claude-dev-env 2.9.0 → 2.10.0

package/hooks/blocking/test_eli11_reply_enforcer.py

@@ -15,16 +15,16 @@ if _HOOKS_ROOT not in sys.path:
 import eli11_reply_enforcer
 from hooks_constants.eli11_reply_enforcer_constants import (
     MAXIMUM_BULLET_LINE_COUNT,
-    MAXIMUM_OVERPACKED_LINE_COUNT,
-    MAXIMUM_REPLY_WORD_COUNT,
-    MAXIMUM_WORDS_PER_LINE,
-    MINIMUM_ENFORCED_WORD_COUNT,
+    MAXIMUM_OVERPACKED_LIST_LINE_COUNT,
+    MAXIMUM_WORDS_PER_LIST_LINE,
     TARGET_BULLET_LINE_COUNT,
 )
 from hooks_constants.text_stripping import strip_code_and_quotes
 
-SAFE_LINE_WORD_COUNT = MAXIMUM_WORDS_PER_LINE // 2
-OVERPACKED_LINE_WORD_COUNT = MAXIMUM_WORDS_PER_LINE + 5
+SAFE_LINE_WORD_COUNT = MAXIMUM_WORDS_PER_LIST_LINE // 2
+OVERPACKED_LIST_LINE_WORD_COUNT = MAXIMUM_WORDS_PER_LIST_LINE + 5
+LONG_REPORT_WORD_COUNT = 240
+SHORT_REPLY_WORD_COUNT = 30
 
 
 def build_filler_prose(word_count: int) -> str:
@@ -45,17 +45,13 @@ def build_prose_block(line_count: int, words_per_line: int) -> str:
 
 
 def build_reply_of_exactly(total_word_count: int) -> str:
-    """Return a filler reply holding exactly the requested words, no line overpacked.
-
-    ::
-
-        in:  22  ->  three lines of 10, 10, and 2 filler words
+    """Return a filler reply holding exactly the requested words, no list line overpacked.
 
     Args:
         total_word_count: How many countable words the whole reply carries.
 
     Returns:
-        A newline-joined reply whose every line stays under the per-line cap.
+        A newline-joined reply whose every line stays under the list-line word cap.
     """
     all_lines = []
     remaining_word_count = total_word_count
@@ -66,63 +62,34 @@ def build_reply_of_exactly(total_word_count: int) -> str:
     return "\n".join(all_lines)
 
 
-SHORT_REPLY_WORD_COUNT = MINIMUM_ENFORCED_WORD_COUNT // 2
-OVERLONG_REPLY_WORD_COUNT = MAXIMUM_REPLY_WORD_COUNT * 2
-JUST_OVER_WORD_CAP_WORD_COUNT = MAXIMUM_REPLY_WORD_COUNT + 1
-UNDER_FLOOR_WORD_COUNT = MINIMUM_ENFORCED_WORD_COUNT - 1
-
 SHORT_REPLY = build_reply_of_exactly(SHORT_REPLY_WORD_COUNT)
-OVERLONG_REPLY = build_reply_of_exactly(OVERLONG_REPLY_WORD_COUNT)
-UNDER_FLOOR_REPLY = build_reply_of_exactly(UNDER_FLOOR_WORD_COUNT)
-JUST_OVER_WORD_CAP_REPLY = build_reply_of_exactly(JUST_OVER_WORD_CAP_WORD_COUNT)
-THREE_OVERPACKED_LINE_REPLY = build_prose_block(
-    MAXIMUM_OVERPACKED_LINE_COUNT + 1, OVERPACKED_LINE_WORD_COUNT
+LONG_REPORT_REPLY = build_reply_of_exactly(LONG_REPORT_WORD_COUNT)
+REQUESTED_FULL_REPORT = (
+    "Outcome: the audit is complete.\n\n"
+    + build_reply_of_exactly(LONG_REPORT_WORD_COUNT)
 )
-TWO_OVERPACKED_LINES = build_prose_block(
-    MAXIMUM_OVERPACKED_LINE_COUNT, OVERPACKED_LINE_WORD_COUNT
+THREE_OVERPACKED_LIST_LINE_REPLY = build_bullet_block(
+    MAXIMUM_OVERPACKED_LIST_LINE_COUNT + 1, OVERPACKED_LIST_LINE_WORD_COUNT
 )
-TWO_OVERPACKED_LINE_REPLY = (
-    f"{TWO_OVERPACKED_LINES}\n{build_filler_prose(SAFE_LINE_WORD_COUNT)}"
+TWO_OVERPACKED_LIST_LINE_REPLY = build_bullet_block(
+    MAXIMUM_OVERPACKED_LIST_LINE_COUNT, OVERPACKED_LIST_LINE_WORD_COUNT
 )
-SEVEN_BULLET_JUST_OVER_FLOOR_REPLY = build_bullet_block(
+SEVEN_BULLET_REPLY = build_bullet_block(
     MAXIMUM_BULLET_LINE_COUNT + 1, SAFE_LINE_WORD_COUNT
 )
-LONG_FORM_OVERLONG_REPLY = f"Long form: the audit report follows.\n\n{OVERLONG_REPLY}"
-BOLD_LONG_FORM_OVERLONG_REPLY = f"**Long form:** the audit report follows.\n\n{OVERLONG_REPLY}"
-QUOTED_LONG_FORM_OVERLONG_REPLY = f"> Long form: the audit report follows.\n\n{OVERLONG_REPLY}"
-HEADING_LONG_FORM_OVERLONG_REPLY = f"# Long form: the audit report follows.\n\n{OVERLONG_REPLY}"
-
+SIX_BULLET_REPLY = build_bullet_block(
+    MAXIMUM_BULLET_LINE_COUNT, SAFE_LINE_WORD_COUNT
+)
 INSTRUCTION_LINE = "Run the migration script."
-INSTRUCTION_LINE_WORD_COUNT = len(INSTRUCTION_LINE.split())
-
 ACTION_WITHOUT_STEPS_FIRST_REPLY = (
-    f"{build_reply_of_exactly(MAXIMUM_REPLY_WORD_COUNT // 2)}\n\n"
-    f"{INSTRUCTION_LINE}\n\nMerge the branch."
+    f"{build_reply_of_exactly(40)}\n\n{INSTRUCTION_LINE}\n\nMerge the branch."
 )
 ACTION_WITH_STEPS_FIRST_REPLY = (
     "1. **Run** the migration script.\n"
     "2. **Merge** the branch.\n\n"
