The Owned Surface Is One Protocol
Firebrand's primary planes are not 'local versus every cloud vendor.' They are two classes of model behind one native protocol: a local model on an operator-selected daemon, and a cloud-class model on that same daemon after the operator has signed in there. The engine still talks only to the daemon. The egress fence does not open because a tag ended in a cloud suffix.
A cloud-class model is detected as a tag suffix, a stub small enough to be a pointer, or an empty format field. That classification flips the leg to first-class cloud: different identity, different memory, different context window, same wire. A /model flip across that boundary has to re-resolve all four and write the change down. Resume trusts the latest recorded state, not the first session_start.
The Daemon Will Silently Lie If You Let It
The daemon's default context size is a VRAM guess. Its default truncate drops history from the front and keeps smiling. A prompt that does not fit can still return tokens, and the model will invent the missing prefix. Firebrand therefore always sends its own context length, forbids silent truncate, and treats a real overflow as a typed stop. Local legs take the model's native context whole unless the operator clamps it. A runner that accepted a fifteen-thousand-token prompt inside a one-thousand-token window is why the engine, not the runner, owns the cap.
Compatibility Is Not the Center
OpenAI-shaped servers, subscription backends, and key-auth brains exist as legs. They are how the family reuses logins it already opened elsewhere. They are not the cleanroom, and they are not the first research track. If those legs vanished, the two Ollama planes would still have to carry a coding turn.