Pilas de objetivos y retroceso

Una pila de objetivos

goal_stack = []

def push(goal):
    goal_stack.append(goal)

def pop():
    return goal_stack.pop() if goal_stack else None

def current():
    return goal_stack[-1] if goal_stack else None

# --- demo ---
push('write the report')
push('collect the data')
print(f'Current goal: {current()}')
print(f'Popped: {pop()}')
print(f'Current goal after pop: {current()}')

Example: Refactoring Workflow

push('Refactor auth')
while current():
    g = current()
    if g == 'Refactor auth':
        if not done('add_oauth_dep'):
            push('Add OAuth dependency')
        elif not done('replace_login'):
            push('Replace login endpoint')
        else:
            pop()   # done
    elif g == 'Add OAuth dependency':
        # ...actually do the work...
        mark_done('add_oauth_dep')
        pop()

Retroceso

Backtracking Example

def attempt(goal):
    for approach in [oauth, saml, custom_token]:
        try:
            return run_with(goal, approach)
        except FailedException:
            continue
    raise NoApproachWorked(goal)

Seguimiento de los enfoques probados

{
    'goal_id': 'refactor-auth',
    'tried': ['oauth', 'saml'],
    'remaining_options': ['custom_token']
}

El LLM como selector de objetivos

next_goal = llm.invoke(f'Current goal: {current()}. Done so far: {done_list}. What sub-goal next?').content
push(next_goal)

Precondiciones y efectos

action = {
    'name': 'replace_login',
    'preconditions': ['oauth_dep_installed'],
    'effects': ['login_endpoint_replaced']
}

Bucle de planificación y actuación

Simbólico híbrido + LLM

Puntos muertos

if no_progress_in_n_steps(5):
    pop_to_parent()
    try_different_approach()

Limite la profundidad del retroceso

MAX_BACKTRACKS = 10
if backtrack_count >= MAX_BACKTRACKS:
    escalate_to_human()

Muestre los retrocesos a los usuarios

Activador del retroceso

Repaso