-    f"{build_reply_of_exactly(MAXIMUM_REPLY_WORD_COUNT // 2)}"
-)
-AT_FLOOR_ACTION_REPLY = (
-    f"{build_reply_of_exactly(MINIMUM_ENFORCED_WORD_COUNT - INSTRUCTION_LINE_WORD_COUNT)}"
-    f"\n\n{INSTRUCTION_LINE}"
-)
-SEVEN_BULLET_REPLY = build_bullet_block(
-    MAXIMUM_BULLET_LINE_COUNT + 1, SAFE_LINE_WORD_COUNT + 3
-)
-SIX_BULLET_REPLY = build_bullet_block(
-    MAXIMUM_BULLET_LINE_COUNT, SAFE_LINE_WORD_COUNT + 5
-)
-FENCED_CODE_REPLY = f"{SHORT_REPLY}\n\n```python\n{OVERLONG_REPLY}\n```\n"
-BLOCKQUOTE_REPLY = f"{SHORT_REPLY}\n\n> {build_filler_prose(OVERLONG_REPLY_WORD_COUNT)}\n"
-TABLE_REPLY = "{}\n\n{}\n".format(
-    SHORT_REPLY,
-    "\n".join(
-        f"| {build_filler_prose(MAXIMUM_WORDS_PER_LINE)} | cell |" for _ in range(15)
-    ),
+    f"{build_reply_of_exactly(40)}"
 )
-
+SINGLE_LINE_INSTRUCTION = INSTRUCTION_LINE
 ALL_NARRATIVE_OPENER_LINES = (
     "Open questions remain about the stripper edge cases.",
     "Run time stays under one second on the package suite.",
@@ -130,10 +97,10 @@ ALL_NARRATIVE_OPENER_LINES = (
 )
 NARRATIVE_OPENER_REPLY = "{}\n\n{}".format(
     "\n".join(ALL_NARRATIVE_OPENER_LINES),
-    build_reply_of_exactly(MINIMUM_ENFORCED_WORD_COUNT),
+    build_reply_of_exactly(40),
 )
 INSTALL_WITHOUT_STEPS_FIRST_REPLY = (
-    f"{build_reply_of_exactly(MINIMUM_ENFORCED_WORD_COUNT)}\n\n"
+    f"{build_reply_of_exactly(40)}\n\n"
     "Install the package from the registry."
 )
 DO_THINGS_WITH_STEPS_FIRST_REPLY = (
@@ -141,12 +108,24 @@ DO_THINGS_WITH_STEPS_FIRST_REPLY = (
     "1. Install the package.\n"
     "2. Restart the daemon.\n"
     "3. Save the file.\n\n"
-    f"{build_reply_of_exactly(MINIMUM_ENFORCED_WORD_COUNT)}"
+    f"{build_reply_of_exactly(40)}"
 )
 DO_THINGS_WITHOUT_STEPS_FIRST_REPLY = (
-    f"{build_reply_of_exactly(MINIMUM_ENFORCED_WORD_COUNT)}\n\n"
+    f"{build_reply_of_exactly(40)}\n\n"
     "Do 3 things: install, restart, save."
 )
+FENCED_CODE_REPLY = f"{SHORT_REPLY}\n\n```python\n{LONG_REPORT_REPLY}\n```\n"
+BLOCKQUOTE_REPLY = f"{SHORT_REPLY}\n\n> {build_filler_prose(LONG_REPORT_WORD_COUNT)}\n"
+TABLE_REPLY = "{}\n\n{}\n".format(
+    SHORT_REPLY,
+    "\n".join(
+        f"| {build_filler_prose(MAXIMUM_WORDS_PER_LIST_LINE)} | cell |"
+        for _ in range(15)
+    ),
+)
+LONG_PROSE_PARAGRAPHS_REPLY = "\n\n".join(
+    build_filler_prose(MAXIMUM_WORDS_PER_LIST_LINE + 10) for _ in range(4)
+)
 
