The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes model PR history and framework source code (vLLM, SGLang) to influence its autonomous patching decisions. This creates a surface where adversarial instructions in external documentation or code comments could redirect the agent's behavior.
Ingestion points: Reads model-pr-optimization-history and source code from vLLM and SGLang to populate history notes as described in SKILL.md (Phase 0.5).
Boundary markers: The instructions do not specify any delimiters or safety warnings to distinguish between data and instructions when reading these external sources.
Capability inventory: The skill has the ability to patch source code (Python, C++, CUDA, Triton) and execute shell commands for benchmarking and profiling.
Sanitization: There is no mention of filtering or validating the ingested content.
[DYNAMIC_EXECUTION]: The skill autonomously modifies and executes vLLM source code to improve performance.
Evidence: The RLCR loop involves patching files in csrc/, vllm/, and vllm/compilation/, then running vllm serve to benchmark the modified code.
Computed Paths: It executes scripts/setup-rlcr-loop.sh from a path determined by the $HUMANIZE_RUNTIME_ROOT environment variable.
[COMMAND_EXECUTION]: The skill runs various performance benchmarking and profiling tools.
Evidence: Executes vllm serve, python -m sglang.launch_server, trtllm-serve serve --backend pytorch, and various internal scripts like scripts/query.py and scripts/setup-rlcr-loop.sh.

vllm-sota-humanize-loop