This script automates commit message generation using AI. It analyzes the current git diff, sends it to an Ollama model (tavernari/git-commit-message:reasoning), and formats the output neatly.

git-gen-commit

#!/usr/bin/env bash
set -u
set -o pipefail

# =======================================================
# Minimalist Colors
# =======================================================
BOLD=$'\033[1m'
RESET=$'\033[0m'
FG_RED=$'\033[31m'
FG_GREEN=$'\033[32m'
FG_YELLOW=$'\033[33m'
FG_BLUE=$'\033[34m'
FG_CYAN=$'\033[36m'
FG_GRAY=$'\033[90m'

# =========================
# Minimalist UI Helpers
# =========================
log_ok()   { [ "$ONLY_MESSAGE" == "false" ] && echo -e "${FG_GREEN}[OK]${RESET} $1"; }
log_warn() { [ "$ONLY_MESSAGE" == "false" ] && echo -e "${FG_YELLOW}[WARN]${RESET} $1"; }
log_err()  { echo -e "${FG_RED}[ERROR]${RESET} $1"; }
log_info() { [ "$ONLY_MESSAGE" == "false" ] && echo -e "${FG_BLUE}>${RESET} $1"; }

# =========================
# Flags
# =========================
ONLY_MESSAGE=false
VERBOSE=false
HELP=false
UPDATE=false
# MODEL_VARIANT may be set via environment variable GIT_GEN_COMMIT_MODEL.
# Allowed values: default, mini, pro
MODEL_VARIANT="${GIT_GEN_COMMIT_MODEL:-default}"
DIFF_CONTEXT=5 # Default diff context lines
PREV_COMMITS_WINDOW=20 # Default number of previous commit titles to fetch for context

# Switched to a while loop for better argument parsing (to handle flags with values)
while [[ $# -gt 0 ]]; do
  arg="$1"
  case "$arg" in
    --only-message) ONLY_MESSAGE=true; shift ;;
    --verbose)      VERBOSE=true; shift ;;
    -h|--help)      HELP=true; shift ;;
    --update)       UPDATE=true; shift ;;
    --model)
      if [[ -z "$2" ]]; then
        log_err "The --model flag requires a value: default|mini|pro"
        exit 1
      fi
      case "$2" in
        default|mini|pro) MODEL_VARIANT="$2"; shift 2 ;;
        *) log_err "Unknown model variant: $2. Use default|mini|pro"; exit 1 ;;
      esac
      ;;
    --context) # Flag for diff context
      if [[ -z "$2" || ! "$2" =~ ^[0-9]+$ ]]; then
        log_err "The --context flag requires a numeric value."
        exit 1
      fi
      DIFF_CONTEXT="$2"
      shift 2 # Shift past the flag and its value
      ;;
    --prev-commits) # New flag to customize the number of previous commits for context
      if [[ -z "$2" || ! "$2" =~ ^[0-9]+$ ]]; then
        log_err "The --prev-commits flag requires a numeric value."
        exit 1
      fi
      PREV_COMMITS_WINDOW="$2"
      shift 2 # Shift past the flag and its value
      ;;
    *) # Unknown option
      shift
      ;;
  esac
done

# =========================
# Model Configuration
# =========================
MODEL_SP_CHANGE="tavernari/git-commit-message:sp_change"
MODEL_SP_COMMIT="tavernari/git-commit-message:sp_commit"
case "$MODEL_VARIANT" in
  mini)
    MODEL_SP_CHANGE="tavernari/git-commit-message:sp_change_mini"
    MODEL_SP_COMMIT="tavernari/git-commit-message:sp_commit_mini"
    ;;
  pro)
    MODEL_SP_CHANGE="tavernari/git-commit-message:sp_change_pro"
    MODEL_SP_COMMIT="tavernari/git-commit-message:sp_commit_pro"
    ;;
  default|*)
    ;; # keep defaults (8B)
esac

# =========================
# Help
# =========================
if [ "$HELP" == "true" ]; then
  cat <<'EOF'
Gen Commit (Ollama) - Minimalist

Usage:
  git-gen-commit [flags]

Flow:
  1) Analyzes staged files.
  2) Generates a summary for each file change.
  3) Synthesizes a final commit message.
  4) Presents a compact menu to commit, edit, regenerate, or discard.