 
 def run_hook_with_payload(payload: dict) -> subprocess.CompletedProcess:
@@ -171,8 +150,8 @@ def test_blocker_uses_shared_strip_code_and_quotes() -> None:
     assert eli11_reply_enforcer.strip_code_and_quotes is strip_code_and_quotes
 
 
-def test_short_reply_passes_through() -> None:
-    """A reply under the enforced word floor is never judged."""
+def test_short_correct_reply_passes_through() -> None:
+    """A concise outcome reply passes without padding."""
     completed_process = run_hook_with_message(SHORT_REPLY)
     assert completed_process.returncode == 0
     assert completed_process.stdout == ""
@@ -188,84 +167,56 @@ def test_empty_message_passes_through() -> None:
 def test_stop_hook_active_flag_passes_through() -> None:
     """A re-entrant Stop invocation never blocks again."""
     completed_process = run_hook_with_payload(
-        {"last_assistant_message": OVERLONG_REPLY, "stop_hook_active": True}
+        {
+            "last_assistant_message": ACTION_WITHOUT_STEPS_FIRST_REPLY,
+            "stop_hook_active": True,
+        }
     )
     assert completed_process.returncode == 0
     assert completed_process.stdout == ""
 
 
-def test_reply_over_word_cap_emits_block() -> None:
-    """A reply past the word cap blocks with the count and the cap named."""
-    completed_process = run_hook_with_message(OVERLONG_REPLY)
-    assert completed_process.returncode == 0
-    parsed_response = json.loads(completed_process.stdout)
-    assert parsed_response["decision"] == "block"
-    assert str(OVERLONG_REPLY_WORD_COUNT) in parsed_response["reason"]
-    assert str(MAXIMUM_REPLY_WORD_COUNT) in parsed_response["reason"]
-
-
-def test_reply_just_over_word_cap_emits_block() -> None:
-    """A reply one word past the cap blocks on length alone."""
-    completed_process = run_hook_with_message(JUST_OVER_WORD_CAP_REPLY)
-    assert completed_process.returncode == 0
-    parsed_response = json.loads(completed_process.stdout)
-    assert parsed_response["decision"] == "block"
-    assert (
-        f"{JUST_OVER_WORD_CAP_WORD_COUNT} words, over the "
-        f"{MAXIMUM_REPLY_WORD_COUNT}-word cap"
-    ) in parsed_response["reason"]
-
-
-def test_reply_under_word_floor_passes_through() -> None:
-    """A reply one word under the floor is never judged."""
-    completed_process = run_hook_with_message(UNDER_FLOOR_REPLY)
+def test_long_requested_report_passes_without_magic_prefix() -> None:
+    """A thorough report passes with no Long form: prefix and no word ceiling."""
+    completed_process = run_hook_with_message(REQUESTED_FULL_REPORT)
     assert completed_process.returncode == 0
     assert completed_process.stdout == ""
 
 
-def test_reply_at_the_word_floor_is_judged_for_action_first() -> None:
-    """A reply of exactly the floor word count earns its action-first violation."""
-    completed_process = run_hook_with_message(AT_FLOOR_ACTION_REPLY)
+def test_long_prose_paragraphs_are_not_overpacked() -> None:
+    """Plain paragraphs over the list-line word cap do not block."""
+    completed_process = run_hook_with_message(LONG_PROSE_PARAGRAPHS_REPLY)
     assert completed_process.returncode == 0
-    parsed_response = json.loads(completed_process.stdout)
-    assert parsed_response["decision"] == "block"
-    assert "put the steps first" in parsed_response["reason"]
+    assert completed_process.stdout == ""
 
 
-def test_reply_just_over_word_floor_is_judged() -> None:
-    """A bullet-heavy reply just over the floor earns its bullet violation."""
-    completed_process = run_hook_with_message(SEVEN_BULLET_JUST_OVER_FLOOR_REPLY)
+def test_single_line_instruction_passes_through() -> None:
+    """A one-line imperative does not require a numbered list."""
+    completed_process = run_hook_with_message(SINGLE_LINE_INSTRUCTION)
     assert completed_process.returncode == 0
-    parsed_response = json.loads(completed_process.stdout)
-    assert parsed_response["decision"] == "block"
-    assert (
-        f"cut findings to {TARGET_BULLET_LINE_COUNT} bullets"
-        in parsed_response["reason"]
-    )
+    assert completed_process.stdout == ""
 
 
-def test_three_overpacked_lines_emit_block() -> None:
-    """One line past the overpacked-line cap blocks on one idea per line."""
-    completed_process = run_hook_with_message(THREE_OVERPACKED_LINE_REPLY)
+def test_three_overpacked_list_lines_emit_block() -> None:
+    """List lines past the overpacked-list cap block on one idea per bullet."""
+    completed_process = run_hook_with_message(THREE_OVERPACKED_LIST_LINE_REPLY)
     assert completed_process.returncode == 0
     parsed_response = json.loads(completed_process.stdout)
     assert parsed_response["decision"] == "block"
-    assert "lines carry too many words - one idea per line" in (
-        parsed_response["reason"]
-    )
+    assert "list lines carry too many words" in parsed_response["reason"]
 
 
-def test_two_overpacked_lines_pass_through() -> None:
-    """Over-packed lines at the cap pass."""
-    completed_process = run_hook_with_message(TWO_OVERPACKED_LINE_REPLY)
+def test_two_overpacked_list_lines_pass_through() -> None:
+    """Over-packed list lines at the cap pass."""
+    completed_process = run_hook_with_message(TWO_OVERPACKED_LIST_LINE_REPLY)
     assert completed_process.returncode == 0
     assert completed_process.stdout == ""
 
 
 def test_bullet_marker_is_not_counted_as_a_line_word() -> None:
-    """A bullet at the per-line cap stays under it once its marker comes off."""
+    """A bullet at the per-list-line cap stays under it once its marker comes off."""
     capped_bullet_lines = build_bullet_block(
-        MAXIMUM_BULLET_LINE_COUNT - 2, MAXIMUM_WORDS_PER_LINE
+        MAXIMUM_BULLET_LINE_COUNT - 2, MAXIMUM_WORDS_PER_LIST_LINE
     )
     all_violations = eli11_reply_enforcer.find_reply_shape_violations(
         capped_bullet_lines
@@ -273,41 +224,14 @@ def test_bullet_marker_is_not_counted_as_a_line_word() -> None:
     assert all_violations == []
 
