Frugal bounded autonomous depth loop

[mvanhorn]

Issue

None.

Version and model info

None.

Problem

Social comparisons say Aider is cleaner/frugal while Cline goes deeper multi-step at huge token cost. Users want more autonomous depth without adopting Cline-scale spend. Architect mode helps editing, but there is no first-class bounded "keep going" loop with hard turn and token ceilings.

Proposed approach

1. CLI flag + command.
2. Loop in base/architect coder after apply, reusing existing send/apply paths.
3. Hard stops: N, token budget, parse failure, user interrupt.
4. Docs tip contrasting with unbounded agent harnesses.

I'm happy to open a PR implementing this if the direction works for you.