The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell scripts (run-team.sh, run-workflow.sh, resume-team.sh) to invoke the claude CLI for managing multi-agent engineering sessions. These scripts pass user-provided goals and feedback as prompts to the agent tier. The lib/lifecycle.sh utility manages the process lifecycle of these commands using ps, awk, and kill to ensure clean teardown of child processes and their descendants.
[DYNAMIC_EXECUTION]: The skill utilizes Claude Code's native dynamic workflow feature through assets/workflows/gpt-engineer-dynamic.js. This JavaScript-based orchestration manages parallel research, planning, building, and verification phases. The run-workflow.sh wrapper manages the execution of this workflow within isolated git worktrees to prevent accidental mutations to the main checkout.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted repository content (source code, documentation, CI logs) and user-supplied engineering goals which are processed by subagents. This creates a vulnerability surface for indirect prompt injection. The skill implements several mitigations, including strict JSON schemas for agent outputs and explicit safety instructions in the workflow prompts to prevent unauthorized commits, pushes, or external network operations.
[EXTERNAL_DOWNLOADS]: The post-edit-lint.sh hook uses npx --no-install to execute the prettier formatter after file modifications, and the bootstrap.sh script verifies the presence of the claude CLI on the system path. These are routine development operations for tool discovery and maintenance.

claude-multi-agent