cas

Claude Agent System

Research deeply. Delegate precisely. Parallelize deliberately. Keep sensitive actions gated.

Claude Agent System (CAS) is a focused Claude Code plugin built around four workflows that are useful in real sessions:

Core capability	Command	Purpose
Deep research	/spectre	Parallel research, synthesis, and independent claim validation
Codex delegation	/gpt-architect	Claude keeps architectural control while a normal Codex MCP handles focused repository work
Explicit parallel execution	/hydra	Deterministic Claude Agent Teams for several coordinated tasks—without Dynamic Workflows
Safety perimeter	/setup-hooks	Verifies secret protection and approval gates for removal, commits, and pushes

The rest of the plugin remains available as an optional toolbox, but these four capabilities define the recommended CAS experience.

Install

Run these commands inside Claude Code:

/plugin marketplace add Kasempiternal/Claude-Agent-System
/plugin install cas

After an upgrade, start a new Claude Code session so the current skill and hook definitions are loaded.

Three execution modes

CAS keeps delegation modes explicit. They are tools for different jobs, not layers to stack indiscriminately.

1. Direct Claude

This is the default. Claude reads, edits, runs tools, and talks to you directly. Use it for normal work that does not benefit from delegation.

2. GPT Architect

Use /gpt-architect when Claude should remain the architect while Codex performs a focused implementation, diagnosis, or review.

/gpt-architect investigate the failing synchronization tests and implement a verified fix
/gpt-architect on
/gpt-architect off

• One-shot form delegates only the supplied task.
• on makes Codex the exclusive delegation backend for the current conversation.
• off restores native Claude teammates immediately.
• New sessions always start OFF; no marker files, background workers, or global mode are created.
• Immediately before delegation, Claude shows Codex → <model>/<effort>: <task> so you can see what is running and why.

GPT Architect uses the separately configured normal Codex MCP. CAS does not install a wrapper service, approval window, watcher, fleet, or quota daemon.

3. Native Claude Agent Teams

Use /spectre or /hydra when the task genuinely benefits from coordinated native Claude teammates.

Agent Teams and GPT Architect are intentionally exclusive. If GPT-A ON is visible in your session, run:

/gpt-architect off

Then invoke /spectre or /hydra. Both skills stop before launching teammates if GPT Architect is still ON, which prevents confusing hook failures and accidental mixed orchestration.

GPT Architect routing

Claude selects both a model and an effort level for every Codex call. The goal is the lowest-cost route that still matches the task’s risk and ambiguity.

Model	Default lane
gpt-5.3-codex-spark	Tiny, near-instant, low-risk edits on an already known surface
gpt-5.6-luna	Mechanical transformations, extraction, classification, and exact structured work
gpt-5.6-terra	Everyday implementation, debugging, tests, and focused code review
gpt-5.6-sol	Architecture, security, concurrency, performance, or ambiguous cross-system work

Effort	Meaning
low	Obvious and primarily mechanical
medium	Bounded work requiring ordinary planning and validation
high	Behavioral work with meaningful edge cases or tradeoffs
xhigh	Difficult work with interacting invariants or substantial uncertainty
max	The hardest single coherent problem
ultra	Explicitly authorized parallel delegation for genuinely independent branches

Typical defaults are Spark low or medium, Luna low or medium, Terra medium or high, and Sol high or xhigh. max is exceptional. ultra is a parallelism mode, not a generic quality switch.

Spark’s fast lane

Spark is deliberately narrow. It is appropriate only when the file or tiny surface is already known, the expected behavior is unambiguous, the change is low risk, and a focused check can validate it.

Do not use Spark for repository discovery, broad diagnosis, synchronization, networking reliability, authentication, permissions, security, privacy, persistence, migrations, dependencies, concurrency, or release behavior. Those go directly to Terra or Sol.

After Spark writes executable code, Claude inspects the diff and validation, then normally requests a separate read-only Terra medium review. A substantive review finding is corrected by Terra or Sol—not by repeatedly pushing Spark beyond its lane.

Safety contract

CAS protects a narrow set of high-consequence actions without interrupting ordinary development.