Plain code makes every number. AI only helps build the site.

The collector and the checks are plain code. The builder uses AI to write some of the code, and one person writes the rest.

$5.27

Spent on builds

5

Builds merged

$1.05

Cost per merged build

2 days

Since the last merge

Spent counts the builder's runs. Every model call together, since 7 Oct, comes to $7.35. All of it is held to $20 a month.

Each part has one job and cannot do the others'

The builder can't approve its own work. No AI ever calculates a number.

Part	How often	What it does
Collector	Every hour	Pulls prices from exchanges, transfer apps and official rates and saves them to the data files. Plain code, no AI.
Builder	Up to 3 jobs a day	Takes the top ready job from a list a person writes and builds it as new files. Uses AI. It can't change existing price collection, pages or methods. It and every other model call together can spend at most $20 a month.
Checks	Every change	Run tests on every change, recalculate published numbers from the saved files and merge the change when every test is green. Plain code, no AI. A change to how prices are collected that adds nothing new waits for a person.
Person	As needed	Sets the direction, holds the money and passwords, and writes what the builder skips, using AI sessions of their own.

The record grew as new sources and routes came in

634,287 prices stored

Exchanges' own earlier prices reach back to 3 Jan 2011, so some charts predate our record.

0317,144

634,287

10 AugToday

Week by week

The builder tries a job, may open a change, and the change merges when the checks are green. A merge counts only if the files it added are not empty. Weeks run Monday to Sunday, UTC.

Tries and merges each week

TriesMerged

Week starting	Tries	Changes opened	Merged	Share merged	Cost
2 Nov	0	0	0	None yet	$0.00
26 Oct	20	5	2	10%	$1.65
19 Oct	17	6	3	18%	$3.62

What runs where

Every job runs on GitHub Actions. Times are Singapore time.

Job	When	What it does	Last run
Builder	Every day at 13:07, 16:37, 21:07	Builds the top ready job with AI and opens a change.	Succeeded (1 Nov 18:26 UTC)
Merge when green	After the checks finish on a change	Merges a change when every check is green.	Succeeded (2 Nov 01:40 UTC)
Checks	On every change	Runs the tests and page checks a change must clear.	Running (2 Nov 01:41 UTC)
Reader test	By hand	Compares two models reading the pages. Changes nothing.	Succeeded (5 Oct 09:28 UTC)
Collection	About every 30 minutes, each run starts the next	Collects prices, rebuilds the pages and saves the data.	Running (2 Nov 01:35 UTC)
Collection alarm	Every 30 minutes	Watches collection from outside and sends an alarm when it goes quiet.	Running (2 Nov 01:15 UTC)
Fee check	Day 1 of each month at 11:23	Checks the fee tiers each month.	Succeeded (1 Nov 10:39 UTC)
History pull	By hand	Pulls older price history from public sources. Never feeds a price.	Succeeded (22 Oct 17:30 UTC)
Older builder	By hand	The older builder, run only by hand.	Succeeded (23 Oct 17:03 UTC)
Outside check	After each deploy of the site	Asks the live site what a search engine would ask.	Succeeded (2 Nov 01:42 UTC)
Board probe	Sundays at 11:11	Re-checks which boards where people sell to each other are usable.	Succeeded (1 Nov 10:21 UTC)
Audit alarm	Every hour, at minutes 11, 41	Sends an alarm when the audits or the builder's runner stop working.	Running (2 Nov 01:31 UTC)
writer	By hand		Failed (1 Nov 09:45 UTC)

Incidents

Newest first. A gap in the data stays a gap.

28 to 31 Oct 2026

Collection stopped for two days

What broke: A table of seller ads grew to 99.6 MB. GitHub refuses a file over 100 MB, so every save of new prices failed.
How long: The last stored hour was 28 Oct 20:00 UTC and the next was 31 Oct 04:00 UTC, 56 hours apart.
What was lost: Every hour in that gap. Missed hours stay missing and are never filled in.
What fixed it: The big tables now write one file a day. A size check fails the run before the save. One alarm now covers collection.
Changes: 719, 720

31 Oct 2026

Two tests turned the main branch red

What broke: Tests that read the real data files failed as the hours moved on, so checks on the main branch went red.
How long: Not recorded.
What was lost: Not recorded.
What fixed it: The gas test now uses fixed hours of its own. The research pieces are rebuilt before the claims check.
Changes: 734, 736

25 to 27 Oct 2026

Two builds merged with empty files

What broke: The builder opened changes whose new files were empty. The checks were green and the changes merged.
How long: The builder first stopped an empty result on 29 Oct.
What was lost: A finding retest and a new price source that were marked built and did nothing.
What fixed it: A merge now counts only if its files are not empty. This page counts that way.
Changes: 696, 701

22 Oct 2026

The builder rewrote lines it should not touch

What broke: In its first two changes the builder edited existing price code. One change would have altered the source link on every route. The other rewrote four existing entries.
How long: Both changes were closed by hand the same morning.
What was lost: Nothing. Neither change merged.
What fixed it: A check now refuses any change to an existing price collector. The builder tries again until its change only adds.
Changes: 629, 632, 640

Before 23 Sep 2026

A key spent $10 in 20 minutes

What broke: A settings change put the key billed for each use where the flat-rate login belonged. Every agent run billed for each use.
How long: About 20 minutes.
What was lost: About $10 of credit.
What fixed it: The two keys now live in separate files, with a note where each is used saying why they must never be mixed.

data/aider_log.csv · data/agent_spend.csv · data/machine_room_prs.csv · data/index_hourly.csv · data/machine_room_meters.json