The Agent Skills Directory

[SAFE]: The skill implements a sophisticated state machine using deft_state.json and loop_log.jsonl. It includes a journal-based recovery system (.deft_commit.journal) to handle interrupted writes and ensures that the agent follows a strictly defined, audited sequence of operations.
[COMMAND_EXECUTION]: Shell command execution via subprocess.run and docker run is used exclusively for automation of the model training lifecycle. Container execution is configured with --user to maintain host file permissions and prevent privilege escalation within the results directory.
[EXTERNAL_DOWNLOADS]: Downloads are performed for pre-trained model checkpoints from well-known sources (HuggingFace and NVIDIA NGC). These downloads are mediated by dedicated scripts (fetch_codetr_checkpoint.py and fetch_gdino_checkpoint.py) that verify file sizes and SHA-256 hashes before the binaries are utilized in the workflow.
[DATA_EXFILTRATION]: The skill instructions proactively manage security by requiring credentials (NGC_KEY, HF_TOKEN) to be set in environment variables and forbidding the agent from asking the user for their values or reading them directly.

tao-run-deft-object-detection