All stories
Osaurus Ecosystem Community reports

Osaurus gives Mac agents finer thinking controls and longer tool chains

Version 0.25.31 exposes native reasoning-effort tiers, adds optional response metrics per agent and summarizes growing tool chains between completed batches.

By AITrending2 min read
Sources (6)

Why now The original 0.25.31 GitHub release was published on November 2 at 18:57:51 UTC. Its tagged documentation and linked changes establish a new application release, separate from the October 24 harness integrations.
Project-published Osaurus interface showing a Writer agent, model selector, Sandbox and Folder controls, and a message composer.
Osaurus interface artwork from the English 0.25.31 README. It illustrates the application workspace, not a demonstration of the new reasoning tiers or an independent model-performance test. Source: Osaurus contributors

Osaurus 0.25.31 gives local Mac agents more precise thinking controls and a way to summarize a growing tool chain before the next model request. Published on November 2, the community application’s new release also restores optional inline response statistics per custom agent.

The changes address two practical problems: a model’s declared reasoning options being flattened into a simple on/off choice, and a long-running agent accumulating tool results without triggering the existing summarizer.

Thinking tiers follow the model’s contract

The linked reasoning-control change says bundles declaring both boolean thinking modes and explicit effort tiers previously appeared as On/Off in the editor and model picker. The new handling recognizes the fuller declaration: None disables thinking, Light sends native low effort and High sends native high effort.

Those options depend on the model bundle. They are not a promise that every local or cloud model supports the same controls. The maintainer reports focused tests and real-app checks, while explicitly limiting the performance conclusions that can be drawn from short, different tasks. This release is not evidence of a universal inference speedup.

Summaries can happen during the work

The compaction change runs the existing summarizer between completed tool batches before constructing the next request. It preserves the latest complete batch where possible, and rejects incomplete or mismatched tool-result boundaries. If summarization fails, the continuation ends with the original transcript intact, according to the change description.

A separate Advanced Settings toggle shows response metrics per agent. It exposes existing token counts, timings and decode-rate measurements; it does not introduce a new benchmark or a token-boundary visualizer.

Osaurus is a standalone Mac harness around models, memory and tools. Installation is available through brew install --cask osaurus or the release’s disk image. The tagged README requires Apple Silicon and macOS 15.5 or newer. Its Linux VM sandbox requires macOS 26; earlier supported versions use a different native sandbox with different network and filesystem behavior.

Local models can run offline. Choosing a cloud provider sends requests to that service and brings its access requirements and costs. Custom-agent working folders, shell tools and optional browser or app integrations also determine what the agent can access. The release improves that workspace’s controls; it does not make every configuration local or grant new model capabilities by itself.

Related reading