The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The engineering workflow is designed to ingest and process untrusted data from external sources throughout its multi-phase execution.
Ingestion points: The skill retrieves results from research tools (research skill), web searches (web_search), and GitHub repository issue/code searches (gh) as detailed in references/stage-progression-protocol.md and references/failure-recovery-protocol.md.
Boundary markers: The agent uses structured Markdown templates (e.g., templates/goal-plan-template.md and templates/atomic-task-template.md) and phase-based state transitions to delimit untrusted input from execution commands.
Capability inventory: The skill possesses capabilities for local file-system modification, shell command execution (via the tracker script), and Git operations.
Sanitization: The skill mandates a "Dual-Round Review" gate in Phase 4 (P4) and requires the successful execution of tiered testing (L0-L3) before any code is finalized, providing a robust sanitization mechanism for generated content.
[COMMAND_EXECUTION]: The helper utility scripts/goal-state-tracker.sh executes the git command via subprocess to dynamically identify the project root directory. This is used for locating the state storage directory and plan files within the local environment.
[DYNAMIC_EXECUTION]: The scripts/goal-state-tracker.sh file utilizes a shell-to-Python heredoc wrapper (`exec python3
"$@" <<'PY'`) to execute logic provided as a raw string to the Python interpreter. While used here for unified delivery of a state-tracking utility, this pattern represents dynamic execution of a piped script script.

goal-loop