The server defines 2 tools with complete input schemas and descriptions, but both lack documented output schemas. Tool naming follows verb_noun convention (metrics.refresh, agent.nudge) which aids discovery. Descriptions are adequate (115-139 chars, within the 10-1024 baseline) but lack actionable guidance on recovery, dependencies, and expected return structures. Input schemas are well-formed with enums and descriptions for all parameters. However, the absence of output documentation and missing error guidance represent significant gaps for agent planning. Neither tool includes destructive/idempotent hints or confirmation patterns despite agent.nudge being a WRITE operation. The server declares logging=true but no structured logging patterns are visible in the sample code.

• Add explicit output schema documentation. Example: 'metrics.refresh returns {period: string, timestamp: ISO8601, totalEvents: number, agentCount: number, throughput: {events_per_minute: number}, costs: {total_usd: number}, stageDurations: [{stage: string, avg_ms: number}]}'
• For agent.nudge, add a warning block: 'DESTRUCTIVE: This tool sends messages to agents and may interrupt their current tasks. Use dry_run=true to preview without side effects.' Include confirmation guidance.
• Document error cases with recovery paths: 'If period is invalid, call metrics.get_valid_periods(). If no agents match scope, the response returns empty_agents=[] with available_scopes=[]. If event_log is unavailable, check database connectivity before retrying.'
• Clarify parameter relationships in agent.nudge: 'agentIds further filters agents matching scope. If scope=blocked and agentIds=[A,B,C], only blocked agents in [A,B,C] receive nudges. If agentIds is empty, all agents matching scope are nudged.'
• Add idempotency note: 'metrics.refresh is idempotent, calling it twice with the same period returns the same aggregated metrics. agent.nudge with dry_run=true is idempotent; with dry_run=false, repeated calls may send duplicate messages.'
• For agent.nudge, document the NudgeReport structure in the description: 'Returns NudgeReport with fields: nudged_count (int), succeeded (array of agent_ids), failed (array of {agent_id, error_reason}), dry_run_preview (if dry_run=true).'
• Add per-parameter value guidance. For metrics.refresh period: 'hour: aggregates last 60 minutes (default), day: aggregates last 24 hours, week: aggregates last 7 days. Use based on your analysis window.'

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

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