cc-plugin-codex (sendbird) — Summary

Elevator pitch: A Codex-native plugin (.codex-plugin/plugin.json) that is the semantic inverse of codex-plugin-cc — it runs inside Codex and delegates review/rescue work to Claude Code, rather than running inside Claude Code and delegating to Codex. Published by Sendbird under Apache-2.0. The plugin ships 7 Codex skills ($cc:review, $cc:adversarial-review, $cc:rescue, $cc:status, $cc:result, $cc:cancel, $cc:setup) plus a claude-companion.mjs script that invokes the claude CLI for all work. It uses Codex's native plugin hook system (SessionStart, SessionEnd, Stop, UserPromptSubmit) for job lifecycle management and an unread-result nudge on each user turn. The plugin is described as following "the shape of openai/codex-plugin-cc but runs in the opposite direction." Compared to seeds: this is a mirror of codex-plugin-cc with the host and guest models swapped — a complementary first/second-party plugin pair that together enable bidirectional Claude↔Codex delegation.

cc-plugin-codex (sendbird) — Overview

Origin

By Sendbird, Inc. (jin.ku@sendbird.com). Version 1.2.1, Apache-2.0. 42 stars, 7 forks. Active: last commit 2026-05-25 (most recently updated plugin in this batch). Published to Sendbird's marketplace.

Philosophy (verbatim from README)

"cc-plugin-codex turns Codex into a host for Claude Code work. Codex stays in charge of the thread. Claude Code does the review and rescue work."

"It follows the shape of openai/codex-plugin-cc but runs in the opposite direction."

The design is explicitly positioned as the complement to codex-plugin-cc:

• codex-plugin-cc: Runs inside Claude Code → delegates to Codex
• cc-plugin-codex: Runs inside Codex → delegates to Claude Code

Key Differences from Upstream (codex-plugin-cc)

The README explicitly documents how this fork differs from the upstream:

• Full 4-hook lifecycle (adds UserPromptSubmit for unread-result nudges, missing from upstream)
• --model / --effort flags on all review commands (upstream lacks effort controls)
• --scope auto|working-tree|branch explicit scope flag
• 1M-context model variants (opus-4-7[1m], sonnet-4-6[1m]) as defaults
• Background review via built-in Codex subagent (uses spawn_agent with gpt-5.4-mini + fork_context: false)
• Job ownership tracked per Codex session + --all flag to see cross-session jobs
• GitHub CI on Windows/macOS/Linux

Manifest Data

From .codex-plugin/plugin.json:

• Name: cc
• Version: 1.2.1
• Category: Coding
• Capabilities: Interactive, Write
• Default prompts: "Use Claude Code to review my current changes through $cc:review", etc.

cc-plugin-codex (sendbird) — Architecture

Distribution

• Type: Codex plugin (.codex-plugin/plugin.json)
• Marketplace: sendbird/codex-marketplace
• Plugin ID: cc
• Version: 1.2.1
• License: Apache-2.0

Install

codex marketplace add sendbird/codex-marketplace
# Then install 'cc' from the Sendbird marketplace inside Codex, run $cc:setup once.

# Optional npx helper:
npx cc-plugin-codex install

The plugin copy lives under Codex's plugin cache. Commands resolve through $PLUGIN_ROOT.

Required Runtime

• Codex CLI with hook support
• Node.js 18+
• claude CLI installed and authenticated
• npm install -g @anthropic-ai/claude-code && claude auth login

Directory Tree

.codex-plugin/
└── plugin.json          # Codex plugin manifest

hooks/
├── hooks.json           # 4 Codex hooks
├── session-lifecycle-hook.mjs
├── stop-review-gate-hook.mjs
└── unread-result-hook.mjs

skills/
├── review/SKILL.md
├── adversarial-review/SKILL.md
├── rescue/SKILL.md
├── status/SKILL.md
├── result/SKILL.md
├── cancel/SKILL.md
└── setup/SKILL.md

internal-skills/
└── review-runtime/runtime.md    # Internal reference

scripts/
└── claude-companion.mjs         # Runtime bridge: invokes claude CLI

prompts/
├── adversarial-review.md
└── stop-review-gate.md

schemas/                         # JSON schemas for skill arguments

Key Architectural Difference from codex-plugin-cc

The companion script is claude-companion.mjs (invokes claude CLI) vs. codex-plugin-cc's codex-companion.mjs (invokes codex CLI). The hook system uses Codex's native hook events (SessionStart, SessionEnd, Stop, UserPromptSubmit) vs. Claude Code's hook events.

Target AI Tools

• Primary: OpenAI Codex CLI (as host / orchestrator)
• Peer: Claude Code (as reviewer / rescue worker)

cc-plugin-codex (sendbird) — Components

Skills (7)

Hooks (4 Codex hook events)

The UserPromptSubmit hook is the key differentiator from codex-plugin-cc — it checks for unread Claude Code results and surfaces a notification at the start of each user turn.

Scripts (3)

Internal Skills (1)

• internal-skills/review-runtime/runtime.md — Internal reference document for the review runtime contract (not user-invocable)

Schemas

JSON schemas for skill argument validation in schemas/.

Review Model Defaults

• $cc:review defaults: model=opus (claude-opus-4-7[1m], 1M context), effort=xhigh
• $cc:rescue defaults: model=opus (1M context), effort=xhigh
• Sonnet resolves to claude-sonnet-4-6[1m]; effort defaults to high
• --model and --effort flags accepted on all review/rescue commands

cc-plugin-codex (sendbird) — Uniqueness

differs_from_seeds

The closest comparison is codex-plugin-cc (not a seed, but the direct upstream). Among seeds, the closest is superpowers (plugin architecture, single peer interaction model). This plugin differs from all seeds by being the only framework that runs inside Codex and uses Claude Code as its worker — the reverse of every other Claude-centric framework. The UserPromptSubmit hook-based unread-result nudge is unique in the corpus — no seed uses hook events to surface push notifications about completed background work. The 3-model pipeline (Codex parent → gpt-5.4-mini forwarder child → Claude opus-4-7[1m] worker) is the only three-tier model routing in this batch.

Mirror Relationship with codex-plugin-cc

Most Unusual Feature

The UserPromptSubmit hook that automatically checks for unread Claude Code results at the start of every user turn. No other framework in this batch or the 11 seeds uses a hook event to implement a push notification mechanism for completed background work.

Observable Failure Modes

1. Claude auth expiry: If Claude CLI authentication expires during a background job, the job fails silently. $cc:setup checks auth at setup time but not during job execution.
2. Codex job scope ambiguity: $cc:status by default shows only current-session jobs. Users may miss results from sibling sessions without --all flag.
3. gpt-5.4-mini forwarder errors: If the cheap forwarder child fails (e.g., token limit), the background review fails. No fallback to direct execution.
4. Stop gate zombie: 900s timeout on the Stop hook can leave Codex in a non-completable state if Claude Code hangs.
5. Windows compatibility: README notes Codex CLI Windows support is experimental; the Claude CLI supports Windows natively. Mixed Windows compatibility creates uncertain behavior.