Agentic Framework Auditor in Openclaw Skills scans agentic projects for prompt drift, instruction conflicts, injection exposure, and production-readiness gaps.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agentic-framework-auditor
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install agentic-framework-auditor using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Agentic Framework Auditor?
Agentic Framework Auditor is an Openclaw Skills capability for inspecting the behavioral-control surface of agentic frameworks. It analyzes directories that contain prompts, skills, planners, configs, memory-like files, and agent-facing documentation, then turns that evidence into deterministic audit artifacts instead of changing the target project.
This skill is built for report-only review across Hermes, Codex/OpenAI-style skills, OpenClaw/ClawHub skills, LangGraph, CrewAI, AutoGen, and custom agent stacks. It helps teams catch prompt bloat, contradiction chains, unsafe autonomy, over-enforcement, prompt-injection exposure, and review-integrity risks before they reach production.
Agentic Framework Auditor Use Cases
- Auditing an agentic repo before launch or release
- Reviewing Openclaw Skills, prompt systems, memory files, and planner artifacts for instruction conflicts
- Finding prompt injection, concealment, override, destructive, or exfiltration instructions
- Checking for prompt bloat, over-enforcement, layer drift, and weak review integrity
- Producing deterministic evidence for security, QA, or compliance review
- Running fast operator-edited scans on large repositories
- Generating bounded same-agent review packets only after a deterministic gate passes
How Agentic Framework Auditor Works
- Confirm the target roots, profile, and scan mode; default to
--mode standard --root . --profile autofor the current project. - Run
scripts/agentic_audit.pyto inventory files, classify roles, extract active and example-context instructions, and score behavioral risks. - Read the generated Markdown report first; use the JSON, CSV, and instruction graph only when automation or deeper evidence is required.
- Classify findings with
references/audit-taxonomy.mdand compare framework expectations withreferences/framework-profiles.md. - If same-agent review is needed, create a bounded packet only after the deterministic scan and only when the gate allows it.
- Validate any agent-review response with
scripts/validate_agent_review.py; keep deterministic findings immutable and do not certify safety. - When scanner logic changes, run
scripts/run_self_check.pyand, from this repository,python stress-lab/run_stress_lab.py.
Agentic Framework Auditor Setup
- Open the repository that contains the Openclaw Skills auditor. The scanner uses only Python's standard library, so no extra pip install is required for the core workflow.
- Run a baseline scan from the repository root.
python scripts/agentic_audit.py --mode standard --root . --profile auto
- Tune scope explicitly when you need broader or narrower coverage.
python scripts/agentic_audit.py --mode full --root . --exclude node_modules --exclude .git
python scripts/agentic_audit.py --mode shallow --root . --operator-edited-only
python scripts/agentic_audit.py --config audit.json
- If you need same-agent review, provide the reviewer identity and keep home-directory expansion opt-in.
python scripts/agentic_audit.py --mode standard --root . --agent-review --agent-reviewer-id "Codex task <specific-id>"
python scripts/agentic_audit.py --mode standard --root . --profile hermes --include-profile-home
- After changing scanner logic, run regression and stress checks.
python scripts/run_self_check.py
python stress-lab/run_stress_lab.py
Agentic Framework Auditor Data Schema & Taxonomy
Primary artifacts
| Artifact | Purpose | Metadata / contents |
|---|---|---|
<prefix>_report.md |
Human-readable audit report | Audit ID, scope summary, top findings, coverage gaps, and remediation notes |
<prefix>_findings.json |
Structured findings data | Stable finding IDs, fingerprints, severity, failure codes, provenance, and evidence |
<prefix>_inventory.csv |
File inventory and scope record | Path, role classification, inclusion/exclusion status, and sensitivity flags |
<prefix>_instruction_graph.json |
Instruction relationship graph | Nodes and edges for prompts, skills, configs, hooks, and local instruction layers |
<prefix>_fix_plan.md |
Report-only remediation plan | Classified fixes, priority, and review guidance without auto-editing |
<prefix>_agent_review_gate.json |
Review gating status | Gate readiness, blocking conditions, and packet eligibility |
<prefix>_agent_review_packet.md / .json |
Bounded evidence packet for same-agent review | Selected findings, reviewer identity, and allowed evidence only |
Taxonomy and identifiers
- Audit ID: ties every artifact to one scan run.
- Finding fingerprint: stable identifier for correlating repeated evidence.
- Severity: used to rank behavioral risk and review urgency.
- Failure code: taxonomy label from
references/audit-taxonomy.md. - Provenance: source path, scope mode, and evidence context.
- Remediation type: report-only classification for the fix plan.
- Coverage flags: record truncated files, missing roots, excluded sensitive paths, and other evidence gaps.
Configuration inputs
- CLI flags define scan mode, root scope, exclusions, layer filters, and review gates.
- JSON or flat/simple YAML audit configs can be supplied through
--config. - Framework profile hints come from
references/framework-profiles.md, which helps Openclaw Skills adapt to different agent stacks.
Agentic Framework Auditor Advanced Features
- Report-only default prevents accidental changes to audited projects.
- Clause-by-clause prompt-injection checks and semantic policy scoring improve coverage beyond simple keyword matching.
- Same-agent review gating creates bounded packets only after a deterministic scan passes eligibility checks.
- Profile-aware scope control supports Openclaw Skills, Hermes, LangGraph, CrewAI, AutoGen, and custom frameworks.
- Fast modes like
--operator-edited-onlyand layer-specific flags (--only-skills,--only-prompts,--only-config,--only-role) reduce audit time on large repos. - Sensitive-path skipping helps avoid accidental secret exposure, with explicit opt-in required for sensitive files.
- Instruction graph generation, CSV inventory, Markdown reporting, and fix-plan rendering make the output automation-friendly.
- Immutable deterministic findings preserve review integrity even when adding agent-review observations.
scripts/validate_agent_review.pyenforces response-contract validity, allowed assessments, and field hygiene.- Built-in self-check and stress-lab workflows help keep the scanner reliable as Openclaw Skills evolves.
SKILL.md
Loading
Related Openclaw Skills
Imitation Author
A research-first Openclaw Skills workflow that turns public sources into a portable, evidence-grounded style model and runtime skill.
Yanfu
Yanfu, an Openclaw Skills workflow, rewrites an existing public landing page and 1–5 screenshots into a clear, evidence-backed landing page story without changing product truth.
Cron Job Tester
Cron Job Tester is an Openclaw Skills CLI that simulates cron schedules, previews the next 10 runs, and validates shell scripts before deployment.
Workflow DAG Validator
Openclaw Skills Workflow DAG Validator finds DAG cycles, missing dependencies, and parameter mismatches before bad workflows ship.
Doubao Watermark Removal
Openclaw Skills automates batch Doubao watermark removal for AI images using OpenCV inpainting and configurable rectangular masks.
Compliance Policy Validator
An Openclaw Skills compliance scanner for Kubernetes and Docker that flags SOC2, HIPAA, NIST, and custom policy violations with fix hints.
Featured*
