SKILL.md
$2a
The setup script auto-detects your git root, applies the preset to .claude/settings.json, and creates the safety hook.
# Find the script (checks installed locations)
_RALPHMODE_SCRIPT=""
for _candidate in \
"$(git rev-parse --show-toplevel 2>/dev/null)/.agent-skills/ralphmode/scripts/apply-ralphmode.sh" \
"$HOME/.agent-skills/ralphmode/scripts/apply-ralphmode.sh" \
"$HOME/.claude/skills/ralphmode/scripts/apply-ralphmode.sh" \
"$(pwd)/.agent-skills/ralphmode/scripts/apply-ralphmode.sh"; do
if [[ -f "$_candidate" ]]; then
_RALPHMODE_SCRIPT="$_candidate"
break
fi
done
if [[ -n "$_RALPHMODE_SCRIPT" ]]; then
bash "$_RALPHMODE_SCRIPT" --mode "$MODE"
else
echo "⚠️ apply-ralphmode.sh not found — applying inline (Claude Code only)"
fi
If the script is not found, fall through to STEP 2 (inline application for Claude Code).
STEP 2: Inline application (Claude Code — fallback when script is missing)
Only run this if the script from STEP 1 was not found.
Repo preset (default — for normal development):
python3 - <<'EOF'
import json, os, subprocess
try:
root = subprocess.check_output(['git', 'rev-parse', '--show-toplevel'],
stderr=subprocess.DEVNULL, text=True).strip()
except Exception:
root = os.getcwd()
target = os.path.join(root, '.claude', 'settings.json')
os.makedirs(os.path.dirname(target), exist_ok=True)
try:
existing = json.loads(open(target).read()) if os.path.exists(target) else {}
except Exception:
existing = {}
existing['_ralphmode_previous_permissions'] = existing.get('permissions')
existing['permissions'] = {
'defaultMode': 'acceptEdits',
'allow': [
'Bash(npm *)', 'Bash(pnpm *)', 'Bash(bun *)', 'Bash(yarn *)',
'Bash(python3 *)', 'Bash(pytest *)',
'Bash(git status)', 'Bash(git diff)', 'Bash(git add *)',
'Bash(git commit *)', 'Bash(git log *)', 'Bash(git push)',
'Read(*)', 'Edit(*)', 'Write(*)'
],
'deny': [
'Read(.env*)', 'Read(./secrets/**)',
'Bash(rm -rf *)', 'Bash(sudo *)',
'Bash(git push --force*)', 'Bash(git reset --hard*)'
]
}
with open(target, 'w') as f:
json.dump(existing, f, ensure_ascii=False, indent=2)
print(f'✓ Repo preset applied to {target}')
EOF
Sandbox preset (only for disposable environments):
python3 - <<'EOF'
import json, os, subprocess
try:
root = subprocess.check_output(['git', 'rev-parse', '--show-toplevel'],
stderr=subprocess.DEVNULL, text=True).strip()
except Exception:
root = os.getcwd()
target = os.path.join(root, '.claude', 'settings.json')
os.makedirs(os.path.dirname(target), exist_ok=True)
try:
existing = json.loads(open(target).read()) if os.path.exists(target) else {}
except Exception:
existing = {}
existing['_ralphmode_previous_permissions'] = existing.get('permissions')
existing['permissions'] = {'defaultMode': 'bypassPermissions'}
with open(target, 'w') as f:
json.dump(existing, f, ensure_ascii=False, indent=2)
print(f'✓ Sandbox preset applied to {target}')
EOF
STEP 3: Ensure safety hook exists
HOOK="$HOME/.claude/hooks/ralph-safety-check.sh"
if [[ ! -f "$HOOK" ]]; then
mkdir -p "$(dirname "$HOOK")"
cat > "$HOOK" << 'HOOKEOF'
#!/usr/bin/env bash
CMD=$(echo "$CLAUDE_TOOL_INPUT" | python3 -c \
"import sys,json; print(json.load(sys.stdin).get('command',''))" 2>/dev/null)
TIER1='(rm[[:space:]]+-rf|git[[:space:]]+reset[[:space:]]+--hard|git[[:space:]]+push.*--force|DROP[[:space:]]+TABLE|[[:space:]]sudo[[:space:]]|chmod[[:space:]]+777|\.env|secrets/)'
if echo "$CMD" | grep -qE "$TIER1"; then
echo "BLOCKED: Tier 1 dangerous command detected." >&2
echo "Command: $CMD" >&2
exit 2
fi
HOOKEOF
chmod +x "$HOOK"
echo "✓ Safety hook created: $HOOK"
else
echo "✓ Safety hook exists: $HOOK"
fi
STEP 4: Report to the user
After applying, tell the user:
- Which preset was applied (
repoorsandbox)
- Which file was written (
.claude/settings.jsonpath)
- What the allow/deny list contains
- How to revert:
rm .claude/settings.json(project-local) or restore~/.claude/settings.json(global)
- Restart Claude Code to activate permission changes
When to use this skill
- You want
ralphto iterate without repeated approval popups.
