orchestrate
Pass
Audited by Gen Agent Trust Hub on Jul 5, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses a Python harness script (
harness/orchestrate_run.py) to execute shell commands. These commands are either provided as arguments or retrieved from a local configuration file (executors.local.md). The script allows shell features such as pipes and redirects to support complex CLI workflows for tools likecodexandopencode. - [DATA_EXFILTRATION]: The skill's intended purpose involves sending portions of the codebase to third-party model providers. This inherent risk is mitigated by explicit instructions and guardrails within
SKILL.mdthat command the agent to exclude secrets,.envfiles, API keys, and PII from the data sent to external executors. - [EXTERNAL_DOWNLOADS]: The documentation guides users to install external CLI utilities (
codex,opencode) from official package registries (NPM and Bun). These tools are necessary for the skill's delegation functionality. - [SAFE]: Several robust security controls are integrated into the skill's design:
- Workspace Isolation: The skill mandates the use of isolated git worktrees for any agentic execution, ensuring that external modifications land in a temporary directory for review rather than directly on the user's active branch.
- Verification Protocol: The skill explicitly instructs the agent to never trust the self-reported success of an external executor. Instead, the agent is required to re-run local verification tools (e.g.,
typecheck,lint,test) on the generated code. - Secure Parsing: The Python harness utilizes
yaml.safe_loadfor parsing configuration and brief metadata, preventing potential code execution vulnerabilities during YAML deserialization. - Data Boundary Enforcement: The instructions include specific checks to ensure sensitive files are gitignored and that database credentials are never shared with external models.
Audit Metadata