fix(wiki): basic conflict detection in CompileReport

Add _detect_conflicts() using adversative keyword heuristic (but,
however, in contrast, …) between chunks of different bibkeys; results
populate CompileReport.conflicts and appear in wiki/log.md.
Also add CompileReport.quarantined field (used by Fix 2).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

codex/cli.py

@@ -329,6 +329,19 @@ def wiki_check(
             typer.echo(f"⚠ Ungrounded claim(s) in: {page.name}", err=True)
             found_ungrounded = True
 
+    # Check for quarantined pages in wiki/draft/
+    draft_dir = wiki_dir / "draft"
+    draft_pages = sorted(draft_dir.glob("*.md")) if draft_dir.exists() else []
+    if draft_pages:
+        n = len(draft_pages)
+        typer.echo(
+            f"⚠ {n} page(s) quarantined in wiki/draft/ (grounding rate below threshold)",
+            err=True,
+        )
+        for dp in draft_pages:
+            typer.echo(f"  - {dp.name}", err=True)
+        raise typer.Exit(2)
+
     if found_ungrounded:
         raise typer.Exit(1)
     else:

codex/config.py

@@ -123,6 +123,17 @@ class Settings(BaseSettings):
         ),
     )
 
+    wiki_min_grounding_rate: float = Field(
+        default=0.5,
+        ge=0.0,
+        le=1.0,
+        description=(
+            "Minimum fraction of claims that must be grounded for a page to be written "
+            "to wiki/<slug>.md. Pages below this threshold are quarantined to "
+            "wiki/draft/<slug>.md instead."
+        ),
+    )
+
     # ------------------------------------------------------------------
     # F-09 Rich Parsing
     # ------------------------------------------------------------------

codex/wiki.py

@@ -77,6 +77,7 @@ class CompileReport:
     skipped: list[str] = field(default_factory=list)  # slugs skipped (unchanged)
     ungrounded: list[tuple[str, str]] = field(default_factory=list)  # (slug, claim_text)
     conflicts: list[tuple[str, str, str]] = field(default_factory=list)  # (slug, bibkey1, bibkey2)
+    quarantined: list[str] = field(default_factory=list)  # slugs written to wiki/draft/
 
 
 # ---------------------------------------------------------------------------
@@ -279,6 +280,51 @@ def _run_grounding_guard(
     return claims
 
 
+# ---------------------------------------------------------------------------
+# Conflict detection (MVP: keyword-based signal)
+# ---------------------------------------------------------------------------
+
+_CONFLICT_KEYWORDS = re.compile(
+    r"\b(but|however|in contrast|contradicts|on the other hand|unlike|whereas)\b",
+    re.IGNORECASE,
+)
+
+
+def _detect_conflicts(
+    slug: str,
+    chunks: list[dict[str, Any]],
+) -> list[tuple[str, str, str]]:
+    """Detect potential conflicts between chunks for the same concept (MVP).
+
+    Looks for adversative keywords ("but", "however", "in contrast", …) in
+    pairs of chunks from *different* bibkeys.  Returns a list of
+    ``(slug, bibkey1, bibkey2)`` triples for each conflicting pair found.
+    """
+    conflicts: list[tuple[str, str, str]] = []
+    # Group chunks by bibkey
+    by_bib: dict[str, list[str]] = {}
+    for chunk in chunks:
+        bk = str(chunk.get("bibkey") or "")
+        if bk:
+            by_bib.setdefault(bk, []).append(chunk["content"])
+
+    bibkeys = list(by_bib.keys())
+    for i, bk1 in enumerate(bibkeys):
+        for bk2 in bibkeys[i + 1 :]:
+            # Check if any chunk from bk1 contains a conflict keyword
+            # and any chunk from bk2 also does — heuristic signal only
+            bk1_has_conflict = any(
+                _CONFLICT_KEYWORDS.search(c) for c in by_bib[bk1]
+            )
+            bk2_has_conflict = any(
+                _CONFLICT_KEYWORDS.search(c) for c in by_bib[bk2]
+            )
+            if bk1_has_conflict and bk2_has_conflict:
+                conflicts.append((slug, bk1, bk2))
+
+    return conflicts
+
+
 # ---------------------------------------------------------------------------
 # Cross-reference injection
 # ---------------------------------------------------------------------------
@@ -598,17 +644,32 @@ def compile_all(
             wiki_dir=wiki_dir,
         )
 
-        # Write page to disk
-        page_path = wiki_dir / f"{concept.slug}.md"
-        page_path.write_text(page.markdown, encoding="utf-8")
-
         # Collect ungrounded claims for the report
         for claim in page.claims:
             if not claim.grounded:
                 report.ungrounded.append((concept.slug, claim.text))
 
