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

VisualWebBench — Agent Guide

Seven-subtask multimodal web understanding and grounding benchmark, faithfully implementing VisualWebBench (Apache-2.0). Evaluates ROUGE-L (captions/OCR), F1 (WebQA), and MCQ accuracy (grounding/action). Registered in the suite registry as visualwebbench.

Run

# Direct, from packages/benchmarks/visualwebbench/
PYTHONPATH=packages:packages/benchmarks/eliza-adapter \
  python -m benchmarks.visualwebbench --max-tasks 70

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

Smoke test (no API keys, no HF download)

PYTHONPATH=packages:packages/benchmarks/eliza-adapter \
  python -m benchmarks.visualwebbench --use-sample-tasks --mock --max-tasks 7

--mock echoes task.answer through the oracle agent (always 100 %). Combined with --use-sample-tasks it uses the bundled 7-row JSONL fixture — one row per subtask, no images, no network calls. Scores from this path are not comparable to upstream.

Test the harness

pip install -e ".[dev]"
pytest visualwebbench/tests/ -v

Layout

Path	Role
cli.py	Argument parsing, main() entrypoint
__main__.py	python -m benchmarks.visualwebbench hook
runner.py	Async execution loop across all subtasks
agent.py	Eliza adapter + oracle mock agent
dataset.py	HF streaming loader and JSONL fixture reader
evaluator.py	Per-subtask ROUGE / F1 / MCQ scorers
types.py	Dataclasses and enums (VisualWebBenchTaskType, etc.)
fixtures/smoke.jsonl	7-row offline fixture (one per subtask, no images)
fixtures/local_vlm_real.jsonl	Sample rows for local VLM (eliza-1) testing
tests/test_visualwebbench.py	pytest suite (metric unit tests + runner smoke)

Notes

• Results write to benchmark_results/visualwebbench/<timestamp>/ (gitignored). Output files: visualwebbench-results.json, summary.md, traces/<task-id>.json.
• The _visualwebbench_result locator in registry/commands.py expects visualwebbench-results.json at the root of output_dir.
• Scored by _score_from_visualwebbench_json in registry/scores.py.
• HF dataset: visualwebbench/VisualWebBench, split test, seven configs (one per subtask). Images are lazily fetched and cached as PNG under ~/.cache/elizaos/visualwebbench/images/ by default.
• Use --max-tasks N to cap downloads during development.
• 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.