The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill mandates the persistence of sensitive authentication data to the local filesystem. In the 'Opening Orgs' workflow, the agent is instructed to save the full, unmodified JSON output of the sf org open --json command to a file named org-url-result.json. This output typically includes session-sensitive 'frontdoor' URLs and access tokens. Similarly, the scratch org creation workflow involves writing org details to scratch-org-result.json. Storing these secrets in plain text poses a high risk of unauthorized access or accidental disclosure through version control or shared environments.
[COMMAND_EXECUTION]: The skill relies on executing sf CLI commands via the Bash tool using parameters derived from user input (e.g., <alias>, <name>, <path>, <snapshot-name>). There are no specific instructions provided to validate or sanitize these inputs before they are interpolated into shell commands, which creates a surface for command injection attacks if the agent does not apply its own safety filters during tool invocation.
[PROMPT_INJECTION]: The skill processes external data from scratch org definition files (.json) and incorporates user-provided strings into command arguments without explicit boundary markers or sanitization logic. This creates a surface for indirect prompt injection, where malicious instructions embedded in configuration files or user prompts could potentially influence the agent's behavior, especially given the broad capabilities granted through the Bash tool.

dx-org-manage