The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted legacy game binaries using Ghidra/PyGhidra. This creates a surface where malicious instructions or deceptive strings embedded in a binary could influence the agent's behavior during the iterative reconstruction of C code.
Ingestion points: Binary files processed via pyghidra.open_program(binary_path) in scripts/extract_decomp.py and scripts/pcode_rename.py.
Boundary markers: No explicit markers are present in the scripts to isolate untrusted data extracted from the decompiler from the agent's reasoning instructions.
Capability inventory: The skill possesses the ability to modify local source files ("Apply the single source change") and execute arbitrary build/comparison commands (ninja, make, objdiff).
Sanitization: There is no evidence of sanitization for strings or metadata extracted from the binaries before they are processed by the LLM or added to the iteration ledger.
[COMMAND_EXECUTION]: The skill workflow involves executing various local commands and scripts to automate the matching process. This includes build system calls (ninja, make), comparison tools (objdiff), and local repository scripts (decomp/scripts/run_match.py, decomp/scripts/parse_compare.py). While these are necessary for the skill's primary function, they represent an execution surface for the agent.
[EXTERNAL_DOWNLOADS]: The documentation in references/setup.md instructs users to install the pyghidra package via pip. This is a legitimate library used to bridge Python with the Ghidra reverse engineering suite.

matching-decomp