The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/invoke.py dynamically locates and executes the vendor's engine by searching a list of over 60 predefined hidden directories in the user's home directory (e.g., ~/.claude/skills, ~/.cursor/skills, ~/.codeium/windsurf/skills). It uses subprocess.run to invoke the discovered script (kntnt.py) at a path computed at runtime.
[COMMAND_EXECUTION]: The skill defines a 'verification gate' (Step 6 in SKILL.md) where it instructs the agent to read the project's contributing guide and extract shell commands verbatim. These commands are then executed by subagents during the verification and integration phases. This pattern facilitates the execution of arbitrary, potentially malicious commands if they are present in the repository's documentation.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the processing of external ticket data.
Ingestion points: The skill reads ticket bodies and complete threads, including comments (Step 3 in SKILL.md).
Boundary markers: The subagent briefs (e.g., references/brief.md, references/amend.md) use Markdown headers and bold text to separate instructions from data, but instructions explicitly state that later comments in the thread override previous requirements.
Capability inventory: Subagents are equipped with broad capabilities including shell command execution (gate), file system modifications within worktrees and scratch directories, and interaction with GitHub CLI tools.
Sanitization: There is no evidence of sanitization or escaping of the ticket content; the raw <body> and <thread> data are interpolated directly into subagent briefs.

orchestrate