orbi-cli 0.7.24__tar.gz → 0.7.26__tar.gz

{orbi_cli-0.7.24 → orbi_cli-0.7.26}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: orbi-cli
-Version: 0.7.24
+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-0.7.24 → orbi_cli-0.7.26}/pyproject.toml

@@ -44,7 +44,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "orbi-cli"
-version = "0.7.24"
+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-0.7.24 → 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__ = "0.7.24"
+__version__ = "0.7.26"

{orbi_cli-0.7.24 → orbi_cli-0.7.26}/src/orbi/claim.py

@@ -18,7 +18,8 @@ from pathlib import Path
 from orbi import base_ci_red, check_permission, gitops, ticket_command
 # Imported for its registration side effect: `ai-needs-detail` tickets run
 # the registered commands below, and `/orbi split` (Issue #1927) and
-# `/orbi add-criteria` (Issue #1929) must be in the registry by then.
+# `/orbi amend` (Issue #1929, typed word renamed by #1975) must be in the
+# registry by then.
 from orbi import criteria_command, split_command  # noqa: F401
 from orbi import milestone as milestone_bookkeeping
 from orbi.config import RunnerConfig

{orbi_cli-0.7.24 → orbi_cli-0.7.26}/src/orbi/clarify.py

@@ -133,7 +133,8 @@ CLARIFY_SYSTEM_PROMPT = (
     "state. When the text lacks a fact a proposal needs, or writing one "
     "needs a decision only a person can make, omit the argument and "
     "answer with missing alone.\n"
-    '- ONLY single_outcome failed: suggestion {"kind": "split", '
+    '- single_outcome failed, with or without "acceptance_condition" '
+    'also failing: suggestion {"kind": "split", '
     '"items": [{"result": "...", "acceptance": "...", "after": 1}]}. '
     "One item per new ticket, each result and acceptance stated in the "
     "text: the fewest vertical slices that cover every goal of the "
@@ -215,7 +216,9 @@ class ClarifyVerdict:
 
     "suggestion" is the optional proposal the model attached to a failing
     verdict (Issue #1926): present only when the ticket text was enough to
-    write one, and rendered only when it answers the one missing piece.
+    write one, and rendered only when it answers the failed checks — a split
+    answers `single_outcome` with or without a missing acceptance condition
+    (Issue #1978), a criteria draft an acceptance condition alone.
     "suggestion_dropped" says WHY a proposal the answer carried was
     discarded (Issue #1958) — "misplaced" when the keys sat at the top
     level instead of under "suggestion", "unusable" when the value failed
@@ -244,15 +247,6 @@ SUGGESTION_START = f"<!-- {SUGGESTION_MARKER} kind="
 SUGGESTION_END = "<!-- /orbi/suggestion -->"
 SUGGESTION_START_PREFIX = f"<!-- {SUGGESTION_MARKER}"
 
-# The one missing piece each proposal kind answers: a split plan proposes
-# the split a "single_outcome" failure asks for, a criteria draft the
-# acceptance an "acceptance_condition" failure asks for. Every other
-# verdict renders no proposal.
-SUGGESTION_FOR_MISSING = {
-    "single_outcome": SUGGESTION_SPLIT,
-    "acceptance_condition": SUGGESTION_CRITERIA,
-}
-
 # The one command sentence each proposal kind adds below its edit sentence
 # (Issue #1927 for a split, Issue #1929 for criteria): the reader never
 # copies the lines by hand, and the command reads the SAME section.
@@ -404,17 +398,22 @@ def render_suggestion(suggestion: Suggestion, language_code: str) -> str:
 def verdict_suggestion(verdict: ClarifyVerdict, language_code: str) -> str:
     """The proposal section for one failing verdict, or "".
 
