Recover from malformed agent edits

tests/test_git.py

@@ -25,4 +25,32 @@ def test_summary_includes_untracked_files(tmp_path) -> None:
     summary = workspace.summary()
 
     assert summary.files == ("new.php",)
-    assert summary.added_lines == 1
+    assert summary.added_lines == 1
+
+
+def test_discard_changes_restores_clean_disposable_workspace(tmp_path) -> None:
+    workspace = GitWorkspace(settings(workspace_root=tmp_path), 3)
+    workspace.path.mkdir()
+    subprocess.run(["git", "init"], cwd=workspace.path, check=True, capture_output=True)
+    subprocess.run(
+        ["git", "config", "user.email", "test@example.invalid"],
+        cwd=workspace.path,
+        check=True,
+    )
+    subprocess.run(
+        ["git", "config", "user.name", "Test"],
+        cwd=workspace.path,
+        check=True,
+    )
+    tracked = workspace.path / "tracked.php"
+    tracked.write_text("<?php\nreturn [];\n")
+    subprocess.run(["git", "add", "."], cwd=workspace.path, check=True)
+    subprocess.run(["git", "commit", "-m", "base"], cwd=workspace.path, check=True, capture_output=True)
+    tracked.write_text("<?php\ninvalid\n")
+    untracked = workspace.path / "extra.php"
+    untracked.write_text("<?php\n")
+
+    workspace.discard_changes()
+
+    assert tracked.read_text() == "<?php\nreturn [];\n"
+    assert not untracked.exists()

tests/test_prompt.py

@@ -176,3 +176,27 @@ def test_unresolved_placeholder_task_guides_real_answer_path() -> None:
     assert "exact token map" in task
     assert "HTTP-backed AIService test" in task
     assert "local database-result fallback" in task
+
+
+def test_cross_database_report_guides_staged_source_routing() -> None:
+    report = FailureReport(
+        id=56,
+        chat_message_id=1,
+        source="user_feedback",
+        category="incorrect_answer",
+        user_feedback="Wrong hours",
+        feedback_type="wrong_result",
+        expected_behavior="Join cockpit_config.time_registration to customer manager and ERP positions",
+        prompt="Hours in April",
+        response="0 hours",
+        sql_query="SELECT SUM(hours) FROM LX_AUFGABE",
+        error_message=None,
+        context={"database_type": "erp"},
+    )
+
+    task = build_task(report)
+
+    assert "tables from multiple databases" in task
+    assert "staged-query workflows" in task
+    assert "Do not put an impossible cross-connection join" in task
+    assert "preserving tenant isolation" in task

tests/test_worker.py

@@ -105,6 +105,52 @@ def test_long_stage_emits_heartbeat(caplog) -> None:
     assert "Report 7 agent attempt 1 finished" in caplog.text
 
 
+def test_php_syntax_failure_discards_attempt_before_full_validation() -> None:
+    report = FailureReport(
+        id=56,
+        chat_message_id=1,
+        source="user_feedback",
+        category="incorrect_answer",
+        user_feedback="Wrong hours",
+        feedback_type="wrong_result",
+        expected_behavior="Use time registration",
+        prompt="Hours in April",
+        response="0 hours",
+        sql_query="SELECT 0",
+        error_message=None,
+        context={},
+    )
+    config = settings(mode="pr", dry_run=False, gitea_token="token")
+    worker = AutomationWorker(config)
+    worker.reports = Mock()
+    worker.reports.claim_next.return_value = report
+    worker.agent = Mock()
+    worker.agent.run.side_effect = ["broken PHP", "fixed PHP"]
+    worker.gitea = Mock()
+    worker._run_tests = Mock(return_value=[])
+    workspace = Mock()
+    workspace.path = Path("/tmp/nonexistent-workspace")
+    workspace.branch = "ai/failure-56"
+    workspace.summary.return_value = ChangeSummary(
+        ("config/erp_semantics.php", "tests/Unit/FixTest.php"), 20, 2
+    )
+    workspace.run_args.side_effect = [
+        Mock(returncode=255, stdout="Parse error", stderr="unexpected token"),
+        Mock(returncode=0, stdout="No syntax errors", stderr=""),
+        Mock(returncode=0, stdout="No syntax errors", stderr=""),
+    ]
+    workspace.commit.return_value = "commit-id"
+
+    with patch("lai_automation.worker.GitWorkspace", return_value=workspace):
+        assert worker.run_once()
+
+    assert worker.agent.run.call_count == 2
+    assert "unexpected token" in worker.agent.run.call_args_list[1].args[1]
+    assert "clean repository" in worker.agent.run.call_args_list[1].args[1]
+    workspace.discard_changes.assert_called_once()
+    worker._run_tests.assert_called_once()
+
+
 def test_failed_validation_gets_one_correction_before_pr() -> None:
     report = FailureReport(
         id=13,
@@ -134,6 +180,7 @@ def test_failed_validation_gets_one_correction_before_pr() -> None:
     workspace.summary.return_value = ChangeSummary(
         ("app/Services/Fix.php", "tests/Feature/FixTest.php"), 10, 1
     )
+    workspace.run_args.return_value = Mock(returncode=0, stdout="No syntax errors", stderr="")
     workspace.commit.return_value = "commit-id"
 
     with patch("lai_automation.worker.GitWorkspace", return_value=workspace):
@@ -177,6 +224,7 @@ def test_no_change_response_gets_one_correction() -> None:
         ChangeSummary(("app/Services/Fix.php", "tests/Feature/FixTest.php"), 10, 1),
         ChangeSummary(("app/Services/Fix.php", "tests/Feature/FixTest.php"), 10, 1),
     ]
+    workspace.run_args.return_value = Mock(returncode=0, stdout="No syntax errors", stderr="")
     workspace.commit.return_value = "commit-id"
 
     with patch("lai_automation.worker.GitWorkspace", return_value=workspace):
@@ -219,6 +267,7 @@ def test_test_only_change_gets_correction_before_validation() -> None:
         ChangeSummary(("app/Services/Fix.php", "tests/Unit/FixTest.php"), 15, 1),
         ChangeSummary(("app/Services/Fix.php", "tests/Unit/FixTest.php"), 15, 1),
     ]
+    workspace.run_args.return_value = Mock(returncode=0, stdout="No syntax errors", stderr="")
     workspace.commit.return_value = "commit-id"
 
     with patch("lai_automation.worker.GitWorkspace", return_value=workspace):