alterlab-workflow-orchestration
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references the official
claude-agent-sdkPython package and the@anthropic-ai/claude-agent-sdkNode.js package. These are standard tools for agent orchestration provided by the platform vendor. - [COMMAND_EXECUTION]: Mentions the use of documented environment variables such as
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMSandCLAUDE_CODE_FORK_SUBAGENTto enable platform-specific orchestration features as described in official documentation. - [INDIRECT_PROMPT_INJECTION]: The skill documents patterns for agent delegation and context management. Ingestion points include summaries passed between stages in a pipeline (e.g., in
references/composition-recipes.md). It encourages the use of boundary markers such as turn caps and explicit stop conditions. Its "Adversarial Verification" pattern is a defensive design that uses independent subagents to validate claims, reducing the impact of potentially malicious instructions in processed research data. - [DYNAMIC_EXECUTION]: Includes example implementation code for the Claude Agent SDK. The provided scripts demonstrate standard API usage for defining and querying subagents through the
agentsandqueryparameters without incorporating untrusted inputs or malicious assembly patterns.
Audit Metadata