bullpen/tests/test_grinder_ticket.py

"""Spec for how the grinder reads a ticket: which repo, and which files it may edit.

Both behaviours below cost real delegation rounds on 2026-07-22/23, and both looked like
"the model is incapable" rather than "the harness is broken":

  * @foreman writes the repo as `REPO: boltzmann:/home/mfritsche/src/sic`. The parser only
    understood `repo=<path>`, so @py ground the DEFAULT repo, failed to open the test file,
    and replied with a raw FileNotFoundError traceback. Twice.
  * Target inference matched a hardcoded allowlist of top-level directories. Every repo it
    hadn't been taught about inferred NOTHING, so the model was asked to fix a file it was
    never shown — the same failure had already been papered over three times by appending
    another directory name to the regex.
"""

import importlib.machinery
import importlib.util
import os
import subprocess
import sys

import pytest

GRINDER = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
                       "bin", "bullpen-grinder")
loader = importlib.machinery.SourceFileLoader("grinder_under_test", GRINDER)
spec = importlib.util.spec_from_loader(loader.name, loader)
grinder = importlib.util.module_from_spec(spec)
sys.modules[loader.name] = grinder
loader.exec_module(grinder)


@pytest.mark.parametrize("body,expected", [
    ("please fix repo=/home/mfritsche/src/sic now", "/home/mfritsche/src/sic"),
    ("REPO: boltzmann:/home/mfritsche/src/sic\nFILE: gateway/sicd", "/home/mfritsche/src/sic"),
    ("REPO: /home/mfritsche/src/sic", "/home/mfritsche/src/sic"),
    ("repo: hertz:/opt/thing\n", "/opt/thing"),
    ("no repo mentioned at all", None),
])
def test_repo_is_parsed_from_both_ticket_dialects(body, expected):
    assert grinder._repo_from_body(body) == expected


def test_host_prefix_is_stripped_but_windowsish_colons_are_not_mangled():
    # the grind already runs ON that host, so `host:` is noise; a bare path with no
    # `host:/` shape must survive untouched.
    assert grinder._repo_from_body("REPO: boltzmann:/srv/x") == "/srv/x"
    assert grinder._repo_from_body("repo=/srv/x") == "/srv/x"


@pytest.fixture
def repo(tmp_path):
    """A tiny git repo shaped like sic: a source file the test only names piecewise."""
    subprocess.run(["git", "init", "-q", str(tmp_path)], check=True)
    (tmp_path / "gateway").mkdir()
    (tmp_path / "gateway" / "sicd").write_text("#!/usr/bin/env python3\n")
    (tmp_path / "tests").mkdir()
    (tmp_path / "tests" / "test_proto.py").write_text("x")
    (tmp_path / "README.md").write_text("docs")
    subprocess.run(["git", "-C", str(tmp_path), "add", "-A"], check=True)
    subprocess.run(["git", "-C", str(tmp_path), "-c", "user.email=t@t", "-c", "user.name=t",
                    "commit", "-qm", "init"], check=True)
    return tmp_path


def test_target_inferred_from_a_literal_path(repo):
    src = 'SICD = "gateway/sicd"\n'
    assert grinder._infer_targets(str(repo), "tests/test_proto.py", src) == ["gateway/sicd"]


def test_target_inferred_from_a_piecewise_path(repo):
    # the real sic spec: Path(__file__).resolve().parent.parent / "gateway" / "sicd"
    src = 'SICD = str(Path(__file__).resolve().parent.parent / "gateway" / "sicd")\n'
    assert grinder._infer_targets(str(repo), "tests/test_proto.py", src) == ["gateway/sicd"]


def test_the_test_file_itself_is_never_a_target(repo):
    src = 'import tests.test_proto  # gateway/sicd\n'
    out = grinder._infer_targets(str(repo), "tests/test_proto.py", src)
    assert "tests/test_proto.py" not in out
    assert all(not t.startswith("tests/") for t in out), "never let the grinder edit the tests"


def test_unrelated_repo_files_are_not_offered(repo):
    src = 'SICD = "gateway/sicd"\n'
    assert "README.md" not in grinder._infer_targets(str(repo), "tests/test_proto.py", src)


# --- spec resolution. Ticket 739 (2026-07-23) was a precise, well-formed ticket that @py
# bounced solely because it named the failing TESTS but not the FILE. A coordinator should
# not have to satisfy a regex to get work done; resolve it, or fail with something actionable.

