The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a helper bash script (openmontage.sh) and specialized Python scripts to perform repository health checks, manifest parsing, and project reporting. These tools implement safety measures such as project ID sanitization to prevent path traversal and command injection, and read-only preflight modes to inspect environment readiness without executing production workloads.
[INDIRECT_PROMPT_INJECTION]: The system processes external content including YAML pipelines, Markdown directors, and reference video files, which represent a potential attack surface for indirect prompt injection. However, this is mitigated by the skill's architecture which enforces strict human-in-the-loop approval gates (using awaiting_human checkpoints) before advancing to consequential stages like asset generation or final rendering.
Ingestion points: YAML manifests in pipeline_defs/, Markdown stage directors in skills/pipelines/, and external media files provided for analysis.
Boundary markers: Explicit checkpoint lifecycle states (in_progress, awaiting_human, completed) and manifest-defined human_approval_default flags that pause agent execution.
Capability inventory: File system manipulation (Bash Read/Write/Edit), Python tool registry discovery, and execution of local video composition runtimes (Remotion, HyperFrames, FFmpeg).
Sanitization: Structural validation of manifests via pipeline_inventory.py and mandatory final review protocols involving frame sampling and metadata inspection using ffprobe.

openmontage