skills.rest
seal-harness
seal-harnessOFFICIAL
Explore all skills by this author →
GitHub

external-tools

Delegate implementation and review tasks to external AI CLI tools with cross-model adversarial review.

1|Updated Jun 23, 2026
Search Tags:#git worktree#ai orchestration#codex cli#gemini cli#cost control#code delegation#cross-model review
One-click install
npx skills add https://github.com/seal-harness/seal-harness --skill external-tools-seal-harness
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: external-tools
Source: https://github.com/seal-harness/seal-harness/tree/main/.agents/skills/external-tools
Command: npx skills add https://github.com/seal-harness/seal-harness --skill external-tools-seal-harness

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Single-model coding workflows suffer from blind spots where the same model writes and reviews its own code, and premium model usage drives up costs. This Skill delegates implementation and review tasks to external AI CLI tools (OpenAI Codex CLI, Google Gemini CLI) so that code is always reviewed by a different model, while cheaper models handle routine implementation work. ## Core Features & Use Cases - Cross-Model Adversarial Review: The writer model is always reviewed by a different model (Codex, Gemini, or Claude), with a structured rubric producing PASS/FAIL verdicts and BLOCKING/WARNING findings. - Availability-Aware Escalation: Health checks run per task dispatch, and the orchestrator escalates through a chain (external tool A, external tool B, Claude, user) with retry budgets of up to 5 attempts. - Sandboxed, Budget-Controlled Execution: Each invocation runs in an isolated git worktree with a minimal environment (env -i), timeout enforcement, scope verification, and per-task ($2) and per-session ($20) cost circuit breakers. - Use Case: A team wants to cut LLM costs on routine TypeScript feature work. The orchestrator routes implementation to the cheapest healthy external tool, validates with its own test suite, has a different model review the diff, and only merges after all gates pass. ## Quick Start Ask the agent to delegate the next implementation task to Codex and have Gemini review the resulting diff using the cross-model review rubric.

Frequently Asked Questions about external-tools

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I delegate coding tasks to Codex or Gemini CLI from an AI agent?▼

Run the adapter scripts with the implement command, passing a git worktree path and a self-contained prompt file. The adapter invokes the external CLI in full-auto mode, commits changes on an isolated branch, and returns a JSON envelope with exit code, files changed, and cost.

What is cross-model adversarial code review?▼

Cross-model review means code written by one AI model is always reviewed by a different model, such as Gemini reviewing Codex output. The reviewer receives only the diff, spec, and rubric, and returns PASS or FAIL with BLOCKING or WARNING findings cited by file and line.

Does this work if only one external AI CLI tool is installed?▼

Yes, the escalation model adapts to availability. With one tool it runs a reduced chain with mutual review between that tool and Claude, and with no tools available the skill is a no-op and standard behavior continues unchanged.

How are external tool costs controlled per task?▼

Two circuit breakers apply: a per-task budget defaulting to $2.00 and a per-session budget defaulting to $20.00. When a limit would be exceeded, the adapter is skipped and the orchestrator escalates or falls back to Claude for remaining work.

Why must external tools run in a git worktree instead of the main repo?▼

Worktree isolation prevents concurrent invocations from corrupting each other's state and provides a clean rollback path for failed attempts. Changes merge into the working branch only after validation and cross-model review pass.

What happens when an external AI CLI times out or hits rate limits?▼

Errors are classified into types like timeout, rate_limited, auth_expired, and network_error. Transient failures get backoff and retry, auth errors escalate to the user, and unrecoverable errors skip the adapter to the next tool in the escalation chain.

Related Skills

Explore More Software Engineering Skills→
rohanbhatia27rohanbhatia27
1

second-opinion

Run external LLM code reviews on uncommitted changes, branch diffs, or commits.

Community
Intermediate
TyrizxTyrizx

merge

Merge a topic branch into the base branch and resolve conflicts in the main worktree.

Official
Intermediate
Clouder0Clouder0

self-test

Automates iterative self-testing of code via unit tests, lint, and type checks.

Community
Basic
shichiyoushichiyou

ollama-launch

Configure environment variables and startup scripts to route CLI coding agents through a local Ollama server.

Community
Intermediate