auto-guardian-audit

name	auto-guardian-audit
description	Analyze Auto Guardian bash audit logs and propose safe deterministic allow/block rule improvements. Use when reducing guardian LLM calls or reviewing command approval patterns.

Auto Guardian Audit

Use this skill to inspect guardian-reviewed bash commands and propose safer deterministic rules.

Source log

Read this JSONL file when it exists:

~/.agents/crumbs/auto-guardian/audit.jsonl

Each line is one guardian-reviewed bash command with exact command text, command shape, guardian outcome, user decision, and final outcome.

Workflow

1. Read the audit log.
2. Parse JSONL defensively; skip malformed lines and report count.
3. Group records by:
   • shape.executable
   • shape.normalized
   • shape.features
   • finalOutcome
4. Identify candidates:
   • repeated guardian_allowed commands that are read-only and have no risky shell features
   • repeated user-approved commands that deserve human review
   • repeated denied commands that deserve hard-block tests or rules
5. Output proposals only. Do not edit source or config unless user explicitly asks.

Promotion rules

Recommend deterministic allow rules only when all are true:

• command is read-only, or writes are tightly scoped and intentionally handled by existing path policy
• command has no protected or outside-workspace write risk
• command does not use privilege escalation
• command does not combine network fetch with execution
• command does not depend on ambiguous shell parsing
• similar approved records repeat enough to justify source change
• risky variants can be covered by tests

Do not promote commands with these features unless parser-backed classification proves every segment safe:

• pipeline
• redirection
• command-substitution
• variable-expansion
• globbing
• logical-operator
• subshell-or-group
• heredoc

Output format

Return concise sections:

Audit summary
- records read: N
- malformed lines: N

Safe allow candidates
- command/executable: reason, count, suggested tests

Block candidates
- command/executable: reason, count, suggested tests

Needs human review
- command/executable: concern, count

Recommended next patch
- files/rules/tests to change if user wants implementation