> This is the compact memory for AI assistants (Claude/Cowork). It is short ON PURPOSE so it
> survives a context-window compaction. The DEEP archive is The KAI Codex.md (multi-MB — GREP it,
> never read it whole, never recursive-grep the tree). Flow each session: (1) read THIS file fully,
> (2) grep -a the Codex for the feature you're touching, (3) confirm against the real file.
>
> KEEP THIS FILE CURRENT. When you finish a piece of work: add one line to "Recent work",
> update "Open issues", and (for code) bump the Codex + Cargo.toml. This file is only useful if the
> last AI updated it.
---
fleet. Be conservative with GPU/CPU; prefer CPU/light options; account for other processes.
[internal module], learned/durable weights survive restart) — retrieval per turn, NOT a context window that fills up. This is the model to respect: KAI already solves long-term memory; don't treat it like a stateless chatbot.the fleet, sign in to the dashboard, hear audio, use a GPU, drive two devices, reach [internal service]/LAN.
→ Ship changes guarded behind a flag or with a fallback, and have the owner verify.
(cloudflared). Origin binds [internal service]. Real auth + 2FA + role gate + in-house guard modules exist.
1. Read this file. 2. grep -ain "<feature>" "The KAI Codex.md". 3. Open the file the Codex names.
grep -r / find over [local path] — huge tree, times out at 120 s.grep -a._calFin*), Teller bank sync | oracle-os/shell/desktop.js |_cal via [internal endpoint] | oracle-os/shell/desktop.js (served at /oracle-os/) |_syncKeys, [internal endpoint]), radio, Stripe billing | oracle-core.js, oracle.html |/api/*, static-file allowlist, /ws/games, auth/role gate | [internal module] |[GroqVoice/Leo]); local-TTS hook at LOCAL_TTS_URL | [internal module] |[internal module] |[internal module], [internal module] |oracle-games-pool.js) | oracle-games-*.js |kaiverse.js, kaiverse-graphics.js |Start-KAI.ps1, ecosystem-manager.mjs |[internal module] (NEVER print secret values) |oracle-os/shell/desktop.js._calWiz/_CAL_WIZ_FLOWS, default for +Add, 'All fields (advanced)' escape). Iterate look/flows with owner._calFinPaidOn/_finPaid treat any bill with date < today as PAID (k<today). Owner: 'its thinking when i pay them is when it is.' Want past-unpaid bills to show OVERDUE, not paid; keep the past=settled assumption ONLY for the cash forecast. Confirmed desired behavior, not yet coded._calD) is confirmed correct, bug is in recurrence/statement logic; guessing risks corrupting financial data. File oracle-os/shell/desktop.js._calFinAddCard, handler osFinSetDate/osCalFinQuickAdd.executeVocalSync + LEO_NO_GEMINI_LIVE=1 routes calls off Gemini Live). Owner MUST run Start-KAI.ps1 to load it. Still open: record leo-voice-ref.wav (7-15s) for a true clone (default voice until then); confirm the STT-fallback call path still works + latency acceptable; revert = LEO_NO_GEMINI_LIVE=0.KAI1 ternary format: models/BitNet/Native/*.kai = 30 blocks × 11 tensors + token_embd + output_norm (330 files). Forward pass runs: today's overnight_pipeline.log shows real [BitNet] ... generations.teaching/kai_language_teacher.py → teaching/teacher.log) currently scores ~0/8 (0%) on Level-1 literacy; KAI often returns "(no reply)".neural_weights.bin (7.7GB) + neural_structure.json are ORPHANED (wrong schema, load always failed). Still sit in data/, models/BitNet/, backups. Do not revive without a new extractor+index+loader.[internal module] (candle quantized_llama, MoE-style routing, VRAM-managed/unload-when-idle). Flags: has_native_transformer(), has_dense_expert(), hybrid_fully_wired, native_decode_count — all in [internal module]; [internal endpoint] reports them.[internal module] ("glass window into KAI's head") + GET [internal endpoint] (v9.10.556): what KAI considered but did NOT say, per-cell scores, branches not taken. Advantage over transformers: KAI's intermediate states are literal TEXT, not residual vectors — already verbalizable.