-    Rendered only when exactly one piece is missing, it is the piece this
-    proposal kind answers, and the proposal is usable. Every other stop
-    keeps the pre-#1926 comment byte for byte: no section, no edit
-    sentence.
+    A split is rendered whenever `single_outcome` is among the failed checks,
+    with or without a missing acceptance condition (Issue #1978): every split
+    item carries its own acceptance line, so one split plan answers both
+    failures, and a ticket that should be split gets the split plan. The
+    criteria draft keeps its single-piece rule — it answers an
+    `acceptance_condition` failure alone (Issue #1926). An unusable proposal
+    renders nothing, so every other stop keeps the pre-#1926 comment byte for
+    byte: no section, no edit sentence.
     """
     suggestion = verdict.suggestion
-    if suggestion is None or len(verdict.missing) != 1:
-        return ""
-    if suggestion.kind != SUGGESTION_FOR_MISSING.get(verdict.missing[0]):
+    if suggestion is None or not _usable(suggestion):
         return ""
-    if not _usable(suggestion):
+    if suggestion.kind == SUGGESTION_SPLIT:
+        if "single_outcome" not in verdict.missing:
+            return ""
+    elif verdict.missing != ("acceptance_condition",):
         return ""
     return render_suggestion(suggestion, language_code)
 
@@ -813,9 +812,9 @@ def render_comment(issue: dict, verdict: ClarifyVerdict, run_id: str, *,
     pieces = "\n".join(
         f"- {missing_text(item, language_code)}" for item in verdict.missing
     )
-    # The editable proposal, only when this verdict's one missing piece is
-    # the piece the proposal answers (Issue #1926). Without one the string
-    # is empty, so the comment stays byte for byte as before.
+    # The editable proposal, only when it answers this verdict's failed
+    # checks (Issue #1926, #1978). Without one the string is empty, so the
+    # comment stays byte for byte as before.
     proposal = verdict_suggestion(verdict, language_code)
     # A proposal IS the instruction (Issue #1962): the comment then carries
     # the proposal, its edit sentence and its command sentence, and never

{orbi_cli-0.7.24 → orbi_cli-0.7.26}/src/orbi/criteria_command.py

@@ -1,9 +1,9 @@
-"""`/orbi add-criteria`: put the proposed acceptance criteria on the ticket.
+"""`/orbi amend`: put the proposed acceptance criteria on the ticket.
 
 The clarification gate stops a thin ticket and may attach a `kind=criteria`
 proposal to its stop comment (Issue #1926). This module registers the one
 command that acts on it: a person with write permission replies
-`/orbi add-criteria` and the proposed conditions — **as the comment reads at
+`/orbi amend` and the proposed conditions — **as the comment reads at
 that moment**, so editing it first changes what lands — are appended to the
 ticket body under an acceptance heading that follows the repository's
 `message_language` (`## Acceptance` in English, `## 验收` in Chinese,
@@ -231,12 +231,12 @@ def _criteria_apply(
 
 ADD_CRITERIA_COMMAND = CommandSpec(
     name="add_criteria",
-    subcommand="add-criteria",
+    subcommand="amend",
     arity=0,
     permission="write",
     validate=_criteria_validate,
     apply=_criteria_apply,
-    usage="/orbi add-criteria",
+    usage="/orbi amend",
     description=(
         "append the proposed acceptance criteria to this ticket and queue it"
     ),

{orbi_cli-0.7.24 → orbi_cli-0.7.26}/src/orbi/language_terms.py

@@ -246,8 +246,8 @@ CLARIFY_SUGGESTION_MESSAGES: dict[str, tuple[str, str]] = {
     # names the command that appends these conditions to the ticket and
     # starts it, so the reader never copies them into the body by hand.
     "clarify.suggestion.criteria_reply": (
-        "Reply `/orbi add-criteria` to add these and start.",
-        "回复 `/orbi add-criteria` 即可把这些加入工单并开始交付。",
+        "Reply `/orbi amend` to add these and start.",
+        "回复 `/orbi amend` 即可把这些加入工单并开始交付。",
     ),
     # The language the PROPOSAL VALUES are written in (Issue #1971): the
     # engine's own setting, named in the gate's prompt instead of left to
@@ -343,11 +343,12 @@ SPLIT_MESSAGES: dict[str, tuple[str, str]] = {
     ),
 }
 
