ln-1000-pipeline-orchestrator
Warn
Audited by Gen Agent Trust Hub on Mar 13, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill modifies the agent's environment by installing bash scripts (pipeline-keepalive.sh and worker-keepalive.sh) as Stop and TeammateIdle hooks in .claude/settings.local.json to maintain session persistence during long-running tasks.
- [COMMAND_EXECUTION]: Spawns a background PowerShell process (prevent-sleep.ps1) that uses .NET Add-Type to compile and execute C# code for calling the Win32 API function SetThreadExecutionState.
- [REMOTE_CODE_EXECUTION]: Uses the Task() tool to dynamically spawn and coordinate a hierarchy of worker agents across different pipeline stages, passing them complex instructions and shared context generated from local state.
- [COMMAND_EXECUTION]: Executes various shell commands for workspace and process management, including git worktree operations, file manipulation via cp and rm, and process termination via kill.
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8). Evidence: 1. Ingestion points: Reads content from kanban_board.md, story descriptions, and CLAUDE.md. 2. Boundary markers: Implements markers but lacks strict sanitization for other ingested fields. 3. Capability inventory: Spawns sub-agents, executes system commands, and interacts with the Linear API. 4. Sanitization: There is no evident filtering or validation of ingested story content before it is interpolated into instructions for the worker agents.
Audit Metadata