CLI Cross-Format Testing¶

This page provides reproducible commands for validating all 9 client × model combinations through the argo-proxy gateway. Each CLI tool speaks a different API format; the gateway converts automatically.

graph LR
    CC[Claude Code<br/>Anthropic API] --> AP[argo-proxy]
    CX[Codex CLI<br/>OpenAI Responses] --> AP
    GC[Gemini CLI<br/>Google GenAI] --> AP
    AP -->|Claude models| ANT[Anthropic upstream]
    AP -->|GPT/Gemini models| OAI[OpenAI Chat upstream]

Prerequisites¶

1. Start argo-proxy¶

# Install and start argo-proxy on port 44511
pip install argo-proxy
python -m argoproxy.cli serve --port 44511
# Verify
curl -s http://localhost:44511/health
# Expected: {"status": "healthy"}

2. CLI Tool Configuration¶

Claude CodeCodex CLIGemini CLI

# Set base URL via environment variable
export ANTHROPIC_BASE_URL=http://127.0.0.1:44511

Add to ~/.codex/config.toml:

[model_providers.argo]
name = "Argo Proxy"
base_url = "http://localhost:44511/v1"
env_key = "ARGO_API_KEY"
wire_api = "responses"

export ARGO_API_KEY=your_api_key

export GOOGLE_GEMINI_BASE_URL=http://localhost:44511
export GEMINI_API_KEY=your_api_key

Text Generation Tests¶

Quick validation script¶

Save as test_matrix.sh and run:

#!/usr/bin/env bash
# Cross-format routing matrix test
# Usage: ./test_matrix.sh
set -euo pipefail

PROXY="http://localhost:44511"
PROMPT="What is 7*8? Reply with just the number."

echo "=== Claude Code (Anthropic API) ==="
for model in claude-sonnet-4-20250514 gpt-4.1-nano gemini-2.5-flash; do
    printf "  %-30s → " "$model"
    result=$(ANTHROPIC_BASE_URL="$PROXY" claude -p "$PROMPT" --model "$model" 2>/dev/null | tail -1)
    echo "$result"
done

echo ""
echo "=== Codex CLI (OpenAI Responses API) ==="
for model in claude-sonnet-4-20250514 gpt-4.1-nano gemini-2.5-flash; do
    printf "  %-30s → " "$model"
    result=$(ARGO_API_KEY="${ARGO_API_KEY:-dummy}" codex exec "$PROMPT" -m "argo:$model" 2>/dev/null | tail -1)
    echo "$result"
done

echo ""
echo "=== Gemini CLI (Google GenAI API) ==="
for model in claude-sonnet-4-20250514 gpt-4.1-nano gemini-2.5-flash; do
    printf "  %-30s → " "$model"
    result=$(GEMINI_API_KEY="${GEMINI_API_KEY:-dummy}" \
        GOOGLE_GEMINI_BASE_URL="$PROXY" \
        gemini -m "$model" -p "$PROMPT" 2>/dev/null | tail -1)
    echo "$result"
done

Expected output (all should print 56):

=== Claude Code (Anthropic API) ===
  claude-sonnet-4-20250514       → 56
  gpt-4.1-nano                   → 56
  gemini-2.5-flash               → 56

=== Codex CLI (OpenAI Responses API) ===
  claude-sonnet-4-20250514       → 56
  gpt-4.1-nano                   → 56
  gemini-2.5-flash               → 56

=== Gemini CLI (Google GenAI API) ===
  claude-sonnet-4-20250514       → 56
  gpt-4.1-nano                   → 56
  gemini-2.5-flash               → 56

Individual commands¶

Claude Code

# Claude Code + Claude (passthrough)
ANTHROPIC_BASE_URL=http://127.0.0.1:44511 \
    claude -p "What is 7*8? Reply with just the number." \
    --model claude-sonnet-4-20250514

# Claude Code + GPT (anthropic → openai_chat)
ANTHROPIC_BASE_URL=http://127.0.0.1:44511 \
    claude -p "What is 7*8? Reply with just the number." \
    --model gpt-4.1-nano

# Claude Code + Gemini (anthropic → openai_chat)
ANTHROPIC_BASE_URL=http://127.0.0.1:44511 \
    claude -p "What is 7*8? Reply with just the number." \
    --model gemini-2.5-flash

Codex CLI

# Codex + Claude (openai_responses → anthropic)
ARGO_API_KEY=your_key codex exec \
    "What is 7*8? Reply with just the number." \
    -m argo:claude-sonnet-4-20250514

# Codex + GPT (passthrough)
ARGO_API_KEY=your_key codex exec \
    "What is 7*8? Reply with just the number." \
    -m argo:gpt-4.1-nano

# Codex + Gemini (passthrough)
ARGO_API_KEY=your_key codex exec \
    "What is 7*8? Reply with just the number." \
    -m argo:gemini-2.5-flash

Gemini CLI

# Gemini CLI + Claude (google → anthropic)
GEMINI_API_KEY=your_key GOOGLE_GEMINI_BASE_URL=http://localhost:44511 \
    gemini -m claude-sonnet-4-20250514 \
    -p "What is 7*8? Reply with just the number."

