screen-locker/AGENTS.md

do NOT run tests unless specifically instructed to do so or before committing If tests fail on the same issue twice in a row, STOP and ask the user how to proceed instead of continuing to fix and retry. ALWAYS confirm that the feature you add / bug you fixed behaves as it should by running the program after your changes (not tests!) and inspecting output comparing it with what user wanted, after confirming by yourself ask user if the program behaves as they intended After running tests fix all coverage gaps and issues, do not ignore unless specifically instructed to do so You are NOT done until you install the new version on the phone itself (flutter install --debug from the workout_app directory, then adb shell monkey -p com.kuhy.workout_app to launch).

NEVER fail silently

A failure that says nothing is a bug. The PC's workout sync did nothing for weeks because a missing token returned [] and real errors only logged at INFO — nobody could tell "nothing to do" from "it broke".

• Swallowed exceptions must say why, loudly. Every except must either re-raise or log at warning/error/exception/critical with the concrete reason. debug/info do NOT count — an invisible log is a silent failure with extra steps. Enforced by scripts/check_silent_failures.py (pre-commit hook no-silent-failures); ruff's S110 only catches except: pass, which is the rare case.
• Guard clauses that abort work must log too. if token is None: return is a silent no-op. Say what was missing and what it means.
• Return something meaningful. Prefer a typed result (e.g. _manual_push.PushResult(pushed, record_count, reason)) over a bare None/[]/False, so the caller can react instead of guessing. reason should read like a sentence a human can act on.
• No escape hatch — # noqa is banned repo-wide. If a failure really is benign, log it at warning and move on.

The shutdown hour is derived, never incremented into

/etc/shutdown-schedule.conf is recomputed from scratch by _shutdown_base.reset_to_base_if_new_day on each new day: BASE_HOUR (20) + today's workout hours + the LeetCode hour, capped at 23. Anything that pushes the hour with a plain read-add-write and is not a term of that derivation gets wiped by the next reset -- that is how a 00:02 workout lost its hours on 2026-09-13. Add a new bonus source as a term first, then (optionally) as a live pass stamped in shutdown_base.json so the two do not double-apply. The base is the constant, not the state file: the file once persisted base_*_hour, which made the code's default dead.

The LeetCode hour is read from leetcode-guard's ledger (_leetcode_bonus.py): HMAC-verified credit entries, keyed on LeetCode's own submitted_at, flat +1h/day, and fail closed -- an unreadable ledger or key earns nothing. leetcode-guard itself stays read-only; it never writes the config.