AutoWeight

Repair a local Hugging Face model from Codex or Claude Code. The shared auto-weight skill lets the host agent propose an edit. The local MCP server applies it to a temporary copy, compares it with the original, and exports a complete checkpoint if the selected candidate passes confirmation.

Set up the local runtime

Install auto-weight[hf,mcp] from PyPI, then run auto-weight setup-runtime --input-root /path/to/repair-inputs with your existing input directory. Once it reports ready, the plugin can use the saved interpreter and paths. Install the Python runtime · MCP runtime configuration

Load the plugin

Use the repository tag that matches your PyPI installation. For example, with version 0.5.0, clone the release tag and enter the checkout.

git clone --branch v0.5.0 https://github.com/midhunxavier/auto-weight.git
cd auto-weight

If you installed the runtime from source, use that same checkout. Run the command for your host from the repository root.

Codex. Add this repository as a local marketplace and install AutoWeight. Start a fresh Codex session and invoke $auto-weight. The desktop app can also discover it through the AutoWeight repository marketplace. OpenAI Docs for local marketplaces

codex plugin marketplace add "$PWD"
codex plugin add auto-weight@auto-weight

Claude Code. Load the same plugin directory for a session and invoke /auto-weight:auto-weight. Claude Code plugin setup

claude --plugin-dir "$PWD/plugins/auto-weight"

Ask for a repair

Fix the fenced JSON responses from base and preserve queue extraction. Use development.jsonl and confirmation.jsonl. The improvement group is api. The compatible checkpoint at prior is available as a donor. Export the repaired checkpoint if it passes.

Paths can be relative to your configured input root. Supply at least 32 labeled examples per gated group in each JSONL file. Label improvement examples fix or name the improvement group in your request. The other groups check preservation. The host receives bounded development examples and scores. The server returns aggregate confirmation metrics and the accepted checkpoint path. Hugging Face evaluation files

Use prepare_evals to prepare trusted labeled data and create_run to score the development baseline. Use list_targets and read_tensor to inspect the checkpoint, try_edit to test a proposal, and confirm_edit and get_run to confirm and verify it. Donors are frozen at run creation. Later proposals identify them by ID and may restore, blend, or scale named weights. The runner enforces the attempt budget, evaluates the selected candidate on confirmation once, and keeps reports for discarded edits. Each tool call waits for its bounded local operation, so client timeouts must accommodate the model and dataset size.

For another MCP client, use the installed auto-weight-mcp command as a local stdio server with the saved runtime configuration.

Prepare labeled evaluation data