@pytest.fixture
def files_on_host(monkeypatch):
    """Stub the git-on-the-grind-host lookup; no sic, no network."""
    def setter(listing):
        monkeypatch.setattr(grinder, "_repo_test_files", lambda repo: sorted(listing))
    return setter


def test_explicit_path_wins(files_on_host):
    files_on_host(["tests/other.py"])
    assert grinder._resolve_test_rel("run tests/test_sicd_protocol.py now", "/r") == (
        "tests/test_sicd_protocol.py", None)


def test_the_739_case_bare_test_names_and_a_single_spec_file(files_on_host):
    """The real ticket: names three failing tests, a repo, and no path at all."""
    files_on_host(["tests/test_sicd_protocol.py"])
    body = ("Grind the 3 new red tests green in boltzmann:~/src/sic.\n"
            "TICKET: repo=/home/mfritsche/src/sic\n"
            "1. test_epipe_child_dead_before_payload_write_no_traceback — BrokenPipeError\n"
            "2. test_partial_write_payload_not_truncated — got 7 of 37 bytes\n")
    path, err = grinder._resolve_test_rel(body, "/home/mfritsche/src/sic")
    assert err is None
    assert path == "tests/test_sicd_protocol.py"


def test_bare_filename_is_matched_against_the_repo(files_on_host):
    files_on_host(["tests/test_a.py", "tests/test_b.py"])
    assert grinder._resolve_test_rel("fix test_b.py please", "/r") == ("tests/test_b.py", None)


def test_sole_spec_file_is_used_when_the_ticket_names_none(files_on_host):
    files_on_host(["tests/test_only.py"])
    assert grinder._resolve_test_rel("make the red tests green", "/r") == ("tests/test_only.py", None)


def test_ambiguity_is_never_guessed_and_the_error_lists_candidates(files_on_host):
    files_on_host(["tests/test_a.py", "tests/test_b.py"])
    path, err = grinder._resolve_test_rel("make the red tests green", "/r")
    assert path is None
    assert "test_a.py" in err and "test_b.py" in err, "an error must be actionable"
    assert "SPEC:" in err, "tell the coordinator exactly what to add"


def test_named_file_absent_from_repo_is_reported_with_what_does_exist(files_on_host):
    files_on_host(["tests/test_real.py"])
    path, err = grinder._resolve_test_rel("grind test_imaginary.py", "/r")
    assert path is None
    assert "test_imaginary.py" in err and "test_real.py" in err


def test_no_tests_at_all_still_points_at_testdesigner(files_on_host):
    files_on_host([])
    path, err = grinder._resolve_test_rel("grind it", "/r")
    assert path is None
    assert "@testdesigner" in err, "keep the write-tests-first contract in the refusal"


# --- per-ticket model escalation (markus' "if DeepSeek can't crack one test, hand it to a
# stronger model" policy, made expressible in-room 2026-07-23).

@pytest.mark.parametrize("body,expected", [
    ("grind this  tiers=gpt-oss-120b", "gpt-oss-120b"),
    ("fix it  model=kimi-k3", "kimi-k3"),
    ("ladder tiers=deepseek-v4-flash-dspark,gpt-oss-120b,kimi-k3",
     "deepseek-v4-flash-dspark,gpt-oss-120b,kimi-k3"),
    ("no override here", None),
    ("MODELS=gpt-oss-120b case-insensitive", "gpt-oss-120b"),
])
def test_tiers_parsed_from_ticket_body(body, expected):
    assert grinder._tiers_from_body(body) == expected


def test_tiers_override_does_not_eat_the_word_models_in_prose():
    # a mention like "the models are slow" must not be read as an override
    assert grinder._tiers_from_body("the models are slow today") is None