What problem does it solve?
Delegates implementation and review tasks to external AI CLI tools to reduce cost, eliminate single-model blind spots, and maintain strict safety and validation by having the orchestrator independently verify results.
Core Features & Use Cases
- One-job-per-invocation: external adapters either implement or review, and never self-certify results.
- Cross-model adversarial review: writer and reviewer are different models (Codex, Gemini, Claude) to catch blind spots.
- Safe sandboxing & minimal permissions: isolated git worktrees, env -i minimal environment, and scoped sandboxes for executions.
- Budgeting & escalation: per-task and per-session circuit breakers with availability-aware escalation chains and health checks.
- Use Case: offload small feature implementations or remediation branches to cheaper external models, then run automated CI gates and cross-model reviews before merging.
Quick Start
Use the external-tools skill to run an external model implementer on an isolated git worktree with a local spec, perform cross-model adversarial review, and return a reviewed branch if all gates pass.