Flags:
  --verbose       Show the diff for each file during analysis.
  --only-message  Print only the final commit message and exit. Clean output for scripting.
  --update        Pull the latest Ollama models (standard, mini, pro).
  --model <variant>  Choose model variant: default, mini, pro. Overrides GIT_GEN_COMMIT_MODEL env var.
  --context <n>      Set the number of context lines for the diff (default: 5).
  --prev-commits <n> Set the number of previous commit titles to include as context (default: 20).

Environment:
  GIT_GEN_COMMIT_MODEL  Set default model variant (default|mini|pro). Can be overridden with --model.

EOF
  exit 0
fi

# =========================
# Update models
# =========================
if [ "$UPDATE" == "true" ]; then
  log_info "Pulling latest models (standard, mini, pro)..."
  # Default model (8B)
  ollama pull tavernari/git-commit-message:sp_change >/dev/null || true
  ollama pull tavernari/git-commit-message:sp_commit >/dev/null || true
  # Mini model (4B)
  ollama pull tavernari/git-commit-message:sp_change_mini >/dev/null || true
  ollama pull tavernari/git-commit-message:sp_commit_mini >/dev/null || true
  # Pro model (14B)
  ollama pull tavernari/git-commit-message:sp_change_pro >/dev/null || true
  ollama pull tavernari/git-commit-message:sp_commit_pro >/dev/null || true
  log_ok "All models updated."
  exit 0
fi

# =========================
# Header (only in interactive mode)
# =========================
if [ "$ONLY_MESSAGE" == "false" ]; then
  echo -e "${BOLD}Gen Commit${RESET}"
  echo -e "${FG_GRAY}by tavernari/git-commit-message${RESET}"
  echo -e "${FG_YELLOW}Using ${MODEL_VARIANT} model${RESET}"
  echo
fi

# =========================
# Diff Collection
# =========================
# The -U flag is dynamic using the $DIFF_CONTEXT variable to control the number of context lines in the diff
DIFF="$(git diff --staged -U${DIFF_CONTEXT})"
if [ -z "$DIFF" ]; then
  log_err "No staged changes detected. Run 'git add' first."
  exit 1
fi

# =========================
# Fetch Previous Commits for Context
# =========================
# Detect the main branch (usually 'main' or 'master') by checking the remote origin HEAD
MAIN_BRANCH=$(git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@')
if [ -z "$MAIN_BRANCH" ]; then
  # Fallback to 'main' if no remote HEAD is found
  MAIN_BRANCH="main"
  # Check if 'main' exists, otherwise try 'master'
  if ! git rev-parse --verify "$MAIN_BRANCH" >/dev/null 2>&1; then
    MAIN_BRANCH="master"
  fi
fi

# Get the current branch
CURRENT_BRANCH=$(git rev-parse --abbrev-ref HEAD)

# Fetch the last N commit titles unique to the current branch (not on main branch)
if [ "$CURRENT_BRANCH" != "$MAIN_BRANCH" ]; then
  PREV_COMMITS="$(git log ${MAIN_BRANCH}.. --pretty=format:%s -n ${PREV_COMMITS_WINDOW} 2>/dev/null || true)"
else
  # If on main branch, fetch recent commits
  PREV_COMMITS="$(git log --pretty=format:%s -n ${PREV_COMMITS_WINDOW} 2>/dev/null || true)"
fi

if [ -z "$PREV_COMMITS" ]; then
  PREV_COMMITS="(no previous commits available)"
fi
# Wrap previous commits in a XML-like tag for structured input to the model
PREV_COMMITS_TAG="<previous_commits>\n$PREV_COMMITS\n</previous_commits>"

