Output schemas not documented. Tools like read_memory, get_scan_results, and list_processes have descriptions but no published return type schemas. LLMs cannot plan downstream calls or extract required fields (e.g., address, value fields from memory read responses).
Parameter descriptions lack actionable detail. 'data_type' parameter in scan_value, read_memory, write_memory mentions supported types as inline text ('int8, int16, ...') but these are not enforced as JSON Schema enums. No format constraints (e.g., hex address format, value range). Descriptions are 10-50 chars, below the 50-200 char baseline for A-grade tools.
Document output schemas for all tools in the server definition. For example, read_memory should return {type: 'object', properties: {value: {type: [number, string]}, address: {type: string}, size: {type: integer}}}. This enables agents to plan multi-step chains.
Convert inline data_type descriptions to JSON Schema enums. Instead of 'data_type: {type: string, description: "int8, int16, int32, ...'}', use 'data_type: {type: string, enum: ["int8", "int16", "int32", "int64", "float", "double", "utf8", "utf16", "wide", "bytes"]}'. This prevents hallucinated values.
Expand parameter descriptions to 50-200 characters and include format/range constraints. E.g., 'address: {type: string, description: "Memory address in the target process. Accepts hexadecimal (0x12345678) or decimal (305441400) notation."}' instead of just 'Memory address to read from (hex or decimal)'.
Add recovery guidance to error responses. If attach fails, return {success: false, error: 'Process not found', suggestion: 'Call list_processes to see available targets'}. If scan_next finds no matches, suggest widening the search with scan_changed.
Implement permission gates for destructive operations. Gate write_memory behind a permission check (e.g., require 'write:process' scope). Log all mutations with timestamp, caller ID, process, address, and old/new values.
Add tool annotations to server registration: mark read_* and list_* as readOnlyHint=true; mark write_memory and attach as destructiveHint=true; mark detach/clear_scan as idempotentHint=true.
No error recovery guidance. Tools return success/failure messages (e.g., 'success': true/false) but provide no guidance on retryability, what to do if attachment fails, or how to recover from scan exhaustion. E.g., 'attach' could fail if the process doesn't exist, the error should suggest 'Process not found. Call list_processes to see available targets.'
No permission gates or security auditing. write_memory is a DESTRUCTIVE tool (can corrupt target process) but has no permission checks, audit logging, or confirmation step. No evidence of server-side rate limiting to prevent runaway agents.
Missing tool annotations. No use of readOnlyHint, destructiveHint, or idempotentHint in tool metadata. Agents cannot infer which tools are safe to retry or which will cause side effects without annotations.
Tool definitions appear in test file (automated_test.py) rather than main server.py. Tool registration is not visible in the provided code excerpt. If tools are inferred from test usage rather than explicitly registered in the MCP server, their definitions are brittle and cannot be verified.
Parameter type overloading. 'target' parameter in attach accepts both string (process name) and integer (PID), but the description does not distinguish between them or specify which is preferred. Similarly, 'address' in read_memory/write_memory accepts 'string|integer' (hex or decimal) without format guidance. This forces LLMs to guess representation.
Limited pagination support. get_scan_results accepts a 'limit' parameter but no offset/page/cursor for result navigation. If a scan finds 10,000 addresses and the agent retrieves the first 100, there is no mechanism to fetch the next batch without a new scan.
Split overloaded parameters: change attach(target: string|integer) into two optional parameters: attach(pid: integer, process_name: string) with a constraint that exactly one must be provided. Or provide separate tools: attach_by_pid, attach_by_name.
Add offset/cursor pagination to list_* and get_scan_results. E.g., get_scan_results(limit=50, offset=0) or get_scan_results(limit=50, cursor='next_page_token'). Return {results: [...], total_count: 10000, next_cursor: '...'}.
Verify tool registration in src/frida_game_hacking_mcp/server.py is explicit and complete. All tools should be registered via server.tool() decorator with full schemas, not inferred from test code.
Add a dry_run parameter to write_memory for confirmation workflows: write_memory(..., dry_run=true) returns the hypothetical change without applying it, enabling agents to preview before executing.