Pull Request Explorer

Summary

Restore NAX in Metal v4 builds

Fixes missing NAX kernels in Metal v4 builds by updating macOS deployment target to 26.2, ensuring MLX compatibility.

Health Assessment

• • Quick fix with minimal changes, fast review and merge, indicating low complexity and high confidence.

Summary

Add agent TUI command

Adds a terminal UI for interacting with the agent harness, improving usability across terminals and simplifying agent management.

Health Assessment

• • Long cycle time (~6.7 days) and high review friction indicate complex changes; multiple commits reflect iterative refinement and extensive testing.

Summary

Warn before using deprecated agent models

Adds guardrails to prevent launching older weak model families, hiding them from pickers and warning users, while allowing override with --yes.

Health Assessment

• • The PR required over 3 days for review and merge, with a large code change affecting launch logic, indicating significant complexity and potential risk.
• • No AI assistance was detected in the code or commit history.

Summary

Select Qwen3.5 parser and renderer

Ensures correct parser and renderer for Qwen3.5/Next models, preventing empty think blocks and improving content framing for users.

Health Assessment

• • Quick fix with minimal changes, resolved in a single commit after review, indicating straightforward issue resolution.

Summary

Enable iGPU projector offload with fit padding

Improves GPU memory usage for integrated GPUs, enabling faster vision inference and better performance for users on non‑Metal iGPUs.

Health Assessment

• • The PR took over 6 days to merge, with a single review after a day, indicating a slow but straightforward process with minimal rework.

Summary

Health Assessment

Summary

Health Assessment

Summary

Disable Claude Code Telemetry by Default

Disables telemetry for Claude Code launches, improving privacy and reducing data sent to external services.

Health Assessment

• • Fast review and merge with minimal changes; low risk and small scope.

Summary

Health Assessment

Summary

Remove OLLAMA_EXPERIMENT client2 environment variable

The experiment client2 has been removed, cleaning up environment variables and reducing configuration complexity.

Health Assessment

• • The PR involved a large code removal but was reviewed quickly and merged after a single approval, indicating a straightforward cleanup.

Summary

Update llama.cpp to b9888

Bumps the llama.cpp dependency to the latest b9888 commit, ensuring compatibility and bug fixes.

Health Assessment

• • Quick review and merge with minimal changes indicates low complexity and high confidence in the update.

Summary

Rewrite create functionality for MLX engine

This PR refactors the create pipeline for the MLX engine, simplifying model import and conversion steps, removing image generation, and improving performance and maintainability.

Health Assessment

• • The PR had a prolonged review cycle with many iterations, indicating significant complexity and potential risk. The large code churn suggests a major refactor that may impact stability.

Summary

Enable Flash Attention on CUDA CC 6.x GPUs

Allows native SM60/SM61 kernel compilation, enabling Flash Attention on older GPUs and improving LLaMA model performance.

Health Assessment

• • PR resolved a GPU compatibility issue with minimal changes and a quick review, demonstrating efficient issue handling.

Summary

Use UTF-8-safe file open for Windows paths

Ensures compatibility with Unicode file paths on Windows, preventing errors when loading models.

Health Assessment

• • Quick fix with minimal changes, resolved in under a day with a single review round.

Summary

Remove unsupported ROCm devices

This update cleans up ROCm device support, aligning docs and build presets with current supported hardware, reducing build times and avoiding runtime failures.

Health Assessment

• • Quick review and merge within 3 hours, minimal changes, low risk.

Summary

llama: clean up dead code from llama-server work

Removes obsolete code from the llama-server, reducing maintenance overhead and potential runtime errors.

Health Assessment

• • PR was merged quickly with a single review, indicating a straightforward cleanup of dead code.

Summary

Health Assessment

Summary

Add fallback to standard CUDA for JetPack

Ensures JetPack 7 devices use the correct CUDA build, enabling full GPU utilization and improving model performance.

Health Assessment

• • The PR had a long cycle time due to a delayed review, but the change was small and straightforward, resulting in minimal risk.

Summary

Health Assessment

Summary

Avoid unbounded parallelism in build script

Reduces build times for Darwin builds, improving developer productivity and CI throughput.

Health Assessment

• • Quick fix to improve build performance, minimal changes, merged within the same day.

Summary

mlx: tighten up gemma4 moe loading code

Enables consistent tensor naming across quantized and non-quantized models, improving model compatibility and reducing errors.

Health Assessment

• • Quick review and merge indicate a straightforward change with minimal risk.

Summary

Add minimum Hermes CLI version check for launch

Ensures users have a compatible Hermes CLI version, preventing confusing failures when launching Hermes desktop.

Health Assessment

• • Quick review and single commit indicate a straightforward change with minimal risk.

Summary

Health Assessment

Summary

Bump MLX dependency

Update MLX to a specific commit and adjust tests to align with thread‑affinity model.

Health Assessment

• • Dependency bump with minimal changes; review took over a day, but no rework was required.