Roadmap

The desktop task adapter's broader validation remains open. Measure task completion, incorrect actions, recovery and latency through the configured model connection. Decision models remain candidates until tested on the same held-out tasks.

What comes after v1.113, in rough order. Nothing here is a promise. docs/CHANGELOG.md says what shipped.

No new verb until items 1 and 2 are done.

1. A faster first line from spark read. Put the reading pass in the answer's stream. Send the text early, while the reader types.
2. Grounding scores for every model auto may pick. Score the 30B-A3B on a machine it fits. Find a fix for answers that keep lines the source does not hold.
3. After a reboot, wait up to 30 seconds for the promised address before the server starts. Today it can bind to the Wi-Fi instead.
4. Let use decide. Read spark stats each week, and drop verbs nobody runs. Then two small hints: an alias you keep typing, and a tool you have but do not use.
5. A first new user installs spark from README.md alone and says what broke.
6. Failure tests on a real machine: the engine killed mid-reply, a full disk, no GPU, the server lost mid-answer.
7. spark do knows more "bad option" messages, such as procps's improper ... field descriptor. The man page lines then come back.
8. spark do ?? goes on with the last task, as ?? does at the prompt.
9. The voice answers in the language it heard. spark stats shows the wait before the first sound.
10. Ollama as a tested address for spark client URL. Today only llama.cpp's llama-server is tested.