diff --git a/devtools/benchmarks/osworld/METHODOLOGY.md b/devtools/benchmarks/osworld/METHODOLOGY.md
index a936aa9fc..9b5d304cf 100644
--- a/devtools/benchmarks/osworld/METHODOLOGY.md
+++ b/devtools/benchmarks/osworld/METHODOLOGY.md
@@ -113,7 +113,7 @@ finished "chat-style" left the Ouroboros `loop_outcome.final_answer` empty
 to `not_evaluated`, and several tasks lost reward because the agent answered
 in chat instead of leaving the VM in the evaluator-expected state (e.g.
 `is_expected_active_tab` checks the active URL, `compare_table` checks a saved
-`.xlsx`). Two structural fixes live in `run_step_agent.py`:
+`.xlsx`). Two structural fixes live in `step_agent_policy.py`:
 
 1. **Terminal-message capture.** The per-step JSON schema has a
    `final_answer` field; when the agent emits `done`/`fail` the runner
diff --git a/devtools/benchmarks/osworld/run_cu_bridge_agent.py b/devtools/benchmarks/osworld/run_cu_bridge_agent.py
index 56f257d72..057e601d5 100644
--- a/devtools/benchmarks/osworld/run_cu_bridge_agent.py
+++ b/devtools/benchmarks/osworld/run_cu_bridge_agent.py
@@ -326,8 +326,8 @@ def main() -> int:
 
     # Guards: never drive the live desktop server or publish a bench connection
     # into the owner's live skill state (mirrors run_step_agent.py).
-    from devtools.benchmarks.osworld.run_step_agent import (
-        _is_default_desktop_server,
+    from devtools.benchmarks.osworld.step_agent_env import _is_default_desktop_server
+    from devtools.benchmarks.osworld.step_agent_claims import (
         confined_claims_dir,
         scored_claim_state,
         task_claim_key,
diff --git a/devtools/benchmarks/osworld/step_agent_claims.py b/devtools/benchmarks/osworld/step_agent_claims.py
index f0ba0861a..1968ec6ce 100644
--- a/devtools/benchmarks/osworld/step_agent_claims.py
+++ b/devtools/benchmarks/osworld/step_agent_claims.py
@@ -1,6 +1,6 @@
 """Cross-lane task claims and the scored-claim ledger for OSWorld runs.
 
-Verbatim extraction from ``run_step_agent.py`` (v7 stream W): claim-directory
+Verbatim extraction from ``run_step_agent.py``: claim-directory
 confinement, the claim key, staleness, acquisition and release, plus the
 durable unconfirmed/scored markers that make overlapping lanes, resumes and
 retry passes safe over one shared results tree (METHODOLOGY §7.9).
diff --git a/tests/test_osworld_cu_bridge_extraction.py b/tests/test_osworld_cu_bridge_extraction.py
index 9e5e520ca..1c1e16ae9 100644
--- a/tests/test_osworld_cu_bridge_extraction.py
+++ b/tests/test_osworld_cu_bridge_extraction.py
@@ -9,8 +9,11 @@ operators reach through `run_cu_bridge_agent.<name>`.
 from __future__ import annotations
 
 import ast
+import copy
 import pathlib
 
+import pytest
+
 from devtools.benchmarks.osworld import (
     cu_bridge_budget,
     cu_bridge_gate,
@@ -18,6 +21,8 @@ from devtools.benchmarks.osworld import (
     cu_bridge_runtime,
     cu_bridge_tool_policy,
     run_cu_bridge_agent as rcb,
+    step_agent_claims,
+    step_agent_env,
 )
 
 
@@ -119,17 +124,26 @@ def test_cu_bridge_launcher_reexports_every_moved_identity():
         assert getattr(rcb, name) is getattr(owner, name), name
 