- You are setting up the same repo for Claude Code, Codex CLI, and Gemini CLI.
- You need a shared safety model: repo-only writes, no secrets reads, no destructive shell by default.
- You want a stronger separation between day-to-day automation and true YOLO mode.
Instructions
Step 1: Define the automation boundary first
Before changing any permission mode:
- Pick one project root and keep automation scoped there.
- List files and commands that must stay blocked:
.env*,secrets/**, production credentials,rm -rf,sudo, uncheckedcurl | sh.
- Decide whether this is a normal repo or a disposable sandbox.
If the answer is "disposable sandbox," you may use the platform's highest-autonomy mode.
If not, use the repo-scoped preset instead.
Step 2: Choose one preset per platform
Use only the section that matches the current tool:
- Claude Code: everyday preset first,
bypassPermissionsonly for isolated sandboxes.
- Codex CLI: use the current official approval and sandbox model first; treat older
permissions.allowandpermissions.denysnippets as compatibility-only.
- Gemini CLI: trust only the project root; there is no true global YOLO mode.
Detailed templates live in references/permission-profiles.md.
Step 3: Apply the profile locally, not globally, unless the workspace is disposable
Prefer project-local configuration over user-global defaults.
- Claude Code: start with project
.claude/settings.json.
- Codex CLI: start with project config and repo instructions or rules files.
- Gemini CLI: trust the current folder, not
~/or broad parent directories.
If you must use a user-global default, pair it with a stricter denylist and a sandbox boundary.
Step 4: Run Ralph with an explicit verification loop
After permissions are configured:
- Confirm the task and acceptance criteria.
- Run
ralphor thejeoplan-execute-verify loop.
- Verify outputs before claiming completion.
- If the automation profile was temporary, revert it after the run.
Recommended execution contract:
boundary check -> permission profile -> ralph run -> verify -> cleanup or revert
Step 5: Keep "skip" and "safe" separate
Treat these as different modes:
- Repo automation: minimal prompts inside a bounded workspace.
- Sandbox YOLO: promptless execution in a disposable environment only.
Do not collapse them into one shared team default.
Step 6: Configure Mid-Execution Approval Checkpoints
Static permission profiles (Steps 2–3) reduce friction before a run starts, but they do not stop dangerous operations that arise during execution. Add dynamic checkpoints so that Tier 1 actions are blocked or flagged at the moment they are attempted.
#### Dangerous operation tiers
Tier
Action
Platform response
Tier 1 (always block)
rm -rf, git reset --hard, git push --force, DROP TABLE, sudo, .env*/secrets/** access, production environment changes
Block immediately, require explicit user approval
Tier 2 (warn)
npm publish, docker push, git push (non-force), DB migrations
Output warning, continue only with confirmation
Tier 3 (allow)
File reads/edits, tests, local builds, lint
Allow automatically
#### Platform checkpoint mechanisms
Platform
Hook
Blocking
Recommended pattern
Claude Code
PreToolUse (Bash)
Yes — exit 2
Shell script pattern-matches command; blocks Tier 1
Gemini CLI
BeforeTool
Yes — non-zero exit
Shell script blocks tool; stderr fed to next turn
Codex CLI
notify (post-turn)
No
approval_policy="unless-allow-listed" + prompt contract
OpenCode
None
No
Prompt contract in opencode.json instructions
Principle: Combine static profiles (Steps 2–3) with dynamic checkpoints (this step).
- Platforms with pre-tool hooks (Claude Code, Gemini): use the hook script.
- Platforms without (Codex, OpenCode): rely on
approval_policyand explicit prompt contracts that instruct the agent to outputCHECKPOINT_NEEDED: <reason>and wait before proceeding with Tier 1 actions.
See references/permission-profiles.md for full hook script templates per platform.
Examples
Example 1: Claude Code sandbox run
Use the Claude sandbox preset from references/permission-profiles.md, then run Ralph only inside that isolated repo:
/ralph "fix all failing tests" --max-iterations=10
Example 2: Codex CLI sandbox Ralph run
For sandbox ralph runs, use the CLI flags directly:
codex -c model_reasoning_effort="high" --dangerously-bypass-approvals-and-sandbox -c model_reasoning_summary="detailed" -c model_supports_reasoning_summaries=true
For repo-scoped (non-sandbox) runs, use the config file approach from references/permission-profiles.md:
approval_policy = "never"
sandbox_mode = "workspace-write"
Place this in ~/.codex/config.toml (or a project-local override) and restart Codex before running Ralph.
Example 3: Gemini CLI sandbox or trust-only setup
For sandbox ralph runs, use --yolo mode:
gemini --yolo
For normal repo automation, trust the current project folder with explicit file selection and run the Ralph workflow for that repo only. See references/permission-profiles.md for details.
Best practices
- Default to the least-permissive preset that still lets Ralph finish end-to-end.
- Keep secret denylists and destructive command denylists even when approvals are reduced.
- Use full bypass only in disposable environments with a clear project boundary.
- Record which preset was applied so teammates can reproduce or revert it.
- Re-check platform docs when upgrading CLI versions because permission models change faster than skill content.