Cross-Wire Analysis: Codex × Qwen-Code × Gemini CLI

Deep analysis of wire protocol transformations, null spaces, and feature deltas across four CLI harness implementations.

What This Is

This repo documents the novel cross-wiring of LLM API wire protocols between CLI harnesses:

1. NetBrah/codex (apex/messages-wire-dev) — Anthropic /messages wire hooked into OpenAI Codex's Rust harness
2. NetBrah/qwen-code (dev) — OpenAI /responses wire (and /messages) hooked into Gemini CLI's TypeScript harness
3. google/gemini-cli (main) — Upstream reference, native Gemini generateContent wire
4. openai/codex (main) — Upstream reference, native OpenAI /responses wire

The two novel implementations:

• Codex + /messages: Anthropic models running natively in Codex's Rust harness via the /messages wire protocol
• Qwen-Code + /responses: OpenAI models running in Gemini CLI's fork via the /responses wire protocol (everything transforms through Google's internal format)

Repository Structure

docs/
  architecture/
    00-system-overview.md          # Three-layer model: Models × APIs × Harnesses
    01-wire-protocol-matrix.md     # Complete mapping of all wire transformations
  api-specs/
    anthropic-messages.md          # /v1/messages wire spec
    openai-responses.md            # /v1/responses wire spec  
    gemini-generatecontent.md      # v1beta generateContent spec
  null-space/
    01-gemini-to-messages.md       # Null space: Gemini CLI harness → /messages wire
    02-gemini-to-responses.md      # Null space: Gemini CLI harness → /responses wire
    03-codex-native-vs-messages.md # Null space: Codex /responses vs /messages wire
    04-cross-layer-invariants.md   # Invariants across all three layers
  feature-delta/
    01-gemini-vs-qwen.md           # Features: upstream Gemini CLI ↔ qwen-code
    02-codex-harness-additions.md  # Features added to codex beyond /messages wire
    03-cross-implementation.md     # Cross-comparison matrix (all 4)
sortie-board/
    SORTIE-BOARD.md                # Actionable code ops with priorities
    active-sorties.md              # In-flight sortie branches  
refs/
    [source code snapshots]        # Key files from all repos for evidence

The Three-Layer Model

┌─────────────────────────────────────────────────────────┐
│ LAYER 1: MODELS                                         │
│ Claude Opus/Sonnet │ GPT-4o/o3 │ Gemini Pro │ Qwen     │
├─────────────────────────────────────────────────────────┤
│ LAYER 2: WIRE PROTOCOLS (APIs)                          │
│ /messages          │ /responses │ generateContent       │
├─────────────────────────────────────────────────────────┤
│ LAYER 3: HARNESSES                                      │
│ Codex (Rust)       │ Qwen-Code (TS) │ Gemini CLI (TS)  │
└─────────────────────────────────────────────────────────┘

Each cross-wire implementation creates a transformation layer between Layer 2 and Layer 3, introducing potential null spaces where features/data are lost in translation.

Source Repos

Repo	Branch	Role
NetBrah/codex	apex/messages-wire-dev	/messages wire in Rust harness
NetBrah/qwen-code	dev	/responses + /messages in TS harness
google-gemini/gemini-cli	main	Upstream Gemini reference
anthropics/*	various	Official Anthropic SDKs & tools

Anthropic Ecosystem (Official Repos)

Repo	Stars	Language	Purpose
anthropics/skills	102K	Python	Agent Skills
anthropics/claude-code	82K	Shell	Claude Code CLI
anthropics/anthropic-sdk-python	3K	Python	Python SDK
anthropics/anthropic-sdk-typescript	1.7K	TypeScript	TypeScript SDK
anthropics/anthropic-sdk-go	927	Go	Go SDK
anthropics/anthropic-sdk-ruby	309	Ruby	Ruby SDK
anthropics/anthropic-sdk-java	264	Kotlin	Java SDK
anthropics/anthropic-sdk-csharp	201	C#	C# SDK
anthropics/anthropic-sdk-php	121	PHP	PHP SDK
anthropics/claude-agent-sdk-python	5.7K	Python	Agent SDK (Python)
anthropics/claude-agent-sdk-typescript	1K	TypeScript	Agent SDK (TypeScript)
anthropics/claude-plugins-official	14.6K	Python	Official Plugins
anthropics/claude-code-action	6.5K	TypeScript	GitHub Actions

Key Findings Summary

See individual docs for full analysis. High-level:

• 12 null-space gaps identified across wire transformations
• 23 feature deltas between implementations
• 8 critical sortie items (P0/P1) for immediate action
• 15 portability candidates for cross-pollination between harnesses