# Gemini CLI + GPT (google → openai_chat)
GEMINI_API_KEY=your_key GOOGLE_GEMINI_BASE_URL=http://localhost:44511 \
    gemini -m gpt-4.1-nano \
    -p "What is 7*8? Reply with just the number."

# Gemini CLI + Gemini (google → openai_chat)
GEMINI_API_KEY=your_key GOOGLE_GEMINI_BASE_URL=http://localhost:44511 \
    gemini -m gemini-2.5-flash \
    -p "What is 7*8? Reply with just the number."

Image Understanding Tests¶

Each CLI has a different method for attaching images:

CLI Tool	Image Method	Notes
Codex CLI	`-i path/to/image.png`	Encodes image directly into request body
Claude Code	Read tool (automatic)	Model reads the file via built-in tool
Gemini CLI	`read_file` tool + `-y` flag	Model reads file via built-in tool; requires `-y` for auto-approval and file must be in workspace

Codex CLI image tests¶

# Prepare a test image
cp /path/to/any/image.png /tmp/test_image.png

# Codex + Claude + image
ARGO_API_KEY=your_key codex exec \
    "Describe this image in 5 words or less." \
    -m argo:claude-sonnet-4-20250514 -i /tmp/test_image.png

# Codex + GPT + image
ARGO_API_KEY=your_key codex exec \
    "Describe this image in 5 words or less." \
    -m argo:gpt-4.1-nano -i /tmp/test_image.png

# Codex + Gemini + image
ARGO_API_KEY=your_key codex exec \
    "Describe this image in 5 words or less." \
    -m argo:gemini-2.5-flash -i /tmp/test_image.png

Claude Code image tests¶

# Claude Code + Claude + image
echo "Read the image at /tmp/test_image.png and describe it in 5 words." | \
    ANTHROPIC_BASE_URL=http://127.0.0.1:44511 \
    claude -p --model claude-sonnet-4-20250514 --allowedTools "Read"

# Claude Code + GPT + image (use GPT-5.4+ for reliable results)
echo "Read the image at /tmp/test_image.png and describe it in 5 words." | \
    ANTHROPIC_BASE_URL=http://127.0.0.1:44511 \
    claude -p --model gpt-5.4 --allowedTools "Read"

# Claude Code + Gemini + image
echo "Read the image at /tmp/test_image.png and describe it in 5 words." | \
    ANTHROPIC_BASE_URL=http://127.0.0.1:44511 \
    claude -p --model gemini-2.5-flash --allowedTools "Read"

Gemini CLI image tests¶

Workspace restriction

Gemini CLI can only read files within its workspace directories. Copy the test image into the project directory first. Use -y (YOLO mode) to auto-approve tool calls in headless mode.

# Copy image into workspace
cp /tmp/test_image.png /path/to/your/project/test_image.png

# Gemini CLI + Claude + image
GEMINI_API_KEY=your_key GOOGLE_GEMINI_BASE_URL=http://localhost:44511 \
    gemini -m claude-sonnet-4-20250514 -y \
    -p "Read the image file at $(pwd)/test_image.png and describe what you see in 5 words."

# Gemini CLI + GPT + image
GEMINI_API_KEY=your_key GOOGLE_GEMINI_BASE_URL=http://localhost:44511 \
    gemini -m gpt-4.1-nano -y \
    -p "Read the image file at $(pwd)/test_image.png and describe what you see in 5 words."

# Gemini CLI + Gemini + image
GEMINI_API_KEY=your_key GOOGLE_GEMINI_BASE_URL=http://localhost:44511 \
    gemini -m gemini-2.5-flash -y \
    -p "Read the image file at $(pwd)/test_image.png and describe what you see in 5 words."

# Clean up
rm test_image.png

Conversion Path Reference¶

Client Format	Claude Model Target	GPT/Gemini Model Target
Anthropic Messages	passthrough	anthropic → IR → openai_chat
OpenAI Responses	responses → IR → anthropic	passthrough (responses → openai_chat at upstream)
Google GenAI	google → IR → anthropic	google → IR → openai_chat