elizaos--eliza/packages/benchmarks/tau-bench/CLAUDE.md

Tau-bench — Agent Guide

Vendored implementation of Sierra's tau-bench (Yao et al., 2024): Tool-Agent-User Interaction benchmark across retail (115 tasks) and airline (50 tasks) domains, with pass^k scoring and an LLM judge. Registered in the suite registry as tau_bench.

Run

# Direct, from this directory — full 165-task suite, pass^4 (paper default)
python -m elizaos_tau_bench --agent-model gpt-4o

# With a non-OpenAI agent; keep openai for user-simulator and judge
python -m elizaos_tau_bench \
    --agent-provider anthropic --agent-model claude-3-5-sonnet-latest \
    --user-provider openai --user-model gpt-4o \
    --judge-provider openai --judge-model gpt-4o-mini

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

Smoke test (no API keys)

# Deterministic mock agent — no LLM calls, no keys required
python -m elizaos_tau_bench --mock --use-sample-tasks

Test the harness

# One-time install (from this directory)
pip install -e ".[dev]"

# Run the pytest suite
pytest packages/benchmarks/tau-bench/ -v

Layout

Path	Role
elizaos_tau_bench/cli.py	CLI entrypoint (python -m elizaos_tau_bench)
elizaos_tau_bench/runner.py	Main execution loop (TauBenchRunner)
elizaos_tau_bench/judge.py	LLM judge (gpt-4o-mini, falls back to substring)
elizaos_tau_bench/pass_k.py	Unbiased pass^k estimator
elizaos_tau_bench/types.py	TauBenchConfig, TauBenchReport DTOs
elizaos_tau_bench/upstream/	Vendored sierra-research/tau-bench source (MIT)
elizaos_tau_bench/compact_fixtures/	Compact DB fixtures for smoke runs
tests/	pytest suite (dataset, pass^k, judge, output contract, smoke)
pyproject.toml	Package metadata; tau-bench console script

Notes

• Results write to benchmark_results/tau-bench/<timestamp>/ (report.json + trajectories.json).
• Scored by _score_from_taubench_json in registry/scores.py.
• Required env vars: OPENAI_API_KEY (agent + user simulator + judge by default). Override each component's provider with --agent-provider, --user-provider, --judge-provider.
• Full retail + airline data is fetched lazily into ~/.cache/elizaos_tau_bench/ on first run. Set TAU_BENCH_DATA_DIR to a pre-populated path, or TAU_BENCH_DATA_MODE=smoke to use only compact fixtures.
• Vendored upstream commit: 59a200c6d575d595120f1cb70fea53cef0632f6b.
• 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.