Chatgpt 5.6 Optimization

# ChatGPT 5.6 Optimization

Use this prompt to audit and simplify a personal Codex setup for GPT-5.6. Work from the real installation, make only justified changes, and verify the effective result.

## Task

Inspect the current Codex client and CLI configuration, then produce the smallest setup that preserves the user's real workflows while improving model routing, context control, permissions, Skills, plugins, and validation.

Do not stop at recommendations when safe in-scope changes can be applied. Ask before removing user-owned data, weakening a security boundary, publishing externally, or changing an intentional workflow.

## Working Rules

- Respond in the user's language and keep identifiers, paths, commands, model IDs, and errors unchanged.
- Read the effective `AGENTS.md` chain and preserve unrelated working-tree changes.
- Prefer the installed CLI's accepted keys and observed runtime behavior. Use current official OpenAI documentation for model availability, context limits, pricing, and time-sensitive Codex behavior.
- Distinguish documented API behavior from Codex subscription or credit behavior. Do not promise a fixed savings multiplier.
- Keep durable rules on one owning surface; do not duplicate full instruction blocks across personalization, `AGENTS.md`, Skills, and project configuration.

## Sources of Truth

- `~/.codex/AGENTS.md`: concise personal execution preferences.
- Repository and nested `AGENTS.md`: project-specific rules and validation.
- `~/.codex/config.toml`: global model, reasoning, permissions, features, plugins, MCP, hooks, and Agent registration.
- `~/.codex/agents/*.toml`: explicit model, reasoning, permissions, and role instructions for custom Agents.
- Project `.codex/config.toml`: genuine project overrides only.
- Skills: reusable bounded workflows.
- Plugins and connectors: managed capabilities and live external data or actions.

## Workflow

### 1. Establish Current Truth

1. Confirm the current directory, Git root, branch, dirty state, Codex version, and effective `AGENTS.md` files.
2. Inspect the relevant paths that exist:
   - `~/.codex/config.toml`
   - `~/.codex/AGENTS.md`
   - `~/.codex/agents/`
   - `~/.codex/skills/` and `~/.agents/skills/`
   - enabled plugins, connectors, MCP servers, hooks, and trusted projects
   - project `.codex/config.toml` only when it affects the active task
3. Check current official documentation when a model, key, price, feature, or client behavior may have changed.
4. Report conflicting configuration layers before editing them.

### 2. Simplify Skills and Plugins

- Prefer built-in or official capabilities when they fully cover the need.
- Preserve personal Skills that encode distinct, proven workflows.
- Keep one clear owner for each stable responsibility; report overlapping or stale packages before removing them.
- Update personal Skills through their recorded source and installer. Do not overwrite system Skills or managed plugin caches manually.
- Remove plugins through supported commands only after confirming they are disabled or no longer used.
- Never classify configuration, Agents, personal Skills, memories, sessions, credentials, state databases, or user-created plugins as disposable cache.
- Keep connectors only when live private data or external actions are part of the user's workflow.

### 3. Configure GPT-5.6 Routing

First inspect the installed model catalog. If the following models are available, use this personal baseline:

- Interactive execution: `gpt-5.6-luna`, high reasoning.
- Planning and final review: `gpt-5.6-sol`, `xhigh`, read-only.
- Bounded execution Agents: `gpt-5.6-luna`, `xhigh`, with the minimum required permissions.
- Main Agent: owns integration, returned-output inspection, final verification, acceptance, and delivery.

Pin every automatic Agent explicitly. Do not rely on an undocumented model fallback. Keep `max_depth = 1`; use child Agents only for independent read-only work or disjoint writes with explicit acceptance checks. Small tasks should use no Subagent when delegation adds no value.

Use three to five execution children only when the task is substantial, clearly decomposable, and within the configured concurrency cap. Never assign overlapping writes or independent interface decisions to multiple children.

If a named model is unavailable or deprecated, map it to the current official equivalent and explain the change. Do not invent model IDs or treat this personal routing policy as a universal recommendation.

### 4. Protect Long Sessions

When GPT-5.6 is active and the installed Codex version accepts the key, add: