The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow that ingests and reconciles data from an external agent (Codex CLI). While this creates a surface for indirect instructions, the skill implements comprehensive defensive measures.
Ingestion points: The skill processes diffs, commit messages, and status reports generated by the Codex lane (managed via the git and codex CLIs).
Boundary markers: Isolation is enforced through git worktrees (e.g., /tmp/${SAFE_TASK}-codex-lane) and specific branch naming conventions to separate untrusted edits from the main repository.
Capability inventory: The primary agent uses terminal and process tools to execute git commands and the Codex CLI, and manages the lifecycle of the Kanban board.
Sanitization: The skill mandates a manual reconciliation checklist, requiring the primary agent to review all changes (git diff), verify safety invariants (especially for prediction-market bots), and independently re-run all canonical tests before accepting any contribution.
[COMMAND_EXECUTION]: The skill uses shell commands to manage isolated workspaces and execute the secondary agent tool.
Evidence: Shell snippets include git worktree add, command -v codex, and codex exec --full-auto. These commands are used for operational isolation and to verify the presence of required local tooling.

kanban-codex-lane