Improve agent repair retries

src/lai_automation/prompt.py

@@ -42,7 +42,8 @@ Repair workflow:
 7. Do not duplicate one rule in service code and semantics config, or add explanatory comments for
     self-evident test setup.
 8. Inspect the final diff for malformed edits, unrelated files, hardcoded report values, and tests
-    that pass without exercising the changed behavior.
+    that pass without exercising the changed behavior. Before retrying an edit, reread the current
+    file and never leave conflict markers or partial methods behind.
 """
 
 
@@ -85,6 +86,11 @@ context builder through a test subclass and assert that its real output contains
 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.
+
+For an unknown-column failure, compare every alias.column in the observed SQL with the table
+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.
 """
 
     return f"""You are fixing failure report #{report.id} in the Laravel AI Chat repository.
@@ -131,7 +137,7 @@ Run focused tests while working. Leave the repository with the fix and regressio
 
 
 def build_retry_task(report: FailureReport, exception: Exception) -> str:
-    failure = redact(str(exception))[-6_000:]
+    failure = redact(str(exception))
 
     return build_task(report) + f"""