Audit Claude Permissions

Scan project-local Claude Code settings files, aggregate permission patterns, and recommend promotions to global configuration.

Workflow Overview

This audit runs in three phases, each as a separate task. Use TaskCreate at the start to create all three tasks, then work through them sequentially with user input via AskUserQuestion.

• Phase 1: Promote to Global — Requires user judgment. Present candidates, get decisions, apply to global config.
• Phase 2: Automated Redundancy Cleanup — Script-driven. After Phase 1 promotions, run cleanup script to remove local permissions now covered by global. Present for sanity-checking, then apply.
• Phase 3: Judgment Calls — Everything requiring user input: security hygiene, policy conflicts, one-off cruft, empty file deletion, moderate-risk items.

Initial Setup

1. Check for symlinked global settings:

readlink -f ~/.claude/settings.json

If the global settings file is a symlink, note the real path. All writes to ~/.claude/settings.json must edit the symlink target, not create a new file that replaces the symlink.

1. Load user preferences from ~/.claude/cc-maintenance.local.md (if it exists). See User Preferences section below. Apply any configured defaults (risk tolerance, auto-cleanup preference).

2. Read global CLAUDE.md from ~/.claude/CLAUDE.md to identify tool preference policies for Phase 3 policy conflict detection.

3. Run the discovery and extraction pipeline:

scripts/discover-settings.sh | xargs scripts/extract-permissions.py

1. Read the actual global settings from ~/.claude/settings.json — compare against the real allow list, not just the static examples in this skill.

2. Create tasks for the three phases:

TaskCreate: "Review and promote permissions to global config"
TaskCreate: "Automated redundancy cleanup"
TaskCreate: "Judgment calls: security, policy, and cruft"

1. Analyze the data and categorize permissions (see Categorization Rules below).

---

Phase 1: Promote to Global

Goal: Identify permissions worth adding to global config and get user approval.

Present Findings

Show a summary table of promotion candidates:

## Promotion Candidates

### Strong Recommendations (safe patterns, multiple projects)

| Permission | Projects | Suggested Global Pattern |
| ---------- | -------- | ------------------------ |
| ...        | ...      | ...                      |

### Moderate Recommendations (review carefully)

| Permission | Projects | Notes |
| ---------- | -------- | ----- |
| ...        | ...      | ...   |

### Cross-Project File Patterns

[If any Read/Write/Edit permissions reference paths outside their project directory
and appear in multiple projects, flag them here. Example: multiple projects have
`Write(~/.config/some-tool/config.json)` - might indicate a shared config worth
adding globally.]

Get User Decision

Use AskUserQuestion to let the user decide:

Question: "Which permissions should I add to global settings?"
Options:
- "Add all strong recommendations"
- "Add strong + moderate recommendations"
- "Let me pick specific ones" (then list individually)
- "Skip - don't add any"

Apply Changes

If user approves additions:

1. Add selected permissions to ~/.claude/settings.json (or symlink target)
2. Respect existing logical groupings (git, nix, gh, etc.)
3. Sort alphabetically within groups
4. Use space-syntax: Bash(cmd *) not Bash(cmd:*)
5. Mark Phase 1 task as completed

---

Phase 2: Automated Redundancy Cleanup

Goal: Remove local permissions now covered by global config. This is mechanical — the script identifies exact matches; user just sanity-checks the list.

Preview Cleanup

Run the cleanup script in dry-run mode:

scripts/discover-settings.sh | scripts/cleanup-redundant.py

Present Findings

Show what would be removed:

## Redundant Permissions

| File      | Permissions to Remove | Remaining |
| --------- | --------------------- | --------- |
| project-a | 5 (ls _, grep _, ...) | 12        |
| project-b | 3 (gh api \*, ...)    | 8         |
| ...       | ...                   | ...       |

**Total:** X permissions across Y files

The script also normalizes any remaining colon-syntax (Bash(cmd:*)) to space-syntax (Bash(cmd *)) when applying.

Get User Decision

If auto_cleanup_redundant: true in user preferences, skip the question and apply directly (still show the summary). Otherwise:

Question: "Should I remove these redundant permissions from local files?"
Options:
- "Yes, clean them up"
- "Show me the full list first"
- "Skip cleanup"

Apply Changes

If user approves:

scripts/discover-settings.sh | scripts/cleanup-redundant.py --apply

Mark Phase 2 task as completed.

