Five tools with complete schemas and descriptions. Naming follows verb_noun pattern (doctor, metrics, audit, insights, submit_pr). Descriptions are detailed (100-250 chars) and explain prerequisites and use cases. All parameters have types and descriptions. However, output schemas are not documented in the visible code, only input schemas are defined. Error handling is present but generic. Tool composition is sound: each has one responsibility, and outputs should chain (site parameter is consistent). Missing: explicit output schema documentation, per-item error details for batch operations, and idempotent/destructive hints.

• Document output schemas for all tools. Define the structure of returned markdown reports, metrics objects, and PR responses so LLMs know what fields to extract and how to chain calls.
• Add tool annotations: mark submit_pr with destructiveHint=true and idempotentHint=false; mark doctor, metrics, audit, insights with readOnlyHint=true. This signals to agents which tools are safe to retry.
• Enhance error messages with recovery guidance. Instead of 'Error: GSC credentials not found', return 'GSC credentials not found. Run doctor first to validate .env setup, or set GOOGLE_APPLICATION_CREDENTIALS.'
• For metrics and insights, document partial-result behavior explicitly. E.g., 'If CF credentials are missing, gsc data is returned; if GSC is missing, cf data is returned; if both are missing, only AI referrals are available.'
• Add precondition validation to submit_pr. Check for uncommitted changes, gh CLI availability, and origin remote before attempting the operation. Return a clear error with remediation steps if any precondition fails.