Single tool 'scan' has a clear verb-noun name and acceptable description (89 chars), but critical gaps undermine quality. Input schema is present with 5 parameters, all typed, but parameter descriptions are extremely terse (5-20 chars each), far below the 72-char baseline. The 'only' parameter description ('comma-separated recognizer labels to run (e.g. EMAIL,IPV4)') violates the no-examples rule and lacks format clarity. No output schema is documented, LLMs cannot infer what fields the response contains or how to chain downstream calls. Error handling is absent from the visible code. The tool is READ_ONLY (good security posture) but lacks any guidance on failure modes, timeouts, or recovery paths. Schema completeness is ~40% (types present, descriptions severely lacking).

• Expand parameter descriptions to 50-100 chars each. For 'min_score': 'Confidence threshold (0.0 - 1.0); only report findings with confidence >= this value. Default 0.5 filters low-confidence false positives.' For 'use_context': 'Enable context-word confidence boosting: when true, nearby words (e.g. 'SSN:' before a number) increase match confidence. Default true.'
• Document output schema explicitly. E.g. 'Returns {findings: [{id, type, confidence, location, context}], summary: {total, by_type}, scan_duration_ms}'. Include field types and what each means.
• Replace 'only' example with enum or pattern. Either enumerate valid recognizer labels (EMAIL, IPV4, SSN, CREDIT_CARD, PASSPORT, DRIVER_LICENSE) or document the format (uppercase alphanumeric, comma-separated).
• Add error handling section to description: 'Fails if path does not exist (suggest checking path), if file is unreadable (permission error), or if scan exceeds 30s timeout (suggest reducing scope or increasing timeout). Retryable on timeout.'
• Add result-limiting guidance: 'Returns up to 1000 findings per scan. For large directories, use 'recursive: false' to scan single directory, or filter by 'min_score' to reduce noise. Findings are sorted by confidence descending.'