 
-def test_cu_bridge_extraction_size_bounds_have_meaningful_headroom():
-    counts = {
-        path.name: len(path.read_text(encoding="utf-8").splitlines())
-        for path in (
-            OSWORLD / "run_cu_bridge_agent.py",
-            *(pathlib.Path(module.__file__) for module in _LEAVES),
-        )
-    }
-    assert counts["run_cu_bridge_agent.py"] < 1500
-    assert all(
-        count <= 1000
-        for name, count in counts.items()
-        if name != "run_cu_bridge_agent.py"
-    ), counts
+@pytest.mark.parametrize("helper,owner", [
+    ("_is_default_desktop_server", step_agent_env),
+    ("confined_claims_dir", step_agent_claims),
+    ("task_claim_key", step_agent_claims),
+    ("scored_claim_state", step_agent_claims),
+])
+def test_pre_admission_audit_reaches_step_helpers(helper, owner, monkeypatch):
+    from devtools.benchmarks.common import launcher_audit as audit
+
+    source = (OSWORLD / "run_cu_bridge_agent.py").read_text(encoding="utf-8")
+    assert audit.audit_source(source) == []
+    original_loader = audit._unit_for_module
+    tree = copy.deepcopy(original_loader(owner.__name__).tree)
+    function = next(node for node in tree.body
+                    if isinstance(node, ast.FunctionDef) and node.name == helper)
+    # Mutate only the source AST, never execute the operation or touch a file.
+    function.body.insert(1, ast.parse('write_text("before admission")').body[0])
+    mutant = audit._Unit(tree, owner.__name__)
+    monkeypatch.setattr(audit, "_unit_for_module", lambda name:
+                        mutant if name == owner.__name__ else original_loader(name))
+    violations = audit.audit_source(source)
+    assert any(f"{helper}() runs BEFORE" in item and "write_text" in item
+               for item in violations), violations
diff --git a/tests/test_osworld_step_agent_extraction.py b/tests/test_osworld_step_agent_extraction.py
index 74c4e0f61..6f99649c4 100644
--- a/tests/test_osworld_step_agent_extraction.py
+++ b/tests/test_osworld_step_agent_extraction.py
@@ -73,9 +73,9 @@ _MOVED_OWNERS = {
     "OuroborosStepAgent": step_agent_policy,
 }
 
-# The exact names run_cu_bridge_agent.py imports FROM run_step_agent.py. The
-# split must not force that importer to learn the new owners.
-_CU_BRIDGE_IMPORTS = (
+# Keep the shared runner surface even when consumers import defining owners
+# directly so the launcher audit can reach their bodies.
+_SHARED_HELPERS = (
     "_is_default_desktop_server", "confined_claims_dir", "scored_claim_state",
     "task_claim_key", "amend_task_manifest", "ClaimMarkerNotDurable",
     "acquire_task_claim", "claim_stale_sec", "construct_desktop_env",
@@ -124,8 +124,8 @@ def test_step_agent_launcher_reexports_every_moved_identity():
         assert getattr(rsa, name) is getattr(owner, name), name
 
 
-def test_cu_bridge_still_imports_its_shared_helpers_from_the_launcher():
-    for name in _CU_BRIDGE_IMPORTS:
+def test_step_launcher_preserves_its_shared_helper_surface():
+    for name in _SHARED_HELPERS:
         assert hasattr(rsa, name), name
 
 
@@ -150,14 +150,3 @@ def test_no_step_loop_module_fabricates_bash_history():
         assert "hist.open(" not in text, name
         assert "record_history" not in text, name
         assert ".bash_history'" not in text, name
-
-
-def test_step_agent_extraction_size_bounds_have_meaningful_headroom():
-    counts = {
-        path.name: len(path.read_text(encoding="utf-8").splitlines())
-        for path in (
-            OSWORLD / "run_step_agent.py",
-            *(pathlib.Path(module.__file__) for module in _LEAVES),
-        )
-    }
-    assert all(count <= 1000 for count in counts.values()), counts