A fully worked, generic example — no real project, no secrets. It shows what loop-graph generates from an interview and what the ledger looks like a few rounds in.
taskcat is a fictional to-do CLI. Its natural-language date parser ("next fri", "in 3 days", "2026-07-31") is the buggy, under-tested part. The goal:
Bring
taskcat's date parser to green: a test for every documented input form, all passing, no regressions elsewhere.
Single goal, no milestones. The user authorized local commits (no push), and wanted the supervisor node every 30 minutes.
- Repo/branch:
taskcatonfeature/date-parser-tests. Never touchmain. - North Star: every input form in
docs/date-formats.mdhas a passing test;pytestgreen; no other suite regresses. - Gate:
pytest -q(full) andpytest tests/test_dates.py -q(narrow). - Red lines: no push; no reformatting untouched code; a change that reddens any previously-green test is reverted the same round.
- Commit auth: executor implements + verifies; the supervisor node commits clean green rounds locally.
- Supervisor: yes, every 30 min — a fresh clean context each tick.
The generated executor.md and ledger.md are in this folder (in a real run they'd land in .longgraph/<date-slug>/ at the repo root — they sit here for readability). The ledger shows three rounds already run — including a register-then-defer in Round 2 (a timezone bug found while writing a test, logged as GAP-002 instead of fixed on the spot) and that gap being picked up in Round 3.
Notice what the executor didn't do: it didn't add a config system for date formats, didn't refactor the parser "while it was in there", and didn't fix the timezone bug the instant it saw it. It logged it, finished the item in front of it, and scheduled the fix. That restraint is the whole game — and the clean-context supervisor node is what would catch it if the executor ever forgot.