Config Stack Audit

A comprehensive, prompt-driven audit of the entire Claude Code configuration at ~/.claude/ and all project-level .claude/ directories. Produces a scored report with delta tracking across runs and a self-improving heuristics loop.

1. Modes

Parse the arguments passed to /config-audit:

First-run handling: Before starting, check if ~/.claude/config-audit-reports/ contains any previous reports. If not, this is a baseline audit — skip delta comparison in Section 5 and note "Baseline audit — no prior data for comparison" in the Delta section of the report.

2. Inventory Phase

Runs for ALL modes. Build the complete config inventory before any analysis.

Execute these tool calls (parallelize where independent):

Batch 1 — Global config (all parallel):

1. Glob("~/.claude/rules/*.md") — list all global rules
2. Glob("~/.claude/agents/*.md") — list all global agents
3. Glob("~/.claude/skills/*/") — list all global skill directories
4. Glob("~/.claude/commands/*.md") — list all global commands
5. Read("~/.claude/settings.json") — extract hooks, permissions, MCP config
6. Read("~/.claude/mcp.json") — extract MCP server registrations

Batch 2 — Project-level config (all parallel): 7. Glob("~/projects/**/.claude/settings.json") — all project settings 8. Glob("~/projects/**/.claude/rules/*.md") — all project rules 9. Glob("~/projects/**/.claude/agents/*.md") — all project agents 10. Glob("~/projects/**/.claude/skills/*/") — all project skills 11. Glob("~/projects/**/.claude/commands/*.md") — all project commands 12. Glob("~/projects/**/CLAUDE.md") — all CLAUDE.md files

Adjust the ~/projects/** glob root to wherever you keep your repositories if it differs.

Store the file lists and counts. These are the inputs for all dimension analyses.

Token discipline:

• Use Glob/Grep for discovery — never find or ls via Bash
• Only Read file contents when comparison requires it (e.g., duplication checks)
• Read first 100 lines max per file during comparison
• Use parallel Agent calls (model: haiku) for independent dimensions to keep main context clean
• Budget: --quick ~20-30K tokens, --full ~40-60K tokens

2a. Project Scope Policy

Not every repo earns full config scrutiny. Scoping prevents coverage dimensions (GAP/SEC) from inflating with every throwaway experiment — a recurring false signal (see heuristic H-003). Apply these tiers:

Exemption mechanism: A project is out of scope for the dimensions listed in its marker if EITHER:

• it contains an .audit-ignore file at its root (format: exempt_dimensions: GAP, SEC — comma-separated dimension codes, or ALL), OR
• it lives under a scratch/experiments root (e.g. ~/projects/experiments/).

During the inventory phase, record exempt projects separately and list them under "Suppressed Findings" in the report (e.g., "3 projects exempt by .audit-ignore"). Do not generate GAP/SEC findings for exempt projects. New projects start compliant if your project scaffolding sets up the deny-ruled settings.json by default.

3. The 7 Audit Dimensions

Each dimension produces: (a) findings with severity and deterministic IDs, (b) a score 1-10.

Dimension Dependency Graph

Inventory Phase (must complete first)
     │
     ├──► Dim 1: Duplication  ─┐
     ├──► Dim 2: Security     ─┤  Batch A — launch as parallel Agents (haiku)
     ├──► Dim 3: Staleness    ─┤
     └──► Dim 4: Consistency  ─┘
                                │
                                ▼  (wait for Batch A results)
     ├──► Dim 5: Coverage Gaps ─┐
     ├──► Dim 6: Optimization  ─┤  Batch B — launch as parallel Agents (haiku)
     └──► Dim 7: Innovation    ─┘

• --quick: Run dims 1-3 from Batch A only. Skip dim 4 and all of Batch B.
• --dimension N: Run only that dimension (still requires inventory).

Dimension 1: Duplication

What to check: Identical or diverged copies of rules, agents, skills, and commands across the global ↔ project layers.

How to check:

1. For each file type (rules, agents, commands), extract filenames from the global inventory and each project inventory.
2. Find same-named files across layers.
3. For each match, Read first 100 lines of both files.
4. Classify as:
   • IDENTICAL — content matches → recommend consolidating to global
   • DIVERGED — same name, different content → recommend syncing or documenting the fork
   • INTENTIONAL FORK — file contains # audit:ignore or # audit:ignore:DUP → suppress

Scoring rubric:

Finding ID format: DUP-{project-slug}-{dir}-{filename} Example: DUP-devenv-rules-security-md

Dimension 2: Security

What to check: Hook scripts with risky patterns, missing deny rules, exposed paths, overly broad permissions.

