Four tools with complete schemas and descriptions. Tool names follow verb_noun pattern (delegate, resume, list_models, check_codex). Descriptions are substantive (80-150 chars). All parameters have types and descriptions. Input schemas use proper JSON Schema with enums for constrained values (reasoning_effort, sandbox, priority). However, output schemas are not documented, LLMs cannot infer what fields to expect from responses. Error handling is present but generic (textResult/errorResult). No tool annotations (readOnlyHint/destructiveHint). delegate and resume are high-risk WRITE operations but lack confirmation/dry-run patterns. Parameter validation is thorough (validateDirectory, validateWorkingDir, validateAddDirs) but happens at runtime, not schema-time.

• Document output schemas for all four tools. E.g., delegate should return {thread_id: string, status: 'success'|'failure', output: string, duration_ms: number, token_usage: {inputTokens, outputTokens, ...}}. This enables LLM to chain results to resume() and extract key fields.
• Add destructiveHint: true to delegate and resume in tool registration. Mark list_models and check_codex with readOnlyHint: true. This signals to agents which tools are safe to retry.
• Enhance error responses with structured classification. Instead of textResult(message, true), return {content: [{type: 'text', text: message}], isError: true, metadata: {errorType: 'retryable'|'user_fixable'|'fatal', suggestedAction: '...'}}.
• Add minLength: 1 constraint to working_dir schema (or make it non-optional and document the default). Prevent empty-string edge case at schema validation time, not runtime.
• Document add_dirs constraints: maxItems: 10 (or appropriate limit), and add to description: 'Each directory must be an absolute path and must exist. Maximum 10 directories.'
• Add idempotentHint: true to list_models and check_codex (read-only, safe to retry). This helps agents optimize retry logic.