The Agent Skills Directory

[PROMPT_INJECTION]: The skill's description uses directive language ("ALWAYS invoke this skill") and specific trigger phrases designed to override the agent's default task-handling logic to ensure the skill is prioritized for development workflows.
[COMMAND_EXECUTION]: The skill automates the execution of several system commands including git for branch and worktree management, axp for running experiments, jq for parsing trace logs, and gh for managing pull requests.
[REMOTE_CODE_EXECUTION]: The workflow generates dynamic configuration files and shell scripts (setup and test phases) that are executed within isolated sandboxes during the experiment loop.
[SAFE]: The skill implements a robust "human-in-the-loop" security model, explicitly requiring the use of AskUserQuestion to obtain user consent before performing potentially dangerous operations such as running token-consuming experiments, modifying the git state, or pushing code to remote repositories.
[SAFE]: Includes an internal evaluation harness (eval/ directory) with a mock axp fixture script used to verify the skill's operational logic in a controlled environment.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface by reading and processing the output of tools and agent traces to calculate experiment metrics.
Ingestion points: Reads experiment results and trace logs from the .axp/runs/ directory (e.g., run.json, tests/*.json).
Boundary markers: None implemented for the ingested JSON data processed via jq filters.
Capability inventory: The skill can execute commands, manage files, and perform network operations through the gh CLI.
Sanitization: Relies on structured data parsing (jq) rather than executing ingested content directly.

axp-build-loop