Critical failures across naming, descriptions, and schemas. 20 of 26 tools (77%) have empty input schemas with no visible properties, automatic 0 for schema. Tools 5-26 (action tools) have generic 1-2 word descriptions ('Preview X action', 'Apply X action') far below the 10-1024 char baseline and providing zero context for LLM selection. Tool names like 'preview_application_assignment' and 'apply_application_assignment' are paired but descriptions do not explain when to use preview vs apply or what data flows between them. First 4 tools (resolve_job_scope, confirm_job_scope, get_job_scope, get_recruiting_capabilities) have better structure with documented parameters, but even these lack output schema documentation. No tool documents return types or field structures. Error handling guidance is absent across all tools.

• Expand all 22 action tool descriptions from 1-2 words to 50-150 chars. Example: 'preview_application_assignment' → 'Preview the effect of assigning this application to a user or team. Returns a summary of changes without modifying state. Call this before apply_application_assignment to confirm the action.'
• Document input schemas for all 22 action tools. Specify required parameters (e.g., application_id, user_id, team_id) with types and descriptions. Currently empty schemas force LLMs to guess what data to pass.
• Add output schema documentation to all 4 job-scope tools. Example for resolve_job_scope: 'Returns {scope_handle: string, job_id: number, job_title: string, confirmation_token: string, candidates_count: number}' so LLMs know what fields to extract for downstream calls.
• Add error recovery guidance to all tools. Example: 'If scope_handle is invalid, call resolve_job_scope() with the job title or requisition number. If confirmation_token expires, re-run resolve_job_scope() to get a fresh token.'
• Document the preview→apply workflow. Add to all preview tool descriptions: 'Call this first to validate the action. If the preview succeeds, pass the returned confirmation_token to the corresponding apply_* tool to execute the change.' Add to all apply tool descriptions: 'Requires a confirmation_token from the corresponding preview_* tool. This ensures the agent reviewed the change before execution.'

Inferred effective spec: 2025-06-18+.

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained