Subconscious skill suggestions, brain tables, and context window cap

Skill suggestions have been reworked to run fully subconsciously. A new SkillManagerAbility extends SkillBuilderAbility with SYSTEM permissions, bypassing policy enforcement and used only by the background processor. The suggestion message processor now operates on the skills_building channel without user-facing summaries, and the skill builder policy now requires an ‘ask’ subagent.

Seven curated bash skills have been added — system-health-check, media-conversion, archive-management, process-inspector, network-diagnostics, disk-cleanup, and certificate-checker — all framed for personal assistant use. The total curated skill count rises from 66 to 73.

A self-unregistering service worker targets stale PWA cleanup for visitors who cached old JavaScript before a previous update. The no-op worker purges all caches, claims clients, then unregisters itself.

The brain’s day view now correctly isolates only today’s UTC hours, preventing yesterday’s data from leaking into future slots. The token usage table below the chart has been expanded to show full time-slot coverage — 24 hours for day, 7 days for week, all days for month, and monthly aggregates for lifetime — with zeros for empty slots. The table columns adjust between Date/Hour, Input Tokens, and Output Tokens depending on the window.

A 200,000-token cap has been introduced to prevent request payloads from exceeding this limit, regardless of what upstream provider APIs report. The cap is applied via a new constant at both the providers gateway and boot backfill.

• Skill suggestions moved to subconscious processing via SYSTEM-level SkillManagerAbility

• Seven new curated bash skills added, total now 73

• Stale PWA cleanup through self-unregistering service worker

• Token usage table covers all time slots with zeros for empty periods

• Day view in brain shows only today’s UTC hours

• Model context window capped at 200k tokens