Best fit
Reviewing hook declarations and source before relying on them.
// agent security
GuardrailDoctor
Review Claude Code hooks, spot configuration gaps, and plan controlled checks with evidence you can inspect.
Format: Agent skill, not a browser app.
Requirements: A SKILL.md-compatible agent and owner-selected Claude Code settings/source. Python 3.10+ for optional fixture tests. Runtime checks require separate approval and an isolated Claude Code session.
You get: Agent-guided findings, cited source, proposed remediation and a record of runtime checks not yet run.
Package 1.1.0. Marketplace checkout may convert currency. Host/model charges are separate.
Is this the right fit?
It returns
Agent-guided findings, cited source, proposed remediation and a record of runtime checks not yet run.
Limits
Not an autonomous penetration-test runner or a safety certification. Fixture tests do not establish live-host enforcement.
Your first use
Extract the complete guardrail-doctor folder, read SKILL.md, and ask your agent to review selected settings and source without executing hooks. The optional command below runs local fixture tests only.
python tests/test_fixture_hook.pyOlder installations may be named guardraildoctor rather than guardrail-doctor. Back up customizations, inspect both paths and avoid two active copies. Do not automatically overwrite or delete customized installations.
A review request you can use
Ask: Review these selected hook settings and source without executing hooks. Cite findings and mark runtime enforcement NOT_RUN. Expected review structure, not a measured run: - Selected inputs and cited findings - Proposed changes for approval - Runtime checks: NOT_RUN until separately authorized
What changed in 1.1.0
The fixture helper now uses event-specific response handling, abstains rather than granting unmatched actions, and does not execute submitted commands or log raw tool inputs. It is a teaching and testing aid, not a production security boundary.
- Reads
- Owner-selected settings and hook source. Runtime testing must use synthetic data, not real credentials.
- Version
- Package 1.1.0; Agensi platform revision 1.2
- Licence
- Review the package licence and seller terms at checkout; this update does not change them.
Questions
What does GuardrailDoctor do?
Review Claude Code hooks, spot configuration gaps, and plan controlled checks with evidence you can inspect.
What does it need to run?
A SKILL.md-compatible agent and owner-selected Claude Code settings/source. Python 3.10+ for optional fixture tests. Runtime checks require separate approval and an isolated Claude Code session.
Does this prove my agent is secure?
Not an autonomous penetration-test runner or a safety certification. Fixture tests do not establish live-host enforcement.
Companion work
OpenClaw Guardrail Linter covers a different review surface. Skill Injection Scanner reviews a skill package before installation.
Check your agent setup
Review tested hosts, versions and workflow limits. Evidence for another skill does not establish this product’s compatibility; keep the requirements and limitations above in view.