[JenD86]

Summary

A modulo command (installed by pip install -e ., which setup.sh now does) so agents and people can find, add, check and run steering vectors. It never prompts, every subcommand takes --json, and the exit code is 1 when nothing was done.

• search: one search over your registries, the community index and modulo_catalog.yaml. Each hit has a status (local / verified / community / catalog) and a next step.
• list: shows what SteeringAgent would load for a model, and why.
• add: adds vectors from hf:owner/repo[@rev], a huggingface.co URL or a local directory.
  • It pins the commit, inspects each file (format and layer from shape, metadata or filename) and checks it against the model's config.
  • Vectors are registered unverified, with a ±8 range.
  • In a bundle, it only uses paths that name the model (other models' vectors can share its hidden size) and keeps one layer per concept.
  • --as-index-entry and --as-manifest print drafts instead of registering.
• extract, verify (strength sweep, --save-range) and run (the demo, which gains --model and --steer NAME=STRENGTH).
• Bundle manifests (modulo-bundle.yaml): a repo describes its vectors (names, layers, descriptions, sign conventions, test prompts), validated against the index's field rules.
  • Without a manifest, sidecar metadata is read (rotalabs-style metadata.json / layer_N.json, including best_layer).
  • The index keeps holding expanded entries, not pointers to manifests, so all text that reaches a model's prompt shows up in the PR diff.
• Thinking mode off: Qwen3-style chat templates think by default, which spent every generation's budget inside <think> (agent turns, verify samples, extraction). All prompt rendering now passes enable_thinking=False. Qwen2.5 renders byte-identical.
• Smaller changes:
  • The loader accepts dict .pt files ({"vector": ...}).
  • registry.yaml keeps its header comments on save.
  • MODULO_MODEL sets the default model.
  • Index and manifest validation reject unfilled TODO placeholders.

Testing

• CPU: 126 passed, 1 skipped. New tests/test_cli.py covers search, add against a fake hub, manifests, sidecars, local directories and naming.
• H100: all 107 tests passed at the time, including the GPU steering tests, after the loader and thinking-mode changes.
• End to end: modulo verify assistant_axis reproduces the known behaviour (+12 coherent, +15 muddled, −15 florid). add → verify → run --steer works on Qwen3-8B with rotalabs vectors.
• Real repos tried: rotalabs, Butanium assistant-axis, Rachata.

Known limitations

• verify's suggested range only detects repetition, so it suggests ±15 for assistant_axis, where the real positive cap is +12. Read the samples before trusting a range; the docs say so.
• The rotalabs refusal vector is weak in Modulo at any allowed strength. Their own setting is about 70% of the residual norm, with the prompt steered as well. Not addressed here.
• Range calibration is being rethought separately (setting a position within the model's natural range instead of pushing by a fixed amount). Not part of this PR.

🤖 Generated with Claude Code