RPR (Reduce Permissions Requests)

React to the last few commands that prompted for permission and generate narrow, safe allowlist rules for them. This writes to the local, git-ignored .claude/settings.local.json file.

Step 1: Scan Recent Transcript

This skill runs in whatever project the user is working in, not in its own repo, so call its scripts by absolute path — a CWD-relative path like python3 skills/4-occasional/rpr/scan.py only resolves when the project happens to be this skill's own repo. Locate the skill dir once, then keep the current working directory on the project (do not cd into the skill dir — scan.py reads the CWD to find that project's transcript and settings):

SK=""
# `find -L` so a SYMLINKED install (e.g. ~/.claude/skills/rpr -> .../skills/4-occasional/rpr) is traversed.
for root in "$HOME/.claude/skills" \
            "$(git rev-parse --show-toplevel 2>/dev/null)/.claude/skills" \
            "$(git rev-parse --show-toplevel 2>/dev/null)/skills"; do
  [ -d "$root" ] || continue
  hit=$(find -L "$root" -path '*rpr/scan.py' 2>/dev/null | head -n1)
  [ -n "$hit" ] && { SK=$(dirname "$hit"); break; }
done
[ -n "$SK" ] || { echo "rpr: skill dir not found — pass it by absolute path" >&2; exit 1; }

Then run scan.py to find the recent un-allowed commands. If the user specified a number (e.g. /rpr 10), pass it as an argument. If they specified session, pass session. Otherwise, default to 5.

python3 "$SK/scan.py" <N>

Step 2: Analyze the Output

The script returns a JSON report containing the commands that were not covered by existing permissions.allow rules. Review the report carefully:

• Compound Commands: If the script flags a command as compound, warn the user that a partial coverage rule won't silence the prompt because Claude checks the entire line.
• Environment Variables: If the script recommends a wrapper script, do not propose a wildcard rule. Instead, offer to generate the recommended bin/qa.sh (or similar) wrapper script that encapsulates the complex pipeline.
• Dangerous Commands: If the script marks a command as dangerous (e.g., rm, sudo, npm i), it will propose an exact match rule instead of a wildcard. Present this trade-off to the user clearly. You may propose skipping it entirely. Never propose a :* wildcard for these commands. Danger is detected by verb, not position, so option-prefixed forms like git -C path push or npm --prefix app install are still caught.
• Option-Prefixed Subcommands: When needs_exact_match is set — a subcommand-style command carries options before its verb (e.g. git --no-pager log) — the script proposes the exact command rather than a wildcard. A :* prefix built from an option (Bash(git --no-pager:*)) could silently cover destructive siblings (git --no-pager push), so present these as exact-match only, like dangerous commands.

Step 3: Present Preview

Before writing anything, present a clear list to the user of what will be added. For each proposed rule:

1. Show the rule (e.g., "Bash(utils/queue-plan.sh:*)")
2. Provide a one-line rationale.
3. State the safety tier (e.g., Safe Read, Exact Match for Destructive, Wrapper Required).

Show a preview diff of what .claude/settings.local.json will look like. Ask the user to accept all, pick a subset, or edit the rules.

Step 4: Write to Settings

Once the user approves the rules, merge them into .claude/settings.local.json (never .claude/settings.json or the global file unless explicitly requested). Do not touch deny or ask lists.

Use the bundled write_rules.py rather than hand-editing JSON — it preserves existing rules, skips duplicates (so re-running is safe), and handles the legacy allowedTools shape. Critically, hand it the approved rules as a file, not a shell argument: write that file with the editor (Write tool), so rule strings containing quotes, parens, or $ (e.g. Bash(node -e "console.log('ok')":*)) never have to survive shell escaping.

1. Write the approved rule strings as a JSON array to a temp file, e.g. rpr-rules.json:

   ["Bash(git status:*)", "Bash(npm run:*)"]
   

2. Merge it into the project's settings (writes to .claude/settings.local.json in the current project by default; pass a second argument to target a different file):

   python3 "$SK/write_rules.py" rpr-rules.json
   

It prints how many rules were newly added versus already present.

Step 5: Confirm and Revert Instructions

After writing, confirm the changes. Provide the user with a one-line undo instruction: "To undo these changes, simply remove the rules from .claude/settings.local.json."