How to check:

1. Read each settings.json found in inventory. Check permissions.allow — flag Bash(*) in project configs as overly broad.
2. Grep all hook scripts (.js, .sh files referenced in settings.json hooks) for execSync, exec(, eval(, spawn with string concatenation or template literals (injection vectors).
3. For each project with a dependency file (package.json, requirements.txt, pyproject.toml), check if its .claude/settings.json has deny rules for .env*, credentials*, secrets*.
4. Read ~/.claude/mcp.json — check that no tokens, passwords, or API keys appear in plain text (look for patterns: sk-, ghp_, Bearer , base64 strings >40 chars).
5. If a session-observer hook exists (e.g. a continuous-learning observe.sh), check it sanitizes untrusted input before persisting it — grep for sanitiz/truncat/stdin handling.

Scoring rubric:

Finding ID format: SEC-{scope}-{issue-type} Examples: SEC-global-settings-broad-allow, SEC-myproject-no-env-deny

Dimension 3: Staleness

What to check: Broken paths, outdated references, unfilled templates, stale CLAUDE.md dates.

How to check:

1. Grep all files under ~/.claude/ and all CLAUDE.md files for stale machine paths — references to home directories, usernames, or drive paths from a prior machine or OS that no longer exist on the current host. Common shapes:

   • Old Linux/WSL home dirs (/home/<old-user>/…, /mnt/c/Users/<old-user>/…)
   • Old Windows paths (C:\Users\<old-user>\…)
   • Any absolute path under a home dir that fails a test -e check on the current machine

   Build the list of "known old" prefixes from whatever prior setups this config has migrated through, and treat any surviving reference as stale. Remote-scope exception: paths that intentionally document a remote host (SSH targets, VPS/CI runners, docs inside a remote-* skill) are not stale — exclude files whose purpose is remote-host documentation.

2. For each CLAUDE.md, grep for "Last Updated" — parse the date and flag if >30 days old. If the project is a git repo, check git log -1 --format=%ci and flag if >20 commits since the Last Updated date.
3. Grep CLAUDE.md and command files for placeholder text: [PROJECT NAME], [CLIENT NAME], [CLIENT], TODO:, FIXME:.
4. For each settings.json with hooks, extract all script paths from hook command fields. Verify each path exists using test -f.

Scoring rubric:

Finding ID format: STL-{file-slug}-{issue-type} Examples: STL-myproject-claude-md-stale-path, STL-devenv-settings-broken-hook

Dimension 4: Consistency

What to check: Name collisions, contradicting rules, settings mismatches.

How to check:

1. Group all command files by filename across global + all project layers. For groups with >1 file, Read the description frontmatter (first 5 lines) of each. Flag if same name serves different purposes.
2. Group all rules files by filename across layers. Compare first 5 lines (title/heading). Flag same-name-different-purpose.
3. In global settings.json, check hook matchers for overlapping patterns that could cause double-firing on the same event (e.g., two PreToolUse hooks both matching tool == "Bash").
4. Read agent description fields. For agents that say "MUST BE USED for {language} projects", check if there are projects using that language (grep for corresponding dependency files) that don't reference or invoke that agent.

Scoring rubric:

Finding ID format: CON-{item-type}-{name}-{issue} Examples: CON-cmd-review-name-collision, CON-rule-patterns-different-purpose

Dimension 5: Coverage Gaps

What to check: Missing config, unmatched agents/commands, workflow gaps.

How to check:

1. List all project directories under ~/projects/ (excluding archived ones). Apply the Scope Policy (Section 2a): skip any project that contains an .audit-ignore file or lives under a scratch/experiments root. For each remaining (in-scope) project, check existence of: CLAUDE.md, .primeignore, .claude/settings.json. Flag any active project missing these.
2. Compare the global agent roster (filenames without .md) against the global command roster. Flag agents with no corresponding command that users could invoke (e.g., a database-reviewer agent but no /db-review command).
3. Check if these common workflows have commands: deploy, test, lint, format, migrate, backup. Flag gaps.
4. Meta-check: verify the audit skill itself is present in your config.

Scoring rubric:

Finding ID format: GAP-{project-or-item}-{what-is-missing} Examples: GAP-myproject-no-primeignore, GAP-database-reviewer-no-command

Dimension 6: Optimization

What to check: Oversized files, underused hooks, context budget waste.

How to check:

1. For all files in the inventory (rules, skills, commands, agents), count lines. Flag any file >800 lines.
2. In global settings.json, count hooks with matcher "*" per event type. Flag if >3 wildcard hooks on any single event.
3. Count total lines across all ~/.claude/rules/*.md files — this estimates the token cost of rules loaded every session. Flag if >2000 total lines.
4. Read description fields from all skill SKILL.md files. Identify pairs with >80% keyword overlap (potential consolidation candidates). Use simple word overlap comparison, not semantic analysis.
5. For projects with >50 tracked git files, check if .primeignore exists. Flag missing.

