The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/capture_dispatch_result.py executes git status --porcelain using subprocess.run to track file changes in the project repository.
[EXTERNAL_DOWNLOADS]: The skill instructions specify the use of @openai/codex and the Cursor CLI. These are well-known development and AI services.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code and diffs generated by AI builders (Cursor) and passes them to AI reviewers (Codex/Opus). The skill mitigates risks through:
Ingestion points: Project diffs and file content are read into the agent context as described in SKILL.md (Workflow steps).
Boundary markers: Instructions mandate the use of SPEC.md and BRIEF.md to define scope and constraints.
Capability inventory: The orchestrator utilizes shell command execution, file system writes, and network calls to external AI APIs.
Sanitization: The scripts/attest_file.py script is used to verify the integrity of instruction files (briefs) using SHA-256 hashes to detect if the AI builder attempted to modify its own instructions.
[DYNAMIC_CONTEXT_INJECTION]: SKILL.md documents the platform's support for the !command syntax, providing examples like `!`git status and !gh pr view`` to populate development context dynamically.

build-review-loop