Download scripts/record_agent_runs.sh from openenvforge/3ambench: direct link, hf CLI and curl.
- Browser
- Download file 4.51 kB
-
https://huggingface.co/datasets/openenvforge/3ambench/resolve/main/scripts/record_agent_runs.sh
- Command line
-
hf download hf://datasets/openenvforge/3ambench/scripts/record_agent_runs.sh
-
curl -L -o record_agent_runs.sh https://huggingface.co/datasets/openenvforge/3ambench/resolve/main/scripts/record_agent_runs.sh
4.51 kB
| # Record real model runs on a task subset with Harbor (via scripts/harbor_local.sh), then re-export | |
| # space/data/runs.json so the replay Space shows them. This spends model credits; nothing runs by default. | |
| # | |
| # scripts/record_agent_runs.sh -a <agent> -m <model> [-k attempts] [-t af-001,af-009,...] [-n concurrent] | |
| # | |
| # Agents and auth (Harbor reads these from the environment): | |
| # codex subscription: ~/.codex/auth.json (run `codex login` once); this script sets | |
| # CODEX_FORCE_AUTH_JSON=1. Use AUTH=apikey to use OPENAI_API_KEY instead. | |
| # scripts/record_agent_runs.sh -a codex -m openai/<model> -k 2 | |
| # claude-code subscription token from `claude setup-token`, exported as CLAUDE_CODE_OAUTH_TOKEN; this script | |
| # sets CLAUDE_FORCE_OAUTH=1 so an API key in the environment is not used instead. | |
| # CLAUDE_CODE_OAUTH_TOKEN=... scripts/record_agent_runs.sh -a claude-code -m <model> -k 2 | |
| # terminus-2 any LiteLLM model id plus that provider's API key (OPENAI_API_KEY, OPENROUTER_API_KEY, | |
| # GEMINI_API_KEY, ...). | |
| # OPENROUTER_API_KEY=... scripts/record_agent_runs.sh -a terminus-2 -m openrouter/<vendor>/<model> | |
| # | |
| # Default tasks: one per workflow across all tiers (af-001 easy, af-009 medium, af-013 easy, af-017 hard, | |
| # af-021 medium, af-029 hard) - the same tasks as the reference runs, so the leaderboard compares like with like. | |
| # | |
| # harbor_local.sh drops the tasks' `no-network` override so it runs on Docker Desktop (macOS). With network | |
| # access an agent could fetch the public solutions, so treat these as local, not leaderboard-grade, results; | |
| # on a Linux Docker host run `harbor run -p tasks ...` directly to keep the restriction. | |
| # | |
| # Other env: HARBOR (default: harbor on PATH), AF_PROMTOOL / AF_AMTOOL (pinned tools for the export replay), | |
| # PYTHON (a Python with PyYAML; default python3), JOBS_DIR (default dist/replay/agent-jobs). | |
| set -euo pipefail | |
| ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" | |
| AGENT="" MODEL="" ATTEMPTS=1 CONCURRENT=2 | |
| TASKS="af-001,af-009,af-013,af-017,af-021,af-029" | |
| usage() { sed -n '2,27p' "${BASH_SOURCE[0]}" | sed 's/^# \{0,1\}//'; exit "${1:-0}"; } | |
| while getopts "a:m:k:t:n:h" o; do | |
| case "$o" in | |
| a) AGENT="$OPTARG" ;; m) MODEL="$OPTARG" ;; k) ATTEMPTS="$OPTARG" ;; | |
| t) TASKS="$OPTARG" ;; n) CONCURRENT="$OPTARG" ;; h) usage 0 ;; *) usage 1 ;; | |
| esac | |
| done | |
| [ -n "$AGENT" ] && [ -n "$MODEL" ] || { echo "need -a <agent> and -m <model>" >&2; usage 1; } | |
| case "$AGENT" in | |
| codex) | |
| if [ "${AUTH:-subscription}" = "apikey" ]; then | |
| : "${OPENAI_API_KEY:?AUTH=apikey needs OPENAI_API_KEY}" | |
| else | |
| [ -f "$HOME/.codex/auth.json" ] || { echo "no ~/.codex/auth.json; run 'codex login' first" >&2; exit 1; } | |
| export CODEX_FORCE_AUTH_JSON=1 | |
| fi ;; | |
| claude-code) | |
| : "${CLAUDE_CODE_OAUTH_TOKEN:?export CLAUDE_CODE_OAUTH_TOKEN (from 'claude setup-token')}" | |
| export CLAUDE_FORCE_OAUTH=1 ;; | |
| terminus-2) | |
| env | grep -qE '^[A-Z_]*API_KEY=' || echo "warning: no *_API_KEY in the environment for $MODEL" >&2 ;; | |
| *) echo "unsupported agent '$AGENT' (codex | claude-code | terminus-2)" >&2; exit 1 ;; | |
| esac | |
| # stage the chosen tasks in one directory so they run as a single Harbor job | |
| STAGE="$(mktemp -d)/tasks"; mkdir -p "$STAGE" | |
| IFS=',' read -r -a PREFIXES <<< "$TASKS" | |
| for p in "${PREFIXES[@]}"; do | |
| found=0 | |
| for d in "$ROOT"/tasks/"$p"*/; do [ -f "$d/task.toml" ] && { cp -R "${d%/}" "$STAGE/"; found=1; }; done | |
| [ "$found" = 1 ] || { echo "no task matches '$p'" >&2; exit 1; } | |
| done | |
| JOBS_DIR="${JOBS_DIR:-$ROOT/dist/replay/agent-jobs}" | |
| mkdir -p "$JOBS_DIR" | |
| JOB="${AGENT}__$(echo "$MODEL" | tr '/:' '__')__$(date +%Y%m%d-%H%M%S)" | |
| echo "job $JOB: $(ls "$STAGE" | wc -l | tr -d ' ') tasks x $ATTEMPTS attempts -> $JOBS_DIR/$JOB" >&2 | |
| "$ROOT/scripts/harbor_local.sh" "$STAGE" -- -a "$AGENT" -m "$MODEL" -k "$ATTEMPTS" -n "$CONCURRENT" \ | |
| -o "$JOBS_DIR" --job-name "$JOB" -y | |
| # re-export everything the Space shows: reference traces, local Harbor reference jobs, all agent jobs (both tiers) | |
| SOURCES=(--harbor "$JOBS_DIR") | |
| for d in harbor-jobs together-smoke agent-jobs agent-jobs-v02; do | |
| [ -d "$ROOT/dist/replay/$d" ] && [ "$ROOT/dist/replay/$d" != "$JOBS_DIR" ] && SOURCES+=("$ROOT/dist/replay/$d") | |
| done | |
| [ -d "$ROOT/dist/replay/traces" ] && SOURCES=(--traces "$ROOT/dist/replay/traces" "${SOURCES[@]}") | |
| "${PYTHON:-python3}" "$ROOT/scripts/export_runs.py" "${SOURCES[@]}" --out "$ROOT/space/data/runs.json" | |