__pycache__/
*.pyc

# terminal-bench trial outputs (regenerated on every run)
runs/

# Harbor venv (harness dependency, deliberately kept out of pyproject/uv.lock)
.venv-harbor/
jobs/

# SWE-bench: the 500-task dataset is downloadable (`harbor datasets download
# swe-bench/swe-bench-verified`) and every run regenerates its own artifacts.
# The dataset also carries solution/solve.sh (the gold patch) per task, which has
# no business in git.
swebench/dataset/
swebench/*results*.json
# Per-run outputs: predictions, the runner's raw results, and the official
# report. Regenerated every run and persisted to llm_benchmark_* in Postgres —
# committing them means each run dirties the tree.
swebench/predictions*.jsonl
swebench/swebench_*.json
swebench/nubi-code-analyzer.*.json
swebench/logs/
swebench/stratified16.txt
# Generated run reports. The README carries the standing result; a
# point-in-time report goes stale the next time the number moves.
swebench/REPORT-*.md
swebench/REPORT-*.pdf

# benchmark eval reports (regenerated on every run; persisted in DB)
**/report/*.json

# Runtime logs from local benchmark runs — never check these in.
# They contain real user emails (run notification recipients) and
# absolute dev paths from the run host.
*.log

# macOS metadata
.DS_Store
**/.DS_Store
