbormaster

A GPU arbiter daemon mediating VRAM between Ollama (LLM-side) and ComfyUI (image-gen-side) on a single-GPU host. Caymans-resident.

Exposes:

  • OpenAI-compatible /v1/chat/completions (LLM workloads → routed to Ollama)
  • ComfyUI-shaped image generation endpoints (image workloads → routed to ComfyUI)

Routes intelligently based on a state machine with hysteresis + grace periods, so iterative image work doesn't oscillate the underlying model loads.

Status

Design phase. Naming locked, repo seeded, design-anchor tracker incoming (Bosun-routed). No runnable surface yet beyond the placeholder cmd/harbormaster binary; v0.1.0 scope is still being shaped.

Stack

  • Go (single-binary daemon)
  • Configuration-as-data for the routing policy (subject to design)
  • Open-source potential at v0.1.0 cut

Repo conventions

  • main is PR-gated (not jam-shape — design-substantive work)
  • Labels mirror the frankenbit/tmux-tell taxonomy: kind/*, priority/*, size/*, epic/in-progress
  • License: MIT