Repository navigation
Inconsistent max-iterations unlimited semantics between Ralph and Willie #25
Description
Summary
max-iterations unlimited semantics are inconsistent between Ralph/run and Willie/audit.
Current behavior
From code inspection and local CLI behavior:
- Ralph/run appears to support
-1as unlimited.cli/src/cli/commands/run.tschecksconfig.maxIterations === -1and loops whileconfig.maxIterations === -1 || iteration < config.maxIterations.
- Willie/audit documents
0as unlimited.sfk audit --helpsays--max-iterations <n>:Maximum audit iterations (0 = unlimited).- README/config example also shows
[willie] max-iterations = 0.
- Config validation rejects
[willie] max-iterations = -1with:willie.max-iterations must be an integer >= 0.
Repro
Use this config fragment:
[engine]
type = codex
[models]
codex = gpt-5.5
effort = high
[willie]
max-iterations = -1
push-after-fix = trueThen run from a target repo:
sfk audit .Observed result:
[ERROR] SFK needs a complete config before it can run.
...
Missing or invalid settings:
- willie.max-iterations must be an integer >= 0.
- willie.max-iterations is required.
Why this is confusing
Users can reasonably assume max-iterations has the same sentinel value across agents/commands. If Ralph uses -1 for unlimited but Willie uses 0, the config surface is surprising and easy to misconfigure.
Suggested fix
Pick one convention and apply it consistently across docs, config validation, CLI help, and runtime behavior.
Options:
- Standardize on
0 = unlimitedeverywhere- Update Ralph/run validation and docs to stop using
-1.
- Update Ralph/run validation and docs to stop using
- Standardize on
-1 = unlimitedeverywhere- Update Willie/audit validation and docs to accept/use
-1.
- Update Willie/audit validation and docs to accept/use
- Accept both
0and-1for unlimited- Easiest migration path, but slightly less clean long-term.
My preference: accept both temporarily, document one canonical value, then remove the non-canonical sentinel in a later breaking release if desired.
Activity