v0.3.63

Fixes

• An engine joined with grid join --at and no --ctx-size now advertises the context window its own /models states (vLLM's max_model_len). The relay can then compact long conversations for it instead of passing on the engine's context-length refusal. --ctx-size still wins (64e1cbc)

Docs

• docs/cli.md and the openai-error schema describe the error object a non-stream failure now carries for a consumer using a raw token (7c53742)

Full Changelog: v0.3.62...v0.3.63