The Agent Skills Directory

[COMMAND_EXECUTION]: The skill heavily relies on shell command execution via the Bash tool to interact with the GitHub CLI, git, and local project build/test environments (e.g., uv, go, npm, nx).
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from GitHub Actions logs and downloaded artifacts to analyze failures.
Ingestion points: Fetches failed step logs using gh run view --log-failed and reads artifacts from /tmp/ci-debug/ in SKILL.md.
Boundary markers: No explicit delimiters or instructions are used to separate external log content from the agent's internal analysis logic.
Capability inventory: The agent has access to the Bash tool for running reproduction commands and the ability to modify project files via code editing tools.
Sanitization: Logs are truncated to 500 lines, but no further sanitization or escaping of the log content is performed.
[DYNAMIC_EXECUTION]: The skill constructs shell commands for local reproduction by interpolating data extracted from CI logs, such as test file paths and specific test names (e.g., uv run pytest -v <test_file>::<test_name>). This creates an attack surface where maliciously crafted CI output could lead to unintended command execution during the reproduction step.

fix-ci