The Agent Skills Directory

[COMMAND_EXECUTION]: The skill defines a platform hook in hooks/user-hooks.json that executes a local Python script (scripts/compact_rehydration.py) during session start events. This script is responsible for identifying and injecting the correct orchestrator state path into the agent's reasoning context.
[INDIRECT_PROMPT_INJECTION]: The rehydration script parses control.md files within the project's .agents/workflows directory to locate state information. It then generates a context injection that instructs the agent to treat these files as the "durable source of truth."
Ingestion points: The script reads control.md files located in project-specific workflow directories (SKILL.md, scripts/compact_rehydration.py).
Boundary markers: The injected message explicitly defines the file as the source of truth to prevent the agent from relying on potentially stale chat history.
Capability inventory: The skill utilizes thread management tools (create_thread, send_message_to_thread, set_thread_archived) to coordinate tasks.
Sanitization: The scripts/compact_rehydration.py script employs a strict regular expression (^[A-Za-z0-9][A-Za-z0-9._-]*$) to validate the session_id before using it in file path operations, effectively preventing path traversal attempts.

codex-thread-orchestrator