The Agent Skills Directory

[COMMAND_EXECUTION]: The orchestrator agent is required to execute shell commands to verify the outputs of subagents. This includes fixed utility commands like date and git log, as well as arbitrary verification commands defined within the graph's metadata.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from a target repository to define its execution logic.
Ingestion points: A 'Planner' subagent reads the repository and user goals to generate the work graph.
Boundary markers: The skill does not provide specific instructions or delimiters to prevent the subagent from being influenced by instructions embedded in the source code it analyzes.
Capability inventory: The orchestrator has the capability to execute shell commands and write to the ~/.harness/workgraph/ directory.
Sanitization: There is no defined process to sanitize or validate the verification commands generated by the subagent before they are executed by the main agent.
[DYNAMIC_EXECUTION]: The skill utilizes a dynamic execution model where shell commands are generated at runtime by a subagent and then executed by the orchestrator to perform gating checks.
[EXTERNAL_DOWNLOADS]: The installation instructions in the README use npx to fetch the skill from the author's av/skills repository.

workgraph