The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local version control commands (git, sl) and project-specific scripts (scripts/commit-cloud, scripts/checkout) to gather state and perform backups. These operations are limited to the project's environment and are initiated by explicit user invocation.
[DATA_EXFILTRATION]: The skill pushes documented session state to a remote repository (vault). While this involves a network operation, it is the primary intended function of the skill and follows the established committing policy of the project.
[INDIRECT_PROMPT_INJECTION]: The generated handoff files include a 'Resume prompt' intended for future AI sessions. This creates a surface where instructions from one session can influence another.
Ingestion points: Data is read from the local environment via git logs, status, and repository files like templates/README.md.
Boundary markers: The output is structured with markdown headers and the resume prompt is placed within a fenced code block, though no explicit 'ignore instructions' delimiters are used.
Capability inventory: The skill has capabilities to write files, execute subprocesses (git/sl/local scripts), and perform network operations (git push).
Sanitization: The skill includes an explicit instruction to omit secrets and tokens, and the agent is instructed to check claims against live state upon resumption.

handoff