Keep repair context across agent retries

src/lai_automation/prompt.py

@@ -43,6 +43,21 @@ schema/semantics, and fix SQL-generation guidance rather than rewriting one unkn
 to another guessed column. TenantDatabaseService records the explicit source type in
 database.connections.<connection>.tenant_type; use that metadata instead of database-name
 patterns or heuristics.
+"""
+    sql_semantics = ''
+    if report.category in {'incorrect_answer', 'sql_execution'}:
+        sql_semantics = """
+SQL repair guidance:
+When the SQL is syntactically valid but uses the wrong business calculation, prefer adding or
+correcting tracked domain knowledge in config/*_semantics.php so the schema prompt teaches the
+model the right joins, filters, and formula. Do not hardcode one user's question or article ID
+in SQLGeneratorService, and do not add post-generation replacement of guessed SQL.
+
+For regression coverage of prompt guidance, expose the relevant protected schema/business-
+context builder through a test subclass and assert that its real output contains the required
+formula and rules. Do not mock AIService to return the known-bad SQL and then expect prompt
+guidance to rewrite that mocked response. Keep database fixture tests only when production code
+itself deterministically builds and executes the SQL under test.
 """
 
     return f"""You are fixing failure report #{report.id} in the Laravel AI Chat repository.
@@ -80,15 +95,18 @@ Observed error:
 Generated SQL, if relevant:
 {sql or '[not available]'}
 {tenant_scope}
+{sql_semantics}
 
 Run focused tests while working. Leave the repository with the fix and regression test applied.
 """
 
 
-def build_retry_task(exception: Exception) -> str:
+def build_retry_task(report: FailureReport, exception: Exception) -> str:
     failure = redact(str(exception))[-6_000:]
 
-    return f"""The previous implementation failed isolated validation.
+    return build_task(report) + f"""
+
+The previous implementation failed isolated validation.
 
 Correct the existing changes using this test output:
 

src/lai_automation/worker.py

@@ -57,6 +57,7 @@ class AutomationWorker:
                             attempt,
                         )
                         task = build_retry_task(
+                            report,
                             RuntimeError(
                                 "No file changes were applied. Make the smallest concrete code and regression-test edits now."
                             )
@@ -76,7 +77,7 @@ class AutomationWorker:
                         report.id,
                         attempt,
                     )
-                    task = build_retry_task(exception)
+                    task = build_retry_task(report, exception)
 
             summary = workspace.summary()
             metadata["changes"] = asdict(summary)

tests/test_prompt.py

@@ -54,12 +54,53 @@ def test_redact_removes_extended_credentials() -> None:
 
 
 def test_retry_task_contains_failure_without_secrets() -> None:
-    task = build_retry_task(RuntimeError("Missing import; token=secret-value"))
+    report = FailureReport(
+        id=8,
+        chat_message_id=1,
+        source="user_feedback",
+        category="incorrect_answer",
+        user_feedback="Wrong result",
+        feedback_type="wrong_result",
+        expected_behavior="Use a weighted average",
+        prompt="Average sale price",
+        response="Zero",
+        sql_query="SELECT AVG(list_price)",
+        error_message=None,
+        context={"database_type": "erp"},
+    )
+    task = build_retry_task(report, RuntimeError("Missing import; token=secret-value"))
 
     assert "Missing import" in task
     assert "secret-value" not in task
     assert "Do not remove or weaken tests" in task
     assert "unrelated production routing" in task
+    assert "Use a weighted average" in task
+    assert "SELECT AVG(list_price)" in task
+
+
+def test_incorrect_answer_task_guides_semantics_level_repair() -> None:
+    report = FailureReport(
+        id=42,
+        chat_message_id=1,
+        source="user_feedback",
+        category="incorrect_answer",
+        user_feedback="Wrong result",
+        feedback_type="wrong_result",
+        expected_behavior="Sum net values and divide by quantity",
+        prompt="Average sale price for article 9031",
+        response="0.00",
+        sql_query="SELECT AVG(Vk_preis) FROM FK_Artikel",
+        error_message=None,
+        context={"database_type": "erp"},
+    )
+
+    task = build_task(report)
+
+    assert "config/*_semantics.php" in task
+    assert "Do not hardcode" in task
+    assert "protected schema/business-" in task
+    assert "context builder" in task
+    assert "known-bad SQL" in task
 
 
 def test_customer_database_task_explains_tenant_scope() -> None: