/hone

Audit Checklist

Purpose: Category-based audit checklist with PASS/WARN/FAIL criteria for Codex CLI, Antigravity CLI, and Claude Code configuration. Read when: Running the AUDIT phase on ~/.codex/ or ~/.gemini/ configuration files.

Checklist Overview

Judgment Criteria

PASS

• Configuration meets the check requirement completely.
• No action needed.

WARN

• Configuration partially meets the requirement or has minor issues.
• Improvement suggested but not urgent.
• Applies to: P2-P3 items with minor deviations.

FAIL

• Configuration does not meet the requirement.
• Action recommended (P0-P1) or improvement needed (P2-P3).
• Applies to: Missing items, security issues, broken paths.

Priority Definitions

Safety Classification

Category-Specific Audit Procedures

Model Audit (M1-M3)

1. Read model value from config.toml
2. WebSearch for latest Codex CLI supported models
3. Compare and classify (PASS if current, WARN if one version behind, FAIL if deprecated)
4. Check reasoning_effort against user's typical task complexity
5. Verify verbose is not permanently enabled

Trust Audit (T1-T5)

1. Extract all paths from [project_trust]
2. Check each path exists on disk (T1)
3. Flag sensitive-looking paths (auth, finance, secrets) that are trusted (T2)
4. Count total entries and flag if excessive (T5)
5. Check for wildcard patterns and assess specificity (T4)

Feature Audit (F1-F3)

1. Extract all feature flags from config
2. WebSearch for current available flags in latest version
3. Identify missing new flags (F1), deprecated flags (F2), stable but disabled flags (F3)

MCP Audit (C1-C4)

1. Extract MCP server configs
2. Check each binary path exists and is executable (C1)
3. Check for unused servers (C2)
4. Scan for plaintext secrets in args/env (C3)
5. Check server versions against latest (C4)

Rules Audit (R1-R3)

1. Read all files in ~/.codex/rules/
2. Compare rules for duplicates across files (R1)
3. Validate glob patterns (R2)
4. Check for references to removed or changed tools (R3)

AGENTS.md Audit (A1-A3)

1. Read ~/.codex/AGENTS.md
2. Check for clarity and contradiction (A1)
3. Verify priority hierarchy exists (A2)
4. Check for redundant directives (A3)

Instructions Audit (I1-I2)

1. Check ~/.codex/instructions.md exists and is non-empty (I1)
2. Review content for relevance and actionability (I2)

Antigravity CLI Checklist

Gemini Category-Specific Audit Procedures

Antigravity Model Audit (GM1-GM3)

1. Read selectedModel value from settings.json
2. WebSearch for latest Antigravity CLI supported models
3. Compare and classify (PASS if current, WARN if one version behind, FAIL if deprecated)
4. Check model compatibility with the user's API tier (free vs paid)
5. Verify model supports required capabilities (code generation, multimodal input)

Antigravity Safety Audit (GS1-GS2)

1. Read safetySettings array from settings.json
2. Check for BLOCK_NONE on all categories without justification (GS1: FAIL)
3. Check for BLOCK_HIGH_AND_ABOVE on all categories blocking legitimate dev use (GS2: WARN)
4. Verify all harm categories are configured (missing categories inherit defaults)
5. Assess consistency of threshold levels across categories

Antigravity Extensions Audit (GE1-GE4)

1. Extract extension configs from settings.json
2. Check each binary path exists and is executable (GE1)
3. Check for unused extensions (GE2)
4. Scan for plaintext secrets in args/env (GE3)
5. Check extension versions against latest (GE4)

Antigravity Instructions Audit (GI1-GI2)

1. Check ~/.gemini/GEMINI.md exists and is non-empty (GI1)
2. Check project-level GEMINI.md if applicable
3. Review content for relevance and actionability (GI2)
4. Check for contradictions between global and project-level instructions

Antigravity Auth Audit (GA1-GA2)

1. Check that selectedAuthType is configured or GEMINI_API_KEY env var is set (GA1)
2. Scan project files and shell configs for hardcoded GEMINI_API_KEY values (GA2)
3. Never read actual auth tokens or OAuth session files

Claude Code Checklist

Claude Code Category-Specific Audit Procedures

Claude Code Model Audit (CCM1-CCM2)

1. Check for CLAUDE_MODEL in settings.json env block or environment
2. WebSearch for latest Claude Code supported models
3. Compare and classify (PASS if current, WARN if one version behind, FAIL if deprecated)
4. Assess model appropriateness for typical task complexity

Claude Code Permissions Audit (CCP1-CCP4)

1. Read permissions.allow and permissions.deny from both global and project settings
2. Check for overly broad patterns like bare Bash in global allow (CCP1: FAIL)
3. Check for missing deny rules for destructive commands (CCP2: WARN)
4. Compare allow and deny for redundant overlaps (CCP3: WARN)
5. Check for references to tools/servers that no longer exist (CCP4: WARN)

Claude Code MCP Audit (CCS1-CCS5)

1. Extract mcpServers from both global and project settings
2. Check each server command exists and is executable (CCS1)
3. Check for unused/orphaned servers (CCS2)
4. Scan for plaintext secrets — recommend ${env:VAR} syntax (CCS3)
5. Check server package versions (CCS4)
6. Assess global vs project placement appropriateness (CCS5)

Claude Code Instructions Audit (CCI1-CCI4)

