CHANGELOG

Changelog

[Unreleased] — Public Readiness (2026-03-21)

Housekeeping

• Removed personal credentials and university-specific defaults from all config files
• Fixed hardcoded absolute paths in .cortex_launcher.c — now derives project root from executable location
• Removed tracked build artifacts from git: .plasmo/, CortexDaemon.app/, .vsix, IDE configs
• Comprehensive .gitignore update for all artifact types
• Fixed deterministic extension ID via manifest key — install_native_host no longer requires --extension-id
• Auto-detects all Chromium browsers (Chrome, Edge, Brave, Arc, Vivaldi, Opera, Chromium) and installs native messaging for each
• Fixed API endpoint documentation (/api/state → /status/current)
• Updated test counts (47 test files)

Click-to-Start Daemon (2026-03-19)

• One-click "Start Cortex" from browser extension — daemon launches via Terminal.app with camera access
• Multi-layer stop flow: WebSocket SHUTDOWN → HTTP /shutdown → native messaging kill → SIGKILL escalation
• Orphaned process detection via both port (lsof) and process name (pgrep)
• Smart camera selection: probes indices 0-4, post-open re-verification rejects Continuity Camera
• Camera permission request via AVFoundation before opening capture device
• MacBook camera warmup retry loop (4 × 0.5s)

Visual Identity Refresh (2026-03-19)

• Design token system with CSS custom properties
• Ambient somatic engine: particles, aura, weather effects
• Redesigned popup dashboard and Pulse Room new tab page
• Onboarding flow with 3-step setup guide

---

[0.1.0] — Full Revision (2026-03-18)

Comprehensive audit-driven revision across signal processing, LLM pipeline, Chrome extension, LeetCode mode, and new features. 903 tests passing.

Wave 1 — Critical Fixes

Signal & State Engine (1A)

• C-01: Raised SQI validity threshold from 0.1 → 0.4 to reject noisy heart rate estimates
• C-02: Added parabolic peak interpolation to peak_detection.py — eliminates 33ms IBI quantization error at 30 FPS
• C-03: IBI count guard requires ≥5 inter-beat intervals before propagating RMSSD
• C-04: Fixed s6 blending — weighted blend (60% thrashing + 40% switch) instead of max() which double-counted velocity
• C-10: Shutdown detector late hour configurable (default 23:00), all thresholds now constructor parameters
• C-11: Resolved YAML vs Pydantic default discrepancies (complexity_threshold, quiet_mode_minutes)
• O-09: Context-aware blink suppression — attenuates score 0.3× when HR is normal (concentration, not stress)

LLM & Intervention Pipeline (1B)

• C-05: Fixed KeyError in v2 templates by adding extra_context parameter to build_user_prompt()
• C-06: Nuanced staleness check — suppresses intervention only on genuine FLOW recovery (≥3s dwell), checks tab snapshot validity
• W-11: Fixed is_destructive false positive — checks action_type field instead of substring matching on label text
• W-14: Added GET /consent/level and POST /consent/reset API endpoints

Chrome Extension (1C)

• C-08: First-run onboarding page with 3-step setup guide
• C-09: Debounced SW state persistence (500ms trailing timer) to chrome.storage.session — survives service worker restarts
• W-25: Tab-manager snapshot persistence across SW restarts
• Popup shows disconnected banner when daemon not running

LeetCode Mode (1D)

• C-07: Lockout overlay renderer with countdown timer and skip button (no penalty — bandit learns naturally)
• C-12: LeetCodeAdapter now inherits CortexAdapter protocol
• W-21: Zombie detector returns False when blink_rate is None instead of falling through to True
• W-28: Parasympathetic rebound temporal guard — only detects within 5 minutes of last acceptance

Wave 2 — Experience Improvements

LLM Quality & Bandit (2A)

• W-08: Structured JSON response_format added to Azure OpenAI calls
• W-12: LinUCB regularization increased 1.0 → 5.0 for better cold-start exploration
• W-13: Helpfulness reward redistributes 30% weight proportionally when no explicit rating
• Token budget hard cap with 3-pass truncation (terminal → tab titles → code content)
• W-18: Deprecated InterventionTrigger in favor of TriggerPolicy with deprecation warnings

Prompt Engineering & Tab Classifier (2B)

• O-08: New tab_classifier.py — single source of truth for domain-based tab type classification
• W-05: Parser enforces keep on AI assistant and documentation tabs via classifier (not hardcoded URLs)
• W-26: Fixed active recall template biometric contradiction
• W-10: Added normalizer defaults to generic-phrase blocklist
• s6 same-category discount — switching between tabs of the same type gets reduced penalty

Detector Robustness & Config (2C)

• W-06: Removed goal-relevant verbs from rabbit hole stop-word list ("implement", "build", "create", "fix", "add")
• W-15: Removed overlay backdrop dim (transparent instead of 35% opacity)
• W-16: Chrome extension cooldowns sync with daemon config via SETTINGS_SYNC
• W-17: Quiet mode toggle added to Chrome popup
• W-19: Adaptive solution friction with exponential decay (max(10, 60*exp(-t/600))*difficulty_mult)
• W-20: Fixed zombie detector docstring/code mismatch

Wave 3 — Innovation Features

• O-10: Pre-break warning at 80% stress threshold via should_warn()
• O-01: SessionReportGenerator — full biometric study session reports with flow percentage, state transitions, golden hour, 7-day comparison
• O-04: Per-topic difficulty tracking in LongitudinalTracker with stress modifier calibration
• O-07: Modality preference tracking in TabRelevanceTracker for proactive resource switching

Test Coverage

• 903 tests passing (up from ~870 pre-revision)
• New test files: test_session_report.py, test_tab_classifier.py, test_rabbit_hole.py, test_stress_integral.py, test_zombie_detector.py, test_parasympathetic_rebound.py
• Pre-existing failures (not caused by revision): test_activity_tracker.py (redis import), test_redis_store.py (redis import), test_context_to_llm.py (system prompt contains "heart rate" in instructions), test_scripts.py (9 failures)