The server defines 4 tools with Korean-language descriptions and basic parameter schemas. However, critical gaps severely limit LLM usability: (1) All descriptions are in Korean, making them inaccessible to English-speaking LLMs and agents; (2) Parameter descriptions are minimal or missing entirely, LLMs cannot infer intent from bare parameter names; (3) Output schemas are completely undocumented, agents cannot plan downstream operations or extract structured data; (4) No error handling guidance, agents have no recovery path on failure; (5) Tool names lack clarity, 'review_episode' vs 'batch_review' distinction is unclear, and 'compile_brief' is vague about what 'brief' means. These issues violate foundational patterns: tool-description, constrained-input, response-shaper, and recovery-guide. The server appears production-grade internally (file paths, environment config, timeout handling), but the tool interface itself is severely under-specified for agentic use.

• Translate all tool and parameter descriptions to English. Provide a parallel English description block (even if descriptions are authored in Korean, expose an 'en' variant to the MCP client). This is non-negotiable for cross-language agent compatibility.
• Document output schemas for all tools. For review_episode and batch_review, explicitly specify the return structure (e.g., { review_text: string, errors_found: number, recommendations: array }). For check_status, define health status (e.g., { services: { nim_proxy: 'healthy'|'unhealthy', ollama: 'healthy'|'unhealthy', ... } }). For compile_brief, specify the markdown structure returned.
• Add min/max constraints and enum validation to episode_number. E.g., 'episode_number (integer, 1 - 999): The episode number to review. Must be a positive integer within the project range.'
• Clarify the boundary between review_episode and batch_review. Rename to 'review_single_episode' and 'review_episodes_batch', or add explicit guidance: 'Use review_episode for ad-hoc single-episode reviews; use batch_review for scheduled periodic reviews of multiple episodes (P7 cadence).'
• Rename 'compile_brief' to 'generate_episode_context_brief' or 'extract_episode_summary'. Make the transformation clear in the name itself.
• Add recovery guidance to tool descriptions. E.g., 'If the review fails due to service timeout, call check_status() to verify external LLM availability, then retry. If episode_number is invalid, list available episodes via [forthcoming] list_episodes tool.'

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

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