 
-def test_long_form_prefix_exempts_an_overlong_reply() -> None:
-    """The Long form escape hatch clears every reply-shape check."""
-    completed_process = run_hook_with_message(LONG_FORM_OVERLONG_REPLY)
-    assert completed_process.returncode == 0
-    assert completed_process.stdout == ""
-
-
-def test_bold_long_form_prefix_exempts_an_overlong_reply() -> None:
-    """A bold-wrapped Long form prefix still opts the reply out."""
-    completed_process = run_hook_with_message(BOLD_LONG_FORM_OVERLONG_REPLY)
-    assert completed_process.returncode == 0
-    assert completed_process.stdout == ""
-
-
-def test_quoted_long_form_prefix_exempts_an_overlong_reply() -> None:
-    """A blockquoted Long form prefix still opts the reply out."""
-    completed_process = run_hook_with_message(QUOTED_LONG_FORM_OVERLONG_REPLY)
-    assert completed_process.returncode == 0
-    assert completed_process.stdout == ""
-
-
-def test_heading_long_form_prefix_exempts_an_overlong_reply() -> None:
-    """A heading-wrapped Long form prefix still opts the reply out."""
-    completed_process = run_hook_with_message(HEADING_LONG_FORM_OVERLONG_REPLY)
-    assert completed_process.returncode == 0
-    assert completed_process.stdout == ""
-
-
 def test_instruction_lines_without_leading_steps_emit_block() -> None:
     """Imperative instructions buried under prose block with a steps-first message."""
     completed_process = run_hook_with_message(ACTION_WITHOUT_STEPS_FIRST_REPLY)
     assert completed_process.returncode == 0
     parsed_response = json.loads(completed_process.stdout)
     assert parsed_response["decision"] == "block"
-    assert "put the steps first" in parsed_response["reason"]
+    assert "numbered steps" in parsed_response["reason"]
+    assert "Rewrite the reply" in parsed_response["reason"]
 
 
 def test_numbered_steps_in_lead_lines_pass_through() -> None:
@@ -350,11 +274,11 @@ def test_names_imperative_object_accepts_determiners_counts_and_paths() -> None:
 def test_strip_markdown_lead_markers_removes_every_wrapper() -> None:
     """Blockquote, heading, and bold wrappers come off the front of a line."""
     assert eli11_reply_enforcer.strip_markdown_lead_markers(
-        "> **Long form:** the report follows"
-    ) == "Long form:** the report follows"
+        "> **Bold lead:** the report follows"
+    ) == "Bold lead:** the report follows"
     assert eli11_reply_enforcer.strip_markdown_lead_markers(
-        "# Long form: the report follows"
-    ) == "Long form: the report follows"
+        "# Heading lead: the report follows"
+    ) == "Heading lead: the report follows"
 
 
 def test_install_instruction_without_leading_steps_emits_block() -> None:
@@ -363,7 +287,7 @@ def test_install_instruction_without_leading_steps_emits_block() -> None:
     assert completed_process.returncode == 0
     parsed_response = json.loads(completed_process.stdout)
     assert parsed_response["decision"] == "block"
-    assert "put the steps first" in parsed_response["reason"]
+    assert "numbered steps" in parsed_response["reason"]
 
 
 def test_do_three_things_without_leading_steps_emits_block() -> None:
@@ -372,7 +296,7 @@ def test_do_three_things_without_leading_steps_emits_block() -> None:
     assert completed_process.returncode == 0
     parsed_response = json.loads(completed_process.stdout)
     assert parsed_response["decision"] == "block"
-    assert "put the steps first" in parsed_response["reason"]
+    assert "numbered steps" in parsed_response["reason"]
 
