orbi-cli 0.7.25__tar.gz → 0.7.26__tar.gz

{orbi_cli-0.7.25 → orbi_cli-0.7.26}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: orbi-cli
-Version: 0.7.25
+Version: 0.7.26
 Summary: Orbi — local AI development worker: claims GitHub Issues, runs Pi in isolated worktrees, and delivers PRs (GitHub Issues + labels are the only state store).
 Author: xqliu
 License-Expression: AGPL-3.0-only OR LicenseRef-Sustainable-Use-1.0

{orbi_cli-0.7.25 → orbi_cli-0.7.26}/pyproject.toml

@@ -44,7 +44,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "orbi-cli"
-version = "0.7.25"
+version = "0.7.26"
 description = "Orbi — local AI development worker: claims GitHub Issues, runs Pi in isolated worktrees, and delivers PRs (GitHub Issues + labels are the only state store)."
 readme = "README.md"
 requires-python = ">=3.14"

{orbi_cli-0.7.25 → orbi_cli-0.7.26}/src/orbi/__init__.py

@@ -18,4 +18,4 @@ the package would shadow the installed package for every process with
 the checkout root on sys.path. The direct-execution compatibility entry
 is `python3 -m orbi.cli` (development path only).
 """
-__version__ = "0.7.25"
+__version__ = "0.7.26"

{orbi_cli-0.7.25 → orbi_cli-0.7.26}/src/orbi_cli.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: orbi-cli
-Version: 0.7.25
+Version: 0.7.26
 Summary: Orbi — local AI development worker: claims GitHub Issues, runs Pi in isolated worktrees, and delivers PRs (GitHub Issues + labels are the only state store).
 Author: xqliu
 License-Expression: AGPL-3.0-only OR LicenseRef-Sustainable-Use-1.0

{orbi_cli-0.7.25 → orbi_cli-0.7.26}/tests/test_prompt_contract.py

@@ -886,3 +886,104 @@ def test_prompt_md_step_4_does_not_use_pytest_as_completion_evidence():
     text = _text(PROMPT)
     assert "unit tests alone are not completion evidence" in text
     assert "pytest alone is not completion evidence" not in text
+
+
+# --- Issue #1980: a delivery changes only what the ticket asks for -----------
+
+# The beta incident (Issue #1980) had two shapes on the same evening: ticket
+# #260 -> PR #261 deleted a false-red test file, and ticket #262 -> PR #265
+# removed an `exit 1` CI job from .github/workflows/ci.yml, both unmentioned,
+# review passed, both merged. The base-red gate itself works (#254 paused), so
+# the missing piece is the scope rule: a delivery changes only what the ticket
+# asks for and never edits unrelated code, tests or CI configuration to get its
+# own checks green. Both prompts must state it up front; these tests render the
+# real templates (the exact string Pi receives) so the rule cannot silently
+# drift away.
+
+BASE_RED_PROMPT_ITEMS = (
+    # The general scope sentence (the ticket's acceptance 1), not a per-file
+    # list: unrelated code, tests AND CI configuration are all named.
+    ("scope-only-ticket", "change only what the ticket asks for"),
+    ("scope-unrelated-files",
+     "do not edit unrelated code, tests or ci configuration to make your own "
+     "checks pass"),
+    # A check already red at the frozen base belongs to the base, not the
+    # ticket — the Runner's base-red pause decides it.
+    ("scope-base-check-not-mine",
+     "a check already failing at "
+     "`77e7f0ea54d8e9748f2debfd5a8f49739dd89b66` is not this ticket's to fix"),
+    ("base-red-failing-test",
+     "already fails at `77e7f0ea54d8e9748f2debfd5a8f49739dd89b66`"),
+    ("base-red-outside", "outside this ticket"),
+    ("base-red-forbidden-actions", "delete, skip, xfail or edit it"),
+    ("base-red-no-weaken", "do not weaken its assertion"),
+    ("base-red-leave-failing", "leave it failing"),
+    ("base-red-runner-handling", "base-red handling"),
+    # Step 5 forbids committing a change whose tests were seen failing; the
+    # carve-out keeps that rule from stalling the delivery on the base's own
+    # red test (the ticket must still end in a commit).
+    ("base-red-step5-carve-out",
+     "never includes a test that already fails at "
+     "`77e7f0ea54d8e9748f2debfd5a8f49739dd89b66`"),
+)
+
+
+def test_prompt_md_leaves_a_test_that_already_fails_on_the_base_alone():
+    """Issue #1980: the rendered implementer prompt carries the scope sentence
+    — change only what the ticket asks for, never unrelated code, tests or CI
+    configuration to get checks green — and states that a check already
+    failing at the frozen base SHA is outside the ticket, never deleted,
+    skipped, xfail-ed or weakened; it stays failing and the Runner's base-red
+    handling takes it from there."""
+    from orbi import pi_session
+
+    rendered = pi_session.render_prompt(
+        PROMPT.read_text(encoding="utf-8"),
+        {"BASE_SHA": "77e7f0ea54d8e9748f2debfd5a8f49739dd89b66",
+         "BASE_BRANCH": "main"},
+    )
+    text = " ".join(rendered.lower().split())
+    missing = _missing(text, BASE_RED_PROMPT_ITEMS)
+    assert not missing, (
+        f"prompt.md does not leave a base-red test alone (Issue #1980): "
+        f"{missing}"
+    )
+    # Rendered against a real base SHA, the rule names that SHA.
+    assert "77e7f0ea54d8e9748f2debfd5a8f49739dd89b66" in text
+
+
+BASE_RED_REVIEW_ITEMS = (
+    # Acceptance 2: out-of-scope changes made to get checks green are a
+    # blocking finding, and the rule names all three file kinds, so a removed
+    # CI job (the #265 shape) is caught like a deleted test (the #261 shape).
+    ("review-base-red-scope",
+     "out-of-scope changes made to get checks green are a **blocking "
+     "finding (blocker)**"),
+    ("review-base-red-delete", "deleting, skipping or weakening"),
+    ("review-base-red-unrelated", "unrelated existing test"),
+    ("review-base-red-other-files",
+     "editing unrelated code, tests or ci configuration"),
+    ("review-base-red-ci", "disabled ci job"),
+)
+
+
+def test_prompt_review_md_flags_a_deleted_or_weakened_existing_test():
+    """Issue #1980: the rendered review prompt treats out-of-scope changes
+    made to get checks green — a deleted or weakened test, unrelated code or
+    CI configuration, a disabled CI job — as a blocking finding, so the PR
+    that removes the customer's red test or red job is rejected instead of
+    merged."""
+    from orbi import pi_session
+
+    rendered = pi_session.render_prompt(
+        PROMPT_REVIEW.read_text(encoding="utf-8"),
+        {"BASE_SHA": "77e7f0ea54d8e9748f2debfd5a8f49739dd89b66",
+         "BASE_BRANCH": "main"},
+    )
+    text = " ".join(rendered.lower().split())
+    missing = _missing(text, BASE_RED_REVIEW_ITEMS)
+    assert not missing, (
+        f"prompt_review.md does not flag a deleted or weakened existing test "
+        f"(Issue #1980): {missing}"
+    )
+    assert "77e7f0ea54d8e9748f2debfd5a8f49739dd89b66" in text