Server has 7 tools with schemas present but significant quality gaps. Tool descriptions are present (10-50 chars) but lack LLM-optimized detail on WHEN to use each tool and what it returns. Parameter descriptions exist but are minimal (e.g., 'Text prompt for generation'). No output schemas documented. Error handling and recovery guidance absent. Tools mix concerns (e.g., run_children handles parallelism, retries, and sandboxing in one tool). No tool annotations (destructiveHint, idempotentHint). Naming is verb-forward but some tools are vague: 'run_children' and 'aggregate' don't clearly signal what data they operate on or what they return.

• Add output schemas to all tools. Document the structure of returned data (e.g., generate_image returns {file_path, url, dimensions}; run_children returns {completed_tasks, failed_tasks, results}).
• Expand tool descriptions to 100-200 chars. Include WHAT the tool does, WHEN to use it (e.g., 'Call this after creating a manifest'), and WHAT it returns. Example: 'Run Codex research tasks defined in a manifest file. Returns a list of completed task results and any failures. Use this to parallelize research work across multiple topics.'
• Add per-parameter descriptions for all parameters. Current descriptions like 'Research run directory' are too terse. Expand to 'Path to the directory containing the research run (created by codex init). Must exist and contain a manifest.json file.'
• Document error cases and recovery paths. E.g., 'If run_children times out, retry with a smaller parallel value or increase timeout. If a task fails, check the task logs in <run_dir>/logs/<task_id>.log.'
• Add tool annotations (destructiveHint, idempotentHint) to guide agent retry logic. Mark run_children as idempotent (safe to retry), aggregate as idempotent, and update_codex_plugins as destructive (may uninstall plugins).