feat: add BudgetPolicy rules for token/cost/tool-call limits #409
Description
Summary
PolicyEvaluator handles capability checks but not resource consumption limits. Consumers must implement token budgets, tool call limits, and dollar cost caps custom.
Proposal
Add built-in BudgetPolicy rules:
- max_tokens_per_task: int (e.g., 8000)
- max_tool_calls_per_task: int (e.g., 20)
- max_cost_per_task: float (e.g., 0.50 USD)
- max_cost_per_agent_per_day: float (e.g., 10.00 USD)
- Accumulator that tracks consumption within a task/session
- Automatic DENY when budget exceeded, with remaining_budget in response
Why This Matters
Resource consumption governance is table stakes for production agent deployments. LLM API costs can spiral without limits. Every consumer needs token/cost budgets — this shouldn't be reimplemented.
Real-World Evidence
ai-native-team daemon governance.py implements: token_budget (default 8000), max_tool_calls (default 20), per-agent cost tracking in reliability.py. All generic, all reimplemented from scratch.