Imported from elizaOS/benchmarks (suites/nl2repo/AGENTS.md). Install upstream with npx skills add elizaOS/benchmarks --skill nl2repo. Copyright stays with the author.

NL2Repo-Bench — Agent Guide

Long-horizon, 0-to-1 repository generation benchmark (arXiv:2512.12730). An agent receives a natural-language requirements document (start.md) and an empty workspace, and must produce a fully installable, runnable Python library. 104 tasks scored by pytest pass-rate inside per-task Docker eval images. Not registered in the suite orchestrator; run directly via adapter_matrix.py.

Run

# From suites/nl2repo — requires Docker + NL2REPO_AGENT_COMMAND_TEMPLATE
pip install -r requirements.txt
python adapter_matrix.py \
  --task-agent elizaos \
  --model-provider cerebras \
  --model gpt-oss-120b \
  --output /tmp/nl2repo-out \
  --max-tasks 1

# Original OpenHands batch runner (requires openhands Docker images + config.json creds)
python main.py

Smoke test (no Docker, no API keys)

python adapter_matrix.py \
  --task-agent elizaos \
  --model-provider cerebras \
  --model gpt-oss-120b \
  --output /tmp/nl2repo-mock \
  --max-tasks 5 \
  --mock

Test the harness

pip install -r requirements.txt pytest
pytest suites/nl2repo/tests/test_adapter_matrix.py -v

Layout

Notes

• Results write to --output <dir>/result.json (not in git; gitignored by convention).
• Scoring: score = passed / test_case_count per task; aggregate = mean across all 104.
• Eval images: ghcr.io/multimodal-art-projection/nl2repobench/<task>:1.0 (104 images, multi-GB; pulled lazily).
• Agent command is injected via NL2REPO_AGENT_COMMAND_TEMPLATE env var or --agent-command-template.
• --no-docker skips Docker post-processing (generation-only mode; scores are 0, not release-comparable).
• Dataset source and paper: INTEGRATION.md. Upstream repo: github.com/multimodal-art-projection/NL2RepoBench.

⛔ 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 (the elizaOS repo’s scenario-runner: bun 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.