Tracesearch MCP

🚀 Trace Analysis Tool Based on large-text-viewer

The text search backend can be replaced according to your preference, such as rigrep or even direct command-line grep.

Currently, two usage modes are supported:

• Trae Skill Mode (Recommended): Located at .trae/skills/large-text-searcher/, deeply integrated with Trae.
• MCP Server Mode: Supports the traditional MCP protocol and is compatible with IDEs like Kiro.

For the detailed migration guide, please refer to MIGRATION_GUIDE.md.

You need to use your own trace format in search_strategy.md and trace_format.md.

⚠️ Important Note

The "brain folds" of AI are smooth. You can only let it perform mechanized labor for you or expect it to have a sudden flash of inspiration.

Therefore, you should consider a streamlined and standardized trace search method. AI can't do things beyond your own cognition.

AI performs well with simple algorithms, but more complex algorithms will consume a large number of tokens, and the performance may not be good. You need to do more debugging on your own.

For technical communication, you can add WeChat baserker2 and add a note.

[+] New Feature: When the search results exceed 100, the results are meaningless. Intercept them directly in MCP and let the model rethink how to search, so as not to waste tokens and review the results. [+] New Feature: Summarize the prompt words and put them into the skill. [+] Added the ability to go through the magic number prior process. // Integrate pyberchef

Thoughts

• Different models have different preferences when using APIs. Claude prefers to use context, while Gemini prefers to use regular matching. However, both are not very fond of searching for instructions like add a,c,b.
• After multiple thinking rounds, the behavior may mutate. For example, Gemini starts using grep to bypass the specified MCP regular matching.
• Currently, the model doesn't understand the relative relationship between lines. For example, when searching twice consecutively, it can't make a comparison between the two searches.
• Regarding the format issue, currently st____ /ld____ and instruction operations are on two lines. This will waste one search round when the AI is looking, and it may even get lost directly. You can try to put st__ and the instruction on the same line, but it will sacrifice some readability. The same goes for string__ and funcall.
• You can also give hints in the MCP's return to the AI, such as counting. Remind the AI to summarize after 10 searches.
• It seems necessary to limit the AI's free play at the Python layer. Some parameters should not be given to it, as it may lead to extreme thinking.
• The intelligence of the AI itself plays a large part. A less capable AI will only search blindly, while a more powerful one can start from plain text, guess in both forward and reverse directions, and know how to use information like memcpy.
• Add some auxiliary functions, such as magic number search, to let the model know that there are encryption algorithms in certain places. Or enumerate all the libc functions that have appeared to give the AI an overall impression.
• Provide ida mcp to let the model obtain local information from decompiled code.
• Add unidbg mcp to let the model analyze and simulate algorithms through unidbg.
• Tell the AI which search schemes are not rigorous.

    {
      "context": 5,
      "file_path": "/Users/teng/PycharmProjects/pythonProject/shopee_proj/log/record_00.txt",
      "max_results": 10,
      "pattern": "704da53700"
    }
    For example, such search results are extremely likely to cause bias. `max_results` can be set so that the AI can only obtain the results without operating it.
    In this case, maybe `find_prev` and `find_next` can be provided? But then it will return to the issue of taint.

• TODO: Provide a small-range reverse taint for context to improve the quality of context, but this will affect the format. How to make them compatible?
• Upgrade the backend to add line range filtering to search_text.
• Add a bookmark function to let the model mark the function of a certain range of lines, mark some ranges as useless, and affect the filtering results.
• Limit the length of the displayed lines to avoid consuming context too quickly.
• TODO: Identify and handle common encryption algorithms (without magic numbers).
• Regarding the issue of the model's preference for using APIs, try to execute an example for the model to understand.