fix: classify landed file mutations with diagnostics

tests/agent/test_display.py

@@ -1,6 +1,7 @@
 """Tests for agent/display.py — build_tool_preview() and inline diff previews."""
 
 import os
+import json
 import pytest
 from unittest.mock import MagicMock, patch
 
@@ -149,6 +150,27 @@ class TestCuteToolMessagePreviewLength:
         assert path in line
         assert "..." not in line
 
+    def test_write_file_lint_error_result_is_not_marked_failed(self):
+        result = json.dumps({
+            "bytes_written": 12,
+            "lint": {"status": "error", "output": "SyntaxError: invalid syntax"},
+        })
+
+        line = get_cute_tool_message("write_file", {"path": "/tmp/a.py"}, 0.1, result=result)
+
+        assert "[error]" not in line
+
+    def test_patch_lsp_diagnostics_result_is_not_marked_failed(self):
+        result = json.dumps({
+            "success": True,
+            "diff": "--- a/tmp.py\n+++ b/tmp.py\n",
+            "lsp_diagnostics": "<diagnostics>ERROR [1:1] type mismatch</diagnostics>",
+        })
+
+        line = get_cute_tool_message("patch", {"path": "/tmp/a.py"}, 0.1, result=result)
+
+        assert "[error]" not in line
+
 
 class TestEditDiffPreview:
     def test_extract_edit_diff_for_patch(self):

tests/agent/test_tool_guardrails.py

@@ -7,6 +7,7 @@ from agent.tool_guardrails import (
     ToolCallGuardrailController,
     ToolCallSignature,
     canonical_tool_args,
+    classify_tool_failure,
 )
 
 
@@ -131,6 +132,21 @@ def test_success_resets_exact_signature_failure_streak():
     assert controller.before_call("web_search", args).action == "allow"
 
 
+def test_file_mutation_lint_error_result_is_not_a_tool_failure():
+    write_result = json.dumps({
+        "bytes_written": 12,
+        "lint": {"status": "error", "output": "SyntaxError: invalid syntax"},
+    })
+    patch_result = json.dumps({
+        "success": True,
+        "diff": "--- a/tmp.py\n+++ b/tmp.py\n",
+        "lsp_diagnostics": "<diagnostics>ERROR [1:1] type mismatch</diagnostics>",
+    })
+
+    assert classify_tool_failure("write_file", write_result) == (False, "")
+    assert classify_tool_failure("patch", patch_result) == (False, "")
+
+
 def test_same_tool_varying_args_warns_by_default_without_halting():
     controller = ToolCallGuardrailController(
         ToolCallGuardrailConfig(same_tool_failure_warn_after=2, same_tool_failure_halt_after=3)

tests/agent/test_tool_result_classification.py

@@ -0,0 +1,30 @@
+"""Tests for shared tool result classification helpers."""
+
+import json
+
+from agent.tool_result_classification import file_mutation_result_landed
+
+
+def test_write_file_with_nested_lint_error_counts_as_landed():
+    result = json.dumps({
+        "bytes_written": 12,
+        "lint": {"status": "error", "output": "SyntaxError: invalid syntax"},
+    })
+
+    assert file_mutation_result_landed("write_file", result) is True
+
+
+def test_patch_with_nested_lsp_diagnostics_counts_as_landed():
+    result = json.dumps({
+        "success": True,
+        "diff": "--- a/tmp.py\n+++ b/tmp.py\n",
+        "lsp_diagnostics": "<diagnostics>ERROR [1:1] type mismatch</diagnostics>",
+    })
+
+    assert file_mutation_result_landed("patch", result) is True
+
+
+def test_top_level_file_mutation_error_does_not_count_as_landed():
+    result = json.dumps({"success": True, "error": "post-write verification failed"})
+
+    assert file_mutation_result_landed("patch", result) is False

tests/run_agent/test_file_mutation_verifier.py

@@ -166,6 +166,56 @@ class TestRecordFileMutationResult:
         )
         assert agent._turn_failed_file_mutations == {}
 
+    def test_write_file_with_lint_error_counts_as_landed(self):
+        agent = _bare_agent()
+        agent._record_file_mutation_result(
+            "write_file",
+            {"path": "/tmp/a.py", "content": "bad"},
+            json.dumps({"error": "write failed"}),
+            is_error=True,
+        )
+        assert "/tmp/a.py" in agent._turn_failed_file_mutations
+
+        result = json.dumps({
+            "bytes_written": 24,
+            "lint": {"status": "error", "output": "SyntaxError: invalid syntax"},
+        })
+
+        agent._record_file_mutation_result(
+            "write_file",
+            {"path": "/tmp/a.py", "content": "def nope(:\n"},
+            result,
+            is_error=True,
+        )
+
+        assert agent._turn_failed_file_mutations == {}
+
+    def test_patch_with_lsp_diagnostics_counts_as_landed(self):
+        agent = _bare_agent()
+        agent._record_file_mutation_result(
+            "patch",
+            {"mode": "replace", "path": "/tmp/a.py", "old_string": "x", "new_string": "y"},
+            json.dumps({"error": "Could not find old_string"}),
+            is_error=True,
+        )
+        assert "/tmp/a.py" in agent._turn_failed_file_mutations
+
+        result = json.dumps({
+            "success": True,
+            "diff": "--- a/tmp.py\n+++ b/tmp.py\n",
+            "files_modified": ["/tmp/a.py"],
+            "lsp_diagnostics": "<diagnostics>ERROR [1:1] type mismatch</diagnostics>",
+        })
+
+        agent._record_file_mutation_result(
+            "patch",
+            {"mode": "replace", "path": "/tmp/a.py", "old_string": "x", "new_string": "y"},
+            result,
+            is_error=True,
+        )
+
+        assert agent._turn_failed_file_mutations == {}
+
     def test_repeated_failure_keeps_first_error(self):
         agent = _bare_agent()
         agent._record_file_mutation_result(