Context

Implement budget enforcement per DESIGN_SPEC §10.4. Since agents make LLM calls that cost real money, the system must enforce spending limits at multiple levels to prevent runaway costs.

Acceptance Criteria

Budget Controls

• Total monthly budget limit (hard cap)
• Alert notification at 75% of budget consumed
• Critical warning at 90% of budget consumed
• Hard stop at 100% — blocks all further LLM calls
• Per-task spending limit
• Per-agent daily spending limit

Enforcement

• Budget check performed before every LLM call
• Alert/warning emitted at threshold crossings
• Hard stop blocks LLM calls with clear error message
• Auto model downgrade at configurable threshold (e.g., switch from Opus to Sonnet)

Auto-Downgrade Boundary (§10.4)

• Model downgrades apply only at task assignment time, never mid-execution
• An agent halfway through a task cannot be switched to a cheaper model — the task completes on its assigned model
• The next task assignment respects the downgrade threshold
• boundary: "task_assignment" field in auto-downgrade config

Tracking & Reset

• Per-agent spending tracked and queryable
• Per-task spending tracked and queryable
• Monthly budget auto-reset on configured date
• Spending history retained for reporting

Testing

• Unit tests for all threshold behaviors (>80% coverage)
• Test: alert fires at exactly 75%
• Test: hard stop blocks at exactly 100%
• Test: auto downgrade triggers correctly at task boundary (not mid-execution)
• Test: monthly reset works

Dependencies

• Depends on cost tracking (M2, done)

Design Spec Reference

• §10.4 — Budget Controls (auto-downgrade boundary clarification)

Updated 2026-03-06: Added auto-downgrade boundary constraint from §10.4 — downgrades at task assignment only, never mid-execution.