v0.3.0

Added

• Added stop-command cancellation for OpenAI-compatible requests, so commands such as stop and cancel generation cancel matching in-flight local work instead of starting another GPU request.
• Added runtime telemetry support for model-load progress, inference prefill progress, prefill input counts, and transfer bandwidth.
• Added recent_work to /status so fast one-shot jobs remain visible after completion.
• Added a Recent Work table to /dashboard.
• Added runtime adapter/mode fields to work status: runtimeAlias, runtimeAdapter, and runtimeMode.

Changed

• Updated the dashboard to show runtime, adapter, and mode instead of the confusing public type/kind vocabulary.
• Reworked one-shot LoRA/llama-cli work to appear publicly as a llama_cli runtime adapter with one_shot_command mode.
• Stabilized dashboard loading animation and live status rendering.
• Updated doctor and agent-install guidance around coordinated one-shot runtime adapter work.

Fixed

• Fixed dashboard loading progress edge cases where unavailable telemetry could appear as a misleading zero-percent value.
• Fixed dashboard status refresh behavior after the gateway restarts while a runtime is already loaded.