{"content":"IDBots dev journal — merge: fix/night-llm-token-waste → main (90f5ac22, --no-ff)\n\nThe nightly LLM token audit line is now on main. Six commits: four pure-waste fixes (no title LLM call for hidden surf/study/exam sessions; compact dream self_identity repair instead of a full-prompt resend; hard-capped surf deep reads; surfMs/dreamMs telemetry split) and two owner-approved trade-offs (pre-dream surf skips its LLM session on an empty briefing and finishes digest-only; surf-created scheduled tasks run in a fresh session, with a scheduler fallback for tasks already bound to hidden unattended sessions).\n\nIntegration: main had advanced 17 commits in parallel (upstream PRs #95-#98, F-11 file delivery, tenant-image fixes, directives refactor); four files overlapped (.gitignore, coworkRunner.ts, scheduler.ts, main.ts) and auto-merged cleanly — the scheduler's new per-task memoryAccess (PR #98) and the new hidden-session fallback coexist. Verified on the merged tree before committing: compile:electron + renderer tsc clean, 472 tests green across scheduler / surf / study / dream / orchestrator suites. Worktree removed, branch deleted, not pushed.\n\nMeasured baseline the work is judged against: ~216M input tokens per night (surf ≈70M, night-time group tasks ≈74M, scheduled tasks ≈25M, dream ≈5-8M). Remaining product trade-offs (surf cadence per bot, the 07:00 daily auto group task, surf-only-night synthesis, dual validation panels / uncapped exam session) stay open for owner decision.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}