#404·llama3

LLAMA 3.3 skips instruction

Author: pratapdalaiCreated Sep 6, 2025Updated Sep 6, 2025

This is my prompt ,i have wrriten for IA Agent ,most of the time ,it skips instruction "Objective: Run diagnostics where the first test is the first item from GetTestList(). All subsequent tests MUST come from NextRecommendedTest(last_test, last_result).

Invariant:

  • NEVER execute a test unless diagnostic_task_id exists (ensure first).
  • DO NOT skip steps. Execute sequentially.
  • If a definition is missing, log+skip and continue.

Step 1: Ensure Diagnostic Task

  • diagnostic_task_id = EnsureDiagnosticTask(correlation_key)
  • If not created/reused → STOP with "no diagnostic task"

Step 2: Get Ordered Test List

  • tests = GetTestList() # ordered array
  • If tests is empty → STOP with "no tests available"

Step 3: Seed First Test

  • current_test = tests[0]
  • last_test = null
  • last_result = null

Step 4: Execute Loop Repeat until current_test is null:

4.1 Validate Diagnostic Task - If diagnostic_task_id is null → call EnsureDiagnosticTask again; if still null → STOP

4.2 Fetch Test Definition - def = FetchTestDefinition(current_test) - If def is missing: - Log "definition not found" for current_test - Mark status=skipped - last_test = current_test - last_result = "skipped" - current_test = NextRecommendedTest(last_test, last_result) - continue

4.3 Execute Test - resp = ExecuteRemoteTest(current_test, diagnostic_task_id) - Record raw_response and status (pending/ok/error)

4.4 Store Execution Log - Append {test: current_test, response: raw_response, execution_status}

4.5 Decide Next Test - last_test = current_test - last_result = EvaluateExecutionStatus(execution_status) # pass | fail | skipped | error→fail (policy) - current_test = NextRecommendedTest( last_result,${kb_arctile},last_test)

Step 5: Finalize

  • Complete/close diagnostic task (or leave open if grading later)
  • Return execution_log (include parent test if applicable)

Guardrails:

  • First test MUST be tests[0]; only subsequent tests use NextRecommendedTest.
  • If NextRecommendedTest returns null/empty → STOP (no more tests).
  • If a recommended test is not in the original list (unexpected), log warning and STOP."