docs/inference.md

Running Red-vs-Blue Inference

This document covers Dockerized Red-vs-Blue benchmark runs for WarGames.

Build the Inference Image

Build the local benchmark image from Dockerfile.inference:

make build-inference

This creates the wargames-inference image. Rebuild it after changing prompts, environment logic, reward logic, or benchmark code.

Required Environment Variables

For OpenRouter runs, set:

OPENROUTER_API_KEY=...

For OpenCode Go runs, set:

OPENCODE_GO_KEY=...
OPENCODE_GO_URL=...

If the variables are stored in .env, load them into the shell without printing the file:

set -a
source .env
set +a

Run a Single OpenRouter Benchmark

Use evals/run_red_blue_benchmark.py inside the inference image. The script creates a model-specific folder under outputs/ and writes all result files there.

docker run --rm \
  -e OPENROUTER_API_KEY \
  -v "$PWD/outputs:/home/user/app/outputs" \
  -v "$PWD/evals:/home/user/app/evals:ro" \
  wargames-inference \
  python evals/run_red_blue_benchmark.py \
    --models "qwen/qwen3.5-9b" \
    --max-steps 30

The same model is used for Red and Blue in showdown mode.

Run Multiple Models Sequentially

Pass multiple --model values. The harness runs them sequentially to reduce rate-limit pressure.

docker run --rm \
  -e OPENROUTER_API_KEY \
  -v "$PWD/outputs:/home/user/app/outputs" \
  -v "$PWD/evals:/home/user/app/evals:ro" \
  wargames-inference \
  python evals/run_red_blue_benchmark.py \
    --models "meta-llama/llama-3.1-8b-instruct,qwen/qwen3.5-9b" \
    --max-steps 30

Run an OpenCode Go Benchmark

docker run --rm \
  -e OPENCODE_GO_KEY \
  -e OPENCODE_GO_URL \
  -v "$PWD/outputs:/home/user/app/outputs" \
  -v "$PWD/evals:/home/user/app/evals:ro" \
  wargames-inference \
  python evals/run_red_blue_benchmark.py \
    --models "minimax-m2.7" \
    --max-steps 30

Output Files

Each run writes a timestamped folder like:

outputs/docker_openrouter_qwen35_9b_YYYYMMDD_HHMMSS/

The important files are:

• summary.csv: one row per model with actual_steps, final_score, max_reward, avg_reward, and error state.
• steps.csv: one row per environment step with Red command, Red reasoning, reward, termination state, Blue actions, services_affected, and services_restored.
• red_vs_blue.log: full step-by-step terminal log, including Red commands, Blue commands, output, errors, and final status.

Current Game Rules Reflected in Inference

The Red agent can use one direct process-kill command per episode. Later direct kill, pkill, or killall actions are rejected by the environment with PROCESS_KILL_BUDGET_EXHAUSTED.

In phase-2-blue-llm-showdown, Blue receives two repair turns after every Red action. Red reasoning is stored in benchmark outputs, but it is not shown to Blue.

The services_affected and services_restored fields are evaluation-only CSV columns. They are not added to either model prompt.