---

Phase 3: Judgment Calls

Goal: Everything that requires real user judgment — security risks, policy conflicts, stale cruft, and cleanup opportunities. Present all categories together.

Category A: Security Hygiene

Flag permissions that match these patterns:

High Risk (recommend removal):

• Bash(curl *), Bash(wget *) — network exfiltration risk
• Bash(rm *) — can delete any file
• Bash(source *) — executes arbitrary scripts
• Bash(eval *) — arbitrary code execution

Moderate Risk (review):

• Bash(git reset *), Bash(git checkout *) — can discard work
• Bash(pkill *), Bash(kill *) — process termination
• Bash(python *), Bash(python3 *), Bash(node *) — arbitrary code (flag if user hasn't consciously chosen this)

Adjust what counts as "moderate" vs "high" based on the user's risk_tolerance preference if set.

Category B: Policy Conflicts

Read the user's global ~/.claude/CLAUDE.md for stated tool preferences. Flag local permissions that conflict with those policies.

How to detect: Look for patterns like "prefer X over Y", "use X instead of Y", "avoid Y". Then scan all local permissions for uses of the deprecated tool. For example:

If CLAUDE.md says "prefer tool X over built-in Y", flag all Y permissions across local settings as policy conflicts.

Present these as informational — the user may have valid reasons for specific overrides.

Category C: One-Off Cruft

• Hardcoded file paths (e.g., Bash(prettier --write /full/path/to/file.md))
• Incomplete shell constructs (Bash(done), Bash(for file in *.rs))
• Very specific commands with no wildcards that look like debugging artifacts
• Duplicate entries
• Legacy colon-syntax permissions (Bash(cmd:*)) that weren't caught by the cleanup script (e.g., ones without a global equivalent)

Category D: Cross-Project File Access

• Read, Write, or Edit permissions for paths outside the project
• Flag if the same external path appears in multiple projects (potential global candidate)
• Flag broad patterns like Write(~/.config/*) as security concerns

Category F: MCP Tool Permissions

• Flag MCP tools with write/send/delete capabilities (message sending, data deletion, post creation) — these have side effects beyond the local environment
• Flag server-wide wildcards (mcp__server__*) — convenient but auto-permits any future tools added to that server without review
• Suggest consolidation: if all or most tools from a server are individually listed, suggest replacing with the server wildcard (with a note about the trade-off)

Category E: Empty File Deletion

After cleanup, identify settings files where:

• All permissions have been removed (empty allow list or no allow list)
• No other settings exist (no hooks, enabledPlugins, etc.)

Offer to delete these empty files entirely — they serve no purpose.

Present Findings

## Judgment Calls

### Security Hygiene

#### High Risk — Recommend Removal

| Permission     | Project   | Risk              |
| -------------- | --------- | ----------------- |
| `Bash(curl *)` | project-x | Data exfiltration |

#### Moderate Risk — Review

| Permission | Project | Risk |
| ---------- | ------- | ---- |
| ...        | ...     | ...  |

### Policy Conflicts

| Permission      | Projects | Policy                      |
| --------------- | -------- | --------------------------- |
| `WebFetch(url)` | 3        | CLAUDE.md: prefer Firecrawl |

### One-Off Cruft

| Permission                                         | Project   |
| -------------------------------------------------- | --------- |
| `Bash(prettier --write /path/to/specific/file.md)` | project-y |

### External File Access

| Permission                     | Projects   | Notes                           |
| ------------------------------ | ---------- | ------------------------------- |
| `Write(~/.config/tool/config)` | 3 projects | Shared config — consider global |
| `Edit(/etc/hosts)`             | 1 project  | System file — review necessity  |

### MCP Tool Permissions

| Permission                 | Projects | Notes                                       |
| -------------------------- | -------- | ------------------------------------------- |
| `mcp__slack__post_message` | 2        | Side effect: sends messages externally      |
| `mcp__puppeteer__*`        | 1        | Server wildcard — auto-permits future tools |

### Empty Settings Files

| File                                      | Reason                                     |
| ----------------------------------------- | ------------------------------------------ |
| `~/project-z/.claude/settings.local.json` | All permissions removed, no other settings |

Get User Decision

Use AskUserQuestion:

Question: "How should I handle the judgment call items?"
Options:
- "Remove all flagged items"
- "Remove high risk + cruft only"
- "Let me review each category"
- "Skip — keep everything"

If user wants to review categories, ask about each separately.

