AgentBoards Coding-Agent Comparison Kit v1 - 2026-10-21
Requires Node.js 22 or newer. No dependencies or network access required.

These are deliberately small, public acceptance tests. Passing them is not a
production-readiness assessment or an independent leaderboard result.

Run each task separately in a fresh copy of the entire kit and a new agent
session. Give every agent the same prompt, files, permission settings, context,
time limit (10 minutes) and message budget (3 user turns). Do not let one agent
see another agent's output. Keep all tests unchanged. Record all failures and
human interventions. Never include secrets in shared logs.

TASK 1: Invoice quantities (edit invoice.mjs only)
Fix totalCents(items). Input is an array of objects with priceCents and quantity.
Both values must be nonnegative safe integers; otherwise throw RangeError.
Return the sum of priceCents * quantity. Do not mutate the input.
Check: node --test invoice.test.mjs

TASK 2: Pagination (edit pagination.mjs only)
Fix paginate(items,page,size). page is a nonnegative integer, zero-based; size
is a positive integer. Reject invalid page/size with RangeError. Return the
appropriate slice, or an empty array beyond the last page. Do not mutate input.
Check: node --test pagination.test.mjs

TASK 3: Campaign URL (edit campaign-url.mjs only)
Fix cleanCampaignUrl(input). Require an absolute HTTP(S) URL. Remove every query
parameter whose name starts with utm_ (case-insensitive). Preserve other query
parameters, repeated keys, and the hash. Return the URL's serialized string.
Check: node --test campaign-url.test.mjs

Before any edits all three task suites should fail. After solving all tasks:
npm test

Record exact agent version, model ID, effort, OS, dates, permission settings,
prompts and test output, elapsed wall-clock seconds, turns and interventions.
Record the actual billed cost and currency, or null if unavailable (never zero
as a substitute for unknown). Subscription use does not establish per-run cost.
Use results-template.json; status starts as not-run. Do not publish as measured
until actual output and test artifacts exist. Repeat runs before generalizing.
