The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various shell commands through Bash to manage git operations (rebases, worktrees, commits), project builds, and testing. These are restricted to the project's repository context and standard development tools like git and gh.- [DATA_EXFILTRATION]: During the 'Self-Retrospective' phase, the skill reads Claude session logs from the home directory (~/.claude/projects/) to identify friction points. Findings derived from these logs are sanitized to remove project-specific identifiers and can be posted to a user-configured GitHub repository or local directory, subject to a user approval gate.- [EXTERNAL_DOWNLOADS]: The skill's HTML artifacts and plan review viewer reference external frontend libraries (marked, highlight.js, mermaid) from Cloudflare's CDN. These references are documented and include Subresource Integrity (SRI) hashes to prevent unauthorized code modifications.- [INDIRECT_PROMPT_INJECTION]: The workflow involves processing external data such as the codebase, PR comments, and session transcripts. This creates an attack surface where instructions embedded in project files could influence the agent's behavior during planning or review phases. The skill includes some sanitization and formatting logic to mitigate these risks.

dev-workflow