elizaos--eliza/packages/benchmarks/realm/AGENTS.md

REALM-Bench — Agent Guide

Real-World Planning benchmark: 11 problem types (TSP, VRP, DARP, event coordination, disaster relief, JSSP) drawn from arXiv:2502.18836. Vendored upstream task definitions and datasets under upstream/. Registered in the suite registry as realm.

Run

# Direct — all 11 problems, one instance each, via the eliza TS bridge
python -m benchmarks.realm.cli --max-tasks 1

# Subset of problem types
python -m benchmarks.realm.cli --problems P1 P11

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

Smoke test (no API keys)

# Deterministic mock oracle agent, tiny built-in P1 + P11 sample
python -m benchmarks.realm.cli --provider mock --use-sample-tasks

# Full mock run (all vendored instances, mock agent)
python -m benchmarks.realm.cli --provider mock --full-dataset

Test the harness

# One-time install (from packages/benchmarks/realm/)
pip install -e ".[dev]"

# Run tests (from repo root or packages/benchmarks/)
pytest packages/benchmarks/realm/ -v

Layout

Path	Role
cli.py	CLI entrypoint (python -m benchmarks.realm.cli)
runner.py	Async execution loop + _MockREALMAgent (oracle-based mock)
evaluator.py	Per-problem extrinsic scoring (quality, optimality, CSR, …)
solvers.py	OR-Tools oracles: TSP-TW, DARP, JSSP CP-SAT, disaster
dataset.py	Loader; normalises upstream schema variations
disruption.py	Mid-run disruption injection for P4/P7/P8/P9/P10
types.py	RealmProblem (P1–P11), REALMConfig, DTOs
plugin/	Plan-response parsing helpers
upstream/	Vendored from genglongling/REALM-Bench (datasets + evaluation)
tests/	pytest suite (smoke, dataset, runner, solver, env-loader)

Notes

• Results write to ./benchmark_results/realm/<timestamp>/ (gitignored). Result files match realm-benchmark-*.json.
• Scored by _score_from_realm_json in registry/scores.py.
• OR-Tools is optional; install with pip install "elizaos-benchmarks-realm[ortools]" or pass --auto-install-ortools. Without it, P1/P3/P4 use heuristic fallbacks and P11 errors unless the instance has an upper_bound header.
• Solver wall-clock budget: --solver-timeout (default 30 s). Use 120 s for large DMU/TA JSSP instances to reach OPTIMAL.
• Upstream reference: https://github.com/genglongling/REALM-Bench.
• 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.