Recover from malformed agent edits

src/lai_automation/git.py

@@ -5,7 +5,7 @@ import pwd
 import shlex
 import shutil
 import subprocess
-from collections.abc import Mapping
+from collections.abc import Mapping, Sequence
 from dataclasses import dataclass
 from pathlib import Path
 from urllib.parse import urlparse
@@ -48,6 +48,20 @@ class GitWorkspace:
         check: bool = True,
         username: str | None = None,
         extra_environment: Mapping[str, str] | None = None,
+    ) -> subprocess.CompletedProcess[str]:
+        return self.run_args(
+            shlex.split(command),
+            check=check,
+            username=username,
+            extra_environment=extra_environment,
+        )
+
+    def run_args(
+        self,
+        command: Sequence[str],
+        check: bool = True,
+        username: str | None = None,
+        extra_environment: Mapping[str, str] | None = None,
     ) -> subprocess.CompletedProcess[str]:
         environment = dict(os.environ)
         drop_privileges = None
@@ -66,7 +80,7 @@ class GitWorkspace:
             environment.update(extra_environment or {})
 
         return subprocess.run(
-            shlex.split(command),
+            command,
             cwd=self.path,
             env=environment,
             text=True,
@@ -76,6 +90,10 @@ class GitWorkspace:
             preexec_fn=drop_privileges,
         )
 
+    def discard_changes(self) -> None:
+        self._git("reset", "--hard", "HEAD")
+        self._git("clean", "-fd")
+
     def summary(self) -> ChangeSummary:
         self._git("add", "-N", "--all")
         output = self._git("diff", "--numstat").stdout

src/lai_automation/prompt.py

@@ -25,6 +25,9 @@ Repository ownership map:
 - config/cockpit_semantics.php owns device and customer-database semantics.
 - app/Services/AIService.php owns provider calls, pseudonymization boundaries, and response parsing.
 - app/Services/TenantDatabaseService.php owns ERP/config/customer source selection and tenant metadata.
+- Cross-database workflows are orchestrated in ChatController/TenantDatabaseService; one SQL query
+    cannot join tables exposed through different Laravel connections unless the tracked architecture
+    explicitly supports qualified access.
 - app/Services/SQLValidatorService.php owns read-only SQL safety; do not weaken it to accept a fix.
 - app/Http/Controllers/ChatController.php orchestrates routing, execution, answers, and diagnostics.
 - app/Services/FailureTrackingService.php and FailureAutomationService.php own failure handoff.
@@ -44,6 +47,8 @@ Repair workflow:
 8. Inspect the final diff for malformed edits, unrelated files, hardcoded report values, and tests
     that pass without exercising the changed behavior. Before retrying an edit, reread the current
     file and never leave conflict markers or partial methods behind.
+9. PHP config files are executable code. Preserve brackets, commas, quotes, and array nesting; run
+    `php -l` on every changed PHP file before broader tests.
 """
 
 
@@ -91,6 +96,11 @@ For an unknown-column failure, compare every alias.column in the observed SQL wi
 schema and identify the owning table before editing. Teach the exact ownership, join, grouping,
 and real column names in semantics config, then add a reference SQL example for the reported
 question shape. Never replace one missing column with an unverified synonym.
+
+When expected behavior names tables from multiple databases or connections, first inspect source
+routing and existing staged-query workflows. Do not put an impossible cross-connection join only
+in ERP semantics. Implement deterministic orchestration that resolves identifiers in one source
+and passes only the required identifier into the next source, preserving tenant isolation.
 """
     placeholder_guidance = ''
     if '__LAI_' in response:

src/lai_automation/worker.py

@@ -11,7 +11,7 @@ from dataclasses import asdict
 from .agent import AiderAgent
 from .config import Settings
 from .database import FailureRepository
-from .git import GitWorkspace
+from .git import ChangeSummary, GitWorkspace
 from .gitea import GiteaClient
 from .models import FailureReport
 from .policy import evaluate_changes
@@ -98,6 +98,25 @@ class AutomationWorker:
                     raise RuntimeError(
                         "Policy rejected changes: " + "; ".join(policy.errors)
                     )
+                try:
+                    self._lint_changed_php(workspace, summary, report.id, attempt)
+                except RuntimeError as exception:
+                    if attempt >= self.settings.max_agent_attempts:
+                        raise
+                    LOG.warning(
+                        "PHP syntax failed for report %s on attempt %s; discarding malformed attempt",
+                        report.id,
+                        attempt,
+                    )
+                    workspace.discard_changes()
+                    task = build_retry_task(
+                        report,
+                        RuntimeError(
+                            f"{exception}\nThe malformed attempt was discarded. Start again from the clean "
+                            "repository, make smaller edits, and run php -l on each changed PHP file."
+                        ),
+                    )
+                    continue
                 try:
                     metadata["tests"] = self._run_tests(workspace, report.id, attempt)
                     LOG.info("Report %s attempt %s validation passed", report.id, attempt)
@@ -150,6 +169,26 @@ class AutomationWorker:
             except OSError:
                 LOG.exception("Could not clean workspace for report %s", report.id)
 
+    def _lint_changed_php(
+        self,
+        workspace: GitWorkspace,
+        summary: ChangeSummary,
+        report_id: int,
+        attempt: int,
+    ) -> None:
+        for filename in summary.files:
+            if not filename.endswith(".php"):
+                continue
+            result = workspace.run_args(
+                ["php", "-l", filename],
+                check=False,
+                username="coder",
+            )
+            if result.returncode != 0:
+                output = (result.stdout + "\n" + result.stderr)[-4_000:]
+                raise RuntimeError(f"PHP syntax check failed for {filename}:\n{output}")
+        LOG.info("Report %s attempt %s PHP syntax preflight passed", report_id, attempt)
+
     def run_forever(self) -> None:
         LOG.info(
             "Worker started mode=%s model=%s poll_seconds=%s max_attempts=%s tests=%s direct_main=%s",

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):