ouroboros-ai 0.38.3.dev135__tar.gz → 0.38.3.dev136__tar.gz

{ouroboros_ai-0.38.3.dev135 → ouroboros_ai-0.38.3.dev136}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ouroboros-ai
-Version: 0.38.3.dev135
+Version: 0.38.3.dev136
 Summary: Specification-first workflow engine for AI coding agents. Works with Claude Code and Codex CLI.
 Author-email: Q00 <jqyu.lee@gmail.com>
 License-File: LICENSE

{ouroboros_ai-0.38.3.dev135 → ouroboros_ai-0.38.3.dev136}/src/ouroboros/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.38.3.dev135'
-__version_tuple__ = version_tuple = (0, 38, 3, 'dev135')
+__version__ = version = '0.38.3.dev136'
+__version_tuple__ = version_tuple = (0, 38, 3, 'dev136')
 
 __commit_id__ = commit_id = None

{ouroboros_ai-0.38.3.dev135 → ouroboros_ai-0.38.3.dev136}/src/ouroboros/mcp/tools/authoring_handlers.py

@@ -204,10 +204,29 @@ def _normalize_interview_answer(answer: str) -> str:
 
 
 def _is_interview_completion_signal(answer: str | None) -> bool:
-    """Return True when the answer explicitly asks to end the interview."""
+    """Return True when the answer explicitly asks to end the interview.
+
+    Only ``[from-user]`` and prefix-less answers represent human intent to close.
+    The auto driver's ``_feature_acceptance_answer`` echoes the LLM question into
+    its answer text, which can accidentally include phrases like "no remaining
+    ambiguity" and trip the shortfall branch. Gate the heuristic by prefix so
+    ``[from-auto]`` / ``[from-code]`` / ``[from-research]`` answers — which carry
+    facts or auto-generated fillers, not user intent — never enter completion.
+
+    If a new ``[from-<source>]`` prefix is introduced in
+    ``skills/interview/SKILL.md`` (or the LLM-facing prompt at
+    ``bigbang/interview.py``), audit whether that source represents direct human
+    intent and update the guard below — otherwise it is silently default-denied,
+    which is safe (the user can still close with a prefix-less ``"done"``) but
+    may surprise callers.
+    """
     if answer is None:
         return False
 
+    stripped = answer.lstrip().lower()
+    if stripped.startswith("[from-") and not stripped.startswith("[from-user]"):
+        return False
+
     normalized = _normalize_interview_answer(answer)
     if not normalized:
         return False

ouroboros_ai-0.38.3.dev136/tests/unit/mcp/tools/test_interview_completion_signal_prefix_guard.py

@@ -0,0 +1,105 @@
+"""Regression tests for the prefix-aware completion-signal gate.
+
+The auto driver's ``_feature_acceptance_answer`` echoes the LLM question text
+into its answer. When the LLM emits closing-mode phrasing such as "no remaining
+ambiguity", the echoed phrase used to trip ``_is_interview_completion_signal``
+inside ``InterviewHandler.handle`` and drive the shortfall branch, which in
+turn produced a ``Cannot record answer - the previous round is already
+answered`` deadlock in ``ooo auto``.
+
+These tests pin the contract: only ``[from-user]`` and prefix-less answers
+represent human intent to close. ``[from-auto]`` / ``[from-code]`` /
+``[from-research]`` answers MUST return False regardless of their text.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from ouroboros.mcp.tools.authoring_handlers import _is_interview_completion_signal
+
+
+@pytest.mark.parametrize(
+    "answer",
+    [
+        "done",
+        "DONE",
+        "  done  ",
+        "complete",
+        "no ambiguity remains",
+        "ready for seed generation",
+        "[from-user] done",
+        "[from-user][refined] No remaining ambiguity",
+    ],
+)
+def test_human_typed_completion_intent_still_signals(answer: str) -> None:
+    """Raw user input and ``[from-user]`` answers preserve the original heuristic."""
+    assert _is_interview_completion_signal(answer) is True
+
+
+@pytest.mark.parametrize(
+    "answer",
+    [
+        "[from-auto][conservative_default] Acceptance for r3 must cover the requested "
+        "behavior; there is no remaining ambiguity to reduce.",
+        "[from-auto][assumption] Assume a single local user; ready for seed generation.",
+        "[from-auto] done",
+        "[from-code] auth uses JWT; no remaining ambiguity to reduce",
+        "[from-code][auto-confirmed] Python 3.12, FastAPI (pyproject.toml)",
+        "[from-research] Stripe rate limit is 100 rps; ready for seed generation",
+    ],
+)
+def test_driver_and_factual_prefixes_never_signal_completion(answer: str) -> None:
+    """Driver and fact-source prefixes never represent user closure intent.
+
+    Regression for the ``ooo auto`` deadlock where the auto driver's
+    acceptance template echoed the LLM question into the answer, accidentally
+    matching the human-intent heuristic and shunting the round into the
+    shortfall branch (state.rounds left with an unanswered placeholder).
+    """
+    assert _is_interview_completion_signal(answer) is False
+
+
+def test_none_answer_returns_false() -> None:
+    """A None answer must short-circuit before any string slicing."""
+    assert _is_interview_completion_signal(None) is False
+
+
+def test_empty_string_returns_false() -> None:
+    """An empty string has no completion intent."""
+    assert _is_interview_completion_signal("") is False
+
+
+def test_leading_whitespace_does_not_bypass_prefix_guard() -> None:
+    """Whitespace before the prefix must not allow ``[from-auto]`` to trip the heuristic."""
+    answer = "   [from-auto] no remaining ambiguity"
+    assert _is_interview_completion_signal(answer) is False
+
+
+@pytest.mark.parametrize(
+    "answer",
+    [
+        "[From-Auto] no remaining ambiguity",
+        "[FROM-AUTO] done",
+        "[From-Code] done",
+        "[FROM-RESEARCH] ready for seed generation",
+    ],
+)
+def test_prefix_guard_is_case_insensitive(answer: str) -> None:
+    """Prefix case must not weaken the guard.
+
+    If a future caller (or typo) emits ``[From-Auto]`` instead of the canonical
+    lowercase form, the guard must still block — otherwise the same shortfall
+    trigger reopens under a trivial case variation.
+    """
+    assert _is_interview_completion_signal(answer) is False
+
+
+def test_negation_still_blocks_human_completion_intent() -> None:
+    """Negations applied to a [from-user] prefix must keep completion=False.
+
+    Regression for the heuristic itself: the prefix guard widens the surface
+    that passes through to the heuristic; ensure the negation rule still wins.
+    """
+    assert _is_interview_completion_signal("[from-user] not done") is False
+    assert _is_interview_completion_signal("[from-user] don't close") is False