| 1 | # 🤖 TodoMVC bakeoff on Carloku — design |
| 2 | |
| 3 | Status: draft for Paul's approval, 2026-10-23. Reviewed by Astra (builder check). |
| 4 | |
| 5 | ## Why |
| 6 | |
| 7 | CARLOS apps don't have to be written in Go, and the CARLOS Platform runs |
| 8 | any binary that keeps its process contract. This bakeoff shows that with |
| 9 | working apps. It's driven by curiosity and acts as a showcase. It isn't |
| 10 | meant to pick a winner or feed a decision. |
| 11 | |
| 12 | We build the same app, TodoMVC, in nine language setups. Each one renders |
| 13 | HTML on the server and layers JavaScript on top, and all of them are |
| 14 | deployed to the Carloku **bakeoff** account (`bfj`). A report page on the |
| 15 | same account compares them. |
| 16 | |
| 17 | ## Goals |
| 18 | |
| 19 | - Nine working TodoMVC apps, live on Carloku, one per setup: |
| 20 | 1. Go, without Rastrillo |
| 21 | 2. Go, with Rastrillo |
| 22 | 3. Rust |
| 23 | 4. Zig |
| 24 | 5. OCaml |
| 25 | 6. Haskell |
| 26 | 7. Swift |
| 27 | 8. Nim (replaced Mojo, Paul's call 2026-10-24: Mojo can't make a |
| 28 | single-file binary; see `docs/toolchains.md`) |
| 29 | 9. Spinel (a Ruby compiler that produces native binaries) |
| 30 | - Every app passes the same behaviour tests with JavaScript on and with |
| 31 | it off. |
| 32 | - A report page that puts honest numbers side by side: what each build |
| 33 | cost in tokens, how readable the code is, how fast it responds, how it |
| 34 | holds up under load, and the other measures listed below. |
| 35 | |
| 36 | ## Non-goals |
| 37 | |
| 38 | - Choosing a "best" language. The report presents numbers and doesn't |
| 39 | crown anyone. |
| 40 | - Tuning one app beyond the others after the build. Every app gets the |
| 41 | same build process and the same review-and-fix round, and nothing more. |
| 42 | - Rich features beyond TodoMVC: no sign-in, no sharing, no sync between |
| 43 | devices. |
| 44 | |
| 45 | ## Anti-goals (failure even if everything technically works) |
| 46 | |
| 47 | - **A report that looks rigged.** One app gets a different UI, a different |
| 48 | database, a bigger instance, more tuning, extra review rounds, or quietly |
| 49 | drops part of the spec. |
| 50 | - **An app missing from the report without comment.** If a setup can't be |
| 51 | built or can't run on Carloku, the report says so plainly and says why. |
| 52 | - **Working around Carloku.** If the platform lacks something an app needs |
| 53 | (for example, a binary that needs files beside it), we file a request |
| 54 | against the platform and report the app as blocked on it. We don't |
| 55 | build workarounds. (Paul's standing rule.) |
| 56 | |
| 57 | ## Decided with Paul (2026-10-23) |
| 58 | |
| 59 | - **One shared UI.** Every app uses TodoMVC's standard markup and CSS |
| 60 | (`tastejs/todomvc-app-template`, `todomvc-app-css`). For the bakeoff |
| 61 | this overrides the CARLOS skill's default that Rastrillo apps use |
| 62 | Rastrillo's own UI components. The Go + Rastrillo app uses the TodoMVC |
| 63 | CSS too. |
| 64 | - **Server-rendered with progressive enhancement.** No client-side app |
| 65 | framework. The page is HTML from the server and still works with |
| 66 | JavaScript turned off. |
| 67 | - **Each agent chooses its own stack.** Web server, router, templating, |
| 68 | SQLite binding and so on are each agent's own best recommendation for |
| 69 | its language. |
| 70 | - **The builder is Opus 5.5 for every app.** Astra reviews every app |
| 71 | through the codex CLI, and the same Opus agent fixes what the review |
| 72 | raises. |
| 73 | - **Spinel is already supported on Carloku.** The Spinel builder starts |
| 74 | from lpay's working setup (`paulca/lpay`, directory `spinel/`): tep |
| 75 | (a Sinatra-style layer that drives Spinel), the `tep-carlos.patch` for |
| 76 | the socket contract, and the zig-based cross-build to a static |
| 77 | aarch64-musl binary in `BUILD-arm64.md`. Mojo is expected to work if it |
| 78 | compiles to a single binary. |
| 79 | - **Arm64 builds:** try docs-t1 first (cross-compiling or an emulated |
