The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_dcode.py uses subprocess.run to execute the local dcode CLI tool and various fallback runner scripts (run_claude.py, run_codex.py, run_pi.py) to process user prompts.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from the workspace and interpolates it into prompts for an agentic CLI tool.
Ingestion points: The build_prompt function in scripts/run_dcode.py reads content from file paths provided via the --prompt-file and --session-file arguments.
Boundary markers: While the script adds descriptive headers to delineate injected content (e.g., 'Prior conversation context to continue from:'), the documentation in references/runtime-reference.md explicitly warns that these are 'soft constraints' and that the read-only overlay is not a hard security boundary.
Capability inventory: The skill invokes the dcode CLI, which is an agentic tool capable of performing workspace modifications and file operations.
Sanitization: There is no sanitization, escaping, or schema validation performed on the content loaded from external files before it is injected into the prompt.
[DYNAMIC_EXECUTION]: The scripts/run_dcode.py script dynamically modifies sys.path to load sibling modules and utility scripts (runner_jobs, skill_paths) based on the detected skill layout.
[EXTERNAL_DOWNLOADS]: The documentation in references/runtime-reference.md suggests installing the dcode CLI prerequisite from a well-known service using the command curl -LsSf https://langch.in/dcode | bash.

dcode-runner