1. Check ~/.claude/CLAUDE.md exists and is non-empty (CCI1)
2. Check project-level CLAUDE.md and .claude/CLAUDE.md if applicable
3. Review content for relevance and actionability (CCI2)
4. Scan for API keys, tokens, or credentials in instruction files (CCI3)
5. Check for contradictions between global and project instructions (CCI4)

Claude Code Commands Audit (CCK1-CCK2)

1. List files in ~/.claude/commands/ and <project>/.claude/commands/
2. Verify each is valid Markdown (CCK1)
3. Scan for hardcoded secrets in command templates (CCK2)

Claude Code Hooks Audit (CCH1-CCH2)

1. Read hooks configuration from project settings.json
2. Check that hook commands are executable and paths exist (CCH1)
3. Scan for plaintext secrets in hook commands and environment (CCH2)
4. For detailed design or debugging, switch to Hone's explicit hook* recipes

Claude Code Auth Audit (CCA1-CCA2)

1. Check that ANTHROPIC_API_KEY env var is set or OAuth is configured (CCA1)
2. Scan settings.json and CLAUDE.md files for hardcoded API keys (CCA2)
3. Never read ~/.claude/credentials.json or auth session files

Claude Code Env Tuning Audit (CCE1-CCE5)

1. Check env block in settings.json for performance-critical variables
2. If MCP servers are configured, verify ENABLE_TOOL_SEARCH is auto:0 for context savings (CCE2)
3. Check CLAUDE_AUTOCOMPACT_PCT_OVERRIDE — default 95% causes late compaction; 70-80% recommended (CCE1)
4. If CLAUDE_CODE_MAX_OUTPUT_TOKENS is unset, note default 32K may truncate large outputs (CCE3)
5. If alwaysThinkingEnabled: true, verify MAX_THINKING_TOKENS is calibrated for cost/quality balance (CCE4)
6. Review telemetry/privacy env vars for unnecessary data sharing (CCE5)

Claude Code Display & Config Hygiene Audit (CCD1-CCD5)

1. Check if outputStyle is set for consistent output behavior (CCD1)
2. Check attribution uses new format (not deprecated includeCoAuthoredBy) (CCD2)
3. Verify cleanupPeriodDays is explicitly set (CCD3)
4. Verify autoUpdatesChannel is set, prefer stable for production use (CCD4)
5. Verify respectGitignore is true (CCD5)

Claude Code Sandbox Audit (CCB1-CCB2)

1. Review whether sandbox settings exist and are appropriate for security posture (CCB1)
2. If sandbox enabled, check sandbox.filesystem.denyRead blocks sensitive paths (CCB2)
3. Check sandbox.network.allowedDomains is scoped appropriately
4. Note: Sandbox adoption should be gradual — recommend ask-first for initial enablement

Capability Detail (SKILL.md excerpt)

• core_audit: Audit ~/.codex/, ~/.gemini/, ~/.claude/ config files against best practices; fetch official docs (WebSearch/WebFetch), perform PASS/WARN/FAIL gap analysis, and emit Before/After diff proposals with P0-P3 priority + safe/ask-first/risky safety labels
• codex_audit: Codex CLI config.toml audit — trust levels, feature flags, AGENTS.md/rules/instructions.md, MCP server config, and the wire_api = "chat" hard-error check (deprecated since Feb 2026)
• antigravity_audit: Antigravity CLI (agy) audit — settings.json model/auth/theme, GEMINI.md (progressive disclosure via @file.md imports, line-count thresholds), safety threshold appropriateness, extension/skill/plugin config (accessibility, secrets, version currency)
• claude_code_config_audit: Claude Code (~/.claude/) audit — settings.json permissions (allow/deny patterns, wildcard detection), instructions (CLAUDE.md quality, line-count density, advisory-vs-hook triage), custom slash commands, model/auth settings, settings hierarchy conflict detection (user/project/local/managed override resolution + managed-settings.d/ drop-in merge order)
• claude_code_hooks_audit: Structural + security audit of Claude Code hooks — exit code correctness (0/2), permissionDecision: "deny" usage for security-critical gates, non-interactive (-p) mode coverage gaps, HTTP hook URL pattern validation, tighten-only semantic verification (allow cannot bypass deny), 4-handler-type review (command/http/prompt/agent). Design/debugging uses Hone's explicit hook* recipes
• mcp_security_audit: MCP server security posture — per-server PAT least-privilege, tool poisoning risk on metadata integrity, OAuth 2.1 transport compliance (PKCE/client-credentials), token passthrough detection, version pinning, DCR endpoint validation, OAuth discovery URL validation (CVE-2025-6514 mitigation), RFC 8707 resource indicator binding (MCP spec 2026-03-15)
• plugin_audit: Plugin source verification (official vs third-party marketplace), auto-update configuration (third-party auto-update = supply-chain risk), marketplace trust review, plugin permission scope
• rules_and_budget_audit: .claude/rules/ path-scoped rule validation (glob frontmatter syntax + specificity); CLAUDE.md/GEMINI.md instruction-budget waste detection (lint/formatter rule duplication = wasted context)
• prompt_cache_hierarchy_audit: Verify session context layout (tools → system → messages) keeps T-static above T-dynamic per _common/PROMPT_CACHE_HIERARCHY.md; flag cache breakpoints on timestamps / per-request data; verify _common/ load order stability; detect inlined excerpts that should be pointers

Output Routing Table (SKILL.md excerpt)