GOAT 0.1.3 (Kid)

GOAT 0.1.3 makes local-model work easier to set up, easier to follow and more dependable over a long session. It brings model management into the app, keeps large conversations responsive and gives coding workflows clearer recovery when a model or engine has a difficult turn.

This release keeps oMLX 0.6.4 as the qualified baseline.

Your models, in one place

• A new Models area in Settings shows what the active engine can run, with favourites, capability filters, compatibility details and a direct model picker.
• Model menus now put the current model and favourites first. Clear icons distinguish coding, vision, multimodal, reasoning and general models.
• GOAT can apply bounded context and sampling guidance for a model while respecting engine facts and your explicit overrides. Response diagnostics show where the effective settings came from.

Long work stays usable

• Conversations can compact when they run out of context, and /compact lets you start that process yourself. The original messages remain available for recovery and review.
• Large transcripts no longer ask the interface to lay out an entire oversized response at once. Earlier and Later navigation keeps the full conversation reachable, while your reading position stays steady as new output arrives.
• Very large responses and reasoning remain selectable and copyable. Copying the full source preserves Markdown, code fences, Unicode and whitespace.
• A saved Custom engine stays selected when it is temporarily offline, so GOAT reports the real connection state instead of quietly switching to another discovered service.

More dependable coding conversations

• GOAT can retry a transient engine failure when no output has arrived. Once a response has started, it will not silently replay the turn.
• Printed or malformed tool markup is treated as text and never executed as a command.
• File reads can return more useful plain text, search supports globs and regular expressions, and independent read-only tool calls can run together. Edits and commands remain ordered and permission-controlled.
• Approved Pen commands use the selected Xcode toolchain and report clearer permission failures while remaining confined to the Pen.
• oMLX connections show clearer model, memory, queue and output-limit information when the server provides it. A response stopped at the effective output limit offers Continue response.
• Stats now separates server-reported decode speed from GOAT's received-output estimate and records bounded timing diagnostics without saving prompt or response content.

Testing and developer experience

• Tests are organised by the domain that owns the behaviour, including Bleet, Paddock, Pens, Memory and Inference. Website checks are separate from macOS app and embedded-preview tests.
• Focused commands such as make test MODULE=Bleet and make test MODULE=Paddock run the relevant package and hosted coverage together. The release gate rejects empty test selections.
• GOAT.xcodeproj is maintained source. A fresh checkout can open, build, debug and test in Xcode without regenerating the project or installing XcodeGen.
• Shared schemes, domain test plans and the app's dependency lock are committed. make build-signed creates a hardened, timestamped Developer ID build and verifies its publisher, team and release entitlements.

Installation and upgrade

GOAT 0.1.3 requires Apple Silicon and macOS 26 or later. Download the signed, notarized latest release with its manifest and checksums.

An ordinary upgrade replaces the app without resetting chats, engine profiles, memory, permissions or Pens. Finish active work before quitting GOAT to install the update, and keep a backup of local data.

Qualification and known limits

The official Release-channel package is built from the immutable v0.1.3 source. The release workflow verifies the complete Release suite, website and docs, Developer ID publisher identity, hardened runtime, notarization, stapling and artifact checksums. See the release metadata for the exact source revision and artifact hashes.

The previously reported difference between a 37 tok/s server reading and a 2-3 tok/s live display has not been reproduced. This release makes those measurements distinct and easier to diagnose; it does not claim an engine-speed improvement. Model and tool results apply to the tested configurations, including stable oMLX 0.6.4, rather than every compatible server or workload.