recover mixed artifact tool payloads

HARNESS_VERSION

@@ -1 +1 @@
-0.20.12
+0.20.13

src/clean_agent_preview.py

@@ -1603,6 +1603,11 @@ def required_artifact_completion_tool_choice(required_artifacts, offered_schemas
     targets = [str(path or '').strip().rstrip('/') for path in required_artifacts]
     if not targets:
         return None
+    # A required directory commonly contains several files and may need a
+    # programmatic extractor.  Requiring *a* tool call preserves forward
+    # progress without trapping the model in one enormous write_file payload.
+    if len(targets) == 1 and not Path(targets[0]).suffix:
+        return 'required'
     offered = {
         canonical((schema.get('function') or {}).get('name')):
         (schema.get('function') or {}).get('name')
@@ -1688,6 +1693,51 @@ def expand_concatenated_write_calls(calls, *, max_calls=16):
                 if len(items) > max_calls:
                     raise ValueError('too many concatenated write calls')
         except (TypeError, ValueError, json.JSONDecodeError):
+            # Qwen's native parser can leave one complete, grounded writer
+            # object followed by textual <tool_call> blocks in the same
+            # argument.  Execute only that leading writer.  The trailing calls
+            # may depend on reads that have not run yet, so they stay inert and
+            # the next model round can propose them normally with fresh
+            # evidence.
+            try:
+                leading, end = decoder.raw_decode(raw)
+            except (TypeError, ValueError, json.JSONDecodeError):
+                leading, end = None, 0
+            tail = raw[end:].strip() if end else ''
+            markup_openers = tail.count('<tool_call>')
+            markup_closers = tail.count('</tool_call>')
+            bounded_markup = (
+                tail.startswith('<tool_call>')
+                and 1 <= markup_openers <= 32
+                # A provider may truncate the last textual call at its output
+                # limit.  It remains inert; only the complete leading JSON
+                # writer is recovered.
+                and markup_closers in {markup_openers, markup_openers - 1}
+                and not strip_tool_blocks(
+                    tail,
+                    skip_fenced=True,
+                    additional_tool_names=('bash', 'python', 'read_file', 'write_file'),
+                ).strip()
+            )
+            if (
+                bounded_markup
+                and isinstance(leading, dict)
+                and set(leading) == {'path', 'content'}
+                and isinstance(leading['path'], str)
+                and bool(leading['path'].strip())
+                and isinstance(leading['content'], str)
+            ):
+                base_id = str(call.get('id') or 'call_write')
+                expanded.append({
+                    'id': f'{base_id}_0',
+                    'type': 'function',
+                    'function': {
+                        'name': function.get('name', 'write_file'),
+                        'arguments': json.dumps(leading, ensure_ascii=False),
+                    },
+                })
+                recovered_count += 1
+                continue
             expanded.append(call)
             continue
         if not (2 <= len(items) <= max_calls) or not all(

tests/test_clean_agent_preview.py

@@ -684,7 +684,7 @@ def test_required_binary_artifact_forces_python_instead_of_text_writer():
     ]
 
 
-def test_required_directory_artifact_forces_writer_with_descendant_schema():
+def test_required_directory_artifact_allows_any_offered_mutation_tool():
     offered = [
         {'type': 'function', 'function': {'name': 'bash'}},
         {'type': 'function', 'function': {'name': 'write_file'}},
@@ -692,7 +692,7 @@ def test_required_directory_artifact_forces_writer_with_descendant_schema():
 
     assert required_artifact_completion_tool_choice(
         ('/tmp_workspace/results',), offered,
-    ) == {'type': 'function', 'function': {'name': 'write_file'}}
+    ) == 'required'
 
 
 def test_action_promise_response_rejects_future_work_but_not_real_answers():
@@ -1035,6 +1035,38 @@ def test_concatenated_writer_recovery_fails_closed_for_ambiguous_or_large_batche
     assert expand_concatenated_write_calls([too_large]) == ([too_large], 0)
 
 
+def test_mixed_writer_argument_recovers_only_grounded_leading_json_write():
+    from src.clean_agent_preview import expand_concatenated_write_calls
+
+    first = {'path': '/workspace/results/1.tex', 'content': 'grounded table'}
+    trailing = '''
+<tool_call>
+<function=bash>
+<parameter=command>sed -n '2,3p' /workspace/paper.tex</parameter>
+</function>
+</tool_call>
+<tool_call>
+<function=write_file>
+<parameter=path>/workspace/results/2.tex</parameter>
+<parameter=content>ungrounded table</parameter>
+</function>
+</tool_call>
+'''
+    source = {
+        'id': 'mixed', 'type': 'function', 'function': {
+            'name': 'write_file',
+            'arguments': json.dumps(first) + trailing,
+        },
+    }
+
+    recovered, count = expand_concatenated_write_calls([source])
+
+    assert count == 1
+    assert len(recovered) == 1
+    assert recovered[0]['id'] == 'mixed_0'
+    assert json.loads(recovered[0]['function']['arguments']) == first
+
+
 def test_malformed_writer_handoff_uses_descendant_file_not_required_directory():
     from src.clean_agent_preview import malformed_write_handoff_target
 
@@ -5527,7 +5559,7 @@ async def test_native_stream_reserves_remaining_budget_for_required_artifact(mon
     ]
     contract = resolve_full_inventory_contract(schemas=schemas, policy=ToolPolicy())
     raw = [chunk async for chunk in stream_preview(
-        endpoint_url="http://test", model="deepseek-flash",
+        endpoint_url="http://test", model="test",
         messages=[{"role": "user", "content": "Research and create the requested output."}],
         headers={}, turn_contract=contract, session_id="test", owner="test",
         disabled_tools=set(), tool_policy=ToolPolicy(), workspace="/tmp/workspace",
@@ -5562,16 +5594,16 @@ async def test_native_stream_reserves_remaining_budget_for_required_artifact(mon
     )
     assert "write_file" in request_contract["offered_tools"]
     assert "web_search" not in request_contract["offered_tools"]
-    assert request_contract["tool_choice"] is None
+    assert request_contract["tool_choice"] == "required"
     reserved_names = [tool["function"]["name"] for tool in requests[12]["tools"]]
-    assert reserved_names == ["write_file"]
+    assert reserved_names == ["bash", "python", "write_file"]
-    assert "tool_choice" not in requests[12]
+    assert requests[12]["tool_choice"] == "required"
-    directory_path = requests[12]["tools"][0]["function"]["parameters"]["properties"]["path"]
+    directory_path = requests[12]["tools"][2]["function"]["parameters"]["properties"]["path"]
     assert re.search(directory_path["pattern"], "/tmp_workspace/results/out.md")
     assert not re.search(directory_path["pattern"], "/tmp_workspace/results")
     assert "create one or more files inside" in requests[12]["messages"][-1]["content"]
     assert "Do not pass the directory itself as a file path" in requests[12]["messages"][-1]["content"]
-    assert "bash" not in executed
+    assert "bash" in executed
     assert any(
         event.get("type") == "tool_output" and event.get("tool") == "write_file"
         and not event.get("error") for event in events

tests/test_harness_version.py

@@ -8,7 +8,7 @@ ROOT = Path(__file__).resolve().parents[1]
 def test_original_harness_version_is_canonical_semver() -> None:
     version = (ROOT / "HARNESS_VERSION").read_text(encoding="utf-8").strip()
 
-    assert version == "0.20.12"
+    assert version == "0.20.13"
     assert re.fullmatch(r"\d+\.\d+\.\d+", version)