Two tools with complete input schemas and descriptions, but significant gaps in output documentation and error handling. Tool names use verb_noun pattern (codex_handoff, claude_handoff) which is clear. Descriptions are adequate (100-150 chars) but lack WHEN/WHY guidance. Input schemas are well-structured with nested objects and enums for network access. However, no documented output schemas, no error recovery guidance, and no indication of idempotency or side-effect handling beyond the description text mentioning 'evidence gathering'. Parameters are well-typed but lack examples of acceptance_criteria format or timeout behavior on edge cases.

• Add explicit output schema documentation: return {success: bool, evidence: string, side_effects: string[], output_text: string, error?: string}. This enables LLMs to extract and chain results.
• Expand descriptions to 100-150 chars with WHEN/WHY: 'Delegate a task to Codex (not Claude) when you need code generation with write access to project files. Returns evidence of side effects and acceptance check results.'
• Document acceptance_criteria format with an example: 'Array of strings describing success conditions, e.g. ["code compiles", "tests pass", "no new warnings"]'.
• Add error recovery guidance: 'If timeout occurs, retry with longer timeout_seconds. If permission denied, check requested_scope.writable_paths. If acceptance criteria fail, review evidence and adjust goal.'
• Clarify idempotency: 'Calling with the same goal and project_dir is idempotent, returns cached result if available. Use a unique goal string to force re-execution.'