UAT: Autonomy guardrail cost budgets are never enforced during plan execution #7972

New issue

Open

opened 2026-04-12 16:16:05 +00:00 by HAL9000 · 0 comments

HAL9000 commented

2026-04-12 16:16:05 +00:00

Owner

Summary

Autonomy guardrails expose check_tool_budget, check_budget_hierarchy, and record_plan_cost_to_session, and the service wires in CostBudgetService for session/org limits.
PlanExecutor and RuntimeExecuteActor never call these APIs, so tool execution never consults plan/session/org budgets.
As a result, plans blow past configured cost limits with no warning or block, violating the specification’s “Cost Controls” requirements.

Steps to reproduce

Configure guardrails for a plan (e.g. AutonomyGuardrails(tool_budget=1.0) or by wiring CostBudgetService + associate_plan_with_session).
Execute the plan via PlanExecutor/RuntimeExecuteActor while issuing tool calls that should exceed the budget.
Observe that execution proceeds without any denial or budget events.

Expected result

Before or during each tool invocation, the executor should call into AutonomyGuardrailService.check_tool_budget / .check_budget_hierarchy so that plans exceeding their budget are blocked and audit entries are recorded.

Actual result

_enforce_guardrails (src/cleveragents/application/services/plan_executor.py, ~800-818) only checks wall-clock and step counts. No cost checks are performed, and RuntimeExecuteActor.execute simply loops decisions without touching the guardrail service. The budget-specific methods in AutonomyGuardrailService are unused.

Technical notes

AutonomyGuardrailService.check_tool_budget and .check_budget_hierarchy are defined around lines 154-210 and 529-586 respectively but never referenced outside tests/benchmarks.
PlanExecutor is the natural integration point—after resolving each decision/tool we should consult the guardrail service and update budgets via record_plan_cost_to_session.
Until that integration exists, cost budgets configured through guardrails or the cost budget service have zero effect in production.

Automated by CleverAgents Bot
Supervisor: UAT Testing | Agent: uat-tester

## Summary - Autonomy guardrails expose `check_tool_budget`, `check_budget_hierarchy`, and `record_plan_cost_to_session`, and the service wires in `CostBudgetService` for session/org limits. - `PlanExecutor` and `RuntimeExecuteActor` never call these APIs, so tool execution never consults plan/session/org budgets. - As a result, plans blow past configured cost limits with no warning or block, violating the specification’s “Cost Controls” requirements. ## Steps to reproduce 1. Configure guardrails for a plan (e.g. `AutonomyGuardrails(tool_budget=1.0)` or by wiring `CostBudgetService` + `associate_plan_with_session`). 2. Execute the plan via `PlanExecutor`/`RuntimeExecuteActor` while issuing tool calls that should exceed the budget. 3. Observe that execution proceeds without any denial or budget events. ## Expected result - Before or during each tool invocation, the executor should call into `AutonomyGuardrailService.check_tool_budget` / `.check_budget_hierarchy` so that plans exceeding their budget are blocked and audit entries are recorded. ## Actual result - `_enforce_guardrails` (src/cleveragents/application/services/plan_executor.py, ~800-818) only checks wall-clock and step counts. No cost checks are performed, and `RuntimeExecuteActor.execute` simply loops decisions without touching the guardrail service. The budget-specific methods in `AutonomyGuardrailService` are unused. ## Technical notes - `AutonomyGuardrailService.check_tool_budget` and `.check_budget_hierarchy` are defined around lines 154-210 and 529-586 respectively but never referenced outside tests/benchmarks. - `PlanExecutor` is the natural integration point—after resolving each decision/tool we should consult the guardrail service and update budgets via `record_plan_cost_to_session`. - Until that integration exists, cost budgets configured through guardrails or the cost budget service have zero effect in production. --- **Automated by CleverAgents Bot** Supervisor: UAT Testing | Agent: uat-tester

Rows
Columns