Single tool 'ask_codex' has a complete input schema with 4 parameters (task, context, resume_session_id, workdir), all with type definitions and descriptions. Description is 287 chars, within the 10-1024 baseline. However, output schema is not documented, the tool returns (response_text, session_id) but this is not declared in the tool definition visible in the code. Parameter descriptions are adequate but lack format constraints (e.g., no guidance on workdir format, no enum for sandbox modes). Error handling is minimal, no recovery guidance for SSH failures, timeouts, or Codex CLI errors. The tool performs a single, well-defined action (delegate to Codex), but lacks production-grade error classification and user-fixable guidance.

• Add explicit output schema to types.Tool definition: {type: 'object', properties: {response: {type: 'string', description: 'Codex response text'}, session_id: {type: 'string', description: 'Session ID for resuming conversation'}}, required: ['response']}
• Implement error classification: wrap SSH/Codex errors with categorization (retryable: timeout/connection; user-fixable: invalid workdir; fatal: auth failure) and recovery hints.
• Add sandbox parameter as enum: {type: 'string', enum: ['read-only', 'workspace-write'], description: 'Sandbox mode restricting Codex execution. Default: Codex CLI default. read-only: no filesystem writes; workspace-write: writes allowed in workspace.'}
• Document workdir format: 'Absolute path on remote Mac (e.g., /Users/username/project). Must exist and be readable by the SSH user.'
• Clarify session resumption: 'Session ID returned by a previous ask_codex call. Pass to continue the same conversation. Sessions expire after 24 hours. If stale, omit resume_session_id to start fresh.'