# Category 3 (terminal/CLI agent): fixed test tasks v1 (2026-10-14)
Each task runs in a fresh copy of fixtures/tN, which is a git repo. The agent gets ONE non-interactive prompt with no human follow-up. PASS = `python3 -m pytest -q` exits 0 afterward, and nobody edited the files by hand.
- T1 fix-failing-test: "The tests in test_stats.py fail. Fix the bug in stats.py so all tests pass. Do not modify the tests."
- T2 implement-from-tests: "Add a slugify(s) function to text.py so that test_text.py passes. Do not modify the tests."
- T3 multi-file-rename: "Rename calc_total to compute_total in cart.py and update every caller (checkout.py) so test_cart.py passes. Do not modify the tests."
Model for the zero-cost run: qwen2.5-coder-3b-instruct Q4_K_M GGUF (sha256 724fb256…730b7), run locally with Ollama 0.40.4 on CPU. A 3B local model is far weaker than the frontier models these tools normally use, so treat these results as a check of the harness and workflow, not of the tool's best-case capability.