The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires and executes multiple command-line tools, including git, ai-handoff, ai-coord, and macOS system utilities pbcopy and pbpaste to manage state and publish handoff commands.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from repository instructions, task surfaces, and working-tree states to generate handoff files for subsequent agent sessions. Malicious content within a repository could be propagated into the handoff instructions without adequate isolation.
Ingestion points: Ingests content from repository-level instructions, task-related files, and the current Git working-tree state.
Boundary markers: Does not specify the use of delimiters or clear separation between the ingested untrusted data and the generated instructions for the receiving agent.
Capability inventory: Possesses local file-write access, CLI tool execution capabilities, and access to the system clipboard.
Sanitization: The instructions do not include specific steps to sanitize or escape repository content before interpolating it into the handoff draft.
[DYNAMIC_EXECUTION]: The skill resolves local directory paths for other skills (e.g., ~/.agents/skills/codex-handoff) and embeds these computed absolute paths into handoff files. It explicitly instructs the receiving session to load and execute the skill from that computed path during its initialization.

task-handoff