How to use Codex to add Paris nights without moving flights
Codex used Finite’s live plan to add Paris nights within fixed-flight and budget limits; a person approved one revision with a receipt, without buying travel.
devpost.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Using the live plan’s WebMCP tools, explore adding three Paris nights while preserving fixed international flights and the stated minimum budget room. Read the current revision, compare distinct feasible directions and show exact effects. Wait for the person’s choice, apply only that approved revision once, and report the saved receipt without implying a booking.
Reach people exploring AI agents
See placements →What happened
The public demonstration completed one approved trip-plan revision retaining AUD 910 of room; no hotel purchase, payment or completed journey is claimed.
Reported by Benji Hart — DevpostHow the agent did it
- Codex read the live plan and the request to add Paris nights while preserving flights and budget room.
- Finite checked combinations and returned distinct feasible directions for the person to compare.
- The person approved one direction, and the system saved one revision update with a persistent receipt.
Where a person stepped in
The person selected and confirmed the final plan change; Codex could prepare and explain alternatives without granting itself approval or transaction authority.
Benji Hart’s Finite demonstration used Codex to revise a trip instead of generating another disconnected itinerary. The request added three nights in Paris while keeping international flights fixed, respecting the total budget and retaining at least AUD 500 of unallocated room.
Codex accessed the same accepted plan shown on the page through WebMCP. Finite performed the calculations and constraint checks, evaluating 26 combinations, rejecting eight and grouping the remaining options into five distinct directions. The person chose the option protecting financial breathing room, reviewed its precise effect and confirmed it.
The accepted plan advanced from revision one to revision two, retained AUD 910 of room and saved a before-and-after receipt that survived reload. Revision checks, approval challenges and idempotent updates kept retries or stale sessions from silently applying a different change.
The implementation also separated absolute values from deltas after testing exposed ambiguity around a value of three. Codex handled language and orchestration; deterministic code controlled money, constraints and saved state, while the human chose the final direction. This is an executed public planning demonstration, not evidence of purchased hotels or completed travel. The source explicitly distinguishes research, verification, booking and payment rather than treating a saved revision as a transaction.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Devpost · Benji Hart
Frequently asked questions
Who performed the constraint calculations?
Finite’s deterministic planning kernel; Codex supplied language, research and orchestration.
Who selected the accepted direction?
The person chose and confirmed it on the visible planning surface.
Did saving the plan purchase travel?
No. Research, verification, booking and payment remain separate states.
More details
About this account
Codex read the live plan and the request to add Paris nights while preserving flights and budget room. Finite checked combinations and returned distinct feasible directions for the person to compare.
Dated firsthand public demonstration opened. Counts and currency amounts are the source’s demonstrated scenario, not current travel prices. Codex tool execution is explicit; saved planning state is not a purchase. Suggested prompt written by Agents In Use; no original task prompt was found in the verified text.
Credit: Benji Hart — Devpost
Source published 3 Sept 2026
Source checked 14 Oct 2026
Tools and topics
Codex · WebMCP · Finite · Cloudflare D1
trip revision · human approval · persistent receipt
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.