Loading ToolBench
FragGate FG-0.1: a local kernel against tool fragmentation and model hallucination.
FragGate exposes 5 tools with minimal descriptions and incomplete schemas. Tool names lack action verbs (runtime.ping, registry.list, registry.verify, claim.check, runtime.receipt). Descriptions are extremely terse (10-50 chars), providing no context for when/why to use each tool. Input schemas exist but lack parameter descriptions entirely, the 'claims' array in claim.check has no itemSchema, and 'assertion' in runtime.receipt has no guidance on format. No output schemas documented. No error handling guidance. Tools appear to be infrastructure/kernel operations rather than user-facing agent tools, which explains the sparse documentation, but even kernel tools need clarity for LLM integration.
Accept or refuse a claim bundle.
List bound tools in the hashed registry.
Ask whether a tool name is real (bound).
Liveness, version, ledger tip, registry digest.
Hash an operator assertion.
Tool names lack action verbs. 'runtime.ping', 'registry.list', 'claim.check' use namespace.operation style instead of verb_noun (e.g., 'check_claims', 'verify_registry_entry'). LLMs rely on verb-first naming to infer intent.
Descriptions are critically short (10-50 chars). 'Liveness, version, ledger tip, registry digest' for runtime.ping is a list of outputs, not a description of what the tool does or when to use it. Minimum 50 chars recommended; these fail to explain purpose.
No parameter descriptions. 'claims' array in claim.check has no itemSchema or description of what constitutes a valid claim. 'assertion' in runtime.receipt has no format guidance (string, JSON, hex?). LLMs cannot infer parameter semantics from names alone.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | 2026-07-28+ | v2 |
No output schemas documented. Tools return data but LLMs have no schema to understand response structure. This breaks downstream tool chaining and forces LLMs to guess field names.
No error handling guidance. Tools provide no recovery hints (e.g., 'If claim verification fails, try registry.verify() first'). Agents have no path forward on failure.