Cryptographic module identification is the starting point of post−quantum cryptographic migration in the banking and financial industry, and its results form the basis for compiling a cryptographic asset inventory and prioritizing migration. In view of the large scale, language diversity, deeply nested wrapper layers and complex dependency chains of banking information systems and operating system source code, an LLM−based cryptographic module identification tool is investigated in this paper. Taking the whole codebase and its dependency information as input, the tool completes identification through the collaboration of five agents responsible for data ingestion and preprocessing, call graph construction, cryptographic semantic identification, trustworthy reasoning and reflection, and result aggregation and visualization, and outputs evidence units traceable to source statements. The tool is implemented on the open−source Qwen3−8B model without parameter fine−tuning. On 400 file−level samples constructed from the source code of the Anolis OS kernel, it achieves a precision of 97.50% and a recall of 96.06%. In a comparative experiment against an open−source pattern−matching cryptographic discovery tool, the proposed tool identifies 76 of the 84 genuine cryptographic modules missed by the reference tool; ablation experiments show that the combination of call graph construction and trustworthy reasoning and reflection raises recall from 20.69% to 96.06%. These results indicate that semantic−level criteria extend the scope of cryptographic module identification from implementations explicitly visible in the code text to system−level code organized through registration and encapsulation.