| 80 | arm64 container), and fall back to tito-studio over Tailscale. |
| 81 | - **After the one fix round:** remaining correctness gaps ship as they |
| 82 | are, and the report shows them. Security findings get a second fix |
| 83 | round, counted separately, because the apps are public. |
| 84 | - **Downloads approved:** GHC (via ghcup) and Mojo (via pixi) may be |
| 85 | installed under `~` on docs-t1. |
| 86 | - **Everything lives on the bakeoff account** (`bfj`): nine apps and the |
| 87 | report page. |
| 88 | |
| 89 | ## What every app must do (the shared brief) |
| 90 | |
| 91 | Each builder gets the same written brief. It contains: |
| 92 | |
| 93 | ### Behaviour |
| 94 | |
| 95 | - The TodoMVC app spec (`tastejs/todomvc/app-spec.md`), with these changes |
| 96 | because the server holds the data: |
| 97 | - **Persistence:** in SQLite on the server, not in `localStorage`. |
| 98 | - **Routes:** real paths `/`, `/active` and `/completed`, not `#/` |
| 99 | fragments. The selected filter survives a reload because it's in the |
| 100 | URL. |
| 101 | - Editing state isn't persisted (same as the spec). |
| 102 | - **With JavaScript off,** everything works through plain forms and links: |
| 103 | add, toggle, toggle all, edit, delete, clear completed, and the three |
| 104 | filters. Each action is a POST that redirects back to the current |
| 105 | filter (post/redirect/get). Editing without JS uses a link to an edit |
| 106 | view of that item, such as `/?edit=<id>`, rendered with the edit field |
| 107 | open. |
| 108 | - **With JavaScript on,** the page behaves like the reference TodoMVC: |
| 109 | double-click to edit, Enter saves, Escape cancels, blur saves, an empty |
| 110 | title deletes, toggles happen without a full page reload, and the |
| 111 | counter and the "Clear completed" button update in place. The JS sends |
| 112 | the same POSTs and swaps in server-rendered HTML or applies a small |
| 113 | JSON reply (the builder chooses). There's no build step and no |
| 114 | framework. Plain ES modules are allowed, and so is vendored VanJS (the |
| 115 | one reactive library CARLOS sanctions). |
| 116 | - **Who owns a list:** the whole app is one instance. Each visitor gets |
| 117 | their own list through an anonymous, random, `HttpOnly`, `SameSite=Lax` |
| 118 | cookie set on the first visit. There's no sign-in. Forms are protected |
| 119 | against cross-site posts (by SameSite plus an Origin check, or a CSRF |
| 120 | token; the builder chooses). |
| 121 | |
| 122 | ### Platform contract |
| 123 | |
| 124 | - One binary for `linux-arm64` (the platform default; `carlos deploy |
| 125 | --target`), started as `<bin> --socket <path> --db <path>`. It listens |
| 126 | on the unix socket, has no `$PORT`, and serves `GET /healthz` and |
| 127 | `GET /api/version`. |
| 128 | - **Linked statically, or with only libc.** Instances run sandboxed and |
| 129 | see only their own build, so the builder must not assume any shared |
| 130 | library exists on the box. `carlos vet --boot` (a local check that the |
| 131 | binary starts and answers `/healthz`) must pass before the first deploy. |
| 132 | - Reads `CARLOS_HOST` when it needs its own hostname. Never parses the |
| 133 | socket path. |
| 134 | - Uses SQLite in WAL mode. Migrations only ever add. |
| 135 | - The app starts quickly after hibernation (sleeping when idle) and shuts |
| 136 | down cleanly on SIGTERM. |
| 137 | - Configures the build probe (`carlos apps create --build-probe |
| 138 | /api/version`) so `carlos deploy` confirms that the new build is what's |
| 139 | actually serving. |
| 140 | |
| 141 | ### Repo shape for each app |
| 142 | |
| 143 | - `apps/<slug>/` in this bakeoff repo. The slugs are `go`, `go-rastrillo`, |
| 144 | `rust`, `zig`, `ocaml`, `haskell`, `swift`, `mojo` and `spinel`. |
| 145 | - `AGENTS.md` first. It holds the one rule and the tenancy declaration: |
| 146 | one instance, anonymous lists by cookie, trust class C (plaintext todo |
| 147 | text, chosen and written down). |
| 148 | - A `README.md` with how to build, run locally, test and deploy. One |
