crossfire

Actor-Critic collaboration for Claude + Codex
Two AI partners walk into your repo. One draws the map, one carries the toolbox, both argue like a championship debate team.

English | 中文

Why Crossfire?

What happens when you lock Claude and Codex in a room and tell them to agree on your architecture? Surprisingly good code.

crossfire is a Claude Code skill that runs an end-to-end Actor-Critic pipeline pairing two AI models from different training distributions. Claude Code (Opus 4.6) plays Architect and Reviewer — the strategic thinker with the whiteboard. Codex CLI (GPT-5.4) plays Executor and Auditor — the heavy lifter who codes fast and double-checks everything like a suspicious airport inspector.

The core insight: heterogeneous models catch each other's blind spots better than homogeneous teams (arXiv:2602.03794). Phase 0 is couples therapy for AIs. Phase 2 is trust issues as a service. Your codebase is the winner.

How It Works

The pipeline has four phases — plan, execute, review, report:

1. Phase 0 — PLAN (Claude): Explore codebase → draft blueprint → adversarial debate with Codex → lock final plan
2. Phase 1 — EXECUTE (Codex): Implement exactly from the locked blueprint, with Windows defense layer pre-processing
3. Phase 2 — REVIEW (Both): Claude initial review → deterministic checks (tests/linters) → Codex final audit → cross-review
4. Phase 3 — REPORT: Structured report + automatic git commit

Workflow Tiers

Level	Pipeline	Best For	Review Depth
L1	Skip Phase 0 → EXECUTE → quick review	Small edits, known fixes, low ambiguity	Basic
L2	Full 4-phase pipeline	Default for most development work	Strong (3-round cap)
L3	Full pipeline + mandatory cross-review	Architecture changes, risky refactors, critical releases	Maximum

When file count and line count point to different levels, the higher level wins.

Quick Start

Prerequisites

• Claude Code v1.0+
• Codex CLI v0.115.0+ — npm install -g @openai/codex
• Configure Codex model in ~/.codex/config.toml:

  model = "gpt-5.4"
  model_reasoning_effort = "xhigh"

Installation

git clone https://github.com/PlutoLei/crossfire.git
mkdir -p ~/.claude/skills/crossfire/references
cp crossfire/SKILL.md ~/.claude/skills/crossfire/
cp crossfire/references/*.md ~/.claude/skills/crossfire/references/

First Command

/crossfire code: Add a parse_date function to src/utils.py supporting ISO 8601 and Chinese date formats

Templates

Nine templates for common task types:

Template	Level	Description	Example
code	L1/L2	New feature implementation	/crossfire code: Add parse_date to utils.py
bugfix	L2	Root-cause fix + regression guard	/crossfire bugfix: Fix confusion matrix colors
refactor	L2	Improve structure, preserve behavior	/crossfire refactor: Extract dedup logic
test	L1	Add or harden automated tests	/crossfire test: Unit tests for process_one_paper
review	L1	Focused Codex code review	/crossfire review: Review generate_essay_en.py
audit	L2	Dual-source parallel audit (Claude + code-reviewer agent)	/crossfire audit: Audit TextMamba3D changes
optimize	L2	Performance or resource optimization	/crossfire optimize: Async batch requests
architect	L3	System-level architecture design	/crossfire architect: Design figure compositor
research	L2	Implement from research architecture (auto inject-plan)	/crossfire research: Implement per architecture_proposal.md

Syntax: /crossfire <template>: <description> or /crossfire L2: <description>

Flags: --no-debate (skip Phase 0 debate) · --no-audit (skip Codex final audit) · --inject-plan <dir> (inject research outputs)

Configuration

Codex CLI

Key settings in ~/.codex/config.toml:

model = "gpt-5.4"
model_reasoning_effort = "xhigh"

[windows]
sandbox = "elevated"

Windows Defense Layer

All Codex calls go through mandatory pre-processing on Windows:

Protection	Rule	Phases
Path normalization	All paths converted to forward slashes	All
cd injection	First line of every Codex prompt: cd <normalized-abs-path>	All
Blueprint embedding	≤200 lines inline in prompt; >200 lines via file reference	Phase 1
Output integrity fallback	Detect missing/truncated files → Claude completes via Write tool	Phase 1

Architecture Details

For the complete pipeline breakdown — including Phase 0 sub-steps, multi-layer review, fix-reaudit loops, and skill delegation strategy — see Architecture Documentation.

Credits

• Actor-Critic pipeline for Claude Code × Codex CLI
• Inspired by heterogeneous multi-agent debate
• Banner & diagrams generated with PaperBanana + Google Gemini
• MIT Licensed