Costs Panel
The Costs panel provides real-time visibility into LLM spend across all governed agents. Detect runaway spending, forecast budget exhaustion, and identify optimization opportunities.Cost Velocity Chart
The primary widget shows current burn rate (USD/minute) with a real-time sparkline:Features
- Updated every 5 seconds via WebSocket
- Velocity threshold line shown as a horizontal reference
- Configurable time windows: 1 hour, 6 hours, 24 hours
- Alert state: When burn rate exceeds the configured threshold, the widget turns red with a pulsing border
Velocity Alert State
When aCOST_VELOCITY_ALERT fires:
- Current rate vs configured threshold
- Active throttle status (if action is “throttle”)
- Which agents are contributing most to the velocity
Budget Exhaustion Forecast
A countdown widget showing when budgets will be exhausted at current spend rate:Gauge Color Coding
Forecast Projection
A line chart showing:- Solid line: Actual cumulative spend over time
- Dashed line: Linear budget allocation (expected spend)
- Projected line: Extrapolation showing when the lines cross (exhaustion point)
Budget Scopes
Toggle between views:- Session: Current session budget
- Daily: Today’s budget
- Monthly: Monthly budget
Per-Agent Cost Attribution
A breakdown table showing which agents are responsible for spend:Breakdown Views
- By Agent: Individual agent costs
- By Role: Aggregated by role
- By Model: Which models cost the most
- By Provider: Cost per provider (OpenAI, Anthropic, etc.)
- By Time: Hourly/daily cost breakdown
Cost Optimization Suggestions
An actionable panel showing where money can be saved:Categories
Context Bloat Alert
Flags requests where input tokens exceed 50% of the model’s context window:Refresh Cadence
Optimization suggestions update daily based on the rolling 7-day usage window. This avoids performance impact from constant recomputation.Retry Chain Costs
A table of the most expensive retry chains:
Shows:
- Top 20 most expensive retry chains
- Whether the retry cost cap terminated the chain
- Final outcome (did the retry succeed?)
Model Routing Log
When cost-aware model routing is configured, a log shows routing decisions:
Includes:
- Total savings from routing this month
- Routing frequency (% of requests that hit the ceiling)
- Warning if routing frequency > 50% (“Consider adjusting your ceiling”)
Real-Time Updates
All cost widgets update in real-time via WebSocket:- Burn rate: Every 5 seconds
- Budget gauge: After each request
- Agent attribution: Every 30 seconds
- Optimization suggestions: Daily
- Velocity alerts: Instant (push)