Loading ToolBench
Enterprise Security, Caching, and Trajectory Validation Gateway for AI Agents. Provides AST command security scanning, suffix-delta prompt caching alignment, and trajectory drift measurement.
AgentSentry exposes 3 tools with basic schemas but critical gaps in descriptions, parameter documentation, and error handling. Tool names lack clear action verbs (agentsentry_* prefix is organizational, not functional). Descriptions are present but vague, 'suffix-delta prompt reordering' and 'Levenshtein trajectory similarity' are jargon-heavy and don't explain WHEN or WHY an agent should call these tools. Parameter descriptions are minimal or missing (tool_schemas array has no item schema; baseline_trace/candidate_trace lack context on expected format). No output schemas documented. Error handling is absent, tools return success/failure but no recovery guidance. Schemas are visible in mcp_server.py but incomplete per JSON Schema best practices.
Applies suffix-delta prompt reordering to maximize provider prefix cache hits and reduce token invocation cost by ~50.56%.
Computes Levenshtein trajectory similarity distance between baseline and candidate agent tool call traces.
Recursively parses shell command AST to intercept RCE breakouts, pipe injection, and path traversals with 0.00% false positive rate.
Tool names lack action verbs. 'agentsentry_scan_command', 'agentsentry_align_prompt', 'agentsentry_measure_drift' are prefixed with organizational namespace but lack clear verb_noun structure (e.g., 'validate_command', 'optimize_prompt', 'compute_drift'). LLMs cannot infer intent from organizational prefixes.
Descriptions are jargon-heavy and lack WHEN/WHY context. 'Applies suffix-delta prompt reordering to maximize provider prefix cache hits' assumes LLM knows what suffix-delta is. No explanation of when to call this vs. other prompt optimization tools, or what problem it solves for the agent.
Parameter descriptions are missing or incomplete. 'tool_schemas' is an array with no item schema or description of expected structure. 'baseline_trace' and 'candidate_trace' lack format guidance (are these tool names? IDs? full call objects?). LLMs cannot construct valid inputs without this detail.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | <=2025-11-25 | v2 |
No output schemas documented. Tools return JSON objects (allowed/decision/reason/confidence) but LLMs have no formal schema to parse responses or plan downstream calls. Baseline rubric requires 100% of A+ tools to have documented return types.
Error handling provides no recovery guidance. scan_command returns 'BLOCKED' with a reason, but no suggestion for what the agent should do next (retry with modified command? escalate? abort?). No error categorization (retryable vs. fatal).