Features
-
HEIC and HEIF images — Photos and screenshots from iPhones arrive as HEIC, which nothing in the image stack could decode: the composer showed a broken preview and the image was dropped before it reached the model. Craft Agents now transcodes HEIC/HEIF to JPEG when you attach it, so previews, thumbnails and the model all see a format they can read. Fixes #1099.
-
OpenCode Go provider — OpenCode Go is now a branded provider preset in connection setup, alongside the other OpenAI-compatible presets: pick it from the dropdown and it pins to the right endpoint, with its name and icon filled in. Model IDs are left for you to edit, since its lineup changes over time. Fixes #668.
Improvements
-
Mention sources and skills in Chat Starters — The Chat Starters prompt field in Settings → AI is now the same rich editor as the composer: type
@to pick a workspace skill or source, and it shows as a badge instead of raw[source:…]text. Clicking a custom starter now enables the sources and invokes the skills it mentions, which it previously ignored, and saved starters appear without reopening Settings. -
Configurable output cap for custom endpoints — Custom endpoints capped every model's output at 8192 tokens, so long sessions couldn't compact and failed with "generation hit the token cap and the summary is incomplete". A new endpoint-wide max output tokens (
customEndpoint.maxTokens) lets you raise it; set it to ~13K or more to give custom-endpoint compaction the same budget as the built-in providers. The default is unchanged. Fixes #1100.
Bug Fixes
- "Compact context at" now applies to Pi sessions — The compaction-window setting added in 0.15.2 reached only Claude sessions; Pi sessions ignored it and kept compacting near the model's full context window. Pi now honors the setting, mapping your chosen window to a compaction trigger for the active model.