[internal module] = the Hyperdimensional Probe (Bronzini et al. arXiv [internal service].25045) — a learned 2-layer MLP pulling a transformer's residual stream into KAI's 16384-dim ternary VSA basis. Trained weights on disk: [internal module] (167MB), [internal module] (137MB). Training half = [internal module]; blend = blend_mapper_with_state. Looks OPT-IN (loaded via a path arg in main.rs), not blended into the live serving path by default — verify before claiming it's ON.[internal module], bitnet_inference.rs, bitnet_llama.rs; warehouse/dispatch = language_warehouse.rs; probe = neural_mapper.rs + training.rs; introspection = mind_trace.rs; data loop = overnight_pipeline.py, teaching/kai_language_teacher.py.oracle-core.js?v=12.10.901 (never bumped) → browser served cached old file. Bumped ?v→12.10.902, moved button to a prominent full-width 'Listen to Leo Radio' under the Codex title (search-row spot was off-screen on mobile), sw→v43. LESSON: bump the ?v= cache-bust in oracle.html when editing an oracle-*.js, not just the file.oracle-core.js, next to the 📖 Read button — the Codex reader is in oracle-core.js, NOT desktop.js). Opens /leo-radio.html. Hard-refresh + sw v42.[internal endpoint] to GUEST_API_EXACT in [internal module]; needs dashboard restart) + an Oracle OS app tile (leoradio in OS_APPS + iframe render case in osRenderAppBody, desktop.js) that opens /leo-radio.html in a window with an 'Open full page' fallback (hard-refresh + sw v41 to apply)./leo-radio.html + POST [internal endpoint] in command-center-server.mjs): listen-only continuous Leo monologue, no mic. Uses [internal module] askGoogle (Gemini + google_search grounding on GEMINI_API_KEY_LEO) as brain+live-search, Kokoro (:8123) for voice; per-session rolling context in globalThis.__leoRadio; page prefetches next segment for seamless playback. APPLY = .\Start-Dashboard.ps1 then open /leo-radio.html; needs Kokoro :8123 up + Gemini key.LEO_NO_GEMINI_LIVE=0 in [internal module] (owner confirmed the Gemini key works again). Gemini Live is the all-in-one STT+LLM+TTS path (voice=Fenrir), bypassing the contended Groq pipeline + Chatterbox (6GB GPU) + skipped ElevenLabs (past_due). NEEDS full .\Start-KAI.ps1 fleet restart to apply (bot env, no hot-reload). NOTE: this fixes the Discord voice Leo; the dashboard 'Oracle line' call still runs [internal module] (Groq-based) and would need Gemini wired in separately if that's the surface the owner means.camA/proj2, mouse-look yaw of you.face, W/S forward-back + A/D strafe relative to facing, forward aim reticle; floor drawn as a projected quad, zone as a projected polygon). File-only change, script ?v bumped to 977 → just hard-refresh. oracle-games-bounce.js.oracle-games-bounce.js, id bouncebr): isometric ant-scale bounce-house battle royale, sky-drop spawn, bouncing pellets, parachute power-up crates that change pellet bounce, shrinking air-zone, AI bots. Registered via GC.register; script-tagged in oracle.html + allowlisted in command-center-server.mjs. APPLY = .\Start-Dashboard.ps1 (server allowlist changed) THEN hard-refresh._calMonthZoomHtml, click the month title): month Income/Bills/Saved/Net tiles, in/out bars, projected end-of-month balance, per-pay-period budget meters, and a 'where it goes' category breakdown. oracle-os/shell/desktop.js.it.verified; future = parenthesized estimate). _flowTo(day,vo) helper. Groundwork for paystub/bank-statement verification pipeline (owner's future automation). Owner still needs to set real Checking/Savings balances (currently $0). oracle-os/shell/desktop.js._calSyncPull re-rendered every tick; now gated by _calSig change + never rebuilds an open _calZoom popup. Owner says the blink is ALSO workspace-wide (other apps/settings/popups) = separate OS-level re-render to find next. oracle-os/shell/desktop.js..calCell hover-lift, today number in an accent pill (.calTodayNum), weekend tint, refined weekday header + pay-week card, more