# =========================
# Utils
# =========================
# Utility to colorize diff output for better readability in verbose mode
colorize_diff() {
  while IFS= read -r line; do
    if [[ $line == "+"* ]]; then printf "  ${FG_GREEN}%s${RESET}\n" "$line";
    elif [[ $line == "-"* ]]; then printf "  ${FG_RED}%s${RESET}\n" "$line";
    else printf "  ${FG_GRAY}%s${RESET}\n" "$line"; fi
  done <<< "$1"
}

# Utility to split the full diff into per-file chunks.
# This processes the diff line-by-line, starting a new chunk on each 'diff --git' line.
split_diff() {
  local diff_content="$1" chunk=""; chunks=()
  while IFS= read -r line; do
    if [[ "$line" =~ ^diff\ --git\  ]]; then
      if [ -n "$chunk" ]; then chunks+=("$chunk"); fi; chunk="$line"$'\n'
    else chunk+="$line"$'\n'; fi
  done <<< "$diff_content"
  if [ -n "$chunk" ]; then chunks+=("$chunk"); fi
}

# Utility to detect and choose a timeout command (timeout or gtimeout).
# This is used to prevent Ollama from hanging indefinitely.
choose_timeout_cmd() {
  if command -v timeout >/dev/null 2>&1; then echo "timeout"; return 0; fi
  if command -v gtimeout >/dev/null 2>&1; then echo "gtimeout"; return 0; fi
  return 1
}

# =========================
# Summarization per File
# =========================
split_diff "$DIFF"
total_chunks=${#chunks[@]}
log_info "Analyzing ${total_chunks} staged files..."
[ "$ONLY_MESSAGE" == "false" ] && echo

ACCUMULATED_CHANGES=""; ALL_CHANGES=""
# Prepend previous commits to the accumulated changes for initial context
ACCUMULATED_CHANGES="$PREV_COMMITS_TAG"
for chunk in "${chunks[@]}"; do
  file_display="$(echo "$chunk" | head -n 1 | sed -e 's#.* b/##')"
  # Include previous commits and accumulated file summaries as context for each file analysis
  INPUT_FOR_CHUNK="<previous_changes>\n$ACCUMULATED_CHANGES\n</previous_changes>\n<diff>\n$chunk\n</diff>"

  change_desc="$(ollama run "$MODEL_SP_CHANGE" "$INPUT_FOR_CHUNK" 2>/dev/null || true)"
  change_desc="$(printf "%s" "$change_desc" | sed 's/^[[:space:]]*//; s/[[:space:]]*$//')"
  [ -z "$change_desc" ] && change_desc="(no description returned)"
  
  if [ "$ONLY_MESSAGE" == "false" ]; then
    echo -e "${FG_GREEN}${BOLD}${file_display}${RESET}"
    [ "$VERBOSE" == "true" ] && colorize_diff "$chunk"
    echo -e "${FG_CYAN}→ ${change_desc}${RESET}"; echo
  fi

  # Wrap the change description with a <diff file="..."> tag so ALL_CHANGES includes filenames
  wrapped_change="<diff file=\"${file_display}\">\n${change_desc}\n</diff>"
  if [ -n "$ACCUMULATED_CHANGES" ]; then
    ACCUMULATED_CHANGES+=$'\n'
  fi
  ACCUMULATED_CHANGES+="$wrapped_change"
  ALL_CHANGES+="$wrapped_change"$'\n'
done

# =========================
# Final Message Generation
# =========================
FINAL_COMMIT_MESSAGE=""
generate_final_commit() {
  log_info "Synthesizing final commit message..."

