[field 2026-10-13] token-burn: add EUR cost, agent working time, human prompts and per-round attribution #296

Problem

bin/token-burn.sh counts tokens per model / day / stage, but the questions we were asked for a reference report were money and time per round and per feature. We had to write ad-hoc scripts over the transcripts (and the agent's permission classifier rightly blocked raw transcript listing, so they had to print aggregates only).

What was needed, all derivable from the same transcripts:

• EUR cost at configurable per-model rates (input, cache write, cache read, output), including sub-agents on cheaper models
• agent working time: sum of main-thread gaps ≤ N minutes, excluding gaps that end at a human prompt
• human prompt count (a proxy for developer effort)
• attribution to windows (setup / map / round N / wrap-up) by timestamp boundaries, or by story markers

Our result, for reference: €142.63 total over 20.5 agent hours, 2,747 model calls, 33 human prompts; one-time start (setup + map + test stack) €34.87; ≈ €9.40 per feature delivered after that.

Proposal

Extend token-burn with --rates <file>, --windows <file> (or round markers from the build log), and an agent-time/prompt column; output aggregates only. Feed the result into status and into report templates.

Field context: a 127-module production Mendix 11.12 app (17,500 model units), 11 backlog features delivered in 4 rounds with Claude Code + mxcli v0.24 + this toolkit. Details anonymized.

🤖 Generated with Claude Code