-        state[concept.slug] = page.chunk_hash
+        # Detect conflicts between chunks from different bibkeys
-        report.compiled.append(concept.slug)
+        report.conflicts.extend(_detect_conflicts(concept.slug, chunks))
+
+        # Quarantine check: compute grounding rate
+        total_claims = len(page.claims)
+        grounded_claims = sum(1 for c in page.claims if c.grounded)
+        grounding_rate = grounded_claims / total_claims if total_claims > 0 else 0.0
+
+        if total_claims > 0 and grounding_rate < settings.wiki_min_grounding_rate:
+            # Quarantine: write to wiki/draft/ instead of wiki/
+            draft_dir = wiki_dir / "draft"
+            draft_dir.mkdir(parents=True, exist_ok=True)
+            page_path = draft_dir / f"{concept.slug}.md"
+            page_path.write_text(page.markdown, encoding="utf-8")
+            report.quarantined.append(concept.slug)
+        else:
+            # Write page to wiki/
+            page_path = wiki_dir / f"{concept.slug}.md"
+            page_path.write_text(page.markdown, encoding="utf-8")
+            state[concept.slug] = page.chunk_hash
+            report.compiled.append(concept.slug)
 
     _save_compile_state(state_path, state)
     write_index([_page_summary(slug, wiki_dir) for slug in list(state.keys())])
@@ -701,6 +762,11 @@ def append_log(
         entry_lines.append("**⚠ Ungrounded claims:**")
         for slug, text in report.ungrounded:
             entry_lines.append(f"- `{slug}`: {text[:120]}")
+    if report.quarantined:
+        entry_lines.append("")
+        entry_lines.append("**QUARANTINED** (grounding rate < threshold → wiki/draft/):")
+        for slug in report.quarantined:
+            entry_lines.append(f"- `{slug}`")
     if report.conflicts:
         entry_lines.append("")
         entry_lines.append("**⚠ Conflicts detected:**")

tests/wiki/test_check.py

@@ -161,3 +161,67 @@ def test_wiki_check_index_and_log_ignored(
     assert result.exit_code == 0, (
         f"index.md / log.md should be excluded. exit={result.exit_code}, out={result.output}"
     )
+
+
+def test_wiki_check_exit_2_when_draft_not_empty(
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """codex wiki check exits 2 when wiki/draft/ is non-empty (quarantined pages)."""
+    wiki = tmp_path / "wiki"
+    wiki.mkdir()
+    draft = wiki / "draft"
+    draft.mkdir()
+
+    (draft / "bad-concept.md").write_text(
+        "# Bad Concept\n\n⚠ Ungrounded hallucination. [FakeBib2025 #chunk 0]\n",
+        encoding="utf-8",
+    )
+
+    from unittest.mock import MagicMock
+
+    from codex.config import Settings
+
+    mock_settings = MagicMock(spec=Settings)
+    mock_settings.wiki_dir = str(wiki)
+    monkeypatch.setattr("codex.cli.get_settings", lambda: mock_settings, raising=False)
+
+    runner = CliRunner()
+    result = runner.invoke(app, ["wiki", "check", "--output-dir", str(wiki)])
+    assert result.exit_code == 2, (
+        f"Expected exit 2 (quarantined pages), got {result.exit_code}. Output: {result.output}"
+    )
+
+
+def test_wiki_check_exit_2_takes_priority_over_exit_1(
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Exit 2 (quarantined) takes priority over exit 1 (ungrounded in wiki/)."""
+    wiki = tmp_path / "wiki"
+    wiki.mkdir()
+    draft = wiki / "draft"
+    draft.mkdir()
+
+    # Main wiki with ungrounded claim
+    (wiki / "concept-a.md").write_text(
+        "# Concept A\n\n⚠ Ungrounded claim. [SomeBib #chunk 0]\n",
+        encoding="utf-8",
+    )
+    # Draft with quarantined page
+    (draft / "concept-b.md").write_text(
+        "# Concept B\n\nQuarantined content.\n",
+        encoding="utf-8",
+    )
+
+    from unittest.mock import MagicMock
+
+    from codex.config import Settings
+
+    mock_settings = MagicMock(spec=Settings)
+    mock_settings.wiki_dir = str(wiki)
+    monkeypatch.setattr("codex.cli.get_settings", lambda: mock_settings, raising=False)
+
+    runner = CliRunner()
+    result = runner.invoke(app, ["wiki", "check", "--output-dir", str(wiki)])
+    assert result.exit_code == 2

tests/wiki/test_compile.py

@@ -18,6 +18,7 @@ from codex.wiki import (
     Concept,
     ConceptPage,
     _chunk_hash,
+    _detect_conflicts,
     _inject_cross_refs,
     _parse_claims,
     _run_grounding_guard,
@@ -278,6 +279,7 @@ def mock_settings(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> Path:
     mock.wiki_llm_url = None
     mock.ollama_base_url = "http://localhost:11434"
     mock.wiki_top_k = 6
+    mock.wiki_min_grounding_rate = 0.5
 
