← Back to brief

reference token optimization 2026-05-09

memory · reference_token_optimization_2026-05-09.md

Token Optimization Applied — 2026-05-09

Trigger: Joseph reported Claude Code usage stopping with $50+ still available on 2026-05-09. Audit showed 6 high-frequency Cowork tasks driving ~88% of session burn (~260 of ~294 runs/day).

Applied (durable): Via mcp__scheduled-tasks__update_scheduled_task API. This path goes through the task manager, so the changes survive Claude Code restarts and hot-reloads. (Earlier attempts to edit scheduled-tasks.json directly were silently reverted by Claude Code's in-memory state.)

New cadence (canonical — do not auto-revert)

TaskOld cronNew cronWas/now per day
sot-reconciler/15 /45 96 → 32
engagement-reconciler/30 0 *48 → 24
qb-quarterback/30 6-22 0 6-22 *34 → 17
pwa-refresh/30 7-22 0 7-22 *32 → 16
live-ea/20 9-18 1-50 9-18 * 1-530 → 10
dq-realtime-monitor/30 9-18 0 9-18 *20 → 10
Total260 → 109
Net cut: ~151 sessions/day, −51%.

Why this is safe

These changes reduce CHECKING frequency, not the work each run does:

For anything time-critical (re-auth, manual triggers), the launchd bash poller pipeline (extract_trigger_runner.py, session_keep_alive.py, pwa-periodic-refresh) handles it in seconds with zero Claude tokens.

Do NOT auto-revert

If a future audit or skill flags these as "reduced from prior cadence — restore?", do not restore. This was an explicit token-budget decision by Joseph. The descriptions on each task were also rewritten to match the new cadence so reading them confirms the intent.

To re-tighten if business need changes, edit mcp__scheduled-tasks__update_scheduled_task — never patch the JSON file directly.

Companion docs