The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates the creation and management of experiment arms by executing git and gh (GitHub CLI) commands through the Python subprocess module. These operations include generating nonce commits, pushing experiment branches, and managing draft Pull Requests. The use of list-based arguments for these commands mitigates typical command injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill interacts with the GitHub API via the gh CLI to monitor the status of workflow runs and retrieve job logs. These network operations are strictly limited to fetching metadata and results necessary for the experiment's statistical evaluation.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources, specifically GitHub API responses and local experiment result logs, to compute performance statistics.
Ingestion points: GitHub API check-run and workflow-run data is ingested by scripts/ci_arm_driver.py; experiment samples are read from JSONL files by scripts/arm_stats.py and scripts/ci_arm_metrics.py.
Boundary markers: Not applicable, as the skill performs programmatic data processing in standalone scripts rather than interpolating raw data into LLM prompts.
Capability inventory: The skill maintains the ability to read and write files within the project directory and execute shell commands via git and the gh CLI.
Sanitization: Input to system commands is handled using list arguments, and incoming data is parsed as structured JSON, preventing the execution of embedded instructions.

experiment-arms