commit-msg

name	commit-msg
description	Generate a commit message for the staged changes using this exact format: subject line with optional issue number, then ### Changes: body with bullet points per file/component. Use when: asked to commit, write a commit message, or stage and commit changes.
license	MIT
metadata	{"author":"shaharb","version":"1.0.0"}

Commit Message Generator

Generate commit messages matching the pybandits project conventions.

Format

Short description

### Changes:

* Add/Update/Fix/Remove <filename_or_module> — one-line summary
  - sub-detail if needed (method name, parameter name, reason)
* Add/Update/Fix/Remove <filename_or_module> — one-line summary
  - sub-detail

Rules

1. Subject line: Short description. Keep under 72 characters.

2. Body starts with ### Changes: — blank line before the header.

3. One bullet per logical unit of change (file, class, or method group). Bullet marker is *. Sub-bullets use - (two spaces + dash).

4. Verb choices — pick the most precise:

   • Add — new file, class, method, parameter, or test
   • Update — modification of existing behaviour
   • Fix — bug fix
   • Remove — deletion
   • Rename — identifier or file rename
   • Refactor — structural change with no behaviour change

5. Name the thing explicitly: file path (pybandits/model.py), class (BaseBayesianNeuralNetwork), method (_calibrate_output_bias()), or parameter (calibrate_output_bias). Don't write vague bullets like "various improvements".

6. Sub-bullets explain the why or the constraint, not just the what. Good: - guards against reinitialisation on subsequent updates via bias_calibrated flag Bad: - adds the flag

7. No period at end of bullets.

8. No Co-authored-by line unless explicitly asked.

Example

Add calibrate_output_bias option

### Changes:

* Add calibrate_output_bias field to BaseBayesianNeuralNetwork — fires once on first _update() to set output-layer bias
  - guarded by bias_calibrated flag that resets on cold_start
* Add _calibrate_output_bias() method implementing one-shot calibration
* Remove logit clipping in forward pass (jnp.clip on linear_transform) — calibrated bias makes it unnecessary
* Rename _min_sigma → _numerical_eps to reflect use beyond VI sigma floors
* Update TestBiasScale → TestBiasStd and all related test parameters to match renamed param

Workflow

1. Run git diff --staged (or read the diff already in context).
2. Group changes by logical unit — don't list every line, list every intent.
3. Draft subject line: verb + noun phrase, issue number if available.
4. Write ### Changes: bullets in order: source changes first, test changes after, config/version last.
5. Review: every bullet names a concrete artifact and a precise action verb.