ios-development-agents

iOS Development Framework

Eight iOS specialties, a StoreKit skill and fifteen workflows for Claude Code and Codex. Swift, SwiftUI/UIKit UI, architecture/persistence, platform integration, testing, performance, security and delivery.

Version 6.1.0 keeps specialist knowledge in portable skills and generates client-specific agents from the same source. Models inherit the parent session; Context7 and delegated agents are optional. The framework follows the target project's actual compiler, SDK, deployment target and architecture.

Codex

Requires Python 3.11+ for the installer and a current local Codex client with skills support. From a checkout of this repository:

python3 scripts/framework.py install-codex --project "/path/to/ios-app"

This copies ten skill packages into the app's .agents/skills/. It preserves existing project instructions/configuration and refuses conflicting files before writing. Identical installations are safe to repeat. To update a customized or older installation, compare the conflicting files before replacing them.

In that app's Codex session:

$ios-swiftui Build an accessible profile screen using this project's patterns.
$ios-swift Fix the actor-isolation errors in UserService.swift.
$ios-platform Review background refresh cancellation and notification routing.
$ios-storekit Fix entitlement restoration after subscription renewal.
$ios-workflow analyze Sources/ --focus security
$ios-workflow migrate coredata swiftdata --strategy incremental --backup

Skills can also activate automatically from a matching request. For optional custom delegated agents, install with --with-agents:

python3 scripts/framework.py install-codex --project "/path/to/ios-app" --with-agents

This also copies eight standalone agent TOML files into .codex/agents/ for current Codex releases. Ask explicitly to use a specialist or parallel agents. Without this option, the skills work directly in the main session. /ios:* names are Claude commands; they are not registered Codex slash commands.

Claude Code

/plugin marketplace add CarolaneLFBV/ios-development-agents
/plugin install ios@ios-development

/ios:implement "profile screen with editable avatar" --with-tests
/ios:analyze Sources/ --review
/ios:test UserService --generate

For local development, start Claude Code from the repository with claude --plugin-dir .. Plugin agents remain named swift-specialist, swiftui-specialist, architecture-specialist, performance-specialist, testing-specialist, security-specialist, devops-specialist, and platform-specialist. Commands work directly unless delegation is requested; Claude may also select an agent from its description. The existing advisory Swift hook is Claude-only.

Shared repository instructions

AGENTS.md is the single source for maintaining this repository. Codex reads it as project guidance. Claude Code can read it directly from v2.1.277, depending on its Project instructions setting and other CLAUDE.md/CLAUDE.local.md files. The root CLAUDE.md contains only @AGENTS.md for sessions that need the import; there are no separate legacy core/workflow instruction files under .claude/. These repository instructions are not installed into an app by the Codex installer. Codex instructions, Claude instruction loading.

Workflows

Options express workflow intent, not a compiled CLI parser or model reasoning settings. Preview/audit-only/dry-run modes preserve source. Read-only analysis and release preparation do not post reviews, upload builds or submit releases.

Specialty boundaries

The UI role owns SwiftUI/UIKit screens and their bridges. Platform integration owns lifecycle, background scheduling, notifications, incoming links and extensions. Architecture owns offline data, cache/retry policy and synchronization. Swift owns language and URLSession mechanics, security trust/privacy decisions, and DevOps builds/signing. Purchases and entitlements use ios-storekit; release preparation remains a delivery workflow. Existing roles are retained with these explicit boundaries.

UIKit, data/network and platform references load only for matching work. Generated agents rebase support links to their installed location; keep the corresponding skill packages alongside the agents when distributing them.

Maintenance

python3 scripts/framework.py generate
python3 scripts/framework.py check
python3 -m unittest discover -s tests -v
python3 scripts/framework.py budget

Generation, validation and installation use the Python standard library. Optional budget --tokens uses tiktoken's o200k_base as a text-size proxy and can download its dictionary on first use. It does not measure billed tokens or task effectiveness.

Audit comparisons use docs/audit-baseline.json when it is available locally; without it, the budget command reports current sizes. Local audits and notes in docs/ are ignored and are not required to install or validate the framework.

evals/cases.json contains review tasks and observable criteria with small fixtures. Run each prompt in an isolated copy using its listed skills. Compare framework versions with the same model, settings, inputs and tools; record actual findings, executed checks, wall time and reported token usage under local docs/. Use pass/fail/not-run per criterion and repeat cases before drawing conclusions. Fixture tests and catalog validation alone do not measure model quality.

See the minimal example. MIT: LICENSE.