The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes handoff documents stored in the .agents/handoffs/ directory to resume tasks, creating a potential surface for indirect prompt injection. However, the skill provides an extensive 'Author Authority' framework in REFERENCE.md that instructs the agent to verify the author's identity and permissions (e.g., via GitHub/GitLab APIs) before acting on instructions. It explicitly mandates halting and seeking human intervention if unauthorized content appears to contain instructions.
Ingestion points: Documents in .agents/handoffs/ (e.g., 0001-refactor-auth-layer.md).
Boundary markers: The skill uses a fixed markdown template (templates/handoff.md) with specific headers like ## Goal, ## Context, and ## Next steps to structure data.
Capability inventory: Access to Read, Write, Edit, Grep, Glob, and Bash (for git operations and jq processing).
Sanitization: Implements a multi-forge authorization check and strictly forbids the inclusion of secrets or signatures in the document body.
[COMMAND_EXECUTION]: The skill executes a bundled shell script (templates/resolve-config.sh) to parse configuration from .tituskirch-skills.json. It also utilizes standard git commands (git log, git rev-parse, git fetch, git push) to manage handoff state within the repository. These operations are consistent with the skill's stated purpose of session persistence and state management.
[DATA_EXFILTRATION]: The skill is designed to push work-in-progress notes to remote repositories. To prevent accidental data exposure, the skill includes a 'Secret-free' guardrail that explicitly warns agents not to include tokens, API keys, credentials, or environment variables in the handoff documents.

handoff