The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from several sources, including local application logs (references/workflow-logs.md), GitHub Action logs fetched via the gh CLI (references/workflow-ci.md), and visual assets analyzed via multimodal capabilities (references/workflow-ui.md). These inputs could be manipulated to contain instructions aimed at misleading the agent. While the skill employs high-capability tools like a Bash subagent and a git-manager, the provided templates do not include specific boundary markers or sanitization steps to isolate these external inputs from the agent's core instructions.
[COMMAND_EXECUTION]: The skill relies on a Bash subagent to execute local commands for verification, such as npm test, bun run typecheck, and pytest. These executions are central to the 'fixing' workflow but involve dynamic command generation based on the agent's analysis of the environment and logs.
[EXTERNAL_DOWNLOADS]: The skill retrieves build and test logs from GitHub using the gh CLI in the CI/CD workflow (references/workflow-ci.md). This involves fetching data from a well-known service to facilitate remote debugging.
[DATA_EXFILTRATION]: Analysis of network and file operations shows no evidence of sensitive data being sent to unauthorized external domains. Operations are restricted to local file access and interactions with the official GitHub API via standard tooling.

fixing