Auto Commit Messages

R5 — Auto-commit each accepted edit with a descriptive message

Requirement: Auto-commit each accepted edit with a descriptive, attributed commit message.

Sourced from: Aider (git integration).

Status in atomic-forge: Met (templated, not LLM-generated) — verified against code 2026-08-29. sandbox.py::commit() auto-commits every accepted repair round with a descriptive, structured message (e.g. "forge: repair {file} (round N, green/best-effort, diff N)", or "forge: revert round N (failures A -> B)" on auto-revert). Templated messages carry more precise, verifiable detail (exact round/diff/failure counts) than an LLM-authored summary would reliably include — deprioritizing the LLM-generation upgrade below; the requirement's actual intent (auto-commit + descriptive) is already satisfied.

State of the art

Real but tangential literature — commit-message generation is a well-studied NLG task, just not central to forge's repair-loop value proposition:

• Automated Commit Message Generation with Large Language Models: An Empirical Study and Beyond (arXiv:2404.14824) — LLM-authored messages win human preference in ~78% of evaluated samples over prior state-of-the-art, despite mixed results on BLEU/ROUGE-L (i.e. n-gram metrics undersell how good these actually are to a human reviewer).
• An Empirical Study on Commit Message Generation Using LLMs via In-Context Learning (arXiv:2502.18904) — in-context learning (no fine-tuning) already outperforms prior specialized commit-message models, meaning this requirement is cheap to satisfy well with the same model forge already calls, no dedicated training or tooling needed.
• Brevity is the Soul of Wit: Condensing Code Changes to Improve Commit Message Generation (arXiv:2509.15567) — feeding a condensed diff (not the raw diff) improves message quality; relevant if forge generates the commit message from the same diff object patch.py already produces.

Implication for atomic-forge

Low effort, low risk: this can likely be satisfied by prompting the same model already in the loop with the accepted diff, no architecture change required. Worth confirming current behavior before treating it as a gap.

What needs to be done (to beat the competition)

1. Confirm current state first. Check whether forge already auto-commits accepted patches; if git integration exists but lacks generated messages, this is a small addition, not new plumbing.
2. Condense the diff before prompting for a message, per arXiv:2509.15567 — feed a summarized change description (files touched, symbols changed, verdict) rather than the raw unified diff, which that paper shows improves message quality.
3. Use in-context examples, no fine-tuning, per arXiv:2502.18904 — a handful of good example (diff-summary → message) pairs in the prompt is sufficient; don't over-invest here relative to the repair-loop work.
4. Keep the Co-authored-by / attribution trailer consistent with forge's existing commit conventions so generated commits are indistinguishable in provenance tracking from any other forge-made commit.

Implementation plan

Phase 1 — confirm current behavior (~0.5 day)

• Grep the codebase for existing git-commit logic (likely near checkpoint.py or the CLI entrypoint) and confirm whether commits happen at all today, and if so, whether messages are already descriptive.

Phase 2 — diff condenser (~1 day, skip if Phase 1 shows nothing to build on)

• Add a small summarizer that turns an accepted patch's unified diff into a condensed change description (files touched, symbols added/changed/removed, verdict), per arXiv:2509.15567 — this is a pure-text transform, no model call needed for the condensing step itself.

Phase 3 — message generation (~1 day)

• Prompt the already-in-loop model with the condensed diff plus 2–3 in-context example (summary → message) pairs, per arXiv:2502.18904.
• Auto-commit with the generated message plus forge's standard attribution trailer.

Phase 4 — validation (~0.5 day)

• Spot-check generated messages on 10–15 benchmarks/ runs for accuracy (does the message match what actually changed) before enabling by default.