Subconscious skill building, brain polish, 200k context cap

Skill suggestion has been reworked to run fully in the background (TKT-675). A new SkillManagerAbility extends SkillBuilderAbility with SYSTEM=True, bypassing policy enforcement so only the background processor can drive skill curation. SkillSuggestionMessageProcessor now runs on a dedicated CHANNEL=skills_building, delegates to skill_manager, and drops the user-facing post_turn summary, store, cooldown, and WebSocket broadcast. get_previous_messages() now returns ‘’ so each run is independent, the skill_builder policy switches subagent from allow to ask, and phase-4 invariants plus feature tests were updated to match.

Seven curated bash skills were added to the catalog (TKT-610 framing, personal-assistant focus): system-health-check, media-conversion, archive-management, process-inspector, network-diagnostics, disk-cleanup, and certificate-checker. The curated set now spans 66→73 skills and 221 total entries after rebuilding skills.sqlite.

A self-unregistering service worker was shipped to clean up stale PWA caches for clients that visited before TKT-613. On install it purges all caches, claims existing clients, then unregisters itself — a one-shot no-op that stops old JS from being served.

The brain’s token-usage views got a substantial pass under TKT-673. _buildDaySlots now keys off today’s UTC date prefix instead of aggregating by hour suffix, so the day table shows only today’s hours. The usage table beneath the chart now renders full time-slot coverage: 24 rows per day, named weekdays for the week view, every day of the current month, and monthly aggregates from first data for the lifetime view. Empty slots show zero, and a data table with Date, Input Tokens, and Output Tokens sits below the chart, switching the first column header to ‘Hour’ on day/hour windows.

Finally, the model context window is now hard-capped at 200k tokens (TKT-674). A MAX_CONTEXT_WINDOW = 200_000 constant is applied via min() at both the Providers gateway and the boot backfill, so no request payload can exceed 200k regardless of what the upstream provider reports.

• New SkillManagerAbility (SYSTEM=True) bypasses policy, used only by background processor

• SkillSuggestionMessageProcessor moved to CHANNEL=skills_building with independent runs (empty previous_messages) and no user-facing summary

• 7 curated bash skills added (system-health-check, media-conversion, archive-management, process-inspector, network-diagnostics, disk-cleanup, certificate-checker) — 66→73 skills, 221 entries

• Self-unregistering service worker purges stale PWA caches for pre-TKT-613 clients

• Brain token-usage table now shows full time-slot coverage (24h/day, named weekdays, full month, monthly lifetime) with zeros for empty slots

• MAX_CONTEXT_WINDOW = 200_000 applied at Providers gateway and boot backfill to cap every request payload