The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill executes an embedded multi-line Python script (REMOTE_SNIPPET) on remote machines using python3 -c via the fleet-manager tool. This script performs file system operations like writing and deleting JSON state files.
[DYNAMIC_EXECUTION]: scripts/agents.py uses ctypes.CDLL() to dynamically load shared libraries at runtime, which allows the execution of arbitrary code from external binaries.
[DYNAMIC_EXECUTION]: scripts/agents_copy_channel.py dynamically modifies sys.path to import fleet_manager and fleet_remote from a location computed at runtime relative to its own path.
[COMMAND_EXECUTION]: The coordination script scripts/agents.py executes shell commands using subprocess.run(), which is used to interface with the local file system, git, and other orchestration tools.
[INDIRECT_PROMPT_INJECTION]: The coordinator agent ingests and parses reports and status updates from various sub-agents (threads). This creates a surface where a compromised or adversarial sub-agent could attempt to inject instructions into the coordinator's context via its report files or status lines.
Ingestion points: Reads threads/<id>/report.md, inbox/new/ events, and PR metadata.
Boundary markers: Uses brief templates to define sub-agent scope and specific line prefixes (PR:, STATUS:, BLOCKED(HUMAN):) for parsing, but largely relies on the sub-agent's compliance.
Capability inventory: Includes file system writes, local subprocess execution, and remote execution via the fleet-manager sibling skill.
Sanitization: Employs init - to prevent injection from the user's initial goal, but sub-agent reports are parsed using line-based patterns without comprehensive sanitization.
[PERSISTENCE]: The skill implements a persistent monitoring mechanism through library/wake.sh, which is scheduled to run every 30 seconds to reconcile project state and notify the coordinator of updates across turns.

agents