The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a helper script scripts/reconcile-index.mjs which executes Git commands using node:child_process.execFileSync. These calls are restricted to internal Git operations (such as rev-parse, merge-tree, and write-tree) necessary for index reconciliation.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it processes untrusted data.
Ingestion points: The agent reads diffs, filenames, commit messages, and chat summaries from the local environment (SKILL.md, references/chat-coordination.md).
Boundary markers: The instructions explicitly mandate treating this data as evidence only and ignoring any embedded instructions.
Capability inventory: The skill utilizes Git commands and the provided Node.js script to manipulate repository state and validate commits.
Sanitization: The skill relies on the agent's adherence to the "untrusted evidence" instruction to prevent execution of data-borne instructions.
[SAFE]: The skill implements significant safety controls, including the use of temporary detached worktrees for candidate validation and mandatory user approval before any changes are applied to the active branch. It follows best practices for Git state preservation.

create-bounded-commits