Guardian-Agent / docs /production-roadmap.md
Prince9868's picture
Hackathon submission: Guardian OpenEnv
5743bc2
|
Raw
History Blame Contribute Delete
12 kB

Guardian Agent Production Roadmap

This document turns the product-direction guidance into an execution backlog mapped to the current codebase.

The intended production shape is:

  • the browser extension is the primary product surface
  • the website supports onboarding, trust history, demo flows, and operator visibility
  • Guardian uses a shared staged detection pipeline instead of separate demo logic
  • risky actions stay behind explicit user approval gates

Product Focus

The first production use case should be travel checkout protection:

  • hotels
  • airlines
  • vacation rentals

Why this focus:

  • hidden fees are common and expensive
  • urgency and scarcity patterns are frequent
  • users understand the pain immediately
  • the extension can deliver value without needing full autonomous purchase behavior

P0: Trustworthy Core

These items should happen before broad rollout.

P0.1 Shared Detection Pipeline

Goal:

  • make Quick scan, Mission, and extension analysis use one production detection pipeline

Current files:

Tasks:

  • extract a shared analyzeCheckoutContext() service
  • separate the pipeline into stages:
    • page classification
    • signal extraction
    • risk reasoning
    • recommendation
  • make demo and mission endpoints thin wrappers around that shared service

Definition of done:

  • one source of truth for manipulation detection
  • one source of truth for confidence scoring
  • one source of truth for non-commerce gating

P0.2 Non-Commerce and Page-Type Detection

Goal:

  • reduce false positives before any model or trust score is shown

Current files:

Tasks:

  • add page types:
    • non-commerce
    • product
    • cart
    • checkout
    • post-purchase
  • require stronger evidence before running full detection on unknown domains
  • make extension auto-activation depend on page type, not just heuristics

Definition of done:

  • GitHub, docs, blogs, and normal browsing do not trigger shopping analysis
  • cart and checkout pages activate reliably

P0.3 PII Redaction

Goal:

  • never ship sensitive checkout data to the backend unintentionally

Current files:

Tasks:

  • strip or mask:
    • emails
    • phone numbers
    • addresses
    • payment details
    • loyalty numbers
    • traveler names
  • add a browser-side sanitizer before payload submission
  • add backend log redaction for request bodies

Definition of done:

  • extension only sends checkout-relevant snippets
  • backend logs are safe by default

P0.4 Deterministic Rules Before Model Calls

Goal:

  • use AI where judgment is needed, not where rules are enough

Current files:

Tasks:

  • add deterministic detectors for:
    • pre-checked add-ons
    • countdown text
    • low-stock phrasing
    • hidden fee labels
    • decline button prominence issues
  • only call the model for higher-level synthesis and ambiguous cases

Definition of done:

  • obvious issues do not depend on LLM output
  • model usage becomes cheaper and more stable

P0.5 Confidence and Uncertainty

Goal:

  • never fake certainty

Current files:

Tasks:

  • add field-level confidence to major findings
  • add an explicit uncertain state when evidence is weak
  • distinguish:
    • deterministic findings
    • AI-inferred findings
    • estimated totals

Definition of done:

  • user can tell what Guardian knows vs. what it estimates

P1: Make the Extension Real

These items make the extension feel production-ready.

P1.1 SPA Handling and Dynamic Checkout Updates

Current files:

Tasks:

  • listen for SPA route changes
  • reset stale overlay state when the page context changes
  • add mutation observers for dynamic checkouts

Definition of done:

  • Guardian reacts correctly on Amazon, airline flows, and React/Next/Vue sites

P1.2 Stronger DOM Targeting

Current files:

Tasks:

  • build element scoring from:
    • role
    • label
    • position
    • visibility
    • clickability
    • checked state
  • improve DOM highlighting with evidence labels
  • prefer visible actionable controls over generic text nodes

Definition of done:

  • highlights land on the actual suspicious controls, not just nearby text

P1.3 Supervised Browser Workflow

Current files:

Tasks:

  • model the browser workflow explicitly:
    • inspect
    • compare
    • strip upsells
    • ask for approval
  • keep approval gates for:
    • account creation
    • credential submission
    • merchant switching
    • purchases

Definition of done:

  • Guardian acts like a supervised assistant, not an unsafe autobuyer

P1: Backend Hardening

P1.4 Auth, Rate Limits, and Timeouts

Current files:

Tasks:

  • add auth for non-local mode
  • add per-IP or per-user rate limits
  • add request timeouts
  • add retries and circuit breakers for model dependencies

Definition of done:

  • backend remains stable under failure and misuse

P1.5 Observability

Current files:

Tasks:

  • add request ids across frontend, extension, and backend
  • log model latency and failure rates
  • log false-positive feedback events
  • log extension activation success

Definition of done:

  • major failures can be debugged from logs and traces

P2: Trust System and Data Layer

P2.1 Domain Evidence History

Current files:

Tasks:

  • accumulate repeated evidence for each domain
  • track:
    • pattern counts
    • fee estimates
    • user feedback
    • scan timestamps
  • expose score explanations on trust pages

Definition of done:

  • trust score becomes explainable and longitudinal

P2.2 Score Versioning

Current files:

Tasks:

  • assign a scoring version to every trust result
  • explain why a domain score changed over time

Definition of done:

  • users and operators can audit trust changes

P2: Frontend and UX

P2.3 Show Evidence, Why, and What To Do

Current files:

Tasks:

  • label each output as:
    • deterministic
    • AI-inferred
    • estimated
  • show user action guidance on every warning
  • keep uncertainty visible

Definition of done:

  • product feels honest, not overconfident

P2: Testing

P2.4 Extension and Backend Test Corpus

Current files:

  • tests
  • new fixture directories should be added under tests/fixtures

Tasks:

  • create a fixture corpus for:
    • hotel checkout
    • airline upsells
    • vacation rentals
    • non-commerce pages
  • add regression cases for:
    • false positives
    • false negatives
    • non-commerce misclassification

Definition of done:

  • the team can add a broken example and keep it fixed

P2.5 Extension E2E

Suggested new folder:

  • artifacts/guardian-extension/tests

Tasks:

  • add end-to-end tests for:
    • popup -> content script -> backend flow
    • SPA route changes
    • non-commerce suppression
    • DOM highlighting

Definition of done:

  • extension behavior can be validated automatically before release

P2: Infra and Deployment

P2.6 CI and Release Discipline

Current files:

Tasks:

  • add CI for:
    • lint
    • typecheck
    • backend tests
    • extension tests
    • Docker build
  • version the extension and backend API intentionally together
  • add environment separation:
    • local
    • staging
    • production

Definition of done:

  • every release is reproducible and tested

Compliance and Legal

This should be handled continuously, not as a late-stage task.

Requirements:

  • present Guardian as user-side assistance, not bot abuse
  • do not automate purchases without explicit approval
  • do not silently create accounts or submit credentials
  • publish clear privacy and retention rules

Recommended Order of Execution

If only five items are started next, do these first:

  1. P0.1 shared detection pipeline
  2. P0.2 non-commerce and page-type detection
  3. P0.3 PII redaction
  4. P2.4 fixture corpus and regression tests
  5. P1.4 backend hardening and observability