Scoring rubric:

Finding ID format: OPT-{target}-{issue-type} Examples: OPT-observe-sh-wildcard-hook, OPT-rules-total-lines-2500

Dimension 7: Innovation

What to check: Opportunities for new skills, commands, or agents based on observed patterns.

How to check:

1. Check if ~/.claude/homunculus/instincts/ exists and has files (this is the layout used by the continuous-learning-v2 skill). If so, read instincts with confidence ≥ 0.7. Cross-reference against ~/.claude/homunculus/evolved/ to find high-confidence instincts not yet evolved into skills.
2. Check if a legacy learned-patterns directory exists (e.g. ~/.claude/homunculus/legacy-learned-md/) with .md files. These are legacy patterns not yet promoted to full skills. List them.
3. Compare the agent roster against the command roster — identify agent capabilities that have no user-facing command. These are potential new commands.
4. For each active project (non-archived) that is a git repo, run git -C {path} log -20 --oneline 2>/dev/null to scan for recurring workflow patterns (e.g., repeated manual steps that could be automated).
5. Synthesize inputs 1-4 into 3-5 concrete suggestions. Each suggestion must have: name, purpose, rationale grounded in the inputs, estimated complexity (Low/Medium/High).

Handle gracefully if the instincts or learned directories are empty or don't exist — still produce suggestions from inputs 3-4.

Scoring rubric:

Finding ID format: INN-{suggested-item-name} Examples: INN-db-review-command, INN-deploy-workflow-skill

4. audit:ignore Support

Before reporting any finding, check if the target file contains suppression markers:

• # audit:ignore or <!-- audit:ignore --> — suppress ALL findings for this file
• # audit:ignore:DUP or <!-- audit:ignore:DUP --> — suppress only Duplication findings
• # audit:ignore:SEC — suppress only Security findings
• Pattern: audit:ignore:{DIM} where DIM is one of: DUP, SEC, STL, CON, GAP, OPT, INN

Use Grep to check for audit:ignore in a file before generating findings for it.

At the end of the report, include a "Suppressed Findings" section listing the count and files where suppression was applied.

5. Delta Tracking

After generating all findings for this audit:

1. List files in ~/.claude/config-audit-reports/ matching config-audit-*.md.
2. Sort by filename (lexicographic = chronological due to date format).
3. Read the most recent previous report (not the current one being generated).
4. Extract all finding IDs from the previous report — grep for patterns matching (DUP|SEC|STL|CON|GAP|OPT|INN)-[a-zA-Z0-9_-]+.
5. Classify each current finding:
   • NEW — ID not present in previous report
   • PERSISTENT — ID present in both current and previous
   • REGRESSED — ID was absent in previous report but present in the one before that (came back after being resolved)
6. Classify previous-only findings as RESOLVED (present in previous, absent in current).

First-run handling: If no previous report exists, output:

## Delta from Previous Audit
Baseline audit — no prior data for comparison. All findings are new.

6. Heuristic Promotion

After delta tracking:

1. Read ~/.claude/config-hygiene-heuristics.md.
2. List the 3 most recent report files (including the current one being generated).
3. For each finding ID that appears as PERSISTENT in all 3 reports, check if it already has a heuristic entry.
   • If not: create a new Active heuristic entry (H-{NNN}) with the finding pattern, IDs, and dates.
   • If yes: update Seen count and Last seen date.
4. For Active heuristics whose finding IDs do NOT appear in the last 2 consecutive reports, move to Retired with today's date.
5. Cap the Active section at 20 entries. If full, retire the oldest by Last seen date.
6. Write the updated heuristics file.

7. Report Template

Write the report to ~/.claude/config-audit-reports/config-audit-YYYY-MM-DD-HH.md (use current date and hour in 24h format).

# Config Stack Audit — {YYYY-MM-DD-HH}

**Health Score: {N}/70** (prev: {P}/70, delta: {+/-D})

## Executive Summary
- {Most important finding 1}
- {Most important finding 2}
- {Most important finding 3}
- {Trend observation if delta data available}

## Inventory
| Layer | Count |
|-------|-------|
| Global rules | {N} |
| Global agents | {N} |
| Global skills | {N} |
| Global commands | {N} |
| Global hooks (across all events) | {N} |
| MCP servers | {N} |
| Projects with .claude/ config | {N} |
| CLAUDE.md files | {N} |

