How to use Codex to check cart substitutions against budget
Codex built Forage’s shopping workflow; tests found incomplete carts, substitution budget drift and invented attributes, leading to explicit checks and limits.
devpost.com
Original prompt
Shared by the source author. See source notes for context.
Complete ALL items in a multi-item request. Do not stop after 1-2 items.
Reach people exploring AI agents
See placements →What happened
The creator reports a deployed cart-composition product with specific completion, substitution and enrichment safeguards; cheaper production cost and live inventory remain roadmap work.
Reported by Aditya Kalidindi — DevpostHow the agent did it
- Kalidindi defined catalog and cart modules, and Codex implemented tasks from explicit acceptance criteria.
- Tests exposed early completion, substitution budget drift and unsupported material-percentage claims.
- The build added complete-request instructions, mandatory cart-summary rechecks and evidence confidence distinctions.
Where a person stepped in
Kalidindi set architecture and failure criteria and reviewed the shopping behavior; current inventory and actual purchasing were not established by the demonstration.
Aditya Kalidindi built Forage to test whether a shopping agent could compose a complete cart from structured catalog evidence. Codex handled implementation from planning artifacts through small tasks with explicit acceptance criteria. The running product used separate models for catalog enrichment and cart decisions.
Testing exposed concrete failures. The cart agent sometimes stopped after two of three requested items. A substitute could also exceed the budget because the running total was not checked again. The fixes added explicit completion instructions and required a cart-summary check after every substitution.
Enrichment had its own failure: fabricating precise material percentages from vague product descriptions. The design distinguished observed attributes from inferences with confidence scores and anti-hallucination constraints. The interface compared raw keyword results with a composed cart and showed unavailable items, substitutions and catalog gaps.
Kalidindi retained architecture, product decisions and failure criteria while Codex implemented the approved tasks. The account reports a deployed product built through ten commits. It also reports approximately fifty cents per cart and describes lower costs as a roadmap, not an achieved measurement. This is a shopping-tool build and evaluation using catalog data, not a verified live-inventory purchase. Real-time inventory integration remained future work.
Prompt
Original prompt
Source: Devpost · Aditya Kalidindi
Frequently asked questions
Was Codex the running cart agent?
No. Codex built the application; GPT-5.6 Sol ran the cart loop and Terra enriched the catalog.
How was substitution budget drift addressed?
Every substitution required another cart-summary check before proceeding.
Was the lower projected cart cost achieved?
No. The source presents it as an optimization roadmap, not a measured outcome.
More details
About this account
Kalidindi defined catalog and cart modules, and Codex implemented tasks from explicit acceptance criteria. Tests exposed early completion, substitution budget drift and unsupported material-percentage claims.
Original runtime guard-instruction excerpt shared by the author, not the initial Codex coding brief. Firsthand dated build/evaluation opened. No actual order, live stock verification or achieved roadmap cost is inferred. Original prompt excerpt.
Credit: Aditya Kalidindi — Devpost
Source published 21 Jul 2026
Source checked 14 Oct 2026
Tools and topics
Codex · Shopify CSV · GPT-5.6 Sol · GPT-5.6 Terra · cart tools
cart substitutions · catalog evidence · budget recheck
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.