The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/panel_round.py utilizes subprocess.run to execute local programs and scripts. The command paths and arguments are dynamically resolved from the assets/routing.toml configuration file, which maps different roles to specific provider scripts (e.g., .agents/skills/claude-runner/scripts/run_claude.py).
[PROMPT_INJECTION]: The orchestration logic in scripts/panel_round.py is vulnerable to indirect prompt injection. It reads the contents of external files provided via the --context-file argument and interpolates this data directly into the prompts for sub-agents with minimal boundary separation.
Ingestion points: Files read from the file system and passed as context to the panel_round.py script via the --context-file CLI argument.
Boundary markers: Uses simple Markdown headers (e.g., ## Context file: {path}) and triple-dash separators, which are insufficient to reliably prevent models from following instructions embedded within the context data.
Capability inventory: The skill has the ability to execute subprocesses and write logs/artifacts to the .codex_workflow/delivery directory.
Sanitization: There is no evidence of escaping, filtering, or validation performed on the ingested file content before it is processed by the AI models.

collaborative-delivery