| 149 | command for each (a `Makefile` or a script). |
| 150 | - The app's own tests, whatever is normal for its language. |
| 151 | - The Carloku app name is `todo-<slug>`, e.g. `todo-rust`. |
| 152 | |
| 153 | ### What the builder may not do |
| 154 | |
| 155 | - Change the markup or the CSS beyond adding hooks the JS needs, such as |
| 156 | `data-` attributes or wrapping each toggle in a `<form>`. The rendered |
| 157 | page must look identical to the reference. |
| 158 | - Add caching layers, a CDN or edge tricks that the other apps don't |
| 159 | have. Each app may set ordinary HTTP cache headers on its static |
| 160 | assets. |
| 161 | - Spend effort on performance tuning beyond what an experienced developer |
| 162 | would write the first time. "Write it the way you'd normally write it |
| 163 | well" is the instruction. |
| 164 | |
| 165 | ## How the apps get built |
| 166 | |
| 167 | 1. **Shared tooling first, built by the orchestrator (this session)** |
| 168 | before any builder starts, so every app is judged by the same tools: |
| 169 | - The brief, as one file every builder reads. |
| 170 | - A **behaviour test suite** adapted from TodoMVC's Cypress tests |
| 171 | (`tastejs/todomvc/cypress`). It runs against any URL, once with JS |
| 172 | on and once with JS off. The JS-off run uses forms only and skips |
| 173 | the parts that are JS-only, such as double-click editing (replaced |
| 174 | by the edit link). Builders run it locally against their own app. |
| 175 | - A **load test script** and a **metrics collector** (see "Metrics"). |
| 176 | 2. **Toolchains installed** per user, under `~` (`rustup`, Zig tarball, |
| 177 | `opam`, `ghcup`, Swift's static Linux SDK, Mojo through `pixi`, |
| 178 | Spinel built from source). GHC and Mojo each take several GB, so we |
| 179 | ask Paul before downloading them (machine rule). |
| 180 | 3. **Nine Opus 5.5 builders run in parallel**, each in its own headless |
| 181 | `claude` session with JSON output so token use is recorded. Each |
| 182 | starts from a fresh context with the same brief and the CARLOS skills |
| 183 | (`carlosframework/skills`), plus Rastrillo's own `SKILL.md` for |
| 184 | Go + Rastrillo. Each builds, tests locally, then deploys to `bfj` with |
| 185 | `carlos deploy` and checks it on the live URL. |
| 186 | 4. **Astra reviews each app** through `codex exec` against the brief, the |
| 187 | spec and the code. The review is limited to correctness, spec |
| 188 | compliance and security. It doesn't ask for performance tuning. |
| 189 | 5. **One fix round.** The same Opus builder session fixes what Astra |
| 190 | raised, then redeploys. Only one round, for every app. The one |
| 191 | exception is security: if Astra still finds a security problem after |
| 192 | that round, there's a second round for security fixes only, with its |
| 193 | tokens reported separately. |
| 194 | 6. **Blind readability scoring by Fable.** Each app's code is shown with |
| 195 | the language named but nothing else identifying, and scored against a |
| 196 | fixed rubric. Astra doesn't score readability because it already |
| 197 | reviewed the code. |
| 198 | 7. **Measurement and report** (below). |
| 199 | |
| 200 | All local servers bind to `127.0.0.1`. Scratch space goes in `$TMPDIR` |
| 201 | and is removed at the end. |
| 202 | |
| 203 | ## Metrics |
| 204 | |
| 205 | All of them are collected by scripts in this repo and can be re-run. |
| 206 | |
| 207 | | Measure | How | |
| 208 | |---|---| |
| 209 | | Tokens to build | From each builder's session log: input, output, cache read and cache write, plus turns and wall-clock time. The first build and the fix round are reported separately, and so is the Astra review. | |
| 210 | | Correctness | Behaviour tests passed, JS on and JS off, against the live URL. Spec gaps found by review. | |
| 211 | | Readability | Fable's rubric score (naming, structure, how easy it is to follow one request end to end, error handling, comments), with a short written justification. | |
