adapters/gitgoodbench/README.md

GitGoodBench Adapter Results

Overview

Adapter: GitGoodBench (merge conflict resolution scenario, GitGoodBench Lite) Adapter PR: https://github.com/harbor-framework/adapters/pull/78 (migrated from https://github.com/harbor-framework/harbor/pull/3327) Original benchmark: https://github.com/JetBrains-Research/git-good-bench at b11ae6c, run through the reconstructed reference harness shipped with the adapter (parity/reference-harness.patch); the released baseline agent is not runnable. Agent: mini-swe-agent 2.4.6 (litellm 1.91.1), identical flags on both sides, commands run with /bin/sh Agent model: deepseek-v4.1-flash (2077AI parity gateway, chat completions; served as deepseek-v4-1-flash-260910) Verifier: exact match against the ground-truth merge tree (no LLM judge) Tasks: 60 Trials: 3 per side, ground-truth merge commit hidden on both sides

Results Summary

Original Parity Results (reference harness, --hide-ground-truth)

Trial1	Trial2	Trial3
22/60 = 36.67%	25/60 = 41.67%	26/60 = 43.33%

Mean 40.56 ± 2.00 (sample SEM); sample std 3.47.

Harbor Parity Results (Daytona)

Trial1	Trial2	Trial3
24/60 = 40.00%	26/60 = 43.33%	25/60 = 41.67%

Mean 41.67 ± 0.96 (sample SEM); sample std 1.67.

Run ranges overlap. Per language: original Java 25.0% / Kotlin 46.7% / Python 50.0%, Harbor 25.0% / 50.0% / 50.0%. 52 of the 60 tasks have the same outcome in all three trials on both sides; no task differs by more than one trial.

Oracle

Original side 60/60 (oracle/original/results.jsonl), Harbor side 60/60 with no exceptions (oracle/harbor/, Daytona, run on 2026-11-11 on the registered tasks with the current verifier).

Directory Structure

gitgoodbench/
├── README.md
├── config.yaml                      # Harbor job config of the parity run (3 attempts, Daytona)
├── oracle/
│   ├── harbor/                      # Harbor oracle job: config.json, result.json, job.log, and per trial
│   │                                # result.json + verifier/test-stdout.txt (60 trials)
│   └── original/results.jsonl       # reference harness oracle, one record per sample
├── original_parity/
│   ├── results.jsonl                # reference harness results, one record per sample and run
│   ├── config-shard-{0..3}.json     # harness configuration of the four shard processes
│   └── trial{1,2,3}/<sample_id>/    # per run: prompt.md, mini-swe-agent.trajectory.json,
│                                    # evaluation.json, result.json
├── harbor_parity/
│   ├── config.json, result.json, job.log   # the Harbor job (n_attempts: 3)
│   └── trial{1,2,3}/<task>__<id>/   # the job's trials split by attempt (attempt order = start time
│                                    # per task): config.json, result.json, agent/trajectory.json (ATIF),
│                                    # verifier/details.json, verifier/test-stdout.txt
└── results_collection/
    ├── result_original_trial{1,2,3}.json   # per-sample outcomes of each original-side run
    └── result_harbor_trial{1,2,3}.json     # per-task rewards of each Harbor attempt

Notes

• The Harbor job ran as one job with n_attempts: 3; harbor_parity/trial{N} groups its trials by attempt. One trial (duckduckgo-android-merge-0003) failed with a Daytona 502 while its sandbox was created and was re-run with harbor jobs resume; the job therefore reports 180 trials and no exceptions.
• On the original side three runs that a host sleep had turned into agent timeouts were removed and re-run (all three unsolved before and after); original_parity/results.jsonl holds the final records.
• Each trial directory keeps the files needed to check a run: the agent trajectory (Harbor's ATIF agent/trajectory.json on the Harbor side, mini-swe-agent's native mini-swe-agent.trajectory.json on the original side, both with the full message history), the prompt or config, the verifier or harness evaluation output and the result. Derived duplicates (mini-swe-agent's text dumps and, on the Harbor side, its native trajectory, trial.log, lock.json, artifacts/manifest.json, verifier/reward.txt) are omitted; the oracle trials keep result.json and verifier/test-stdout.txt.
• The parity gateway API key was redacted from 25 Harbor trajectories ([REDACTED-API-KEY] in agent/trajectory.json): mini-swe-agent records its full model configuration, including the key Harbor passes to it. No other credentials are present.
• The Harbor parity trials were graded in October 2026. On 2026-11-11 the verifier was changed to take the tree and the history from commit objects it has checked against their ids, instead of from Git's view of the repository (replacement refs, grafts or a commit-graph could stand in for the commits on the branch). None of the 180 Harbor runs created a replacement ref, a grafts or shallow file or a commit-graph, so the current verifier grades them the same; the oracle here was re-run with it.
• Leakage scans of all 360 trajectories (in the adapter PR, parity/results/*-leakage-scan.jsonl) show no run using the ground-truth merge commit. An earlier non-aligned run on a full clone of the original harness is documented in the adapter's parity/README.md and is not part of this upload.