Coding Agent Spend
Every night, prices your coding agents' token usage by type, compares repositories on cost per merged PR, finds where the tokens go, and opens the fix as a pull request.
How it works
- Runs when: every night, over the last seven days of coding-agent sessions.
- Prices correctly: each token type at its own rate. Cache reads are most of the tokens and a tenth of the price. The total must match the tool's own cost figure within 3%, or it stops and says so.
- Compares fairly: repositories by cost per merged pull request, not by total spend.
- Finds the sink: where tokens went, by tool and by path, and the largest sink in each repository.
- You get: a pull request per fix: an ignore rule for generated or vendored code, or a settings change benchmarked on the team's own tasks first, with pass rate, cost and time for every setup tested.
What it needs
- Datadog: read coding-agent usage
- GitHub: read merged pull requests, open fixes
- Slack (optional): post the summary