preserve output budget for artifact body recovery

src/clean_agent_preview.py

@@ -4393,10 +4393,11 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                         retrievals=successful_web_retrievals,
                         search_limit=2,
                     )
-                round_max_tokens = (
-                    min(request_max_tokens, 4096)
-                    if artifact_body_handoff_target else request_max_tokens
-                )
+                # Thinking providers can consume several thousand tokens before
+                # emitting the requested body.  Reusing the configured output
+                # budget avoids a reasoning-only response that leaves the
+                # required artifact unwritten.
+                round_max_tokens = request_max_tokens
                 request = {'model': model, 'messages': request_messages, 'temperature': temperature,
                            'max_tokens': round_max_tokens, 'stream': True,
                            'stream_options': {'include_usage': True},

tests/test_clean_agent_preview.py

@@ -613,7 +613,7 @@ async def test_repeated_off_contract_calls_recover_via_required_artifact_body(mo
             'completion_requirements': {
                 'required_artifacts': ['/workspace/output.html'],
             },
-        }, max_rounds=5,
+        }, max_tokens=8192, max_rounds=5,
     )]
 
     assert [block.tool_type for block in executed] == ['write_file']
@@ -622,6 +622,8 @@ async def test_repeated_off_contract_calls_recover_via_required_artifact_body(mo
         if message.get('role') == 'assistant' and message.get('tool_calls')
     )
     assert prior_tool_turn['reasoning_content'] == 'reasoning-one'
+    assert requests[2]['max_tokens'] == 8192
+    assert requests[3]['max_tokens'] == 8192
     assert executed[0].content == (
         '/workspace/output.html\n<html><body>Recovered</body></html>'
     )