Skip to content
anchor

Working until a check passes

Models stop early and say they’re done. /until makes a command, not the model, decide when the work is finished.

› /until dotnet test
After each turn anchor runs `dotnet test` and keeps working until it exits 0 (at most 5 rounds, or until a round changes no files).


› fix the failing parser tests
  ...the model works...
  ✗ check failed (round 1 of 5): dotnet test
  ...the model gets the failing output and works again...
  ✓ check passed: dotnet test

How it works

Section titled “How it works”

After each turn, anchor runs your check command in the working directory.

  • Exit code 0: the work is done.
  • Anything else: the model gets the check’s output as its next message and another round. The output is masked for secrets and cut to its last 8,000 characters, where test runners put the failures and the summary.

The loop stops when:

  • the check passes,
  • a round changes no files, because the check would fail the same way again (this also covers the model stopping to ask you something, so you can just answer), or
  • 5 rounds have run.

When it stops without passing, anchor says why:

  check still failing and the last round changed no files: dotnet test

Ctrl+C cancels the loop like any turn.

In the REPL

Section titled “In the REPL”
Command Effect
/until <command> Set the check. It applies to every message until cleared.
/until Show the current check.
/until off Clear it.

A message and all its rounds count as one turn, so /undo reverts the whole loop.

In scripts

Section titled “In scripts”
Terminal window
anchor -p --yolo "fix the failing tests" --until "dotnet test"

--until needs -p. If the check never passes, anchor exits with code 4. With --json, each run of the check emits a check event and the final result line has a check field: passed, no_changes or out_of_rounds.

About the check command

Section titled “About the check command”

The check is your own command, like !cmd, so it runs without an approval prompt and isn’t subject to the shell policy. It has a 10-minute timeout.

Good checks are fast, deterministic and exit non-zero on failure:

  • dotnet test, npm test, cargo test, go test ./..., pytest -q
  • make lint && make test
  • ./scripts/verify.sh