The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill solicits and handles sensitive information including environment credentials (email, password), auth tokens, and TOTP seeds for CI bypass. While it follows security best practices by recommending storage in .env files and requiring opt-in approval before pushing secrets to GitHub via the gh CLI, the active solicitation of these secrets represents a risk surface.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external OpenAPI specifications and project context files to generate test code and infrastructure configuration.
Ingestion points: Reads markdown files from the .context/ directory, project configuration files, and remote OpenAPI spec URLs in references/adaptation-workflow.md.
Boundary markers: Implements a mandatory 9-phase workflow. Phases 0-2 (analysis and planning) are write-protected. Implementation (Phases 3-9) only begins after the user provides explicit approval of a generated plan file (.context/reports/adapt-framework-plan.md).
Capability inventory: Performs file writes across the repository, executes local scripts via bun run, and manages GitHub Secrets via the gh CLI.
Sanitization: The skill relies on its planning phase and user review of the generated plan to mitigate risks from untrusted data ingestion.
[EXTERNAL_DOWNLOADS]: The skill syncs API specifications from remote URLs or GitHub repositories using bun run api:sync in Phase 4.

adapt-framework