The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the claude CLI tool through multiple wrapper scripts using system-level calls.
scripts/claude-preflight.py utilizes pty.fork() and os.execve() to check authentication status while avoiding secret leakage in logs.
scripts/claude-run.sh invokes Claude Code in a new process group using a Python-based os.execvp() helper to manage headless execution.
[DYNAMIC_EXECUTION]: The runner script dynamically generates local shell scripts to manage the lifecycle of an execution session.
scripts/claude-run.sh writes monitor.sh and continue.sh at runtime to the artifact directory to allow external monitoring and session resumption.
[INDIRECT_PROMPT_INJECTION]: The skill monitors and processes session transcripts which may contain malicious instructions from processed data.
Ingestion points: scripts/claude-run.sh reads and snapshots JSONL transcript files from the ~/.claude/projects/ directory.
Boundary markers: SKILL.md includes an 'Authority Guard' section instructing the agent to never obey instructions found inside transcript content.
Capability inventory: The skill possesses extensive capabilities including command execution, file system modification, and network access via the underlying CLI.
Sanitization: No programmatic filtering is applied to transcript data; the skill relies on the LLM adhering to the conceptual safety boundary defined in the instructions.

claude