-# The `/orbi add-criteria` command (Issue #1929): its rejections and the ONE
-# receipt it leaves. Every entry is display text; the machine contract (the
-# markers, the labels) stays in English and outside the catalog. The heading
-# is display text too: it follows the message language (Issue #1971), while
-# the marker that makes the append idempotent is the same in every language.
+# The `/orbi amend` command (Issue #1929, typed word renamed by #1975):
+# its rejections and the ONE receipt it leaves. Every entry is display
+# text; the machine contract (the markers, the labels) stays in English
+# and outside the catalog. The heading is display text too: it follows the
+# message language (Issue #1971), while the marker that makes the append
+# idempotent is the same in every language.
 ADD_CRITERIA_MESSAGES: dict[str, tuple[str, str]] = {
     "criteria.reason_no_suggestion": (
         "the newest stop comment carries no acceptance-criteria proposal; "
@@ -362,10 +363,10 @@ ADD_CRITERIA_MESSAGES: dict[str, tuple[str, str]] = {
     ),
     "criteria.reason_not_criteria": (
         "the newest stop comment proposes `{kind}`, not acceptance "
-        "criteria; `/orbi add-criteria` only adds a `kind=criteria` "
+        "criteria; `/orbi amend` only adds a `kind=criteria` "
         "proposal",
         "最新一条停止评论的提案是 `{kind}`，不是验收条件；"
-        "`/orbi add-criteria` 只添加 `kind=criteria` 提案",
+        "`/orbi amend` 只添加 `kind=criteria` 提案",
     ),
     "criteria.receipt_title": (
         "**Added the proposed acceptance criteria to this ticket.**",

{orbi_cli-0.7.24 → orbi_cli-0.7.26}/src/orbi_cli.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: orbi-cli
-Version: 0.7.24
+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-0.7.24 → orbi_cli-0.7.26}/tests/test_clarify.py

@@ -1031,6 +1031,47 @@ def fake_pi_answer(answer: dict) -> str:
     )
 
 
+@pytest.mark.parametrize("language_code", ["en", "zh"])
+def test_the_real_runner_proposes_the_split_for_a_bundled_ticket(
+    clone, tmp_path, monkeypatch, language_code,
+):
+    """Acceptance 1 (Issue #1978) at the real entry point: a ticket that
+    fails `single_outcome` AND `acceptance_condition` (the evidence shape:
+    xqliu/orbi-e2e-2609260042#253) is stopped with the split plan, not with
+    the generic repair block. `/orbi split` reads that very section."""
+    fix_run_id(monkeypatch, RUN_ID)
+    answer = {
+        "missing": ["acceptance_condition", "single_outcome"],
+        "suggestion": SPLIT_ANSWER["suggestion"],
+    }
+    install_fake_pi(monkeypatch, tmp_path, fake_pi_answer(answer))
+    comments: list[str] = []
+    labels = {ISSUE_NUMBER: [delivery_labels.READY_LABEL]}
+    install_fake_gh(monkeypatch, comments, labels)
+
+    result = runner.process_issue(
+        THIN_ISSUE,
+        gate_config(clone, tmp_path, message_language=language_code),
+        REPO,
+    )
+
+    assert result == runner.IssueResult("needs-detail", None)
+    assert len(comments) == 1
+    body = comments[0]
+    assert (FIXTURES / f"split.{language_code}.md").read_text() in body
+    # The section the comment carries is exactly what `/orbi split` parses.
+    assert clarify.parse_suggestion(body) == SPLIT_PROPOSAL
+    assert language.text(
+        language_code, clarify.SUGGESTION_REPLY[clarify.SUGGESTION_SPLIT],
+    ) in body
+    # One instruction: no generic repair block beside the proposal.
+    assert language.text(
+        language_code, "clarify.repair", ready_label="ai-ready",
+    ) not in body
+    assert labels[ISSUE_NUMBER] == [delivery_labels.NEEDS_DETAIL_LABEL]
+    assert not worktree_for(clone, RUN_ID).exists()
+
+
 @pytest.mark.parametrize("language_code", ["en", "zh"])
 @pytest.mark.parametrize("answer, kind", [
     (SPLIT_ANSWER, "split"),
@@ -1382,6 +1423,20 @@ def test_the_rendered_proposal_matches_its_fixture(kind, language_code):
     assert block.rstrip().endswith(tail)
 
