AI cost dashboard

The Cost dashboard in the AI Cockpit shows what your AI is actually spending — broken out per agent, per turn, and per round.

What's on the screen

Per-day spend strip

A horizontal strip of the last 14 days, each bar colored by agent. Tap a day to drill in.

Per-agent table

For the selected day:

A high $/turn on accounting (e.g. > $4) usually means the agent ran many rounds, the prompt didn't cache well, or a non-trivial chunk of work was retried.

Turn drill-down (gh#414)

Tap a row to expand into per-turn rows. Each turn shows:

• The operator's prompt (first 80 chars)
• Total cost for the turn
• Rounds breakdown — accounting_round_1, accounting_round_2, … each row showing the model, input tokens, output tokens, cache hits, cost
• Per-round scope: which tools were called, with success/failure

This is the surface that gh#414 introduces. Earlier the dashboard showed bare accounting_round_N rows with no context — now each round has its operator prompt + per-round tool scope.

Common patterns to look for

• Cache-write churn — repeated rounds with low cache hit % means the agent rebuilt its system prompt each round. The Sub-Agent registry caches by (agentName, version); if you edit the prompt mid-day, the cache invalidates and you'll see the % drop.
• Long tool loops — if accounting_round_12 keeps showing up, the agent is hitting its maxToolRounds cap. Either lower it (cheaper but answers may be incomplete) or raise it (better answers, more cost).
• Wrong-model selection — if Haiku is doing OK on invoice math (rare), the dropdown lets you A/B; if Opus is over-spending on simple lookups, swap to Sonnet on the offending agent.

Export

There's a CSV export button at the top right for finance review — downloads the selected date range's per-day per-agent totals.

Implements: gh#414 (turn-grouped drill-down via metadata enrichment). Related: gh#412 (cost analysis + 6 cost-reduction options), gh#415 (Agents editor where you change the dials this dashboard measures).