Cost Report

Summarize local spend by day, model, and session from the metrics log the harness's own Stop hook writes. This is a report, not a metering system: it reads what the tracker wrote, it doesn't write that file itself.

Where the data lives

The cost-tracking Stop hook appends one JSON object per session-stop to a local metrics log. Each row is a cumulative snapshot for that session, so the report takes the latest row per session id and sums across sessions — summing every row would multiply-count.

Row schema: { timestamp, session_id, transcript_path, model, input_tokens, output_tokens, cache_write_tokens, cache_read_tokens, estimated_cost_usd }

What this does

1. Check the metrics log exists. If not, the tracker hasn't run yet — it only writes on a session's first Stop event, so a brand-new install has no data until then.
2. Reduce rows to the latest snapshot per session and aggregate.
3. Present a compact report (today / yesterday / total / by-model / last-7-days), or export recent rows as CSV on request.

Rely on the tracker's precomputed estimated_cost_usd values; don't re-estimate pricing from raw tokens in the report itself — that's the tracker's job, not the report's.

Report shape

1. Summary: today, yesterday, total, session count.
2. By model: models ranked by total cost.
3. Last seven days: date and cost.

Anti-pattern quick reference

Pre-flight checklist — run before delivering any cost report

• Rows reduced to latest-per-session before aggregating, not summed raw.
• Missing/empty log handled with a plain explanation, not an error.
• Cost figures come from the tracker's own estimated_cost_usd, not re-derived here.
• Report includes today/yesterday/total, by-model, and last-7-days sections.