Skip to content

CHANGELOG.md

Latest commit

 

History

History
85 lines (69 loc) · 4.4 KB

CHANGELOG.md

File metadata and controls

85 lines (69 loc) · 4.4 KB

Changelog

All notable changes to this project will be documented in this file.

[Unreleased]

Fixed — Severity Scoring & Alexa Alert Tuning (2026-02-14)

  • Severity scoring now trusts AI baseline — rule engine starts from AI's risk level instead of always starting from 0, preventing false escalation of routine events
  • Removed aggressive scoring for normal behaviors (looking around, standing, walking) which were incorrectly bumping LOW events to MEDIUM/HIGH
  • Added score reduction for routine behavior keywords (walking, standing, passing)
  • Known face reduction increased from -3 to -4 for stronger suppression
  • Delivery type reduction increased from -1 to -2
  • Alexa now only announces medium/high/critical — HA automation already had the filter but false severity escalation was triggering it on routine events

Added — 3-Pillar Alert Architecture (2026-02-14)

  • Pillar 1: Structured AI Output — Vision model (Ollama qwen2.5vl:7b) outputs structured JSON with subject identity/description, behavior, risk level/confidence/reason, event type, action
  • Pillar 2: Rule-Based Severity Scoring — Deterministic score_severity() engine adjusts AI risk using time of day, camera zone, home mode, behavioral keywords, known faces. Score thresholds: 0-2=LOW, 3-4=MEDIUM, 5-6=HIGH, 7+=CRITICAL
  • Pillar 3: Professional WhatsApp Formatter — _format_whatsapp_alert() builds structured messages with emoji severity, organized sections (EVENT/SUBJECT/BEHAVIOR/RISK/CONTEXT/ACTION/MEDIA/ESCALATION)
  • Smart media decisions — decide_media() selects snapshot/clip/monitoring based on risk level: LOW=snapshot, MEDIUM=snapshot+15s clip, HIGH=+30s+monitoring, CRITICAL=+60s+monitoring
  • WhatsApp alerts include snapshot image at top and video clip at bottom (single message)

Added — Local Vision AI with Ollama (2026-02-14)

  • Local Ollama VLM analysis using qwen2.5vl:7b on Mac M4 Mini
  • Base64 image encoding sent directly to Ollama /api/generate endpoint
  • Automatic fallback to OpenAI GPT-4o-mini if Ollama fails
  • Structured prompt with camera context, zone, time, home mode, known faces
  • JSON output parsing with 4 strategies: JSON: prefix, markdown code fence, bare JSON, embedded regex

Added — Enhanced JSON Decision Parsing (2026-02-14)

  • 4-strategy parse_decision_json() parser handles multi-line, code-fenced, and embedded JSON
  • Handles both flat ({risk:"low"}) and structured ({risk:{level,confidence,reason}}) formats
  • Smart _fallback_decision() extracts type from text keywords when JSON parsing fails

Added — WhatsApp Delivery Improvements (2026-02-14)

  • WhatsApp filtered to medium/high/critical only — low-risk goes to HA/logs only
  • "DELIVERY MODE" forwarding instruction prefix for OpenClaw agent
  • Clip download from Frigate API saved to ai-clips/
  • Single message delivery: snapshot MEDIA at top, formatted text, clip MEDIA at bottom
  • Action execution runs BEFORE WhatsApp so clip is available for attachment

Added — Descriptive Alexa TTS (2026-02-14)

  • make_tts() generates full security briefings with severity, subject, behavior, risk reason

Added — Structured HA Notifications (2026-02-14)

  • Severity emoji in title, importance levels mapped to risk
  • Structured MQTT payload with behavior, subject, zone, home mode, media fields

Added — Runtime Config Template (2026-02-14)

  • config/bridge-runtime-config.json.example with all settings documented

Added — Control Panel Updates (2026-02-14)

  • Rebranded to "Drishtik Control Panel"
  • Missing UI settings added (MQTT, cooldown, zones, user management)
  • SKILL.md editor in UI

Added — Phase 2: HA Tool Execution (2026-02-13)

  • Bridge executes HA service calls based on decision action field
  • Action handlers: lights, speaker/Alexa, clip save, alarm siren

Added — Phase 1: Decision Engine (2026-02-13)

  • Structured JSON decision parsing
  • MQTT payload enriched with type, confidence, action, reason

Fixed

  • JSON parsing failures (qwen2.5vl outputs multi-line/code-block JSON)
  • Clip attachment on WhatsApp (single message pattern)
  • Alexa announcements now descriptive instead of generic "Alert"
  • Redacted all sensitive data from repo configs

[1.0.0] - 2026-02-07

Added

  • Full Frigate -> OpenClaw -> HA pipeline
  • Interactive installers and prereq checks
  • Home Assistant automation YAML
  • OpenClaw installer scripts
  • Redacted examples and security policy