The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a recovery mechanism via scripts/compact_rehydration.py that reads workflow state from the local file system and injects instructions into the agent's session context.
Ingestion points: The script scripts/compact_rehydration.py reads control.md and child state files within the .agents/workflows/codex-thread-orchestrator/ directory structure.
Boundary markers: The script does not wrap the injected context in specific delimiters, although it instructs the agent that the file is a "durable source of truth, not chat history."
Capability inventory: The skill manages thread creation (create_thread), message dispatching (send_message_to_thread), and maintains a local file-based state protocol.
Sanitization: The script uses a regular expression (SAFE_ID) to validate thread_id strings and verifies the existence of a parent marker file before trusting a directory as a valid workflow root.

codex-thread-orchestrator