The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/calculate_cost.py calls AutoTokenizer.from_pretrained with the trust_remote_code=True parameter. This setting permits the execution of arbitrary Python code bundled within the downloaded model repository. While model IDs are validated against a static pricing table, enabling remote code execution is a high-risk pattern that circumvents security defaults for model-specific tokenization scripts.
[REMOTE_CODE_EXECUTION]: The skill facilitates the download and execution of remote code through the transformers library whenever a model requires a custom tokenizer. This creates a dependency on the integrity of third-party model repositories hosted on the Hugging Face Hub.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from user-provided JSONL files, establishing a surface for indirect prompt injection.
Ingestion points: The dataset argument in scripts/calculate_cost.py is used to open and read external files.
Boundary markers: The script does not use delimiters to isolate user-provided content from the processing instructions.
Capability inventory: The skill possesses file read capabilities and the ability to execute remote code during model loading.
Sanitization: No validation or filtering is applied to the dataset rows before they are passed to the tokenizer.

tinker-training-cost