Cost Per Issue

The Cost Per Issue metric tracks what it costs to move work items through your SDLC workflow. The Factory dashboard uses it for cost trends, cost per station, and cost efficiency ratios.

How the live number is produced

Costs enter the system on factory events: when a session completes, the emitted factory_events row carries costUsd (and token counts) in its payload. The metrics pipeline processes each event at ingest time and upserts additive factory_metrics rows:

1. Event ingest - processEventForMetrics runs for every factory event.
2. Cost upsert - if the event carries costUsd > 0, the pipeline adds it to the cost_per_issue metric for that station, in both hourly and daily buckets (provider + token counts land in the bucket metadata).
3. Read path - the dashboard's Cost panel calls GET /api/factory/metrics?metricType=cost, which aggregates those buckets via getCostMetrics.

There is no hourly cron: aggregation happens on ingest, plus an operator-triggered re-aggregation endpoint (POST /api/factory/aggregate) that rebuilds buckets from the raw event log.

Reading cost data

Dashboard view

The Factory dashboard's cost panel shows:

• Average cost per issue - mean of the cost_per_issue metric buckets in the window
• Cost per station - summed cost grouped by station (development, qa, …)
• Cost trend - daily cost series
• Human vs agent activity split - the ratio of human-actor to agent-actor factory events (an event-count ratio, not a dollar figure)

API access

curl -H "Authorization: Bearer $RENSEI_API_KEY" \
  "https://app.rensei.ai/api/factory/metrics?metricType=cost&timeRange=30d"

Response shape (CostMetrics):

{
  "avgCostPerIssue": 1.23,
  "costPerStation": [
    { "station": "development", "cost": 0.62 },
    { "station": "qa", "cost": 0.38 }
  ],
  "costTrend": [
    { "date": "2026-05-03", "cost": 1.18 },
    { "date": "2026-05-04", "cost": 1.25 }
  ],
  "humanTimePct": 12.5,
  "agentTimePct": 87.5
}

There is no GET /api/factory/cost-per-issue endpoint and no per-issue / per-session cost drill-down API. Cost is attributed per station bucket, not per individual issue, on the live path.

What counts toward cost

Today the only cost source is what session completion events report as costUsd - in practice, LLM token cost computed from the model pricing catalog. Sandbox execution dollars and human review time are not costed:

• Sandbox provider charges (e2b, Modal, …) are not metered into costUsd.
• Human review sessions carry no hourly rate; humans only influence the event-count activity split above.

If model pricing is missing from the catalog for a provider, its sessions contribute 0 - check the Model Catalog before trusting cost figures.

The dormant per-issue rollup module

A second, more precise implementation exists in platform/src/lib/factory/cost-per-issue.ts (+ cost-per-issue-job.ts): a per-issue rollup with a two-leg cost model -

• Token leg - input/output tokens × catalog pricing
• Compute leg - session wall-clock hours × a flat COMPUTE_HOURLY_RATE_USD ($0.08/hr, the amortised per-hour cost of one managed runner slot)

It is coded and unit-tested but not wired: no route exposes getCostPerIssue, no cron runs the recompute job, and its cost_per_issue table has no production rows. Treat it as infrastructure that may be promoted - or removed - rather than a feature you can call.

Cost optimization strategies

• Route simple tasks to cheaper models - Haiku vs Sonnet via routing profiles
• Improve specification clarity - fewer agent loops and rework sessions per issue
• Watch the trend line - a rising per-issue trend usually means more rework or a more expensive model mix

Next Steps

• To configure model pricing, see Model Catalog & Routing
• To set cost caps and budgets, see Cost and Caps
• To understand cost trends, see Cost Breakdown
• To calculate cost efficiency (yield/cost), see Agentic DORA