This example shows how to use Token Saver Loop when your worker is a coding agent that can edit files and run commands.
Improve a small CLI feature without letting the worker roam through the whole repository.
Reviewer prompt:
Read token-saver-kit/START_HERE.md and act as reviewer only.
Create a T2 worker round to add one focused CLI option.
Limit the worker to src/token_saver_loop/cli.py, src/token_saver_loop/core.py,
and tests/test_cli.py. Require python -m unittest tests.test_cli.
- Reads
token-saver-kit/LATEST_WORKER_PROMPT.md. - Changes only the allowed files.
- Runs the required test command.
- Writes:
worker_report.jsonworker_log.mdtests.txtdiffstat.txtround_status.json
The reviewer reads the compact evidence first:
Review the latest worker evidence in token-saver-kit and decide the next step.
Then the reviewer spot-checks:
- whether the file limit was respected
- whether tests passed
- whether
acceptance.*.validatedis true for the important behavior - whether risks or open questions require a same-tier fix
The expensive reviewer model does not need to perform the file search, edit attempts, command retries, or progress narration. It reviews the final evidence packet and only expands into diffs when needed.