## Dimension Scores
| # | Dimension | Score | Prev | Delta | Top Finding |
|---|-----------|-------|------|-------|-------------|
| 1 | Duplication | {S}/10 | {P}/10 | {+/-D} | {finding ID} |
| 2 | Security | {S}/10 | {P}/10 | {+/-D} | {finding ID} |
| 3 | Staleness | {S}/10 | {P}/10 | {+/-D} | {finding ID} |
| 4 | Consistency | {S}/10 | {P}/10 | {+/-D} | {finding ID} |
| 5 | Coverage Gaps | {S}/10 | {P}/10 | {+/-D} | {finding ID} |
| 6 | Optimization | {S}/10 | {P}/10 | {+/-D} | {finding ID} |
| 7 | Innovation | {S}/10 | {P}/10 | {+/-D} | {finding ID} |

## Findings by Dimension

### 1. Duplication
| ID | Severity | Finding | Files | Action |
|----|----------|---------|-------|--------|
| {DUP-xxx} | {CRITICAL/HIGH/MEDIUM/LOW} | {description} | {file paths} | {recommended action} |

### 2. Security
| ID | Severity | Finding | Files | Action |
|----|----------|---------|-------|--------|

### 3. Staleness
| ID | Severity | Finding | Files | Action |
|----|----------|---------|-------|--------|

### 4. Consistency
| ID | Severity | Finding | Files | Action |
|----|----------|---------|-------|--------|

### 5. Coverage Gaps
| ID | Severity | Finding | Files | Action |
|----|----------|---------|-------|--------|

### 6. Optimization
| ID | Severity | Finding | Files | Action |
|----|----------|---------|-------|--------|

### 7. Innovation
**Suggested New Skills/Commands:**
| ID | Name | Purpose | Rationale | Complexity |
|----|------|---------|-----------|------------|
| {INN-xxx} | {name} | {one-line purpose} | {grounded in inventory/instinct data} | {Low/Med/High} |

## Delta from Previous Audit
### New Findings ({count})
{Bulleted list of new finding IDs with one-line descriptions}

### Resolved ({count})
{Bulleted list of finding IDs from previous report that are now gone}

### Regressed ({count})
{Bulleted list of finding IDs that were previously resolved but returned}

### Persistent ({count})
{Bulleted list of finding IDs present in both reports — heuristic promotion candidates}

## Suppressed Findings
{N} findings suppressed by `audit:ignore` markers.
{If N > 0, list: file path → suppressed dimension(s)}

## Heuristic Promotions
{List any findings promoted to ~/.claude/config-hygiene-heuristics.md this cycle}
{List any heuristics retired this cycle}
{If none: "No promotions or retirements this cycle."}

## Recommended Actions (Priority Order)
1. **[{severity}]** {specific action} — fixes {finding ID(s)}
2. **[{severity}]** {specific action} — fixes {finding ID(s)}
3. ...

8. --fix Mode (Auto-Remediation)

When --fix is passed, after generating and presenting the report summary, iterate through findings and offer auto-remediation for automatable issues.

Automatable Fix Categories

Stale paths (STL findings) — search-and-replace using a project-local patterns registry. Maintain a small table mapping each known-old path prefix to its current equivalent, and apply most-specific first so a broad prefix doesn't partially rewrite a more specific one. Fill the registry in from the prior machines/usernames this config has migrated through; the shape is:

Apply in order (most specific first) to avoid partial replacements. Remote-scope exception: do NOT rewrite home-dir paths inside files that intentionally document remote Linux hosts (SSH/VPS/CI docs, remote-* skills) — those are intentional. When new stale patterns are discovered during an audit, suggest adding them to this registry.

Identical copies (DUP findings classified as IDENTICAL) — offer to delete the project-level copy since global is authoritative.

Missing directories (GAP findings for missing .primeignore or config-audit-reports) — offer to create them.

Broken hook paths (STL findings where hook scripts don't exist at the referenced path) — offer to update the path in the relevant settings.json if the script exists at a correctable location.

Fix Flow

For each automatable finding:

1. Present the proposed change with before/after preview
2. Ask for user confirmation: "Apply this fix? (yes/no/skip all)"
3. On approval: apply the change using Edit tool
4. Log the fix in the report under ## Auto-Remediation Log

Non-automatable findings are listed as "Manual action required" with pointers:

• Security gaps → "Run a deeper security review"
• Consistency issues → "Review and resolve manually"
• Coverage gaps → "Run /optimize-context in {project}" or "Create CLAUDE.md for {project}"
• Innovation suggestions → "Run /brainstorm to explore further" or "Run /evolve to promote instincts"

Auto-Remediation Log (appended to report)

## Auto-Remediation Log
| # | Finding ID | Action Taken | File Modified |
|---|-----------|-------------|---------------|
| 1 | {STL-xxx} | Replaced stale path | {file path} |
| 2 | {DUP-xxx} | Deleted identical copy | {file path} |