On Friday, November 14, Timeline Murder was an idea and an empty repo. By Tuesday the Android build was in a closed test on Google Play and the iOS build was sitting in App Review. Google took hours. Apple is still thinking about it. If Apple weren't a turtle, you could already play it on your iPhone.
This post covers what we built, what it cost, and which parts we didn't have to build because the backend already had them. "We" is me and Anne. Coding agents wrote most of the code; we decided what the game should be, read the diffs and played it a lot.
The game
Timeline Murder is a logic crime puzzle. One house, five epochs: 1724, 1924, 1962, 1987 and 2046. Every case is a floor plan of the house with rooms, furniture and six people, one of them the victim. The clues say things like "Barnaby sat in an armchair" or "Eleanor stood next to a bookshelf". You place everyone on the plan until all clues hold, then make your accusation.
Story mode runs in seasons. A season belongs to an epoch, has episodes with a running plot, scene images, music and voiced lines, and ends with a finale. There is also free play, a daily case, a multiplayer race (with a ghost opponent when nobody else is online) and co-op where friends share clues.
The numbers
From the repo, as of November 18:
- 92 commits and 68 merged pull requests
- Commits per day: 4 on the 14th, 44 on the 15th, 23 on the 16th, 21 on the 17th
- About 64,900 lines of TypeScript and TSX
- 108 test files
From the production database (read replica, counts only):
- 5 epochs, 7 seasons, 50 episodes, 57 cases
- 2,598 season variants and 59 asset packs
- 7,397 events since November 15
- 98 job runs
We also have accounts, played cases and one purchase. Nearly all of that is us and our testers, so I won't pretend it's traction.
What the content cost
Every scene image, episode text and voiced line came out of an AI pipeline that runs as a server job. Each model call is recorded as an event with the model, token usage, latency and the cost OpenRouter reports. So the bill is a database query:
| What | Model | Calls | Cost | Avg. time |
|---|---|---|---|---|
| Images | gemini-3.1-flash-image | 309 | $21.01 | 10.2 s |
| Texts | gpt-6-luna | 876 | $5.79 | 117.6 s |
| Speech | gpt-audio-mini | 1,477 | $0.34 | 1.2 s |
| Total | 2,662 | $27.14 |
The OpenRouter key for the pipeline shows $27.60 for the same days, so the recording is close enough to trust.
Images are 77% of the money. Texts take the longest by far: a season plot has to stay consistent across episodes, keep the clues solvable and match the epoch, and the model spends almost two minutes per call getting there. Speech is almost free.
$17 of trying models before the pipeline existed
The table above starts on November 16. The two days before that don't show up in our database, because there was nothing to record into yet. We tried image models, generated the first title images by hand, tested a few voices and let a music model write the theme. That ran on a different OpenRouter key:
- Image model: $15.10
- Text model: $0.95
- Music: $0.52
- Two speech models: $0.42
About $17 in total. From the first prompt to five epochs of finished content, we spent roughly $45 on AI generation. That excludes the coding agents, which are a separate bill.
What we didn't build
This is the part I care about most. Timeline Murder runs on Kumiko, our TypeScript backend framework. The game code is the puzzle engine, the content pipeline and about ten game features (timeline, progress, multiplayer, friends, moderation, purchases and so on). Everything below came from the framework and needed configuration, not code:
Sign-in. Sign in with Apple, Google and an email link. Apple insists on Sign in with Apple if you offer any other social login, so this had to be there before submission. It was one feature to enable.
Account deletion. Apple also wants users to delete their account inside the app. The framework's user data rights feature already does deletion and export, so the app just links to it.
Push notifications. Push notifications go out through the framework's push channel. We wrote the messages, not the delivery.
Background jobs. Season generation runs as a job on a separate worker. When a job gets delivered twice, it resumes the unfinished season instead of starting over. That fix was ours; the job runner and the worker setup were not.
AI connections and cost tracking. The model setup lives in an admin screen: an AI connection to OpenRouter and default models for text and image. Changing the image model is a form edit, not a deploy. Every call goes through the AI foundation package, which is where the cost events in the table come from.
Admin screens. Reviewing a season before release, regenerating a single scene, giving testers early access to episodes: all of that sits in the admin UI the framework builds from the feature definitions.
Event sourcing and audit. Every change is an event. When a player says their progress vanished, we can see exactly what happened. That is also why the content numbers above were one query away.
Tests against a real stack. The integration tests start the server with a real Postgres and talk HTTP to it. Twenty test files use that stack. When an agent breaks the purchase flow, a test goes red before a human notices.
The coding agents were fast because the hard and boring parts were already done and documented. They wrote game logic and wired features. They didn't invent auth.
Getting into the stores
A good part of the week went into the stores, not the game. A few things that bit us:
Product IDs. We set up subscriptions in App Store Connect first, with descriptive IDs of 52 characters. Google Play caps subscription product IDs at 40 characters. So Android got short IDs, and the shared API contract now knows which IDs belong to which platform and rejects a purchase that names the wrong one. Purchases are verified on the server against Apple and Google before anything unlocks.
User-generated names. Players pick display names and play against each other. App Review guideline 1.2 asks for a way to report and block other users when there's user-generated content. We built reporting and blocking on day two, before the first submission, which beat waiting for a rejection.
The closed test that stayed old. Our build tooling submitted every Android build to the internal track as a draft. The closed test track, the one testers actually use, kept serving an older build for a day before we noticed. One line in the submit config fixed it. Checking which build the testers have, rather than which one you uploaded, is the lesson.
Store assets in two languages. Screenshots for iPhone and iPad, a feature graphic, store texts, an age rating questionnaire and review notes, all in German and English. Screenshots come from a script that drives the app through ten scenes per device and locale.
What's next
The Android closed test is open. The iOS build is in review, and judging by our past apps that takes one to two weeks. The website already has the daily case to play in the browser.
You don't need either store to play. The daily case runs in the browser at timelinemurder.cosmicdriftgamestudio.com/en, no account needed. It's the same case for everyone, and the server checks every time before it lands on the leaderboard.
So: who's faster? Solve today's case, hit share and paste your time and rank in the comments. And if you have an idea for the game, a sixth epoch, a new kind of clue, a mode we haven't thought of, put it there too.


Top comments (0)