Summary

The audit-harness framework addresses critical challenges in AI agent development: ensuring adherence to audit rules and mitigating context loss. It transforms audit requirements from mere suggestions into enforced format constraints using a novel three-layer defense system. By leveraging audit records as persistent external memory, the framework enables robust context recovery, preventing repeated mistakes and facilitating continuous improvement through automated daily reports and self-correction loops. This is particularly beneficial for developers building auditable and reliable AI agents.

How It Works

The core of audit-harness is its three-layer defense architecture. Layer 1 employs hardcoded skills (e.g., /start, /end) for 100% reliable audit logic. Layer 2 enforces an [AUDIT] output format constraint, aiming for ~90% reliability by making audit a mandatory format rather than a behavioral suggestion. Layer 3 acts as a fallback session wrapper, with the /end skill verifying audit completeness and alerting on omissions. Context recovery is managed by loading historical audit records on session start, proactive calls to /recover during context compression, and cross-session state restoration, prioritizing user corrections, task state, conclusions, and environment configuration. Three automatic hooks (PostToolUse, Stop, UserPromptSubmit) capture tool operations, archive audit data, and inject session reminders, ensuring continuous audit data generation.

Quick Start & Requirements

Installation is streamlined via a single command: bash /path/to/audit-harness/install.sh. The script supports a "smart mode" for automatic global installation and project initialization, or specific modes like --global, --init, and --auto. After installation, users initiate tasks with /start "your task description" and conclude with /end. The framework installs core components, hooks, and configurations globally (~/.claude/) and per-project ($PROJECT/.claude/). No specific non-default prerequisites like GPUs or CUDA versions are mentioned.

Highlighted Details

• Three-Layer Defense: Integrates hardcoded skills, [AUDIT] format constraints, and session wrappers for comprehensive audit enforcement.
• Context Recovery: Utilizes audit records as persistent memory to restore user corrections, task state, conclusions, and environment configuration across sessions or during context compression.
• Automatic Hooks: PostToolUse, Stop, and UserPromptSubmit hooks automate the capture, archiving, and persistence of audit information.
• Customizable Configuration: audit_config.py allows project-specific adaptation of alert rules, core scripts, and prompt templates without altering the core engine.
• Record Types: Supports Data, Change, and Conversation records at varying granularities for detailed auditing.

Maintenance & Community

The provided README does not contain specific information regarding notable contributors, sponsorships, community channels (e.g., Discord/Slack), or roadmaps.

Licensing & Compatibility

The project is licensed under the MIT license, which generally permits broad usage, including commercial applications and integration into closed-source projects.

Limitations & Caveats

The [AUDIT] output format constraint (Layer 2) is noted as being approximately 90% reliable, suggesting potential edge cases where strict adherence might not be guaranteed. The README does not detail unsupported platforms or known bugs.