elizaos--eliza/packages/benchmarks/webshop/CLAUDE.md

WebShop — Agent Guide

elizaOS adapter for the WebShop benchmark (Yao et al., NeurIPS 2022). Evaluates web-interaction agents on product search and purchase tasks using Princeton-NLP's upstream Gym environment, reward function, and 12k human-written instructions. Registry id: webshop.

Run

# Direct, from this directory (small 1k-product profile, bridge mode)
python -m elizaos_webshop --profile small --bridge --max-tasks 50

# Through the suite orchestrator (resolves provider/model, stores results)
python -m benchmarks.orchestrator run --benchmarks webshop --provider <p> --model <m>

Fetch the upstream data first if not already present:

python scripts/fetch_data.py --profile small   # ~9 MB, 1k products
python scripts/fetch_data.py --profile full    # ~2 GB, 1.18M products

Smoke test (no API keys, no downloads)

python -m elizaos_webshop --use-sample-tasks --mock --max-tasks 3

Uses the bundled ~6-product sample catalog with the deterministic mock agent. No external dependencies beyond a working install.

Test the harness

pip install -e ".[dev]"
python -m spacy download en_core_web_sm
pytest packages/benchmarks/webshop/ -v

Tests are auto-skipped when heavy deps (spaCy model, torch, thefuzz, bs4) are absent, so pytest still passes cleanly in a fresh checkout.

Layout

Path	Role
elizaos_webshop/cli.py	CLI entrypoint (--mock, --bridge, --profile, --use-sample-tasks)
elizaos_webshop/runner.py	Orchestration loop across tasks
elizaos_webshop/environment.py	Adapter over upstream WebAgentTextEnv; BM25 fallback
elizaos_webshop/evaluator.py	Reports Score + SR following the paper
elizaos_webshop/eliza_agent.py	Mock agent driving the upstream env
elizaos_webshop/dataset.py	Loads upstream JSONs, resolves train/test split
elizaos_webshop/types.py	Typed observation / step / report shapes
upstream/web_agent_site/	Vendored Princeton-NLP Flask sim + Gym env (unmodified)
scripts/fetch_data.py	Downloads catalog and instruction files
tests/	pytest smoke suite

Notes

• Results write to benchmark_results/webshop/<timestamp>/ (gitignored): webshop-results.json, webshop-summary.md, webshop-detailed.json.
• Scored by _score_from_webshop_json in registry/scores.py.
• Reward function is upstream's web_agent_site.engine.goal.get_reward (TF-IDF / fuzzy match over title, attributes, options, price) — identical to the published paper.
• spaCy en_core_web_sm is required at runtime; install once with python -m spacy download en_core_web_sm.
• Full background: README.md.

⛔ NON-NEGOTIABLE — evidence, trajectories & real end-to-end tests

The binding, repo-wide standard is AGENTS.md. Read it. Nothing in this package is done until it is proven done — a reviewer must confirm it works without reading the code, from the artifacts you attach. This applies to every feature, fix, refactor, and chore here. "Tests pass" is not proof; "CI is green" is not proof.

• Record AND read model trajectories. Capture the actual inputs and outputs of the model from a live LLM — not the deterministic proxy, not a mock: the prompt, the providers/context, the raw model output, every tool/action call, and the result. Then open the trajectory and review it by hand. A captured-but-unread trajectory is not evidence (packages/scenario-runner/bin/eliza-scenarios run <scenario> --report <out>).
• Real, full-featured E2E — no larp. Every feature ships detailed end-to-end tests that drive the real path end to end. Not the happy "front door" only: cover error paths, edge/empty/invalid input, concurrency, roles/permissions, and adversarial input. A test that asserts against a mock/stub/fixture standing in for the thing under test does not count. If the real model/device/chain/connector/account is hard to reach, make it reachable — that is the work, not an excuse to mock. If the existing tests here are shallow or mocked, fixing them is part of your change.
• Screenshots + logs at every phase, plus a complete walkthrough video/run-through of the entire feature or view, start to finish (bun run test:e2e:record).
• Manually review every artifact the change touches — never just the green check: client logs (console + network), server logs ([ClassName] …), the model trajectories in and out, before/after full-page screenshots, and the domain artifacts listed below for this package.
• No residuals. No shortcuts. The goal is not "done" — it is everything done. Clear every blocker by the hard path: build the real architecture, stand up the real model/device/service, actually test it. Never leave a TODO, a stub, a stepping-stone, or a "follow-up." When unsure, research thoroughly, weigh the options, and ship the best, highest-effort, production-ready version. Keep going until every possibility is exhausted.

Artifacts → attached inline in the PR (MP4 video, JPG screenshots, logs in <details>); attach each evidence type or explicitly mark it N/A with a reason — never leave it blank. If develop moved and changed behavior, re-capture evidence; stale proof is worse than none.

Capture & manually review for this package — benchmark / eval suite:

• A real-model run (not the mock/smoke fixture) producing the score-report JSON, with the numbers inspected and the provider/model recorded.
• The per-item trajectories the harness captured, spot-reviewed for correctness — a green harness run over mock fixtures is not a result.
• The provider matrix actually exercised, and the scoring math validated against a known case.
• Failure / timeout / partial-output handling in the harness itself.