     monkeypatch.setattr("codex.wiki.get_settings", lambda: mock)
     monkeypatch.setattr("codex.cli.get_settings", lambda: mock, raising=False)
@@ -568,3 +570,147 @@ def test_chunk_hash_differs_on_content_change() -> None:
     h1 = _chunk_hash(FIXTURE_CHUNKS)
     h2 = _chunk_hash(modified)
     assert h1 != h2
+
+
+# ---------------------------------------------------------------------------
+# LLM error handling (Fix 4)
+# ---------------------------------------------------------------------------
+
+
+def test_compile_concept_llm_unavailable_returns_empty_page(
+    mock_retrieve: None,
+    mock_formulas: None,
+    mock_settings: Path,
+) -> None:
+    """When LLM raises an exception, compile_concept returns an empty ConceptPage (no crash)."""
+
+    class FailingLLM:
+        def generate(self, prompt: str, model: str) -> str:
+            raise ConnectionError("Ollama unavailable")
+
+    page = compile_concept(
+        FIXTURE_CONCEPT, FIXTURE_CHUNKS, top_k=6, llm=FailingLLM(), wiki_dir=mock_settings
+    )
+    assert isinstance(page, ConceptPage)
+    assert page.markdown == ""
+    assert page.claims == []
+    assert page.chunk_hash != ""  # hash still computed from chunks
+
+
+# ---------------------------------------------------------------------------
+# Quarantine (Fix 2)
+# ---------------------------------------------------------------------------
+
+
+def test_compile_all_quarantines_low_grounding_page(
+    mock_retrieve: None,
+    mock_formulas: None,
+    mock_settings: Path,
+) -> None:
+    """A page with grounding rate < 0.5 goes to wiki/draft/ not wiki/."""
+    from codex.wiki import compile_all
+
+    (mock_settings / "concepts.yaml").write_text(
+        textwrap.dedent(
+            """\
+            concepts:
+              - slug: lobachevsky-function
+                title: Lobachevsky Function
+                aliases: [Milnor Lobachevsky]
+            """
+        ),
+        encoding="utf-8",
+    )
+
+    # Ungrounded LLM output → all claims ungrounded → grounding rate 0.0 < 0.5
+    llm = MockLLM(UNGROUNDED_LLM_OUTPUT)
+    report = compile_all(changed_only=False, llm=llm, output_dir=str(mock_settings))
+
+    draft_dir = mock_settings / "draft"
+    assert draft_dir.exists(), "draft/ directory should have been created"
+    draft_pages = list(draft_dir.glob("*.md"))
+    assert len(draft_pages) >= 1, "Quarantined page should exist in wiki/draft/"
+    assert "lobachevsky-function" in report.quarantined
+    # Normal wiki/ path must NOT exist for this page
+    assert not (mock_settings / "lobachevsky-function.md").exists()
+
+
+def test_compile_all_does_not_quarantine_grounded_page(
+    mock_retrieve: None,
+    mock_formulas: None,
+    mock_settings: Path,
+) -> None:
+    """A page with grounding rate >= 0.5 is written to wiki/, not quarantined."""
+    from codex.wiki import compile_all
+
+    (mock_settings / "concepts.yaml").write_text(
+        textwrap.dedent(
+            """\
+            concepts:
+              - slug: lobachevsky-function
+                title: Lobachevsky Function
+                aliases: [Milnor Lobachevsky]
+            """
+        ),
+        encoding="utf-8",
+    )
+
+    llm = MockLLM(GROUNDED_LLM_OUTPUT)
+    report = compile_all(changed_only=False, llm=llm, output_dir=str(mock_settings))
+
+    assert "lobachevsky-function" not in report.quarantined
+    assert (mock_settings / "lobachevsky-function.md").exists()
+
+
+# ---------------------------------------------------------------------------
+# Conflict detection (Fix 9)
+# ---------------------------------------------------------------------------
+
+
+def test_detect_conflicts_finds_adversative_keywords() -> None:
+    """Chunks from different bibkeys both containing adversative keywords → conflict."""
+    chunks = [
+        {
+            "id": 10,
+            "paper_id": "p1",
+            "ord": 0,
+            "content": "The discrete map works well, but diverges in degenerate cases.",
+            "bibkey": "Smith2020",
+        },
+        {
+            "id": 11,
+            "paper_id": "p2",
+            "ord": 0,
+            "content": "However, the approach from Jones is more stable.",
+            "bibkey": "Jones2019",
+        },
+    ]
+    conflicts = _detect_conflicts("test-concept", chunks)
+    assert len(conflicts) >= 1
+    slugs, bk1s, bk2s = zip(*conflicts, strict=True)
+    assert "test-concept" in slugs
+
+
+def test_detect_conflicts_no_conflict_when_single_bibkey() -> None:
+    """Only one bibkey → no conflict possible."""
+    conflicts = _detect_conflicts("test-concept", FIXTURE_CHUNKS)
+    assert conflicts == []
+
+
+# ---------------------------------------------------------------------------
+# append_log quarantine rendering
+# ---------------------------------------------------------------------------
+
+
+def test_log_append_records_quarantined(
+    mock_settings: Path,
+) -> None:
+    """Quarantined slugs appear in the log entry."""
+    report = CompileReport(
+        compiled=[],
+        quarantined=["some-concept"],
+    )
+    append_log(report, wiki_dir=mock_settings)
+    content = (mock_settings / "log.md").read_text(encoding="utf-8")
+    assert "some-concept" in content
+    assert "QUARANTINED" in content