Skip to content

TaskGuard.md

Latest commit

 

History

History
106 lines (81 loc) · 5.98 KB

TaskGuard.md

File metadata and controls

106 lines (81 loc) · 5.98 KB
Raw

TaskGuard: Feature List

This file is the at-a-glance feature catalogue. For implementation details see README.md; for the full academic write-up see PROJECT_REPORT.md; for the one-page research-positioning brief see RESEARCH_STATEMENT.md.

TaskGuard is framed as a research artifact at the intersection of HCI, applied NLU, and on-device ML, not as a standalone productivity tool.

Core User Features

1. Intent Declaration

  • Users explicitly declare their current task before starting a focus session
  • Supports short natural-language task titles (e.g., "Prepare for OS exam")
  • Optional task description for richer context
  • Category selection: Studying, Coding, Writing, Research, Work, Design, Reading, Other
  • Intent is treated as contextual ground truth, not inferred behaviour

2. Focus Session Management

  • Predefined focus durations (15, 25, 45, 60 minutes) plus a custom minutes input
  • One-click session start; two-step confirmation for early termination
  • Session state persists across browser restarts (chrome.storage.local)
  • Visual feedback via extension badge countdown driven by chrome.alarms
  • Post-session summary view: planned vs. actual minutes, blocked/allowed counts, completion progress bar

3. Context-Aware Website Interception

  • Monitors active tabs during focus sessions (including tabs already open when the session begins)
  • Extracts the registered hostname when a page is accessed
  • Differentiates between permanent allowlist, session-scoped temporary allowlist, and everything else
  • Same website can behave differently under different intents (e.g., YouTube during Study vs. during Research a video essay)

4. Relevance Validation Prompt

  • Triggered when a non-allowlisted site is accessed during a session
  • Displays the user's current task intent for context reinforcement
  • Shows a 0–100 relevance score with a colour-coded label (green / yellow / red) and progress bar
  • User explicitly validates: Allow (with optional permanent allowlisting) or Block
  • Decision is enforced immediately and logged

5. Allowlist Management

  • Pre-approve trusted websites permanently or for the current session only
  • Quick-add chips for common productive domains (Google Docs, Stack Overflow, GitHub, Wikipedia, Google Scholar)
  • Strict regex validation; supports comma-separated bulk add and URL → domain extraction

Relevance Classifier (Implemented)

6. Hybrid In-Browser Relevance Scoring

Four-signal classifier in classifier.js, runs synchronously with no model download:

Signal Weight Idea
Topic Jaccard overlap 50 % Map intent + domain to topic categories via keyword lists; compute set Jaccard
Term-frequency cosine similarity 25 % Build normalised TF vectors over stemmed tokens of intent vs. shared-topic keyword set; compute cosine
Direct domain token hit 15 % Fraction of domain tokens whose stem matches an intent token (equal, prefix, or suffix)
Known-domain bonus 10 % +0.10 if domain is in the curated lookup table and shares ≥ 1 topic with the intent

Raw 0–1 weighted sum is passed through a logistic sigmoid (1 / (1 + e^(−8(x − 0.25)))) and rounded to an integer 0–100.

The cosine component uses term-frequency only — no inverse-document-frequency is computed across a corpus. It is therefore TF cosine similarity, not classical TF-IDF.

7. Explainable Score (Partial)

  • The colour-coded label (likely relevant / uncertain / likely off-task) and the progress bar provide a coarse explanation of the score.
  • Planned: surface the contributing signals (which topics matched, which keywords overlapped) directly on the intercept page.

Data Collection & Research Features

8. Human-in-the-Loop Decision Logging

  • Every allow/block action is logged as labelled data
  • Captures: sessionId, intent, intentCategory, domain, url, decision, addedToAllowlist, userId, ISO timestamp
  • Decisions act as human-validated relevance labels suitable for downstream model training

9. Pseudonymous User Identification

  • Each install generates a random anonymous userId (crypto.randomUUID())
  • No login or personal data required
  • Enables per-user analysis while preserving privacy

10. Local-First Data Storage

  • All session and decision data stored using chrome.storage.local
  • No external servers, no cloud storage, no analytics SDKs
  • Fully offline and privacy-preserving by design

11. Dataset Generation

  • Accumulates intent–description–category–domain–decision tuples over time
  • Enables analysis of context dependency (same site, different outcomes under different intents)
  • Serves as a labelled dataset for relevance modelling and focus-behaviour research

12. Data Export

  • Users can export their interaction data as JSON (sessions + decisions + userId + ISO export timestamp) or CSV (active tab only)
  • Enables offline analysis and research use
  • Maintains informed user consent

13. Analytics Dashboard

  • 7-day stacked bar chart of allow/block decisions
  • Summary stat cards (decisions tab: total / allowed / blocked / unique sites; sessions tab: total / minutes focused / completed / ended early)
  • Sortable tabular history with intent, domain, and decision badge

Planned / Future Work

  • Surface signal contributions. Show which topics matched and which tokens overlapped on the intercept page (full explainability).
  • Fine-tuned ML head. Replace the rule-based hybrid classifier with a lightweight on-device transformer (e.g., distilled MiniLM) via WebAssembly ONNX, using the locally-collected decision logs for fine-tuning.
  • Calendar-driven intent suggestion. Pre-fill the intent field from the user's current calendar event.
  • Per-domain time budgets. Soft caps before hard interception triggers.
  • Cross-device sync. Optional, end-to-end encrypted.
  • Firefox / Edge ports. Edge is already supported; Firefox needs a WebExtensions compatibility pass.