Hooks Permission Request Hook

/hooks:permission-request-hook

Generate a PermissionRequest hook that auto-approves safe operations, auto-denies dangerous ones, and passes everything else through for user decision. A safer, project-aware alternative to --dangerouslySkipPermissions.

When to Use This Skill

Auto Mode vs PermissionRequest Hook

Auto mode (Claude Code 2.1.83+) routes most approve/deny decisions through a classifier model. It overlaps with — but does not replace — a PermissionRequest hook. Choose this skill when you need any of:

• Deterministic, auditable rules — the classifier's probabilistic answer is unsuitable for compliance contexts
• Project-specific deny lists that go beyond the default trust set
• Hard 0% false-positive guarantees for specific commands (the hook is exact, the classifier is not)
• Pre-approve narrow patterns so they skip the classifier round-trip and avoid latency/token cost

Hooks coexist with auto mode — they fire alongside the classifier. See .claude/rules/auto-mode.md for the full auto-mode model and how it interacts with allow rules and subagents.

Context

Detect project stack:

• Lockfiles: !find . -maxdepth 1 \( -name 'package-lock.json' -o -name 'yarn.lock' -o -name 'pnpm-lock.yaml' -o -name 'bun.lockb' -o -name 'poetry.lock' -o -name 'uv.lock' -o -name 'Cargo.lock' -o -name 'go.sum' -o -name 'Gemfile.lock' \)
• Project files: !find . -maxdepth 1 \( -name 'package.json' -o -name 'pyproject.toml' -o -name 'requirements.txt' -o -name 'Cargo.toml' -o -name 'go.mod' -o -name 'Gemfile' \)
• Existing settings: !find . -path '*/.claude/settings.json' -maxdepth 3 -type f
• Existing hooks dir: !find . -maxdepth 2 -type d -name 'scripts'
• jq available: !jq --version
• Existing PermissionRequest hooks: !find . -path '*/.claude/settings.json' -maxdepth 3 -exec jq -r '.hooks.PermissionRequest // empty' {} +

Parameters

Parse these from $ARGUMENTS:

Execution

Execute this workflow:

Step 1: Detect project stack

Identify languages and tooling from the context above.

Language detection:

Report detected stack to user before generating.

Step 2: Generate the hook script

Create the script at scripts/permission-request.sh (or .claude/hooks/permission-request.sh if no scripts/ directory exists).

Use the Script Template from REFERENCE.md. Adapt it by:

1. Including only sections for detected languages (remove {{ if ... }} markers)
2. Including only selected categories if --category flags were provided
3. Removing all template comments ({{ ... }})
4. If --strict is set, include the strict mode catch-all deny at the end

Step 3: Generate the test script

Create scripts/test-permission-hook.sh (or .claude/hooks/test-permission-hook.sh to match the hook location).

Use the Test Script Specification from REFERENCE.md. Include test cases only for detected stacks and selected categories. Remove all {{ ... }} template markers.

Step 4: Configure .claude/settings.json

Read existing .claude/settings.json if it exists. Merge the PermissionRequest hook — preserve all existing configuration.

If a PermissionRequest hook already exists, ask the user whether to:

• Replace the existing PermissionRequest hook
• Add alongside the existing hook (both will run)
• Abort and keep existing configuration

Configuration to merge:

{
  "hooks": {
    "PermissionRequest": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "bash \"$CLAUDE_PROJECT_DIR/scripts/permission-request.sh\"",
            "timeout": 10
          }
        ]
      }
    ]
  }
}

Use timeout: 10 (10 seconds). Use empty matcher "" to match all tools. Adjust path if script is in .claude/hooks/ instead of scripts/.

Step 5: Finalize

1. Make both scripts executable: chmod +x <hook-path> <test-path>
2. Create .claude/ directory if needed for settings.json
3. Run the test script to verify all test cases pass
4. Report summary:
   • List files created/modified
   • Show number of approve rules, deny rules
   • Show test results (pass/fail count)

Post-Actions

After generating the hook:

1. Suggest committing the new files:

   scripts/permission-request.sh
   scripts/test-permission-hook.sh
   .claude/settings.json
   

2. If --strict was NOT used, mention the flag for environments where unknown commands should be denied
3. Explain how to add custom rules — edit the APPROVE/DENY sections of the generated script
4. Remind about CLAUDE_HOOKS_DISABLE_PERMISSION_REQUEST=1 to toggle the hook off temporarily
5. Note that empty matcher "" catches all tools; suggest narrowing to "Bash" if only Bash commands need filtering

PermissionRequest Schema

Input (via stdin)

Output (via stdout)

Agentic Optimizations

Quick Reference