CortexDB 0.9.4 — the new Admin UI

Feature release headlined by a ground-up redesign of the embedded admin console, plus answer-stage determinism fixes, a facts-plane consolidation pipeline, and a new query-shape classifier. No storage format changes, no migration: 0.9.x data directories work as-is. The only new on-disk state is two sidecar files in the data directory (feature_overrides.json, settings_overrides.json) written when you change things from the UI.

The new Admin UI (the headline)

The bundled web console was rewritten from a single-page memories view into a full sidebar application — still zero-dependency, still embedded in the binary (no CDN, works air-gapped), served at / on the API port.

• Pages: Overview, Memories, Console, Artifacts, Conflicts, Observability, Features, Settings, Logs, Data & scopes, Audit & access, Maintenance, Updates.
• Design system v2: bento Overview with live throughput, interactive hand-rolled SVG charts (hover crosshair, tooltips, draw-in animation), count-up KPI tiles, command palette (Ctrl+K), notification center, drawer/modal overlays, mobile sidebar.
• Features: plain-English bundle toggles — Enrichment, Artifacts, Knowledge graph, Deep re-ranking, Time-travel facts, SQ8 vectors, Flat memory mode, Fast bulk upload, Background maintenance, Telemetry. Toggles persist across restarts, apply live where the server re-reads flags per request, and honestly report "restart required" where it doesn't.
• Settings: a curated registry of the ten AI/processor connections (embeddings, enrichment, recall ranking, answers, verifier, reranker, document/audio/video/image processors). API keys are write-only — the UI only ever sees a configured flag and a last-4 hint.
• Logs: live tail of a 2,000-entry in-process tracing ring buffer — no more shelling into the box for a quick look.
• Updates: version + build provenance (git SHA, build date), bundled release notes, and an optional remote update check (CORTEX_UPDATE_FEED_URL).
• One-click restart: persists all view snapshots, exits cleanly, and lets your supervisor (systemd/Docker) relaunch; the UI polls health and reloads itself when the server is back.
• Ops metrics: /v1/admin/metrics now reports a top-level storage breakdown by component, process RSS (Linux), and WAL event counts.

The UI is backed by a new internal API under /internal/ui/* (authenticated, not part of the stable v1 surface, deliberately kept out of the OpenAPI spec; surfaces will be promoted to /v1 as they prove out). Asset URLs are versioned so browsers can never pin a stale UI across server upgrades.

Answer-stage determinism

Two chronic classes of quantitative questions now answer deterministically instead of via LLM re-counting:

• Stated totals win: when the source material explicitly states a cumulative total, the answer stage extracts and returns it rather than re-counting rows (the scan is silent when nothing is stated).
• Provable windowed zero: count questions over a time window with verifiably zero matching events return an exact zero with the window spelled out.
• Optional two-pass quantitative flow behind CORTEX_TWO_PASS_QUANT.
• The Anthropic router speaks the Claude 5 adaptive-thinking request format (thinking.type=adaptive + output_config.effort).

Facts plane

The derived facts layer gained a real consolidation pipeline: a background fact consolidator merges per-session extractions into canonical facts, a fact sweeper compacts oversized ledgers, question-shaped fact keys (with bounded embedding backfill) make cross-session attribute lookups retrievable, and fact ranking is grade-aware. Belief building now reads post-consolidation facts only. consolidation_backlog is reported in layer stats for ops visibility.

New: query-shape classifier

A new cortex-shape-classifier crate classifies incoming recall/answer queries into shapes (enumeration, temporal endpoint, attribute lookup, …) using prototype scoring with a LinUCB bandit and an escalation path, feeding the answer planner. Wired into the server; shadow-mode telemetry ships in this release.

LLM gateway

• Multi-lane compile pool for enrichment models (DeepInfra, Nebius, Modal) with per-lane health rotation.
• Request timeout raised 90s → 300s for dense-session enrichment; worker-thread panic isolation in the facts pipeline.

Compatibility

• No API-breaking changes; /v1 surface is unchanged.
• /internal/ui/* is explicitly unstable and may change without notice.
• Data directories from any 0.9.x release load without migration.