Skip to content

LLAMA 3.3 skips instruction #404

@pratapdalai

Description

@pratapdalai

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."

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions