Pi Codex Core

Everything GPT models need to feel at home in Pi.

Install

pi install npm:@zigai/pi-codex-core

Quick start

Run /login openai-codex and select OpenAI Codex (legacy), then choose an openai-codex model with /model for the full legacy-backed feature set.

Pi's newer /login openai does not authorize Core's ChatGPT-backed web_run, image descriptions, or image generation. Keep the separate legacy login for those features, even when using an openai GPT model. imagegen also retains support for compatible active Responses endpoints with their own credentials.

Tool activation follows model and scope settings; authentication is checked when a tool runs. Local image viewing and apply_patch do not require the legacy login. Native compaction, usage/reset credits, and turn recovery require an active openai-codex model; selecting an openai model does not enable them, and Pi compaction remains the fallback.

Features

Tools

• web_run — Codex-backed web search using the OpenAI Codex search endpoint, supporting cached, indexed, and live modes.
• imagegen — Codex image generation and editing.
• view_image — Native local image viewing.
• apply_patch — Optional port of the Codex native apply_patch editing tool. Enabling it replaces Pi's standard edit tool.

System prompt

Model-aware Codex prompt mode uses matching bundled GPT-6 Astra/Sol/Luna, GPT-5.6 Sol/Terra/Luna, or GPT-5.5 system prompts adapted to Pi's active tools and runtime context.

Compaction

Optional Codex remote compaction v2 replay supports OpenAI Codex response models, with chaining, metadata, fallbacks, world-state injection, auto-compaction, and token-aware context shrinking.

Turn recovery

Resumes transiently failed Codex turns after Pi exhausts standard retry attempts.

Slash command

Use the /codex slash command to manage extension settings and usage:

• Configure extension settings through a dedicated interactive interface.
• View current usage metrics and reset expiration dates.
• Activate available usage resets.

Additional capabilities

• Compatibility for GPT-6 Astra/Sol/Luna and GPT-5.6 Responses Lite during standard turns, image descriptions, and native compaction.
• Streamed reasoning trace controls for GPT Responses models.
• Fast mode option to prioritize response speed.

Configuration

Global settings are stored in ~/.pi/agent/extension-settings/pi-codex-core.json.

{
  "$schema": "./schemas/pi-codex-core.schema.json",
  "scope": {
    "tools": "codex"
  },
  "tools": {
    "webSearch": true,
    "webSearchMode": "live",
    "imageGeneration": true,
    "viewImage": true,
    "viewImageDescriptions": false,
    "applyPatch": "off"
  },
  "prompt": {
    "mode": "codex",
    "personality": "pragmatic"
  },
  "compaction": {
    "enabled": true,
    "auto": true,
    "thresholdPercent": 80
  },
  "recovery": {
    "enabled": true,
    "batchFollowUps": false,
    "maxAttempts": 3,
    "baseDelayMs": 30000,
    "maxDelayMs": 120000
  },
  "openai": {
    "webSearchModel": "current",
    "imageModel": "gpt-image-2",
    "imageDescriptionModel": "current",
    "compactionModel": "current",
    "compactionReasoning": "medium",
    "verbosity": "low",
    "fast": false,
    "showReasoningTraces": true
  }
}

Acknowledgements

This extension builds upon initial work by Igor Warzocha in pi-codex-conversion.

License

MIT