.gitignore

__pycache__/
*.egg-info/
.venv/
reports/*.json
!reports/.gitkeep
# The published run JSON is the source of truth the committed .md projects from.
# tests/test_report_golden.py reads it, so a fresh clone must have it.
#
# This allowlist governs EVERY command that commits a .md, confirm included. A committed .md
# whose .json is ignored is an unfalsifiable artifact: a fresh clone gets the claim without the
# evidence it was projected from, and no golden test can check it has not drifted. If a new
# command starts publishing a report, its JSON belongs here in the same change.
# 2026-07-21: reports/ pruned to the qwen3-1.7b MLX-only run. confirm's and both gemma4
# reports' JSONs were removed, so their allowlist lines went with them -- a `!` line for a
# file that no longer exists is exactly the stale record this allowlist is meant not to keep.
# Re-add a line here in the same change that republishes any of them.
!reports/profile-qwen3-1.7b.json
!reports/prescribe-qwen3-1.7b.json
docs/spikes/*.tmp
reports/spike-pi-agent/*.log
.DS_Store
docs/superpowers/
# Local working guide for AI assistants. Same convention as toolhound: this repo keeps AI
# scaffolding out of its public history. The tracked record of a design decision goes into the
# Chinese design doc, not here.
CLAUDE.md

# v2 Pi-agent spike run artifacts: per-task sandboxes + multi-MB trajectory JSONL.
# Regenerable (`python -m scripts.spike.pi_agent.run`), and the sandboxes contain real
# Exercism *_test.py files. The committed record of these runs is
# docs/spikes/pi-agent-spike.md; the raw trajectories stay local for rescoring.
reports/spike-pi-agent*/