# Orchestrate External AI Coding Agents

Managing several AI coding CLIs requires provider routing, permissions, isolation, and result collection. This guide defines a repeatable Claude Code workflow for coordinated agent execution.

## Install

```bash
npx skillstore add consiliency/orchestration-native-invoke
```

## Metadata

- Status: approved
- Slug: consiliency-orchestration-native-invoke
- Skillstore revision: r1
- Version status: missing
- Tree hash: 633cb4ef5540cffcf9201024f7d8dafb91bc8a6a14b9e424c6ee54c7b8852530
- Author: Consiliency
- GitHub username: Consiliency
- License: MIT
- Repository: https://github.com/Consiliency/treesitter-chunker/tree/main/.ai-dev-kit/skills/orchestration/native-invoke
- Ref: 635f69fb8d2f4e6330ba47a4e5a0fb239c04d110
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: confirmation\_required
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, filesystem
- Quality score: 38
- Quality tier: warning
- Public page: https://skillstore.pages.dev/skills/consiliency-orchestration-native-invoke
- Manifest: https://skillstore.pages.dev/api/skills/consiliency-orchestration-native-invoke/manifest

## Capabilities

- Selects a provider from the requested agent and enabled provider flags.
- Documents background Task invocation for Codex, Gemini, and Cursor.
- Distinguishes read-only and write command modes for each provider.
- Runs providers in parallel and collects results with TaskOutput.
- Recovers authentication through an interactive terminal before retrying.
- Recommends git worktrees and cleanup steps for write isolation.

## Use Cases

- Compare Code Reviews: Run read-only reviews with different providers, then collect their findings for comparison.
- Coordinate Architecture Analysis: Route complex design questions to selected providers and gather independent recommendations.
- Standardize Agent Operations: Define provider permissions, isolation, authentication recovery, and cleanup for team workflows.

## Prompt Templates

### Run a Read-Only Review

```
Use Codex to review [scope]. Check current help first. Keep the sandbox read-only. Do not modify files. Return findings in the response.
```

### Route a Coding Task

```
Choose the best enabled provider for [task]. Explain the selection briefly. Use read-only mode unless changes are explicitly required.
```

### Compare Parallel Analyses

```
Run Codex and Gemini in parallel on [question]. Keep both read-only. Collect both results, identify agreements, and explain important differences.
```

### Isolate an Advanced Edit

```
Use [provider] to implement [change] in a securely created worktree. Review the diff. Do not merge until I approve. Remove only recorded temporary files.
```

## Limitations

- This documentation skill does not include an executable orchestration script.
- Each provider CLI must be installed, authenticated, and permitted before use.
- Several referenced provider cookbooks are outside the included skill files.
- Model names and CLI flags can change, so users must check current help output.

## Best Practices

- Keep read-only mode enabled unless the user explicitly requests file changes.
- Check each installed CLI help output before selecting flags or models.
- Use securely created worktrees and review every diff before merging.

## Anti Patterns

- Do not enable unattended write flags for an unreviewed prompt.
- Do not insert user text directly into a shell command string.
- Do not delete files by wildcard without verifying their provenance.

## Security Audit

- Audited at: 2026-07-19T04:03:35.116\+00:00
- Summary: Most static alerts are false positives caused by Markdown formatting, fixed cookbook references, or controlled worktree paths. Confirmed risks include autonomous external CLI modes, broad wildcard command permissions, cleanup by filename patterns, and prompt sharing with third-party providers. Publication should require narrower permissions, explicit data consent, read-only defaults, and provenance-based cleanup.

## Stats

- Views: 294
- Downloads: 8
- Favorites: 0
- Popularity score: 0