The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a local Python script (scripts/tree.py) to manage the state of the optimization tree. It also orchestrates the execution of user-defined evaluation commands (e.g., python eval.py) to score candidates.
[EXTERNAL_DOWNLOADS]: Reference documentation (references/arbor-upstream.md) provides instructions for cloning and installing the original Arbor tool from the RUC-NLPIR GitHub organization. This is a reference for an upstream research implementation.
[INDIRECT_PROMPT_INJECTION]: The skill implements a multi-agent architecture where a coordinator processes findings generated by subagents, creating a surface for indirect prompt injection.
Ingestion points: The coordinator reads the Observe projection generated from .arbor/tree.json, which includes results and insights returned by subagent executors.
Boundary markers: Evidence is returned in a structured format (dev_score, result, insight, branch_ref) to the coordinator.
Capability inventory: Subprocess execution via Bash, subagent management via the Agent tool, and filesystem access via the Edit and Read tools.
Sanitization: Results are distilled and encapsulated within the tree.py state manager, and experiments are performed in isolated git worktrees to prevent side effects.

arbor