The Agent Skills Directory

[SAFE]: The skill architecture follows best practices for multi-agent coordination, including plan approval phases and explicit authority boundaries for sub-agents.
[COMMAND_EXECUTION]: The skill invokes local shell scripts (scripts/run-codex-handoff.sh and scripts/watch-codex-wave.sh) to manage agent lifecycles. It utilizes the --dangerously-bypass-approvals-and-sandbox flag of the Codex CLI for implementation tasks, which is an expected requirement for its primary function of autonomous code editing within a user-approved scope.
[PROMPT_INJECTION]: The orchestrator uses strict prompt templates for sub-agents to prevent scope creep or redesigns. There are no patterns indicative of attempts to bypass the host agent's safety filters.
[SAFE]: The skill has an indirect prompt injection surface as it processes repository data via sub-agents.
Ingestion points: scripts/run-codex-handoff.sh (stdin) and research agents reading repository files.
Boundary markers: The skill uses 'self-contained, outcome-first' prompts and instructs agents to stay within assigned scopes.
Capability inventory: Implementation agents can edit files and run validation commands via codex exec.
Sanitization: Interaction is governed by structured JSON schemas (result.schema.json).

codex-handoff