Repository navigation
skillopt-sleep: max_tokens_per_night is defaulted and documented but never enforced #313
Description
What happens
skillopt_sleep/config.py defines "max_tokens_per_night": 400_000 under the budgets section, and skillopt_sleep/budget.py has a Budget controller that "stops cleanly when exhausted and reports what it skipped". Nothing connects them on main: the key is never read outside DEFAULTS, and Budget( is never constructed. A night's spend is therefore bounded only by max_tasks_per_night, edit_budget and the number of fan-out skill groups, while the config suggests a token cap.
How to check
Version: main at 343db22 (uv tool install from git), Python 3.14, macOS.
P=$(python -c 'import skillopt_sleep, os; print(os.path.dirname(skillopt_sleep.__file__))')
rg -n 'max_tokens_per_night' "$P" # only config.py DEFAULTS
rg -n 'Budget\(' "$P" # only the class definition in budget.pyThis was also noted in passing in the closed feedback issue #57 ("max_tokens_per_night exists, but no real budget enforcement"); that issue was closed for its other points, and the budget is still not enforced at 343db22.
Request
Either wire max_tokens_per_night (and optionally a wall-clock cap) into the cycle through Budget, stopping before the next model call once it is spent and saying so in report.md, or remove the key from DEFAULTS and the docs so users do not rely on it.
Activity
Yif-Yang commentedon Oct 17, 2026
#317 now implements a documented step-boundary limiter, but the production tool route still schedules an uncached final validation call after the initial step has exhausted the budget. That behavior persists with #251's accounting fix. The review asks for accurate scheduling/evidence handling within the declared step-based policy, not an undisclosed hard-cap redesign. The issue remains open pending that correction, official checks and merge.
Yif-Yang commentedon Oct 18, 2026
Resolved by #317, merged as
1a44050fc53a. The configured budget is now enforced at the documented stage boundaries, including the previously missed uncached final-validation step. Actual backend/optimizer/gate and stored-artifact tests confirm that skipped final validation reuses complete evidence for the final retained documents, including accepted/rejected edits, tool/text routes, timeouts and current accounting behavior.All six exact-head CI jobs and CLA passed; the actual merged main was also re-fetched and independently tested. The issue was automatically closed by the merge. This is a stage-boundary limiter: a started step may finish above the threshold, and backend token counters/estimates are not a per-call hard limit, an elapsed-time limit, or a guaranteed monetary cap.
fix(sleep): enforce max_tokens_per_night in the nightly cycle (micros…