Amorth710/Zero2repo_Trajectory

Zero2Repo Trajectories

Agent trajectories, delivered code and verified reference solutions for the 43 cases of Zero2Repo, a benchmark in which a coding agent receives a product spec and an empty workspace, builds the whole project, and is scored by hidden, deterministic acceptance tests (binary reward 1.0 / 0.0, no LLM judge).

The cases themselves (public spec, interface contract, hidden tests, environment recipe) live in the GitHub repository. This dataset holds what was produced when models solved them.

Contents

• 1032 runs: 8 models × 43 cases × 3 attempts (pass@3), November 2026.
• 43 reference solutions, each scored reward 1.0 against the same hidden tests.

Model	Harness	Runs	pass@1 (mean of 3)	Cases solved in ≥1 of 3
gpt-6-astra	Codex CLI 0.166.0	129	35.7%	18
gpt-6.1-sol	Codex CLI 0.166.0	129	33.3%	16
claude-opus-5-5	Claude Code 2.1.280	129	31.8%	18
claude-sonnet-5-5	Claude Code 2.1.280	129	26.4%	13
grok-4.7-high	Cursor Agent CLI 2026.11.05-86047df	129	17.8%	10
kimi-k3.1-pro	Claude Code 2.1.280	129	15.5%	8
deepseek-v4.1-flash	Claude Code 2.1.280	129	14.7%	7
qwen3.8-max	Claude Code 2.1.280	129	7.8%	5

All runs used effort high and temperature 1 where the harness allows setting it, a 2-hour solve limit, and the same hidden tests. A run whose judge was still executing the hidden tests at the judge's time limit counts as a failure (the delivered code was too slow or hung).

Layout

runs.jsonl                               one metadata line per run
runs/<model>/<case>/run-<k>/
    metadata.json                        model, harness and version, reward, failed tests,
                                         solve time, token usage
    trajectory/agent.log                 the harness's event stream
    trajectory/session/                  the harness's native session files
    deliverable.tar.gz                   text files of the judged workspace at the end of the solve
                                         (dependency trees, build caches and binaries left out;
                                         counts in metadata.json)
reference/<case>/code/                   the reference solution
reference/<case>/verification.json       its judge result on the hidden tests

A few runs have no deliverable.tar.gz: they ended (timeout or error) with an empty workspace, as their metadata.json records.

Token usage is recomputed from the logs: Claude Code reports running per-session totals, so each session's last total is used. All credentials and API endpoint addresses were removed from every file; an endpoint address appears as <api-gateway>.

Notes

• metadata.json records both the requested model and the model name the serving endpoint reported. For some DeepSeek and Qwen runs the endpoint did not report the exact snapshot (deepseek-v4.1-flash-1017, qwen3.8-max-0902), so the snapshot of those runs cannot be confirmed.
• Reference solutions are derived from open-source projects under their own licenses; the identifiers in the public specs are neutralized.
• Grok runs used cursor-agent 2026.11.05-86047df (the installer did not honour the requested pin).