spacing. Visual-only, in _calMonthView + calFxCss. oracle-os/shell/desktop.js..calPop, count-up stat tiles _calFxCountUp, animated meters .calGrowI, staggered side-pane/list reveal .calRiseI = the 'text effect' on day-click) + richer game-like info (status pill, pay-week budget meter = % of check to bills + free-to-save, next-paycheck & next-bill countdowns). CSS via one-time _calFxEnsureCSS. oracle-os/shell/desktop.js._calMoneyC; pay-week + day-cell + chips no longer round to whole $); day analytics POPUP _calDayZoomHtml (In/Out/Saved/Net tiles + bars + grouped item list) opened by 2nd-click on a selected day / side-pane Analytics btn / date header (osCalZoom); modernized _calSidePane (In/Out strip + buttons); mouse-wheel month nav (osCalWheel, Month grid only). NEXT: Week+Day view redesign (eyeball-iterate w/ owner). oracle-os/shell/desktop.js.state/os-appdata/usr_6e6932d1ec1fc7d3/calendar.json: removed duplicate Rent (kept one, 1st-of-month $780 +5d grace), made $309.99 Consumers a one-time catch-up, Mortgage grace 15d, Internet 68.84→68.94. Pay-week totals now realistic (no more $0 weeks). Backups: calendar.json.bak-cleanup-*.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.groq-voice-session.mjs so a dead/slow provider fails fast to the next voice. Applies on fleet restart.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.executeVocalSync (PCM->WAV->ffmpeg) + LEO_NO_GEMINI_LIVE gate to route calls onto it. NEEDS fleet restart; ref wav still missing (default voice). [internal module].oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js.oracle-os/shell/desktop.js._calUserKey (no 'anon' regress) + anon-item rescue. oracle-os/shell/desktop.js.Start-KAI.ps1 Stage 3.76, :5111).chatterbox-server.py).leo.mjs GeminiLiveSessionManager), unless LEO_NO_GEMINI_LIVE=1 -> falls to (2) leo.mjs executeVocalSync (ElevenLabs->Kokoro->edge; Chatterbox wired in v9.10.953). (3) Oracle-OS dashboard voice call ("the Oracle line", [GroqVoice/Leo]) = [internal module] (Whisper->LLM->TTS), a SEPARATE path with its OWN Chatterbox wiring (LEO_LOCAL_TTS=1 + LOCAL_TTS_URL -> POST -> streamPcm, then ElevenLabs/Orpheus). When the owner "calls Leo" on the dashboard, it's path 3. Fixing one path does NOT fix the others..env watcher only restarts specific BOTS for specific keys (e.g. LEO_NO_GEMINI_LIVE -> Gemini). The Dashboard/Oracle-voice process keeps whatever env it started with. So LEO_LOCAL_TTS, LEO_TTS_SKIP_ELEVEN, LOCAL_TTS_URL only take effect after a FULL Start-KAI.ps1 restart. Symptom when stale (seen 2026-09-07 log): [GroqVoice/Leo] still hitting ElevenLabs with tts=9000-37000ms (ElevenLabs is past_due/401 -> slow retries) even though .env has the local-TTS keys set. FIX: full fleet restart + Chatterbox server (:5111) running.open(f,'w').write(open(f).read()...) inline — truncates before the read; emptied Cargo.toml once. Use safe read-then-write.oracle.html / oracle-core.js / oracle-os/** / kaiverse.js. Surgical Python string-replace injections only.Cargo.toml, cache-bust served JS via sw.js SHELL_CACHE, and update THIS file..env changes need a restart of whatever reads them; there is no auto .env watcher. The cloud AI can't restart — the owner runs Start-Dashboard.ps1 (dashboard) or Start-KAI.ps1 (fleet).<script> are BLOCKED by the dashboard's global CSP (script-src-elem 'self' with NO unsafe-inline, set in command-center-server.mjs ~L8968). Symptom: the page renders but every button/handler is dead and NOTHING reaches the server (0 API hits) on ALL devices — easily misread as a mobile cache problem. The console shows Executing inline script violates ... script-src-elem 'self'. FIX: in that page's response handler, set a per-response Content-Security-Policy allowing script-src-elem 'self' 'unsafe-inline' + media-src 'self' data: (mirror the web-proxy handler ~L17771). Needs a dashboard restart. When building any new standalone inline-script page, add this CSP override at the same time as the route.