The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted ticket content (body and discussion threads) from an external tracker and includes it in briefs for subagents (found in references/brief.md, references/amend.md, references/repair.md, references/repaired.md, references/verify.md, and references/wave.md).
Ingestion points: Ticket body and comments (referenced in SKILL.md step 3).
Boundary markers: None. The templates interpolate untrusted content directly into markdown prompts without specific instructions to ignore embedded commands.
Capability inventory: The skill can execute git, gh, and uv commands and manages the lifecycle of multiple subagents.
Sanitization: None mentioned.
[DATA_EXFILTRATION]: The skill accesses sensitive local application data to determine its current operating context.
Evidence: In SKILL.md (Step 3), the skill reads ~/.claude/projects/$CLAUDE_CODE_SESSION_ID.jsonl to extract the permissionMode of the current session to propagate permissions to subagents.
[DYNAMIC_EXECUTION]: The skill dynamically computes paths to find and execute its management engine across various application directories.
Evidence: scripts/invoke.py scans a large number of hidden directories in the user's home folder (such as ~/.cursor/skills, ~/.copilot/skills, and ~/.continue/skills) to find a vendor-specific manager script (kntnt.py) and executes it using uv run.
[COMMAND_EXECUTION]: The skill uses shell commands to interact with the local environment and development tools.
Evidence: scripts/invoke.py uses subprocess.run to call uv and the discovered manager script. SKILL.md indicates the use of git and gh for repository and tracker management.

orchestrate