 
 def test_do_three_things_with_numbered_steps_passes_through() -> None:
@@ -398,13 +322,13 @@ def test_numbered_step_opener_is_a_numbered_step() -> None:
 
 
 def test_more_than_six_bullets_emits_block() -> None:
-    """One bullet past the cap blocks with the cut-to-target-bullets message."""
+    """One bullet past the cap blocks with the put-findings message."""
     completed_process = run_hook_with_message(SEVEN_BULLET_REPLY)
     assert completed_process.returncode == 0
     parsed_response = json.loads(completed_process.stdout)
     assert parsed_response["decision"] == "block"
     assert (
-        f"cut findings to {TARGET_BULLET_LINE_COUNT} bullets"
+        f"put findings in at most {TARGET_BULLET_LINE_COUNT} bullets"
         in parsed_response["reason"]
     )
 
@@ -416,29 +340,29 @@ def test_six_bullets_pass_through() -> None:
     assert completed_process.stdout == ""
 
 
-def test_fenced_code_words_are_not_counted() -> None:
-    """Words inside a fenced code block never push a reply past the cap."""
+def test_fenced_code_does_not_force_shape_block() -> None:
+    """Words inside a fenced code block do not create shape violations alone."""
     completed_process = run_hook_with_message(FENCED_CODE_REPLY)
     assert completed_process.returncode == 0
     assert completed_process.stdout == ""
 
 
-def test_blockquote_words_are_not_counted() -> None:
-    """Quoted lines are the user's words and never push a reply past the cap."""
+def test_blockquote_does_not_force_shape_block() -> None:
+    """Quoted lines are exempt from shape counting."""
     completed_process = run_hook_with_message(BLOCKQUOTE_REPLY)
     assert completed_process.returncode == 0
     assert completed_process.stdout == ""
 
 
 def test_table_rows_are_not_counted() -> None:
-    """Table rows carry reference data and never push a reply past the cap."""
+    """Table rows carry reference data and do not create shape violations alone."""
     completed_process = run_hook_with_message(TABLE_REPLY)
     assert completed_process.returncode == 0
     assert completed_process.stdout == ""
 
 
 def test_urls_are_removed_from_counted_prose() -> None:
-    """A link target is stripped before the words are counted."""
+    """A link target is stripped before prose is judged."""
     prose_text = eli11_reply_enforcer.extract_reply_prose(
         "The draft is at https://github.com/owner/repo/pull/704 now"
     )
@@ -446,12 +370,21 @@ def test_urls_are_removed_from_counted_prose() -> None:
     assert "draft" in prose_text
 
 
-def test_block_response_json_shape() -> None:
-    """The block payload carries the Stop-hook keys and names the escape hatch."""
-    completed_process = run_hook_with_message(OVERLONG_REPLY)
+def test_block_reason_uses_positive_rewrite_language() -> None:
+    """Corrective output tells the model what to write, not only what failed."""
+    reason = eli11_reply_enforcer.build_block_reason(
+        [eli11_reply_enforcer.describe_action_first_violation()]
+    )
+    assert "Rewrite the reply" in reason
+    assert "120-word" not in reason
+    assert "Long form:" not in reason
+    assert "minimum" not in reason.lower()
+
+
+def test_hook_never_forces_padding_on_short_replies() -> None:
+    """A correct short reply is not blocked for being under a word floor."""
+    tiny_reply = "Done."
+    completed_process = run_hook_with_message(tiny_reply)
     assert completed_process.returncode == 0
-    parsed_response = json.loads(completed_process.stdout)
-    assert parsed_response["decision"] == "block"
-    assert parsed_response["suppressOutput"] is True
-    assert parsed_response["systemMessage"]
-    assert "Long form:" in parsed_response["reason"]
+    assert completed_process.stdout == ""
+    assert eli11_reply_enforcer.find_reply_shape_violations(tiny_reply) == []

package/hooks/blocking/test_hedging_language_blocker.py

@@ -1,12 +1,15 @@
 """Tests for hedging_language_blocker hook response shape."""
 
 import importlib.util
+import io
 import json
 import os
 import subprocess
 import sys
 import tempfile
 