 
+@pytest.mark.parametrize("language_code, expected", [
+    ("en", "Reply `/orbi amend` to add these and start."),
+    ("zh", "回复 `/orbi amend` 即可把这些加入工单并开始交付。"),
+])
+def test_the_criteria_proposal_tells_the_reader_to_reply_amend(
+    language_code, expected,
+):
+    """Acceptance 3 (Issue #1975): the rendered stop comment names the typed
+    command word `amend` in both languages."""
+    assert language.text(
+        language_code, clarify.SUGGESTION_REPLY[clarify.SUGGESTION_CRITERIA],
+    ) == expected
+
+
 @pytest.mark.parametrize("language_code", LANGUAGES)
 @pytest.mark.parametrize("kind", sorted(PROPOSALS))
 def test_the_stop_comment_carries_the_proposal_and_the_edit_sentence(
@@ -1444,15 +1499,101 @@ def test_a_stop_without_a_proposal_keeps_the_repair_block(language_code):
     assert clarify.SUGGESTION_START_PREFIX not in body
 
 
+@pytest.mark.parametrize("language_code", LANGUAGES)
+def test_a_bundled_ticket_gets_the_split_proposal(language_code):
+    """Acceptance 1 (Issue #1978): the most common broken shape — several
+    unrelated goals and no acceptance condition — gets the split plan.
+
+    Every split item carries its own acceptance line, so one split answers
+    both failures; the comment still gives ONE instruction (Issue #1962):
+    the proposal and the `/orbi split` sentence, never the generic repair
+    block beside it."""
+    verdict = _failing(
+        ("acceptance_condition", "single_outcome"), SPLIT_PROPOSAL,
+    )
+    body = clarify.render_comment(
+        ISSUE, verdict, RUN_ID, language_code=language_code,
+    )
+    block = clarify.verdict_suggestion(verdict, language_code)
+    assert block
+    assert block in body
+    assert block.startswith(clarify.SUGGESTION_START + "split -->")
+    # Both failed pieces are still named, and before the instruction.
+    for item in ("acceptance_condition", "single_outcome"):
+        assert clarify.missing_text(item, language_code) in body
+        assert body.index(clarify.missing_text(item, language_code)) < (
+            body.index(block)
+        )
+    # The split instruction, and only it: no criteria sentence, no generic
+    # repair block, no `Outcome/Acceptance` template, no follow-up.
+    reply = language.text(
+        language_code, clarify.SUGGESTION_REPLY[clarify.SUGGESTION_SPLIT],
+    )
+    assert reply in body
+    assert language.text(
+        language_code,
+        clarify.SUGGESTION_REPLY[clarify.SUGGESTION_CRITERIA],
+    ) not in body
+    for key in (
+        "clarify.repair", "clarify.repair_outcome",
+        "clarify.repair_acceptance", "clarify.followup",
+    ):
+        assert language.text(
+            language_code, key, ready_label="ai-ready",
+            needs_detail_label=delivery_labels.NEEDS_DETAIL_LABEL,
+        ) not in body
+    assert "```" not in body
+
+
+@pytest.mark.parametrize("language_code", LANGUAGES)
+def test_a_split_is_rendered_whatever_else_is_missing(language_code):
+    """Issue #1978: the renderer maps a verdict that fails
+    `single_outcome` to the split proposal, whatever else also failed —
+    one rule, not a per-pair special case."""
+    verdict = _failing(
+        ("single_outcome", "observable_result"), SPLIT_PROPOSAL,
+    )
+    body = clarify.render_comment(
+        ISSUE, verdict, RUN_ID, language_code=language_code,
+    )
+    assert clarify.SUGGESTION_START + "split -->" in body
+    assert language.text(
+        language_code, clarify.SUGGESTION_REPLY[clarify.SUGGESTION_SPLIT],
+    ) in body
+
+
+@pytest.mark.parametrize("language_code", LANGUAGES)
+def test_an_acceptance_only_stop_still_renders_the_criteria_draft(
+    language_code,
+):
+    """Acceptance 2 (Issue #1978): a stop for `acceptance_condition` alone
+    keeps today's criteria draft and its `/orbi amend` sentence, and never
+    the split instruction."""
+    verdict = _failing(("acceptance_condition",), CRITERIA_PROPOSAL)
+    body = clarify.render_comment(
+        ISSUE, verdict, RUN_ID, language_code=language_code,
+    )
+    block = clarify.verdict_suggestion(verdict, language_code)
+    assert block
+    assert block.startswith(clarify.SUGGESTION_START + "criteria -->")
+    assert block in body
+    assert language.text(
+        language_code, clarify.SUGGESTION_REPLY[clarify.SUGGESTION_CRITERIA],
+    ) in body
+    assert language.text(
+        language_code, clarify.SUGGESTION_REPLY[clarify.SUGGESTION_SPLIT],
+    ) not in body
+    assert "```" not in body
+
+
 @pytest.mark.parametrize("suggestion, missing", [
     # The model declined, or could not propose: nothing is rendered.
     (None, ("single_outcome",)),
     (None, ("acceptance_condition",)),
-    # Kind mismatch: each proposal answers one missing piece only.
+    # Kind mismatch: the split answers `single_outcome` and the criteria
+    # draft `acceptance_condition` alone, never each other's failure.
     (SPLIT_PROPOSAL, ("acceptance_condition",)),
     (CRITERIA_PROPOSAL, ("single_outcome",)),
-    # A second missing piece: the proposal would cover only part of it.
-    (SPLIT_PROPOSAL, ("single_outcome", "observable_result")),
     (CRITERIA_PROPOSAL, ("acceptance_condition", "single_outcome")),
     (SPLIT_PROPOSAL, ("observable_result",)),
     # Unusable proposals.
@@ -1512,6 +1653,22 @@ def test_an_unusable_proposal_keeps_the_stop_comment_byte_for_byte(
     assert "Editing this comment" not in with_proposal
 
