Repository navigation
v0.3.3
A patch release: long jobs no longer stop when the server chokes on a reply, suggested prompts start right away, and chats keep their server slot on llama.cpp.
Fixes
- A failed request no longer ends the turn. When the server can't read the model's reply (llama.cpp answers a malformed tool call with an HTTP 500 or an error in the stream), byob tells the model what went wrong and asks again, up to two more times. Other server errors (5xx, a dropped stream, "Loading model") are tried again after a short pause. Requests that are wrong in themselves (a 4xx such as context too long) still stop at once.
- A broken tool call no longer switches the endpoint to prompt-based tools. Before, a "Failed to parse tool call" error could be mistaken for a server that doesn't support tools.
- The model fixes failed tool calls instead of giving up. The default system prompt now says to read the error and try again; calls you decline are still not repeated.
Interface
- Suggested prompts send on click in an empty chat, instead of only filling in the composer. While an attachment is still uploading, the prompt goes into the composer as before.
Servers with slots (llama.cpp)
- byob sends
id_slotto servers that reporttotal_slots, so each chat keeps the same slot (and its prompt cache) from turn to turn, and subagents run in slots no chat is using instead of evicting the main chat's cache. Endpoints that don't report slots, such as hosted APIs, get nothing extra.