← all documents · raw markdown · 7 KB

HOTFIX NOTES — for the dispatch (Antigravity)

Ryan is away. Code edits were made to the KAI ecosystem and need to be applied to the running server. This note tells you exactly what changed, how to apply it, and how to handle failures. Expect the apply to take a few minutes (engine rebuild + boot).

---

1. ONE COMMAND APPLIES EVERYTHING

powershell -ExecutionPolicy Bypass -File C:\KAI\apply-hotfix.ps1

apply-hotfix.ps1 auto-detects whether the Rust engine needs a rebuild (only if src\*.rs / Cargo.toml is newer than target\release\kai.exe), kills kai.exe first so the binary can't be access-denied, rebuilds, then restarts the fleet via Start-KAI.ps1. If a build fails, it restarts on the old binary so the server stays up and logs the error to C:\KAI\scratch\apply-hotfix.log.

Node/Python-only changes (faster, no compile): add -NoRebuild.
Force a rebuild: add -Force.

After it runs: hard-refresh the dashboard (Ctrl+Shift+R) for oracle.html changes.

---

2. WHAT CHANGED (so you can verify)

| File | Change | Needs |

|------|--------|-------|

| src\main.rs | sleep→world-model learning import fix (crate:: → kai::) | cargo rebuild + engine restart |

| overnight_pipeline.py | training window gate fix — all 5 gate sites now use bounded overnight_window_open() so training only runs 03:00–08:30, never on a daytime boot | restart (Python; the keeper relaunches it) |

| tools\oracle-discord\command-center-server.mjs | /api/vitals now also reads /api/session (so KAI Vitals show data instead of "engine down"); per-user passwords; dotenv path fix | restart Command Center |

| oracle.html | Phoenix-healed KAIVERSE functions (nsPaintHUD, nsOpenNodePanel, nsAdvanceGrowth, nsGrowVal, closeNsEdgePanel); map moved up; live clock; touchpad controls; dashboard metrics (10/9 count, radar, leaderboard) | hard refresh only (served statically) |

| oracle.html (v9.10.1, Jun 23) | Mobile drawer fix — on phones the opened nav/channels drawer was dimmed & dead to touch (every tap hit the scrim and closed it). Stacking-context trap: .shell{position:relative;z-index:1} confined the fixed drawers (.col-left/.col-right, z-index:100) BELOW the body-level .m-backdrop (z-index:90). Fix = added .shell{z-index:auto} inside the @media(max-width:900px) block (next to .rail{display:none}). Small targeted edit, no rewrite. | hard refresh only (served statically) |

---

3. IF SOMETHING FAILS

Compile error (Rust): apply-hotfix.ps1 keeps the server on the OLD binary. Read the error from C:\KAI\scratch\apply-hotfix.log (or run cargo build --release --bin kai to see it), report it verbatim (file:line) so it can be fixed. Do not guess-edit Rust internals.
Access denied during build: the script already kills kai.exe and retries once. If it still fails, open Task Manager → end every kai.exe → re-run the script.
oracle.html looks broken / blank after refresh: check the REAL file (Windows path) ends with </script></body></html> — use Read/Grep, NOT the Linux/bash mount (it serves a stale truncated snapshot and lies). If the real file is truncated, restore from C:\KAI\oracle.html.GOOD.bak (if Ryan made one) or a .bak. Never do a whole-file rewrite of oracle.html — only small targeted edits (that's what caused every truncation).

---

4. VERIFY AFTER APPLYING

1. Engine ticking: curl http://127.0.0.1:3334/api/session → tick advancing.

2. Command Center gated: curl -s -o NUL -w "%{http_code}" http://localhost:3001/api/channels → 401.

3. Pipeline idle during the day (not training): check it's not pegging CPU; it should only run 03:00–08:30.

4. Dashboard (hard-refresh): KAI Vitals show real numbers (not "engine down"); agents count is X/9 not 10/9.

---

*Generated for the dispatch. The single source of truth is apply-hotfix.ps1 — when in doubt, run it and read scratch\apply-hotfix.log.*

---

DISPATCH UPDATE — 2026-06-22 (from Claude/Cowork session, for the coworker)

Summary of work done this session. Nothing here was force-pushed to the live server; most items are pending the fleet rebuild/restart.

1. Kai vitals (frozen/idle, not dying)

Curiosity scale bug fixed in tools/oracle-discord/shared/drive-system.mjs — phi_g (0–5 scale) was clamped and pinning curiosity to 0; now normalized by /5.0. Backup: drive-system.mjs.curiosity-fix.bak.txt. Takes effect on Node fleet restart (no recompile).
rho wired to real FieldState::compute().rho in src/bridge/oracle_server.rs (was hardcoded 0.0).
valence given a headless affective approximation (Drive+NeuralOscillator) in src/bridge/oracle_server.rs (was hardcoded 0.0); reduced model — no amygdala/serotonin/tone. Backup for both Rust edits: oracle_server.rs.heartbeat-vitals-fix.bak.txt. Requires cargo build --release --bin kai.
See C:\KAI\VITALS-FIX-CHECKLIST.md for full detail. NOTE: the live binary timestamps suggest these may already be compiled, but run apply-hotfix.ps1 with -Force to guarantee.
Remaining: lattice (phi_g/chi/cells) only moves with real input (chat/task); hippocampus 210k figure source still untraced (data/hippocampus_status.json not being written).

2. Command Center dashboard login

User was locked out. Token reset in tools/oracle-discord/.env (CC_CONTROL_TOKEN) to a value the user knows. Backup: .env.cc-token.bak.txt. No OS-level env override exists (confirmed). Dashboard process was restarted and respawned cleanly.

3. Leo voice stutter/chopping (NOT internet — confirmed)

Root cause found via log analysis: 27.5% of Leo's turns end at "Turn complete after -1ms" (killed instantly). Primary cause = ungated sendText() injections (esp. the Fleet Status Update at leo.mjs:1321) sending clientContent into Leo's live Gemini session MID-TURN, which interrupts his generation. Secondary = echo leaking through a too-short 320ms half-duplex tail. Also: ~145 Gemini Live session reconnects (GoAway 10-min cap + 1008 aborts) and 760 rate-limit (429) hits.
Fix being applied (Node .mjs, effective on fleet restart): (a) defer context-only sendText while a turn is live, queue on _pendingContext, flush on turnComplete (user replies with turnComplete=true still go through); (b) widen echo tail to env-tunable LEO_ECHO_TAIL_MS default ~1200ms. Files: shared/gemini-live-bridge.mjs, bots/leo.mjs (backups made). Local Discord playout was CLEARED as the cause (clean on long turns).

4. Hotfix script

apply-hotfix.ps1 reviewed and verified safe (build + restart only; no deletions/downloads/cred changes). Note it stops ALL node processes and does a full stack bring-up. Recommend running with -Force to guarantee the Rust engine edits compile.