 
+def test_the_gate_prompt_allows_the_split_when_acceptance_also_fails():
+    """Acceptance 3 (Issue #1978): the split proposal is allowed whenever
+    `single_outcome` fails, with or without a missing acceptance condition,
+    because every split item carries its own acceptance line; the criteria
+    draft keeps its "ONLY acceptance_condition" rule."""
+    prompt = clarify.CLARIFY_SYSTEM_PROMPT
+    assert (
+        '- single_outcome failed, with or without "acceptance_condition" '
+        'also failing: suggestion {"kind": "split", '
+    ) in prompt
+    assert '- ONLY acceptance_condition failed: suggestion {"kind": ' in prompt
+    # The pre-#1978 restriction is gone: a split was proposed only when
+    # `single_outcome` failed alone.
+    assert "ONLY single_outcome failed" not in prompt
+
+
 def test_the_gate_prompt_names_the_configured_output_language():
     """Issue #1971: the proposal's VALUES follow the repository's
     `message_language`, named from the catalog.
@@ -1536,7 +1693,7 @@ def test_the_gate_prompt_guards_the_proposal():
     missing (Issue #1926)."""
     prompt = clarify.CLARIFY_SYSTEM_PROMPT
     for marker in (
-        "ONLY single_outcome failed",
+        'single_outcome failed, with or without "acceptance_condition" ',
         "ONLY acceptance_condition failed",
         "adding no goal or scope the text does not state",
         "no count cap",

{orbi_cli-0.7.24 → orbi_cli-0.7.26}/tests/test_criteria_command.py

@@ -1,5 +1,5 @@
-"""`/orbi add-criteria`: the proposed acceptance criteria start the ticket
-(Issue #1929).
+"""`/orbi amend`: the proposed acceptance criteria start the ticket
+(Issue #1929; renamed from `add-criteria` in Issue #1975).
 
