The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a custom shell script scripts/codex-exec.sh which invokes the codex command-line interface. This script is the primary mechanism for delegating implementation tasks to external models and tools.
[PRIVILEGE_ESCALATION]: The skill instructions and the supporting shell script scripts/codex-exec.sh include a mechanism to explicitly bypass security boundaries. By setting the CODEX_BYPASS environment variable, the agent is instructed to add the --dangerously-bypass-approvals-and-sandbox flag to the tool execution. This allows the delegated process to write files anywhere on the system, bypassing the default workspace-write sandbox restriction.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture for routing and delegating work to sub-agents (Claude and Codex) creates a surface for indirect prompt injection. Untrusted data encountered during tasks like web searching or file reading could be interpolated into prompts passed to downstream executors.
Ingestion points: The PROMPT variable passed to the codex-exec.sh script, which incorporates plan goals and repository paths.
Boundary markers: The instructions recommend using "scope fences" (e.g., "ONLY Phase N") to delimit tasks within the prompt.
Capability inventory: Execution of the codex CLI via shell scripts, use of git worktree for environment management, and execution of bun install for dependency management.
Sanitization: The skill mandates a "Verify" phase where the main orchestrator agent is required to review all diffs and judge them like a contributor PR before merging, which acts as a manual sanitization step.
[DYNAMIC_EXECUTION]: The skill facilitates the dynamic generation and execution of code at runtime. It utilizes the codex tool to implement code slices from specifications and references desplega:script-builder for generating and executing bulk call sequences.

delegate-work