This server has 13 tools with complete input schemas and descriptions visible in src/server.py. However, there are significant quality gaps: (1) Output schemas are entirely undocumented, no tool describes its return structure, which forces LLMs to infer results blindly. (2) Parameter descriptions are minimal and often lack constraints (e.g., 'job_id' is just 'The job ID returned from a submit_* function', no format guidance). (3) Error handling is absent, tools catch exceptions but return generic error dicts with no recovery guidance. (4) Many parameters lack validation hints (e.g., 'rank_threshold' has no range or unit; 'peptide_lengths' format is underdocumented). (5) The server conflates synchronous and asynchronous operations under the same tool prefix pattern, risking LLM confusion. Tool naming is verb-forward and clear (good), but descriptions are too brief (average ~80 chars, well below the 194-char baseline for A+ tools). Overall: the tools ARE defined and schemas ARE present, but the quality is below production baseline, descriptions lack depth, no output contracts, minimal error guidance.

• Add output schema documentation to every tool. Example: 'Returns: {status: string, job_id: string, submitted_at: ISO8601, status: pending|running|completed|failed}'.
• Document HLA allele format and valid values. Either provide an enum (alleles_available_by_class tool) or format hint: 'HLA allele in format HLA-[A|B|C][0-9]{2}:[0-9]{2}'.
• Expand parameter descriptions with constraints. Replace 'Number of lines from end (default: 50, use 0 for all)' with 'Integer in range 0-10000; 0 for all lines, default 50 for recent logs'.
• Add error recovery guidance in descriptions. E.g., predict_peptide_binding: 'If input_file not found, ensure path is absolute or relative to current working directory. Use list_jobs to check available results from prior runs.'
• Implement input validation and return actionable errors. Instead of catching exceptions, validate: file exists, allele matches pattern, rank_threshold in range [0, 100], and return clear messages like 'Invalid allele: HLA-Z99:99. Did you mean HLA-A02:01 or HLA-B07:02?'
• Distinguish sync vs async tools explicitly in names or tool groups. Rename: predict_* → get_*_prediction (sync, ~1s), submit_* → start_* (async, returns job_id). Add 'Synchronous, results available immediately' vs 'Asynchronous, returns job_id, poll with get_job_status' in descriptions.
• Validate file paths to prevent path traversal. Restrict to an allowed directory (e.g., /app/jobs/) and reject paths containing '..' or absolute paths outside the sandbox.
• Document the job lifecycle in list_jobs and get_job_status. Include: 'Job states: pending (queued), running (in progress), completed (ready to fetch), failed (error during execution), cancelled (user-initiated stop). Use get_job_result for completed jobs, get_job_log to debug failed jobs.'

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

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