Skill-Delegate

Inspiration

Coding agents are powerful, but work across different CLI hosts quickly becomes hard to coordinate. Skill-Delegate makes ownership, review, recovery, and cleanup explicit.

What it does

Skill-Delegate routes isolated task and review units between Codex, Claude, Grok, and Agy. File-changing work runs in Git worktrees, durable manifests preserve state, independent reviewers and scorers verify results, and the host inspects and merges the final diff. Approved plans can run as recoverable multi-agent workflows with progress, retry, and failure evidence.

What we built during OpenAI Build Week

The initial core scaffold was created on July 12, 2026. This submission focuses on meaningful extensions built after July 13 with Codex and GPT-5.6: durable session routing and locks, plan-driven develop workflows, automatic parallel DAG orchestration, per-unit difficulty and capacity-aware scheduling, independent scoring and retries, blocked-aware failure handling, safer audit boundaries, adaptive reports, and host-native progress-board mirroring.

How we built it

The product is a standalone Node.js ESM CLI built around Git worktrees, JSON manifests, strict dispatch contracts, and backend adapters. GPT-5.6 supported architecture review, implementation, debugging, regression analysis, and verification. The runtime does not require one specific model; it invokes the CLIs selected by user routes.

Challenges

The hardest problem was preventing silent fallback, duplicate writers, and lost state across hosts. Dirty worktrees, interrupted calls, quota signals, stale sessions, prompt injection, and partial failures all needed explicit durable states. Parallelism also had to respect write-scope conflicts and independent review.

Accomplishments

Skill-Delegate now provides cross-CLI routing, isolated tasks, independent reviews and scorers, recoverable plan execution, explicit capacity feedback, bounded cleanup, and evidence-rich results.

What we learned

Trustworthy orchestration is not about launching the most agents. It is about clear ownership, recovery, and proof at every transition.

What's next

We will simplify installation, expand fresh-host compatibility testing, improve capacity-aware scheduling, and make recovery status easier to understand.

Built With

• gpt-5.6