Supported outcome

Choose or reuse one approved local model provider for a task without silently changing the running provider.

Terminal result: LocalCode returns a verified loopback provider capability and a separate record of why that provider was assigned to the request.

• tested-candidate · macOS 26.5.2 arm64 on a 24 GiB Apple M4 Pro; Qwen2.5-Coder-7B-Instruct-3bit at revision a5ed5c3fb82a68cfd1682b910ac8a492249bbe3e; MLX-LM 0.31.3 and MLX 0.32.0 · Explicit candidate-profile recommendation, approved provider preparation, verified loopback readiness, approved no-mutation reuse, exact status and owned disconnect · docs/releases/0.5.0-rc.1-acceptance.md
• unknown · Other Apple Silicon memory cohorts, Intel macOS, Linux and Windows · Provider recommendation, preparation and reuse · No installed 0.5.0rc1 acceptance run is retained for these environments

Non-goals

• Running or completing the user's coding task
• Claiming the best model for an arbitrary task
• Autonomous model installation, promotion or provider replacement
• Stable-release support
• Remote-machine placement, energy optimization or long-running-agent supervision

Exact recovery boundary

Strategy: rebuild-verified-legacy-source

Disconnect: localcode disconnect

Prior retained version: 0.4.0-rc.2

State backup required: yes

• Disconnect active providers before changing product versions; active-provider migration is not supported.
• Rollback to 0.4.0rc2 requires a verified pre-upgrade state backup and a rebuild from its retained source bundle.
• Cached model artifacts are retained across provider rollback and must be removed separately if no longer wanted.

Exact install command

uv tool install --force --constraints localcode-0.5.0-rc.1-runtime-constraints.txt ./localcode_prototype-0.5.0rc1-py3-none-any.whl