README.md

Sugarfree in AI harnesses & build workflows

The sugarfree CLI is a plain stdin → stdout filter (see ../cli/README.md): deterministic, no network, exits 0 on success. That makes it easy to drop into agent hooks, editor commands, and CI/build steps.

The CLI contract in one line:

cat input | sugarfree [flags] > output      # or: sugarfree FILE, or: sugarfree --clipboard

Exit codes: 0 success, 3 when run with --check and the content would change, non-zero otherwise (bad flags, unreadable input).

Claude Code hooks

See claude-code-settings.example.json for copy-paste snippets. Two patterns:

• Stop hook — after each turn, run sugarfree --clipboard --all so whatever is on the clipboard pastes clean. Pair it with strip-clipboard.sh.
• UserPromptSubmit hook — pipe the submitted prompt through sugarfree --all to strip bold/italic/header noise before Claude reads it.

Point the command at an absolute path to the built binary (e.g. /usr/local/bin/sugarfree) or to strip-clipboard.sh.

Generic agent / harness hook

Any harness that can shell out can call the filter. Strip emphasis from a model's markdown output before handing it to a plain-text sink:

agent --print | sugarfree --all

Or normalize tables into structured lists for downstream parsing:

agent --print | sugarfree --none --tables --table-format yaml

Build / CI workflows

Use --check to gate a build on whether tracked content still carries sugar (exit 3 = would change):

# Fail the job if any committed Markdown still has emphasis markers.
for f in $(git ls-files '*.md'); do
  sugarfree --check "$f" && continue
  echo "::error file=$f::contains formatting sugar — run 'sugarfree --all' on it"
  fail=1
done
[ -z "${fail:-}" ]

Or normalize in place as a pre-commit step:

tmp=$(mktemp); sugarfree --all "$file" > "$tmp" && mv "$tmp" "$file"

The CLI never touches the network and is fully deterministic, so it's safe to run in sandboxed CI and agent environments.