Fix Qwen session tool calls and expose thinking compatibility

harness/tests/test_evidence_pi_restructurer.py

@@ -59,7 +59,7 @@ def test_pi_restructurer_uses_an_ephemeral_no_tools_invocation(tmp_path, monkeyp
     ]
     assert "--no-skills" in argv
     assert argv[argv.index("--extension") + 1].endswith(
-        "/extensions/tht-evidence-json-mode.ts"
+        "/evidence-extensions/tht-evidence-json-mode.ts"
     )
     assert "--system-prompt" in argv
     assert argv[argv.index("--system-prompt") + 1] == "Evidence-only system prompt"
@@ -120,7 +120,7 @@ def test_evidence_authoring_skill_is_loadable_and_declares_the_wire_schema():
 
 def test_evidence_authoring_json_mode_extension_only_rewrites_the_provider_payload():
     extension = (
-        Path(__file__).parents[1] / ".pi" / "extensions" / "tht-evidence-json-mode.ts"
+        Path(__file__).parents[1] / ".pi" / "evidence-extensions" / "tht-evidence-json-mode.ts"
     ).read_text(encoding="utf-8")
 
     assert 'pi.on("before_provider_request"' in extension
@@ -129,6 +129,19 @@ def test_evidence_authoring_json_mode_extension_only_rewrites_the_provider_paylo
     assert "registerTool" not in extension
 
 
+def test_evidence_json_mode_is_not_auto_loaded_into_interactive_sessions():
+    from tht.evidence.authoring import authoring_skill_path
+
+    skill_path = authoring_skill_path()
+    restructurer = PiEvidenceRestructurer("pi", skill_path)
+    extension_path = restructurer._json_mode_extension
+    auto_extensions = skill_path.parents[2] / "extensions"
+
+    assert extension_path.is_file()
+    assert not extension_path.is_relative_to(auto_extensions)
+    assert not (auto_extensions / extension_path.name).exists()
+
+
 @pytest.mark.parametrize("response", [_candidate(), [_candidate()]])
 def test_pi_restructurer_normalizes_bounded_candidate_envelopes(
     tmp_path, monkeypatch, response,