Delegate Skills: Multi-Agent Coding Orchestration & Review
Orchestrate multiple coding agent CLIs—delegate by lane or directly, while you keep the review and the commit.
Evidence shows: README explicitly states relays never commit, committing is the reviewer's job; each implementer skill declares default permissions and read-only modes; delegate-setup requires user approval before writing config; scripts use only Node built-ins, no network calls, no credential reads, no telemetry; dependency security: GitHub Actions pins versions, but no dependency manifest or vulnerability scan evidence. External effects: relays launch implementer CLIs and git, may modify files, but protected by read-only modes and tripwires. Rollback: no automatic rollback mechanism, but diff reviewable. Source attribution: publisher identity unverified, but README provides author info. Deductions: rollback and source attribution evidence thin, publisher unverified.
Evidence shows: README describes unified relay contract and result.json structure, multiple skills claim verification, but static review cannot confirm actual run results. Dependency availability: depends on Node 18+ and git, and external CLIs, but no version pinning or installation guarantees. Failure messages: README mentions exit codes and error handling, but no concrete examples. Deductions: static review cannot verify runtime reliability, external CLI availability not guaranteed.
Evidence shows: README targets orchestrating agents using Skills CLI, provides multiple scenarios (direct delegation, fleet, etc.), and states capability boundaries (e.g., which CLIs support read-only). Trigger precision: skill names clear, but no detailed trigger conditions. Environment fit: supports multiple OS and CLIs, but no full compatibility matrix. Deductions: trigger precision and environment fit details insufficient.
Evidence shows: README provides clear install instructions, examples, and known limitations (e.g., some CLIs lack read-only). Information architecture good, skill files organized. License MIT, but version changelog missing. Maintenance responsibility: no clear maintainer or update policy. Deductions: version changelog and maintenance responsibility evidence thin.
Evidence shows: Output is structured result.json and diff, easy to review. Marginal value in delegating tasks to specialized CLIs while retaining review and commit. Cost-benefit: requires installing multiple CLIs, but reusable. Deductions: static review cannot assess actual efficiency gains.
Evidence shows: README claims align with code structure, but no independent third-party verification. Fact-inference separation: README distinguishes verified status from unverified parts. Deductions: cross-source corroboration insufficient, relies on single repository.
- Publisher identity unverified, treat with caution.
- Static review cannot verify actual runtime behavior; relies on README claims.
- Some implementer CLIs lack read-only mode; be mindful of usage scenarios.
- Version changelog missing, making updates hard to track.
What does this agent do, and when should you use it?
Delegate Skills is a collection of agent skills that let an orchestrating agent (such as Claude Code) delegate coding tasks to separate coding-agent CLIs like Codex, Cursor, Claude Code, and others, while preserving review and commit authority with the orchestrator. The repository includes 18 *-delegate skills, each targeting a specific coding CLI with a unified relay script (`relay.mjs`) that safely launches the CLI, polls for completion, and produces a structured `result.json`. A `delegate-setup` skill discovers installed CLIs and proposes named 'lanes' (e.g., feature, tests, ui) for approval before writing configuration. All relays never commit, make no network calls, are free of telemetry, and are designed to be auditable. The package is distributed via skills.sh and works with the Skills CLI.
This repo provides standardized delegation skills for coding-agent CLIs: aider, agy, claude-code, cline, codex, commandcode, cursor, grok, kimi, opencode, pi, omp, qoder, vibe, copilot, warp, and zcode. Each skill includes a SKILL.md and a relay script (relay.mjs) that constructs arguments, launches the implementer CLI, polls until run completion, and outputs a result.json with status, exit code, touched files, and any session id. delegate-setup uses discover.mjs to probe installed CLIs and proposes a fleet of lanes (e.g., feature, tests, ui) that can be stored globally or per-project after explicit approval. The end-to-end flow is: write a brief -> dispatch via the relay -> wait for result -> review the diff -> run the gates -> commit yourself as reviewer. Relays never commit, never call home, and use only Node built-ins.
