Skip to content

PHASES_1_3_IMPLEMENTATION.md

Latest commit

 

History

History
149 lines (113 loc) · 4.98 KB

PHASES_1_3_IMPLEMENTATION.md

File metadata and controls

149 lines (113 loc) · 4.98 KB

Sally Consumer Extension: Phase 1-3 Implementation

Date: 2026-03-08

This document captures the first implementation pass for making heysalad-ai-shopper safer, smarter, and monetizable without another full architectural reset.

Objective

Ship three things together:

  1. Phase 1: remove unsafe client-side AI usage from the Chrome extension.
  2. Phase 2: move product intelligence to the backend and make model choice configurable.
  3. Phase 3: introduce paid gating for autonomous shopping using Sally subscription status.

What Changed

Phase 1: Security hardening

Files:

Changes:

  • Removed https://api.openai.com/* from Chrome extension host permissions.
  • Removed direct OpenAI calls from the content script.
  • Removed extension-side API key fallback usage in the content script.
  • Kept a local rule-based validation fallback in the content script for final safety checks.

Result:

  • The extension no longer needs a raw OpenAI key.
  • The Chrome Store package no longer requests direct OpenAI network access.

Phase 2: Smarter backend AI routing

Files:

Changes:

  • Added configurable backend model selection:
    • AI_CHAT_MODEL
    • AI_MATCHING_MODEL
  • Updated chat/tool orchestration to use AI_CHAT_MODEL.
  • Updated product matching to use AI_MATCHING_MODEL.
  • Extended /api/match-product to return:
    • bestIndex
    • confidence
    • reason
    • shouldAdd
    • validationConfidence
    • validationReason
    • model
  • Added rule-based rejection of obvious bad matches on the backend, for example:
    • apples vs apple juice
    • tomatoes vs tomato sauce
    • chicken vs chicken stock

Result:

  • AI matching is now server-side.
  • Future model upgrades do not require a Chrome Store release.
  • You can test stronger OpenAI models later by changing worker configuration instead of rewriting the extension.

Phase 3: Paid gating for autonomous shopping

Files:

Changes:

  • Added background worker subscription lookup via Sally API:
    • calls /api/subscriptions/status
    • derives an extension plan object
  • Added an automation entitlement gate before shopping starts.
  • Allowed autonomous shopping only for premium or pro.
  • Added popup plan UI:
    • current extension access tier
    • free vs paid state
    • upgrade CTA
  • Preserved login flow and shopping flow, but now free users are blocked from autonomous shopping with a clear upgrade message.

Result:

  • The monetization boundary is enforced in the extension control path.
  • Free users can still sign in and use the extension UI, but autonomous shopping becomes a paid capability.

Deployment Order

  1. Deploy shopping-agent first.
  2. Verify /api/match-product returns shouldAdd and model.
  3. Update and republish the Chrome extension.
  4. Confirm paid users can start autonomous shopping.
  5. Confirm free users receive the upgrade-required flow.

Recommended Environment Configuration

For shopping-agent:

  • AI_CHAT_MODEL=gpt-4o
  • AI_MATCHING_MODEL=gpt-4o-mini

You can move to a stronger matching model later without changing the extension package.

What This Does Not Solve Yet

This pass does not yet implement:

  • server-signed extension sessions
  • per-user usage quotas or metered billing
  • backend-side cart planning memory by user history
  • ranked alternative suggestions rendered in the popup
  • store-specific structured extraction models

Those are the next production-grade steps after this pass is stable.

Recommended Next Iteration

  1. Add signed extension session tokens from Sally API to shopping-agent.
  2. Add per-plan quotas such as:
    • free: no autonomous runs
    • premium: 10 runs/month
    • pro: unlimited
  3. Add a backend planning endpoint that returns:
    • primary match
    • approved substitutes
    • max price guardrails
    • dietary exclusions
  4. Add analytics on:
    • attempted runs
    • paid conversion prompts
    • product-match rejection rates
    • store-specific failure rates

Summary

This implementation changes the extension from:

  • client-side AI
  • exposed model access
  • no monetization boundary

to:

  • backend AI orchestration
  • configurable models
  • subscription-aware autonomous shopping

This is the right foundation for charging users and improving intelligence safely.