Apply Changes

Edit each affected settings.local.json to remove approved items. Delete empty settings files if approved.

Mark Phase 3 task as completed.

---

Session Wrap-Up

After all phases complete, offer to create or update ~/.claude/cc-maintenance.local.md with session learnings. Include:

• Any risk tolerance preferences expressed during the session
• Tool preference notes mentioned during review
• Freeform notes from user decisions (overrides, exceptions, per-project rules — anything the user said that should carry forward to future audits)
• Session history entry (date, summary of changes)

---

User Preferences

Location: ~/.claude/cc-maintenance.local.md

Read at audit start. If missing, proceed with defaults. At session end, offer to create/update.

Format: YAML frontmatter for structured preferences + markdown body for freeform notes.

---
# All preferences are optional
risk_tolerance: moderate # conservative | moderate | aggressive
auto_cleanup_redundant: true
---
## Tool Preferences
- Prefer Exa/Firecrawl over built-in WebSearch/WebFetch

## Risk Notes
- sqlite3 and op item get considered moderate risk
- uv run python treated as arbitrary code execution

## Notes
- Always keep python3 permissions in the data-science project
- Don't consolidate Firecrawl tools into a server wildcard

## Session History
### 2026-02-09
- Promoted 13 permissions to global
- Removed ~220 cruft permissions from Assistant project

Behavior:

• risk_tolerance adjusts what gets flagged as moderate vs high in Phase 3
• auto_cleanup_redundant skips the confirmation prompt in Phase 2 (still shows summary)
• Tool Preferences section supplements CLAUDE.md policy detection
• Notes section is a freeform catch-all for any user preferences that don't fit the structured sections — overrides, exceptions, per-project rules, anything the agent should remember across audits
• Session History provides context for future audits

---

Scripts

All scripts are in scripts/ relative to this skill directory.

• scripts/discover-settings.sh — Finds all .claude/settings.local.json files across ~ using fd with sensible exclusions (Library, node_modules, .git, etc.). Max depth of 5 for performance.
• scripts/extract-permissions.py — Aggregates permissions from multiple settings files. Outputs JSON with each permission, occurrence count, and list of projects using it. Sorted by count descending.
• scripts/cleanup-redundant.py — Removes permissions from local files that are covered by global config. Handles cross-syntax matching (colon-syntax vs space-syntax). Normalizes remaining permissions to space-syntax on write. Dry-run by default; use --apply to modify files.

Usage examples:

# Extract and aggregate all permissions
scripts/discover-settings.sh | xargs scripts/extract-permissions.py

# Preview redundant permission cleanup (dry-run, default)
scripts/discover-settings.sh | scripts/cleanup-redundant.py

# Actually remove redundant permissions
scripts/discover-settings.sh | scripts/cleanup-redundant.py --apply

---

Categorization Rules

Formatting Rules

Use space-syntax for all permissions: Bash(cmd *) not Bash(cmd:*).

Reasonable Global Candidates

Patterns worth promoting. These operate on local project code or perform read-only operations. Compare against the actual global config, not just this static list — these are examples of the kinds of patterns to look for.

Git Commands (read-only): git branch *, git diff *, git log *, git show *, git status *

File Inspection (read-only): cat *, head *, tail *, ls *, find *, grep *, du *

Build and Check Commands: cargo build *, cargo test *, cargo check *, go build *, go test *, npm run build *, npm run test *, deno check *, deno lint *, xcodebuild *

System Utilities: open *, pbcopy, pbpaste, lsof *, ps *

Nix Commands: nix build *, nix-build *, nix develop *, nix eval *, nix flake *, nix path-info *, nix-prefetch-url *, nh darwin build *

Homebrew (read-only): brew info *, brew search *

GitHub CLI: gh api *, gh issue list *, gh issue view *, gh pr list *, gh pr view *, gh pr diff *, gh pr checks *, gh search *, gh run list *, gh run view *

Wildcards: * --help *, * --version

Pattern Generalization

When promoting, generalize cautiously — only for safe patterns:

Formatting Rules for Global Settings

When adding to ~/.claude/settings.json:

• Respect existing logical groupings (git, file inspection, nix, brew, gh, system utilities, build tools, wildcards, Skills, MCP tools)
• Within each group, sort alphabetically
• Place new permissions in the appropriate group based on command prefix
• Use space-syntax: Bash(cmd *) not Bash(cmd:*)