- A developer wants Codex to implement a mechanical refactor in a specific directory, then reviews the diff and lands the commit personally.
- A team wants to distribute feature, test, and UI work across different coding CLIs (e.g., OpenCode, Codex, Cursor) under a single orchestrator using fleet lanes.
- A power user wants to safely use multiple delegating skills without writing project-wide config until explicitly approved, thanks to content-bound approval.
- A maintainer is testing a new coding CLI and needs to verify its constraints and relay behavior through documented live runs.
- An engineer wants to run a queue of migration tasks through Codex sequentially, reviewing each one before committing.
What are this agent's strengths and limitations?
- Avoids vendor lock-in by supporting 17+ different coding CLIs (Aider, Codex, Claude Code, Cursor, etc.) with a unified interface.
- Relay scripts are zero-dependency, make no network calls, and are telemetry-free, enhancing auditability and trust.
- A consistent delegation loop (brief -> relay -> result -> review -> commit) reduces learning curve and works across orchestrators.
- Fleet lanes and project configuration with explicit approval provide fine-grained control and fail-closed behavior.
- Documentation provides clear security boundaries for each implementer (e.g., which read-only modes are not enforceable) and verification status, aiding adoption decisions.
- Read-only modes are not enforceable for all CLIs (e.g., Grok can write headlessly); relies on tripwires and careful review.
- Some implementing CLIs have caveats like default auto-commit (Aider) that require relay intervention, and platform constraints (e.g., Claude's sandbox unsupported on native Windows, Vibe targets UNIX).
- Some skills (e.g., opencode, vibe, omp) only have contract tests and no end-to-end verified runs, limiting confidence.
- Practical use depends on installing and authenticating multiple external CLIs, adding configuration and maintenance overhead.
- Users need familiarity with advanced git practices (worktrees, isolation) and must watch for Windows-specific shim issues documented for several CLIs.
How do you install or deploy this agent?
Requires Node.js 18+ and git. Browse first: npx skills add amElnagdy/delegate-skills --list. Install the whole package, the setup skill, or one implementer skill, e.g.: npx skills add amElnagdy/delegate-skills, npx skills add amElnagdy/delegate-skills --skill delegate-setup, or npx skills add amElnagdy/delegate-skills --skill codex-delegate. For a specific agent: npx skills add amElnagdy/delegate-skills --skill codex-delegate --agent claude-code. Global install: npx skills add amElnagdy/delegate-skills --global. You must have the target CLI installed and authenticated (e.g., claude login or codex login); each skill documents its own install and login commands.
How do you use this agent?
After installing a delegate skill, ask your orchestrating agent (e.g., Claude Code) to delegate directly: "Use $claude-delegate to have a separate Claude Code session implement the parser fix, then review and commit it." Alternatively, set up a fleet: "Use $delegate-setup to discover my installed implementer CLIs and create a fleet for feature, tests, and UI work." After approving the proposed config, dispatch via lanes like "Use $opencode-delegate --lane feature to implement the billing workflow". The relay runs the implementer CLI, produces a structured result, and you (or your orchestrator) review the diff and commit. All dispatches require the implementer CLI to be authenticated and runnable.
FAQ
Does delegation auto-commit?
--no-auto-commits to disable it.What do I need to install beforehand?
git. For each *-delegate skill, you must install and authenticate the corresponding CLI (e.g., claude, codex) as you would at the terminal; each skill's README provides its own install/login commands. The delegate-setup skill works without any implementer CLI—it discovers what's already installed.What if an implementer CLI has no read-only mode?
--yolo) or Warp that lack a native read-only mode, the relay cannot enforce read-only. It relies on post-run git status and diff as a review aid, but these cannot catch writes outside the repository or ignored files. It is recommended to run these in an isolated worktree or container.Can I use multiple coding agents simultaneously?
--lane. However, all implementer CLIs must be installed and authenticated on that machine, and each run happens locally.Is project configuration safe?
git diff before using them.