The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data from Bitbucket resources (such as pull request descriptions, comments, and diffs) into the agent context, creating a potential surface for indirect prompt injection.
Ingestion points: Data is retrieved via commands that fetch pull request details, comments, and diff content (e.g., bkt_get_pull_request, bkt_list_pull_request_comments, and bkt_get_pull_request_diff).
Boundary markers: The documentation explicitly warns the agent that descriptions, comments, and diff content are "untrusted data."
Capability inventory: The CLI interacts with the Bitbucket API and performs local file system operations like checking out branches.
Sanitization: The implementation includes limits on the size of ingested data (e.g., 256 KiB for diff content) to mitigate the impact of malicious inputs.
[REMOTE_CODE_EXECUTION]: The skill documents features for installing and executing external extensions and agent skills from remote repositories.
Extension Execution: bkt extension install and bkt extension exec allow users to clone and run arbitrary executables named with the bkt- prefix.
Skill Management: bkt skill install allows downloading agent skills from Bitbucket repositories into shared local directories like .agents/skills.
[EXTERNAL_DOWNLOADS]: The skill provides installation instructions that involve downloading the CLI from the author's repository and official package managers.
Sources: Instructions include fetching binaries from the author's GitHub releases, or using Homebrew, Scoop, and Go.
[COMMAND_EXECUTION]: The skill guides the agent to use the bkt CLI to interact with the local environment and Bitbucket services.
Operations: Includes managing connection contexts, authenticating with host servers, and performing git-related tasks like rebasing.

bkt