Loading ToolBench
Trust infrastructure for autonomous AI agents — identity, observability, governance. AISS v2.0 reference implementation with cryptographic audit trails, policy engine, and anomaly detection.
PiQrypt exposes 26 tools across three modules (TrustGate, Vigil, CLI identity). Tool naming follows verb_noun convention (evaluate, simulate, get_policy, approve_decision, etc.), which is strong. Descriptions are present for all tools and range 50-200 chars, meeting baseline expectations. However, parameter schemas are inconsistently detailed: many tools lack explicit type definitions for nested objects (e.g., 'context' in evaluate/simulate is typed as 'object' with no field schema). Output schemas are not documented. Error handling guidance is absent, tools do not indicate what to do on failure or how to recover. Security-critical tools (delete_agent, approve_decision, reject_decision) lack confirmation patterns or permission-gate documentation. Composition is reasonable, tools are single-purpose, but chaining is not optimized (e.g., list_decisions returns decision_id, but get_decision requires decision_id, forcing two calls for full details).
Approve a pending decision in the REQUIRE_HUMAN queue. Allows the action to proceed.
Create a new human principal with SSO authentication and clearance level.
Delete an agent directory and all associated audit data.
Evaluate an action against the Trust Gate policy engine. Returns ALLOW, BLOCK, or REQUIRE_HUMAN decision with reasoning.
Export PDF audit report for an agent with compliance summary.
Export certified .pqz archive for an agent (signed audit trail).
Export memory .pqz archive for an agent (encrypted state snapshot).
Nested object parameters lack field-level schema definitions. 'context' in evaluate/simulate and 'payload' in record_event are typed as 'object' with no documented fields, forcing LLMs to guess structure.
Output schemas are not documented. Tools return data but LLMs cannot plan downstream calls without knowing response structure (e.g., what fields does get_decision return? Does list_decisions include decision details or just IDs?).
Destructive operations (delete_agent, approve_decision, reject_decision) lack confirmation patterns or permission-gate documentation. No guidance on how to prevent accidental execution.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | <=2025-11-25 | v2 |
Export audit journal in compliance format (ANSSI R29, AI Act Art.14).
Get full Vigil Risk Score (VRS) and history for a specific agent.
Retrieve alert journal with optional filtering by severity, agent, or time range.
Retrieve audit journal entries with optional filtering by date range, agent, or action type.
Retrieve details of a specific decision by ID.
Get Vigil installation summary: all agents, global VRS, alert counts, and trust state.
Retrieve the current Trust Gate policy configuration.
Create a new agent identity with Ed25519 keypair. Optionally encrypt with passphrase.
List all registered agents in the local registry.
Rotate an agent's keypair and create rotation attestation linking old and new identities.
Encrypt or re-encrypt an agent's private key with a new passphrase.
List all decisions in the REQUIRE_HUMAN queue with optional filtering.
List all registered human principals with their clearance levels.
Perform manual migration from v1.6.0 to v1.8.4 (updates agent registry structure).
Inject an event into an agent's audit trail (from bridge). Updates VRS and triggers anomaly detection.
Reject a pending decision in the REQUIRE_HUMAN queue. Blocks the action.
Simulate policy evaluation against a hypothetical action without recording a decision. Used for testing policy rules.
Sign an event with an agent's private key and produce a cryptographic signature.
Unlock an agent session interactively (decrypt key and display session info).
No error handling guidance. Tools do not document what errors are possible, whether they are retryable, or what the LLM should do next (e.g., if approve_decision fails due to invalid decision_id, should the agent retry or call list_decisions first?).
Tool names for export functions are ambiguous. 'export_agent_pqz_cert' vs 'export_agent_pqz_memory' vs 'export_agent_pdf' do not clearly distinguish what each exports or when to use each. Consider 'export_agent_audit_cert', 'export_agent_state_snapshot', 'export_agent_compliance_report'.