The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of shell commands and local scripts to manage its lifecycle. It executes python3 .agents/skills/process-inbox/scripts/workspace.py for resolving, starting, and committing tasks. It also invokes platform-specific build commands such as cmake on Windows and Docker-based build scripts (build_debug.sh) on Linux/WSL. While these tools are standard for a developer workflow, they provide a broad surface for command execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data from repository files and task definitions, which creates a surface for indirect prompt injection targeting the subagents it manages.
Ingestion points: Reads task.md, project context files, repository source files (C++, styles, localization), and host-specific adapter files like .grok/ai-workflow-adapter.md.
Boundary markers: Prompts for delegated workers include explicit instructions to act as "leaf workers" and are forbidden from spawning or delegating further, which helps limit the impact of potential injections.
Capability inventory: The skill has the capability to write files, execute shell commands, perform Git operations, and interact with the UI via Computer Use tools.
Sanitization: The skill relies on a "Plan Assessment" phase (Phase 3) where an independent agent reviews the proposed implementation plan for correctness and structural fit, but there is no explicit programmatic sanitization of text values derived from the ingested files before they are included in subagent prompts.
[DYNAMIC_EXECUTION]: The skill dynamically generates prompt files for subagents (logs/phase-*.prompt.md) by populating templates with data gathered during the context-gathering phase. This involves interpolating task-specific content into executable instructions for sub-processes, a form of script generation from templates.

perform-task