+import pytest
+
 HOOK_SCRIPT_PATH = os.path.join(os.path.dirname(__file__), "hedging_language_blocker.py")
 _HOOKS_DIR = os.path.dirname(HOOK_SCRIPT_PATH)
 _HOOKS_ROOT = os.path.join(_HOOKS_DIR, "..")
@@ -30,14 +33,22 @@ CLEAN_MESSAGE = "This is verified by the source document."
 EMPTY_MESSAGE = ""
 
 
-def run_hook_with_message(assistant_message: str) -> subprocess.CompletedProcess:
+def run_hook_with_message(
+    assistant_message: str, *, is_prose_style_enabled: bool = True
+) -> subprocess.CompletedProcess:
     hook_input_payload = json.dumps({"last_assistant_message": assistant_message})
+    environment_by_key = os.environ.copy()
+    if is_prose_style_enabled:
+        environment_by_key["CLAUDE_PROSE_STYLE_ENFORCEMENT"] = "1"
+    else:
+        environment_by_key.pop("CLAUDE_PROSE_STYLE_ENFORCEMENT", None)
     return subprocess.run(
         [sys.executable, HOOK_SCRIPT_PATH],
         input=hook_input_payload,
         capture_output=True,
         text=True,
         check=False,
+        env=environment_by_key,
     )
 
 
@@ -58,6 +69,8 @@ def run_hook_with_patched_search_paths(
         wrapper_file_path = wrapper_file.name
 
     hook_input_payload = json.dumps({"last_assistant_message": assistant_message})
+    environment_by_key = os.environ.copy()
+    environment_by_key["CLAUDE_PROSE_STYLE_ENFORCEMENT"] = "1"
     try:
         completed_process = subprocess.run(
             [sys.executable, wrapper_file_path],
@@ -65,6 +78,7 @@ def run_hook_with_patched_search_paths(
             capture_output=True,
             text=True,
             check=False,
+            env=environment_by_key,
         )
     finally:
         os.unlink(wrapper_file_path)
@@ -80,6 +94,49 @@ def test_user_facing_notice_matches_config_messages_module():
     assert module.USER_FACING_NOTICE == USER_FACING_NOTICE
 
 
+def test_hedging_scan_is_default_off() -> None:
+    completed_process = run_hook_with_message(
+        HEDGING_MESSAGE, is_prose_style_enabled=False
+    )
+
+    assert completed_process.returncode == 0
+    assert completed_process.stdout == ""
+
+
+def test_default_off_emits_privacy_safe_advisory_candidates(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    all_emitted: list[tuple[str, str, str]] = []
+
+    def _record_advisory(
+        matcher_id: str, surface: str, context_text: str, **_kwargs: object
+    ) -> dict[str, object]:
+        all_emitted.append((matcher_id, surface, context_text))
+        return {}
+
+    monkeypatch.setattr(
+        hedging_language_blocker, "emit_advisory_candidate", _record_advisory
+    )
+    monkeypatch.setattr(
+        hedging_language_blocker,
+        "prose_style_enforcement_enabled_in_environment",
+        lambda: False,
+    )
+    monkeypatch.setattr(
+        hedging_language_blocker.sys,
+        "stdin",
+        io.StringIO(json.dumps({"last_assistant_message": HEDGING_MESSAGE})),
+    )
+    with pytest.raises(SystemExit) as exit_info:
+        hedging_language_blocker.main()
+    assert exit_info.value.code == 0
+    assert all_emitted
+    matcher_id, surface, context_text = all_emitted[0]
+    assert matcher_id == "hedging_word"
+    assert surface == "Stop"
+    assert "likely" in context_text
+
+
 def test_hedging_message_emits_block_with_short_user_notice():
     completed_process = run_hook_with_message(HEDGING_MESSAGE)
 