  # Include previous commits in the final input for overall context in commit message synthesis
  local FINAL_INPUT="<all_changes>\n$PREV_COMMITS_TAG\n$ALL_CHANGES\n</all_changes>"
  local MAX_ATTEMPTS=3 ATTEMPT=1
  while [ $ATTEMPT -le $MAX_ATTEMPTS ]; do
    local OUTPUT="" RC=1
    # Attempt to run Ollama with a timeout to prevent hangs.
    # If timeout command is available, use it; otherwise, use a background process with a kill timer.
    if TMO_CMD="$(choose_timeout_cmd)"; then
      OUTPUT="$($TMO_CMD 60 ollama run "$MODEL_SP_COMMIT" "$FINAL_INPUT" 2>/dev/null)"; RC=$?
    else
      local TMP_OUT="$(mktemp)"; ( ollama run "$MODEL_SP_COMMIT" "$FINAL_INPUT" >"$TMP_OUT" 2>/dev/null ) &
      local OLL_PID=$!; ( sleep 60; kill -0 "$OLL_PID" 2>/dev/null && kill -TERM "$OLL_PID" 2>/dev/null; ) &
      local WATCH_PID=$!; wait "$OLL_PID"; RC=$?
      kill "$WATCH_PID" >/dev/null 2>&1 || true; wait "$WATCH_PID" >/dev/null 2>&1 || true
      OUTPUT="$(cat "$TMP_OUT")"; rm -f "$TMP_OUT"
    fi

    OUTPUT="$(printf "%s" "$OUTPUT" | sed 's/^[[:space:]]*//; s/[[:space:]]*$//')"
    if [ "$RC" -eq 0 ] && [ -n "$OUTPUT" ]; then
      FINAL_COMMIT_MESSAGE="$OUTPUT"; return 0
    else log_warn "Generation failed. Retrying ($ATTEMPT/$MAX_ATTEMPTS)..."; fi
    ATTEMPT=$((ATTEMPT+1))
  done
  log_err "Failed to generate commit message after $MAX_ATTEMPTS attempts."; return 1
}

# =========================
# Initial Generation and Output
# =========================
generate_final_commit || exit 1

if [ "$ONLY_MESSAGE" == "true" ]; then
  printf "%s\n" "$FINAL_COMMIT_MESSAGE"; exit 0
fi

# =========================
# Interactive Menu
# =========================
while true; do
  echo "--- Proposed Commit ---"; printf "%s\n" "$FINAL_COMMIT_MESSAGE"; echo "-----------------------"
  read -r -p "$(echo -e "${BOLD}${FG_YELLOW}Choose: ${FG_GREEN}(c)ommit, ${FG_CYAN}(e)dit, ${FG_YELLOW}(r)egenerate, ${FG_RED}(d)iscard > ${RESET}")" final_choice

  case "$final_choice" in
    c)
      TEMP_FILE="$(mktemp)"; printf "%s\n" "$FINAL_COMMIT_MESSAGE" > "$TEMP_FILE"
      if git commit -F "$TEMP_FILE"; then log_ok "Committed."; else log_err "Commit failed."; fi
      rm -f "$TEMP_FILE"; break
      ;;
    e)
      TEMP_FILE="$(mktemp)"; printf "%s\n" "$FINAL_COMMIT_MESSAGE" > "$TEMP_FILE"
      ${EDITOR:-nano} "$TEMP_FILE"
      if git commit -F "$TEMP_FILE"; then log_ok "Committed."; else log_err "Commit failed."; fi
      rm -f "$TEMP_FILE"; break
      ;;
    r)
      echo; generate_final_commit || break; continue
      ;;
    d|q|*)
      log_warn "Discarded."; break
      ;;
  esac
done

Hey mate, thanks for sharing this! I came across on your solution when was looking for a suitable LLM for git commit messages.
Could you please translate your valuable comments in the code to English? :)

Question: Are these really trained models with some dataset, or do these models just have a system prompt?

@andkirby,

Question: Are these really trained models with some dataset, or do these models just have a system prompt?

Yes, it was trained using my dataset and also all weights are open on my HF page https://huggingface.co/Tavernari

Hey mate, thanks for sharing this! I came across on your solution when was looking for a suitable LLM for git commit messages.
Could you please translate your valuable comments in the code to English? :)

I think the latest version is already in English. (It is not true.. sorry..)

@andkirby updated it