wagmi-bench: Instructions file for Codex

AGENTS.md — operating WAGMI Bench

WAGMI Bench is a local, deterministic BTC perpetual-futures survival benchmark for trading agents: it replays recorded market windows, feeds an agent point-in-time observations, and seals decisions, risk checks, fills, ledgers, and metrics into a verifiable evidence bundle. Claim discipline: every result is survival-stress evidence — liquidation survival, drawdown control, funding drag, turnover, rule-following under stress. When writing up a result, never state or imply predictive ability, expected future returns, or adoption; report what survived, what was liquidated, and how it read against fixed baselines on frozen data. Evidence rule: bundles and reports are immutable. Every --output must be a path that does not exist yet, and files inside a bundle are never edited or hand-assembled. If a run was wrong, run again into a new directory.

Setup

Requires Git and uv. Run everything from the repository root.

uv sync --group dev
uv run wagmibench --help

Never activate the venv; every command is uv run wagmibench ..., and the only bare-Python step is the agent server via uv run python ....

Keyless demo (no network, no keys)

uv run wagmibench run --pack fixtures/golden-mini/pack --output bundles/demo
uv run wagmibench report --bundle bundles/demo --output reports/demo
open reports/demo/report.html   # Linux: xdg-open

Committed synthetic fixture, keyless momentum policy — prove the toolchain works here before spending anything.

Real pack + baselines + compare

uv run wagmibench packs list
uv run wagmibench fetch-data --pack covid-black-thursday

for a in buyhold shorthold flat momentum; do
  uv run wagmibench run --pack covid-black-thursday --agent "$a" \
    --output "bundles/covid-$a"
done

uv run wagmibench compare \
  --bundle bundles/covid-buyhold --bundle bundles/covid-shorthold \
  --bundle bundles/covid-flat --bundle bundles/covid-momentum \
  --output reports/covid-compare