The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/patch_eval_image_cap.py implements dynamic source code modification. It executes docker run to cat a library file (cosmos_rl/evaluation/base.py) from a container image, applies a regular expression to modify the limit_mm_per_prompt constant, and writes the resulting code to the local filesystem so it can be mounted back into subsequent container runs.
[DATA_EXFILTRATION]: The script scripts/align_token_usage.py accesses sensitive local data by scanning the ~/.claude/projects/ directory. It reads through session transcript JSONL files to extract and aggregate token usage metrics and assistant message history, which is then written to a state file.
[COMMAND_EXECUTION]: The skill acts as an orchestrator and frequently executes system commands via subprocess.run.
scripts/deft_exec.py is a wrapper used to execute arbitrary shell commands while enforcing a local security policy for air-gapped environments.
scripts/patch_eval_image_cap.py uses subprocess.run to interact with the Docker daemon.
[EXTERNAL_DOWNLOADS]: The skill instructions and scripts facilitate the download of large model weights and container images from external sources.
references/cosmos-reason.md contains instructions for using the Hugging Face CLI (hf download) to fetch models from nvidia/Cosmos3-Nano.
scripts/init_deft_state.py resolves and references container images from NVIDIA's registries.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection as it ingests untrusted data from user-supplied annotation files (proxy_kpi.json, benchmark_kpi.json).
Ingestion points: Annotations are loaded in scripts/check_annotations.py and scripts/validate_sharegpt.py.
Boundary markers: Absent; the skill relies on the expectation that the data follows a specific schema.
Capability inventory: The skill has file-write access and can execute arbitrary shell commands via its orchestration scripts.
Sanitization: The validation scripts check for schema compliance but do not sanitize the natural language content of the prompts or labels processed by the models.

tao-run-deft-aoi-cosmos3