All notable changes to TokioAI are documented here.

Terminal Stability (Zero Keyboard Hangs)

• 3-layer terminal restore: golden state capture → stty sane → show cursor
• _drain_stdin() — clears keyboard buffer after every tool execution
• atexit + signal handlers (SIGINT/SIGTERM/SIGHUP) guarantee cleanup
• Escape key fix: os.read() instead of sys.stdin.read() (no more blocking)

Unlimited Mode by Default

• max_rounds = 0 (unlimited) — no more "Maximo de iteraciones alcanzado"
• max_time = 0 (no timeout) — agent works until task is complete
• Agent engine: MAX_TOOL_ROUNDS = 0, MAX_TOTAL_TIME = 0

15 Slash Commands (respond instantly, no LLM needed)

• /status — Entity + GCP + PiCar overview
• /entity — Entity status (camera, Hailo, FPS)
• /picar — PiCar-X robot status
• /health — Health data with color coding (red if HR>120, SpO2<92)
• /gcp — GCP containers status
• /waf — WAF attack stats (JWT auth)
• /wifi — WiFi defense status
• /coffee — Coffee machine status
• /ha — Home Assistant status
• /see — AI vision analysis
• /logs — Entity logs with colors
• /sitrep — Full situation report
• /threats — Threat assessment
• /drone — Drone status
• /containers — Docker container list

Cost Tracking

• Per-response token cost estimation (input + output)
• Accumulated session costs persisted to ~/.tokio_cli_costs.json
• Supports Claude Opus, Sonnet, and Haiku pricing

Other Improvements

• Tab completion for commands, slash commands, and tool names
• Multi-line input (end line with \)
• Terminal recovery with stty sane if anything goes wrong

Philosophy

• Quantified self / biohacking approach to personal health optimization
• Continuous self-measurement, data-driven health decisions
• Real results: cholesterol 250→103 mg/dL (-58.8%) without statins

Accu-Answer iSaw 4-in-1 Integration

• AI-powered OCR: send photo of iSaw device via Telegram → auto-read and store
• Supported metrics: glucose, cholesterol, hemoglobin, uric acid
• Values displayed on Entity's live panel (LAB section)
• Custom ranges for Accu-Answer (hemoglobin: 11.0-16.0 g/dL)

Health Database

• SQLite storage on Raspberry Pi (health_db.py)
• 1,900+ readings stored across 24 days
• API endpoints: /health/db/latest, /health/db/store, /health/db/history
• 7 tracked metrics: HR, BP, SpO2, glucose, cholesterol, hemoglobin, uric acid

Agent Tools

• health_store — Store lab readings from any interface
• health_full — Combined report (smartwatch + lab values)
• health — Full multi-day history with assessment

AI Health Analysis

• Automatic trend detection and scoring (0-10 per metric)
• Cross-metric correlation (e.g., cholesterol improvement + HR improvement)
• Historical comparison with percentage changes

Safety Proxy (picar_proxy.py — 778 lines)

• REST API on port 5002 (Raspberry Pi 5)
• Speed limiting (max 80), duration limits (max 10s)
• Emergency kill switch
• Rate limiting and command audit log
• Systemd service with auto-restart

Agent Tools (picar_tools.py — 277 lines)

• 36 registered tools for full robot control
• Natural language: "move forward 30cm", "avoid obstacles", "dance"
• Sensor readings: ultrasonic distance, grayscale line detection
• Camera control: pan/tilt, snapshot
• Autonomous modes: obstacle avoidance, line tracking, patrol patterns

Network Access

• LAN: <local-ip>:5002
• Tailscale: <tailscale-ip>:5002

Singleton Protection

• flock + lock file prevents dual-instance (was causing camera conflicts)
• ExecStartPre kills port 5000, ExecStopPost cleans lock file
• Graceful shutdown with SIGTERM/SIGINT handlers

Self-Healing Improvements

• Uses systemctl restart tokio-entity instead of fragile pkill + nohup scripts
• Auto-detects and kills zombie processes
• HA Docker auto-restart if container goes down

Camera Stability

• try/finally with vision.stop() ensures camera release on crash
• __del__ safety net in VisionEngine
• No more "camera busy" errors from zombie processes

SecurityFeed

• Auto-disables if WAF is not configured (no more error spam in logs)

Error Messages

• Actionable errors: "Proxy down — check Raspi", "WiFi not connected to drone"
• Instead of raw httpx exceptions

Variable Timeouts

• takeoff: 20s, patrol: 30s, move: 15s, land: 15s
• Instead of fixed 10s for everything

Medical Image Processing

• Auto-detects Accu-Answer iSaw photos
• Extracts metric values via Gemini Vision OCR
• Stores readings in health database
• Confirms with summary table

Quick Commands

• 7 instant commands: /status, /sitrep, /waf, /health, /threats, /drone, /entity
• WAF command uses JWT authentication

GCP Agent

• 37 registered tools (added: picar, health_store, health_full)
• Unlimited tool rounds and time by default
• Context builder with smart memory injection

Sitrep

• Updated with PiCar-X status and Tailscale IP
• Health DB stats included
• Entity singleton status

• PiCar-X proxy initial deployment
• Health DB endpoints (/health/db/latest, /health/db/store)
• CLI v3.3 with basic slash commands
• Entity display panels for health data

• CLI token cost tracking
• WiFi defense slash command
• Coffee machine slash command
• Home Assistant slash command
• Tab completion for 25 commands

• 9 slash commands (respond without LLM)
• Telegram quick commands with JWT auth for WAF
• CLI entry points: tokioai and tokio-ops

• Native tool use (structured JSON via LLM API)
• 5 LLM providers: Anthropic, OpenAI, Google, OpenRouter, Ollama
• Streaming CLI with escape-to-cancel
• Auto-compaction at 80% context capacity
• Subagent workers for parallel tasks
• Self-healing engine
• 150+ tests