The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/dk.py executes various shell commands to facilitate the decompilation workflow. This includes running compilers, assemblers, object dumpers, and git commands. It also supports execution of user-defined 'hooks' and 'blocks' specified in the project's decomp.toml configuration file.
Evidence: subprocess.run calls in scripts/dk.py for cc, as, git, and hooks.
[DYNAMIC_CONTEXT_INJECTION]: The skill includes configuration templates for AI agent platforms (Claude Code, Codex, OpenCode) that automatically execute dk state when a session starts or resumes. This is used to provide the agent with the current project status and pending tasks automatically.
Evidence: templates/claude-settings.json, templates/codex/hooks.json, and templates/opencode/plugins/dk-state.js all contain hooks to run python3 tools/dk.py state.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves agents reading and processing assembly code and C source files from the project directory. While this provides an attack surface where maliciously crafted code could attempt to influence the agent's behavior, the skill includes clear instructions for agents to restrict their edits to specific markers (INCLUDE_ASM) and follows standard development tool patterns.
Ingestion points: scripts/dk.py reads .s and .c files; workers read docs/worklog/*.md and docs/decomp_guide.md.
Boundary markers: The Worker protocol (templates/agent_protocol.md) explicitly instructs agents to replace only specific INCLUDE_ASM lines and avoid touching neighboring lines.
Capability inventory: The skill can execute git commands and local toolchain binaries (compiler, assembler) via dk.py.
Sanitization: scripts/dk.py uses shlex.split for command line parsing to mitigate argument injection.

matching-decomp