@@ -140,3 +197,65 @@ def test_empty_message_passes_through_with_no_output():
 
     assert completed_process.returncode == 0
     assert completed_process.stdout == ""
+
+
+def test_explicit_unverified_label_in_same_sentence_passes() -> None:
+    completed_process = run_hook_with_message(
+        "This claim is unverified; the deploy is probably blocked."
+    )
+
+    assert completed_process.returncode == 0
+    assert completed_process.stdout == ""
+
+
+def test_i_dont_know_label_in_same_sentence_passes() -> None:
+    completed_process = run_hook_with_message(
+        "I don't know whether the port is probably open."
+    )
+
+    assert completed_process.returncode == 0
+    assert completed_process.stdout == ""
+
+
+def test_supported_probability_without_hedge_word_passes() -> None:
+    completed_process = run_hook_with_message(
+        "The suite reports 0.92 precision on the labeled fixture set in "
+        "test_prose_matcher_advisory.py."
+    )
+
+    assert completed_process.returncode == 0
+    assert completed_process.stdout == ""
+
+
+def test_label_in_one_sentence_does_not_exempt_bare_hedge_in_another() -> None:
+    completed_process = run_hook_with_message(
+        "This claim is unverified. The deploy is probably blocked."
+    )
+
+    assert completed_process.returncode == 0
+    parsed_response = json.loads(completed_process.stdout)
+    assert parsed_response["decision"] == "block"
+    assert "probably" in parsed_response["reason"]
+    assert "explicit uncertainty label" in parsed_response["reason"]
+
+
+def test_bare_probably_still_blocks_with_positive_corrective() -> None:
+    completed_process = run_hook_with_message("The deploy is probably blocked.")
+
+    assert completed_process.returncode == 0
+    parsed_response = json.loads(completed_process.stdout)
+    assert parsed_response["decision"] == "block"
+    assert "probably" in parsed_response["reason"]
+    assert "label that claim unverified" in parsed_response["reason"]
+    assert "AskUserQuestion" in parsed_response["reason"]
+
+
+def test_find_blocking_hedging_terms_is_sentence_scoped() -> None:
+    bare = hedging_language_blocker.find_blocking_hedging_terms(
+        "This claim is unverified. The deploy is probably blocked."
+    )
+    labeled = hedging_language_blocker.find_blocking_hedging_terms(
+        "This claim is unverified; the deploy is probably blocked."
+    )
+    assert bare == ["probably"]
+    assert labeled == []

package/hooks/blocking/test_hook_prose_detector_consistency.py

@@ -3,6 +3,7 @@
 import importlib.util
 import io
 import json
+import os
 import pathlib
 import sys
 from unittest import mock
@@ -192,14 +193,33 @@ def test_written_content_reads_edit_new_string() -> None:
     assert written_content("Edit", edit_input) == "edited body"
 
 
-def _run_main_with_io(input_text: str) -> str:
-    with mock.patch("sys.stdin", io.StringIO(input_text)):
-        with mock.patch("sys.stdout", new_callable=io.StringIO) as mock_stdout:
-            try:
-                hook_module.main()
-            except SystemExit:
-                pass
-            return mock_stdout.getvalue()
+def _run_main_with_io(
+    input_text: str, *, is_prose_style_enabled: bool = True
+) -> str:
+    environment_by_key = os.environ.copy()
+    if is_prose_style_enabled:
+        environment_by_key["CLAUDE_PROSE_STYLE_ENFORCEMENT"] = "1"
+    else:
+        environment_by_key.pop("CLAUDE_PROSE_STYLE_ENFORCEMENT", None)
+    with mock.patch.dict(os.environ, environment_by_key, clear=True):
+        with mock.patch("sys.stdin", io.StringIO(input_text)):
+            with mock.patch("sys.stdout", new_callable=io.StringIO) as mock_stdout:
+                try:
+                    hook_module.main()
+                except SystemExit:
+                    pass
+                return mock_stdout.getvalue()
+
+
+def test_main_is_default_off_for_overstated_hook_module_write() -> None:
+    hook_input = {
+        "tool_name": "Write",
+        "tool_input": {
+            "file_path": "/repo/hooks/hooks_constants/some_blocker_constants.py",
+            "content": _OVERSTATED_MESSAGE_MODULE,
+        },
+    }
+    assert _run_main_with_io(json.dumps(hook_input), is_prose_style_enabled=False) == ""
 
 
 def test_main_blocks_overstated_hook_module_write() -> None: