The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs extensive local shell command execution for reconnaissance and safety. This includes using git for checkpointing (git status, git checkout), grep and glob for file discovery, and wc -l for load balancing.
Evidence: It explicitly identifies and executes the project's native build command (e.g., npm run build or make) to verify changes in Step 6.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local codebase and passes it to specialized sub-agents. This creates a vulnerability surface where malicious instructions embedded in the codebase (e.g., in comments or documentation) could influence the behavior of the agents.
Ingestion points: Codebase files read during Step 3 (Recon) and Step 5 (Deploy).
Boundary markers: The skill uses structured prompt templates for Layer 1 and Layer 2 agents to define objectives, though it lacks explicit sanitization for file content interpolated into these prompts.
Capability inventory: Agents possess file read/write access, shell execution capabilities for build/test tools, and the ability to spawn further agents.
Sanitization: No specific sanitization or escaping of file content is described before passing it to sub-agents.
[DYNAMIC_EXECUTION]: The framework dynamically generates complex prompts and orchestration logic at runtime based on the structure and content of the codebase being analyzed.
Evidence: The skill uses 'Power Levels' to dynamically assign different LLM models (Opus, Sonnet, Haiku) to different layers of the execution tree.

agent-army