respect-guards

Respect Guards

Manage guard scripts that prevent repeating past mistakes. Guards are generated by /oops and live in two locations:

• Global guards: ~/.claude/respect/guards/ — fire in every project
• Local guards: ~/.claude/projects/[ENCODED_CWD]/guards/ — fire only in the current project (ENCODED_CWD = CWD with / and . replaced by -)

Commands

List guards

1. Scan both directories for *.json trigger metadata files:
   • ~/.claude/respect/guards/*.json (global)
   • Local guards: first compute the encoded CWD by running:

     ENCODED_CWD=$(pwd | sed 's/[\/.]/-/g')
     ls ~/.claude/projects/${ENCODED_CWD}/guards/*.json 2>/dev/null
     

2. For each .json file, extract: triggers, lesson, created date
3. Check if corresponding .sh file exists and is executable
4. Display grouped by scope. Truncate lesson text to 50 chars max:

Global Guards
=============
| Guard | Triggers | Lesson | Status |
|-------|----------|--------|--------|
| no-timeout | timeout | Use gtimeout on macOS | active |
| no-friday-deploy | git push, deploy | Don't deploy on Fridays | active |

Local Guards (this project)
===========================
| Guard | Triggers | Lesson | Status |
|-------|----------|--------|--------|
| config-in-config-dir | config.json, config.yaml | Config files go in ./config/ | active |

If no guards exist in either location, display: "No guards installed. Guards are created automatically when you use /oops."

Fix a guard

When the user says "fix the [name] guard" or reports a guard misfiring:

1. Search for the guard in both global (~/.claude/respect/guards/) and local (~/.claude/projects/$(pwd | sed 's/[\/.]/-/g')/guards/) directories
2. Read both [name].sh and [name].json
3. Understand what the guard is trying to prevent (from the lesson and trigger metadata)
4. Ask the user what went wrong (false positive? missed a case? wrong trigger?)
5. Rewrite the guard script and/or trigger file to fix the issue
6. Make the script executable: chmod +x [name].sh
7. Confirm: "Updated guard [name]. It will now [description of new behavior]."

Delete a guard

When the user asks to remove/delete a guard:

1. Search for the guard in both global and local directories
2. Confirm: "Delete guard [name] ([scope])? This will stop checking for [lesson]. (y/n)"
3. On confirmation, remove both [name].sh and [name].json from the appropriate directory
4. Confirm: "Guard [name] deleted."

Test a guard

When the user asks to test a guard:

1. Search for the guard in both global and local directories
2. Read the guard script and trigger file
3. Run the guard with a sample tool input that should trigger it: bash [guards_dir]/[name].sh "Bash" "[sample input matching triggers]"
4. Show the result: exit code and output
5. Run with a sample that should NOT trigger it
6. Report: "Guard [name] fires correctly on [trigger] and passes on [non-trigger]."

Important

• Global guard files are in ~/.claude/respect/guards/
• Local guard files are in ~/.claude/projects/[ENCODED_CWD]/guards/
• Each guard has a .sh (executable script) and .json (trigger metadata) file
• The .sh script receives TOOL_NAME as $1, flattened TOOL_INPUT as $2 (for pattern matching), and raw JSON TOOL_INPUT as $3 (for field extraction)
• Exit 1 with a message = warn the user. Exit 0 = allow the action.
• The .json tools field (optional) controls which tools the guard applies to:
  • Omitted = defaults to Bash/Write/Edit/NotebookEdit
  • Regex patterns like mcp__.*slack.* target specific MCP tools
  • Can mix: ["Bash", "mcp__.*datadog.*"]
• Never modify guard scripts without reading them first