The Agent Skills Directory

[OBFUSCATION]: The instructions direct the agent to locate and decode Base64-encoded payloads hidden within HTML comments inside tool responses (e.g., ). While used for 'transport integrity', this pattern involves acting on hidden, non-human-readable data.
[DYNAMIC_EXECUTION]: The skill uses the decoded Base64 payloads as configuration contracts to spawn and control independent sub-agents (using spawn_agent, OMP Task, or Task/Agent primitives) which operate autonomously to monitor jobs.
[COMMAND_EXECUTION]: The skill executes shell commands including ouroboros tui open for launching a dashboard and git commands (e.g., git checkout -b) for automatic feature branch management.
[DATA_EXFILTRATION]: The agent is instructed to read local configuration files, specifically ~/.ouroboros/config.yaml and project-level CLAUDE.md, to determine behavior and policy settings.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data (YAML seed content and tool output payloads) and possesses high-privilege capabilities such as shell execution and sub-agent spawning, creating an attack surface for instructions embedded in processed data.
Ingestion points: User-provided YAML seeds, tool output comments (job_observer), and local configuration files (CLAUDE.md, ~/.ouroboros/config.yaml).
Boundary markers: The skill requires 'canonical v1 validation' for the Base64 payload, which acts as a structural boundary, but no explicit boundary markers are defined for the primary YAML seed input.
Capability inventory: Sub-agent spawning, shell command execution, git workflow automation, and persistent background job monitoring.
Sanitization: Employs validation logic for the observer contract payload but does not mention sanitization for the main seed specification content.

ouroboros-run