 Every scenario here drives the REAL tick entry (`claim.pick_next_delivery`)
 over the in-memory gh fake: the stopped ticket is picked up by the
@@ -15,13 +15,14 @@ import subprocess
 import pytest
 
 import orbi.criteria_command as criteria_command
-from orbi import claim, clarify, journal, language, ticket_command
+from orbi import claim, clarify, cli, journal, language, ticket_command
 from orbi.config import RunnerConfig
 from orbi.delivery_labels import NEEDS_DETAIL_LABEL, READY_LABEL
 from orbi.resume_scan import ResumeHooks
 from seam import seam
 
 from tests.fakes.github import FakeGh
+from tests.test_cli_milestone import make_world
 
 REPO = "acme/app"
 TICKET = 100
@@ -92,7 +93,7 @@ def stopped_ticket(
 
 
 def reply(gh: FakeGh, *, login: str = "maintainer",
-          command: str = "/orbi add-criteria", permission: str = "write",
+          command: str = "/orbi amend", permission: str = "write",
           number: int = TICKET) -> None:
     gh.comment(number, command, login=login)
     gh.add_collaborator(login, permission)
@@ -537,9 +538,9 @@ def test_the_command_is_registered_with_the_shared_layer():
     spec = ticket_command.TICKET_COMMANDS["add_criteria"]
     assert spec is criteria_command.ADD_CRITERIA_COMMAND
     assert (spec.name, spec.subcommand, spec.arity, spec.permission) == (
-        "add_criteria", "add-criteria", 0, "write",
+        "add_criteria", "amend", 0, "write",
     )
-    assert spec.usage == "/orbi add-criteria"
+    assert spec.usage == "/orbi amend"
     assert spec.aliases == ()
     assert "add_criteria_command_applied" in journal.JOURNAL_EVENTS
     assert "add_criteria_command_failed" in journal.JOURNAL_EVENTS
@@ -581,7 +582,7 @@ def test_apply_refuses_a_ticket_without_a_proposal(monkeypatch):
 
 
 def test_the_terminal_spelling_runs_the_same_append(monkeypatch):
-    """`orbi add-criteria --repo … --issue N` (Issue #1930) is the same
+    """`orbi amend --repo … --issue N` (Issues #1930, #1975) is the same
     execution path: the host login is the one occurrence and the ticket
     itself supplies the proposal."""
     gh = _gh(monkeypatch)
@@ -604,6 +605,49 @@ def test_the_terminal_spelling_runs_the_same_append(monkeypatch):
     assert "comment=cli@" in receipts[0]
 
 
+def test_the_terminal_subcommand_is_orbi_amend(tmp_path, monkeypatch):
+    """Acceptance 4 (Issue #1975): `orbi amend --repo … --issue N` on the
+    host runs the same command through the real CLI entry."""
+    config_path = make_world(tmp_path, milestone=None, repo=REPO)
+    gh = _gh(monkeypatch)
+    stopped_ticket(gh)
+    gh.add_collaborator("orbi-bot", "write")
+
+    exit_code = cli.main([
+        "amend", "--repo", REPO, "--issue", str(TICKET),
+        "--config", str(config_path),
+    ])
+
+    assert exit_code == 0
+    body = gh.issues[TICKET]["body"]
+    assert body.count(APPENDED_MARKER) == 1
+    assert "- The tag survives a reload" in body
+    assert READY_LABEL in labels(gh)
+    assert NEEDS_DETAIL_LABEL not in labels(gh)
+    receipts = success_receipts(gh)
+    assert len(receipts) == 1
+    assert "comment=cli@" in receipts[0]
+
+
+def test_the_old_word_is_not_a_command(monkeypatch, tmp_path):
+    """Acceptance 2 (Issue #1975): the command word was renamed and no alias
+    kept, so the old spelling is an ordinary comment: nothing is written and
+    no receipt is posted."""
+    gh = _gh(monkeypatch)
+    stopped_ticket(gh)
+    reply(gh, command="/orbi add-criteria")
+
+    tick(monkeypatch, tmp_path, gh)
+
+    assert gh.issues[TICKET]["body"] == BODY
+    assert READY_LABEL not in labels(gh)
+    assert NEEDS_DETAIL_LABEL in labels(gh)
+    assert [
+        body for body in comment_bodies(gh)
+        if "orbi-add_criteria-command" in body
+    ] == []
+
+
 def test_the_context_reads_the_ticket_and_its_proposal(monkeypatch):
     """The comment scan hands neither over, so the command reads both from
     the ticket itself; a caller that already has them costs no second read."""

{orbi_cli-0.7.24 → 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