The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture involves ingesting output from external CLI agents, which creates a potential surface for indirect prompt injection.
Ingestion points: Commands such as paneflow read, paneflow status, and paneflow search are used to pull scrollback and agent messages into the conductor's context (e.g., in SKILL.md sections 2 and 3).
Boundary markers: The instructions specify that paneflow read automatically wraps scrollback in <untrusted_terminal_output> tags, providing a clear boundary between the agent's instructions and external data.
Capability inventory: The conductor agent possesses the ability to execute shell commands and dispatch prompts to other panes via paneflow send, including the use of --submit for auto-execution.
Sanitization: The skill contains explicit instructions in Section 5 ('The discipline') directing the agent to treat all peer output as untrusted data and to ignore any commands or instruction-override attempts (e.g., 'ignore your previous instructions') found within that data.
[COMMAND_EXECUTION]: The skill uses the paneflow CLI and standard shell utilities to manage the agent fleet and temporary files.
Context: Uses paneflow ps, paneflow status, paneflow up, and paneflow flow run for orchestration. It also employs mktemp, cat, jq, and rm for handling report files, which are appropriate and standard practices for the skill's defined purpose.

paneflow-conductor