Three tools with complete input schemas and descriptions, but descriptions are terse (10-60 chars) and lack actionable context. Tool names follow verb_noun pattern (screen_, check_) but descriptions don't explain WHEN to use each tool or what distinguishes them. Parameters are typed and described, but lack constraints (enums, ranges). No output schemas documented. Error handling is implicit (PolicyViolationError, ReplayDetectedError raised but not described in tool docs). Security-critical tools (check_payment_policy, check_payment_replay) record state but descriptions don't warn agents of side effects clearly enough.

• Expand tool descriptions to 100-200 chars. Example: 'screen_payment_metadata: Scans payment metadata (URL, description, reason) for PII using regex or NLP. Returns detected entity types and risk level. Call this BEFORE check_payment_policy to redact sensitive data. No side effects.'
• Document output schemas for each tool. E.g., screen_payment_metadata returns {detected_entities: [string], risk_level: 'low'|'medium'|'high', redacted_metadata: object}.
• Add explicit warnings to check_payment_policy and check_payment_replay descriptions: 'WARNING: This tool records state (audit log, replay fingerprint) on every call. Not idempotent, retries may create duplicate records.'
• Constrain numeric and enum parameters. Add to amount_usd: 'Must be >= 0 and <= 1000000 USD.' Add to entities: 'Optional; valid values: PERSON, EMAIL, PHONE, CREDIT_CARD, etc. (see Presidio entity types).'
• Document error cases in descriptions. E.g., check_payment_policy: 'Returns PolicyViolationError if amount exceeds daily_limit_usd or max_per_call_usd. Agent should ask user for approval or reduce amount.'

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry