Make agent validation self-correcting

src/lai_automation/config.py

@@ -43,6 +43,7 @@ class Settings:
     poll_seconds: int
     max_changed_files: int
     max_changed_lines: int
+    max_agent_attempts: int
     stale_claim_minutes: int
     dry_run: bool
 
@@ -75,7 +76,7 @@ class Settings:
             test_commands=_commands(
                 os.getenv(
                     "TEST_COMMANDS",
-                    "composer install --no-interaction --prefer-dist;;php artisan migrate:fresh --force;;php artisan test",
+                    "composer install --no-interaction --prefer-dist;;npm ci;;npm run build;;php artisan migrate:fresh --force;;php artisan test",
                 )
             ),
             test_database_host=os.getenv("TEST_DB_HOST", "host.docker.internal"),
@@ -88,6 +89,7 @@ class Settings:
             poll_seconds=int(os.getenv("POLL_SECONDS", "60")),
             max_changed_files=int(os.getenv("MAX_CHANGED_FILES", "12")),
             max_changed_lines=int(os.getenv("MAX_CHANGED_LINES", "800")),
+            max_agent_attempts=int(os.getenv("MAX_AGENT_ATTEMPTS", "2")),
             stale_claim_minutes=int(os.getenv("STALE_CLAIM_MINUTES", "90")),
             dry_run=mode == "dry-run",
         )
@@ -103,3 +105,5 @@ class Settings:
             raise ValueError("FAILURE_DB_PASSWORD is required")
         if not self.test_database_password:
             raise ValueError("TEST_DB_PASSWORD is required")
+        if not 1 <= self.max_agent_attempts <= 3:
+            raise ValueError("MAX_AGENT_ATTEMPTS must be between 1 and 3")

src/lai_automation/prompt.py

@@ -71,3 +71,17 @@ Generated SQL, if relevant:
 
 Run focused tests while working. Leave the repository with the fix and regression test applied.
 """
+
+
+def build_retry_task(exception: Exception) -> str:
+    failure = redact(str(exception))[-6_000:]
+
+    return f"""The previous implementation failed isolated validation.
+
+Correct the existing changes using this test output:
+
+{failure}
+
+Keep the regression test meaningful. Do not remove or weaken tests, safety checks,
+authentication, authorization, SQL validation, or tenant isolation. Do not commit or push.
+"""

src/lai_automation/worker.py

@@ -12,7 +12,7 @@ from .git import GitWorkspace
 from .gitea import GiteaClient
 from .models import FailureReport
 from .policy import evaluate_changes
-from .prompt import build_task
+from .prompt import build_retry_task, build_task
 
 LOG = logging.getLogger(__name__)
 
@@ -43,15 +43,29 @@ class AutomationWorker:
         try:
             workspace.prepare()
             task = build_task(report)
-            metadata["agent_output"] = self.agent.run(workspace.path, task)
+            metadata["agent_output"] = []
+            for attempt in range(1, self.settings.max_agent_attempts + 1):
+                metadata["agent_output"].append(self.agent.run(workspace.path, task))
+                summary = workspace.summary()
+                metadata["changes"] = asdict(summary)
+                policy = evaluate_changes(summary, self.settings, workspace.path)
+                if not policy.allowed:
+                    raise RuntimeError(
+                        "Policy rejected changes: " + "; ".join(policy.errors)
+                    )
+                try:
+                    metadata["tests"] = self._run_tests(workspace)
+                    break
+                except RuntimeError as exception:
+                    if attempt >= self.settings.max_agent_attempts:
+                        raise
+                    LOG.warning(
+                        "Tests failed for report %s on attempt %s; requesting correction",
+                        report.id,
+                        attempt,
+                    )
+                    task = build_retry_task(exception)
 
-            summary = workspace.summary()
-            metadata["changes"] = asdict(summary)
-            policy = evaluate_changes(summary, self.settings, workspace.path)
-            if not policy.allowed:
-                raise RuntimeError("Policy rejected changes: " + "; ".join(policy.errors))
-
-            metadata["tests"] = self._run_tests(workspace)
             summary = workspace.summary()
             metadata["changes"] = asdict(summary)
             policy = evaluate_changes(summary, self.settings, workspace.path)
@@ -101,6 +115,9 @@ class AutomationWorker:
         results = []
         test_environment = {
             "AGENT_TEST_DB_ENABLED": "true",
+            "APP_ENV": "testing",
+            "APP_KEY": "base64:MDEyMzQ1Njc4OWFiY2RlZjAxMjM0NTY3ODlhYmNkZWY=",
+            "AI_MISTRAL_API_KEY": "automation-test-key",
             "TEST_DB_HOST": self.settings.test_database_host,
             "TEST_DB_PORT": str(self.settings.test_database_port),
             "TEST_DB_NAME": self.settings.test_database_name,