The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill involves the agent parsing and summarizing external data sources, which presents a potential surface for indirect instructions.
Ingestion points: The agent is instructed to read and analyze experiment logs (*.out, *.log) and task artifacts in review-logs.md, triage.md, and validate.md.
Boundary markers: No explicit delimiters or boundary warnings are provided for parsing log content.
Capability inventory: The skill utilizes uv run launch.py for job execution and standard system utilities including find, tail, grep, ls, du, and test for file system and log interactions.
Sanitization: The instructions do not specify sanitization or escaping of log content before the agent generates reports or summaries for the user.
[SAFE]: The skill's use of configuration knobs like trust_remote_code is explicitly documented as a necessary step for specific model architectures (e.g., Mixture-of-Experts or custom models). The instructions provide clear guidance on how and where to apply these flags across different pipeline stages.
[SAFE]: All referenced tools (uv, vLLM, TRT-LLM) and platforms (Hugging Face, arXiv) are industry-standard resources for the declared purpose of model optimization and research.

speculative-decoding