team-worktree

Warn

Audited by Gen Agent Trust Hub on Nov 15, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The scripts/run-teardown.sh script implements a mechanism that reads and executes arbitrary shell commands from a .teamteardown file located in the repository's default branch. The script utilizes sh -c "$line" to run these commands sequentially. While the skill attempts to mitigate risk by only executing the version of the file found on the default branch (e.g., origin/main), this remains a powerful dynamic execution vector that could be abused if the repository's main branch is compromised.
  • [INDIRECT_PROMPT_INJECTION]: The skill's behavior is directed by several external files within the repository environment, creating a surface for indirect injection.
  • Ingestion points: The file docs/plans/<id>/4-repos.md is parsed to determine which repositories the agent will interact with, and .teamteardown is parsed for shell commands.
  • Boundary markers: The skill does not use explicit delimiters or "ignore instructions" markers when reading these files, relying instead on shell-level escaping.
  • Capability inventory: The agent has the capability to create/delete git worktrees, execute arbitrary shell commands (via teardown), and modify repository files.
  • Sanitization: The skill includes some input validation, such as refusing values containing single quotes, newlines, $, or backticks when preparing shell arguments, and using regex-like validation for branch names.
  • [COMMAND_EXECUTION]: The skill relies on multiple bash scripts (add-worktree.sh, run-teardown.sh, sweep-worktrees.sh) that construct and execute shell commands using repository-derived data (branch names, repository paths, PR metadata). The high level of complexity in these shell-based operations increases the risk of command injection if any of the validation steps are bypassed or if unexpected input patterns are encountered during repository metadata processing.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 15, 2026, 12:30 PM