| 212 | | Size of the code | Lines of code split three ways: server code, templates, and JS. Also the number of third-party dependencies. | |
| 213 | | Build | Three build times for the arm64 binary that ships: **cold** (the app and every third-party library compiled from already-downloaded sources; compiler and standard library as shipped), **warm** (libraries already compiled, app rebuilt from clean), and **one-line change** (rebuild after a trivial edit in the request-handling code). Plus binary size. Paul asked for cold and warm, 2026-10-23. Measured on docs-t1 where the app cross-compiles there; otherwise on the machine that built it, with the machine named beside the number and kept out of the ranking. | |
| 214 | | Response time | Time to the first byte for `GET /` with an empty list and with 50 items, from docs-t1, median of repeated runs. | |
| 215 | | Wake time | Time to the first byte on the first request after the instance has gone to sleep. | |
| 216 | | Under load | A ramp for reads (`GET /` on a 50-item list) and for writes (toggling an item): 1, 2, 4 … 1,024 concurrent connections, 10 s per step. Reports peak requests per second, latency at every step, and **where it slows down**: the most connections at which doubling the load still raised throughput by 10% or more. A ramp stops early after two flat doublings or more than 1% errors. Paul asked for this, 2026-10-23, replacing fixed loads at 1, 16 and 64. | |
| 217 | | Memory | Peak and idle memory from the platform, if it exposes them. If it doesn't, we file a request against the platform rather than measure inside the app. | |
| 218 | | Page weight | Bytes of HTML and JS sent for `GET /`. | |
| 219 | |
| 220 | Load tests run from docs-t1 over the public internet, so the numbers |
| 221 | include network time. That's the same for every app, and the report says |
| 222 | so. Every app runs on the same box class and instance size, and we run |
| 223 | the apps one after another, never at the same time. |
| 224 | |
| 225 | ## The report page |
| 226 | |
| 227 | - A static site on `bfj` (`carlos deploy --kind static`), named |
| 228 | `bakeoff`. |
| 229 | - One page with a summary table of every measure for every app, and a |
| 230 | chart per performance measure. Each app gets a short section covering |
| 231 | its stack, notable choices, Astra's main findings, Fable's readability |
| 232 | notes, a link to the live app and a link to its source. |
| 233 | - A "How we measured" section and a "What this does and doesn't show" |
| 234 | section, including the network caveat, Spinel's single-threaded server |
| 235 | and any app that's blocked. |
| 236 | - Built from the metrics files by a script, so a re-run regenerates it. |
| 237 | - Report copy goes through the copy-review gate before it ships. |
| 238 | |
| 239 | ## Known risks |
| 240 | |
| 241 | - **Everything targets `linux-arm64`, and docs-t1 is x86-64.** Rust, Zig, |
| 242 | Go and Swift (static Linux SDK) cross-compile well, and so does Spinel |
| 243 | (lpay's zig recipe). OCaml, Haskell and Mojo are harder. Each tries |
| 244 | docs-t1 first, by cross-compiling or in an emulated arm64 container, |
| 245 | and falls back to tito-studio. If neither works within the effort |
| 246 | limit, that app is reported as blocked. |
| 247 | - **Mojo** may need runtime files beside the binary. The binary kind takes |
| 248 | a single file. If Mojo can't produce one, we file a platform request and |
| 249 | report Mojo as blocked. |
| 250 | - **Spinel** has no threads. lpay's notes flag its aarch64 fiber context |
| 251 | switch as untested at runtime. Whatever concurrency model the Spinel |
| 252 | server ends up with, the report states it next to the load numbers. |
| 253 | - **Load testing shared boxes** can affect other people's apps on the same |
| 254 | box. We keep load runs short, one app at a time. |
| 255 | |
| 256 | ## Effort limit |
| 257 | |
| 258 | Every builder gets the same limit. If an app still doesn't have a local |
| 259 | build passing the behaviour tests after **4 hours of wall-clock time**, |
| 260 | the builder stops and writes down where it got stuck, and the app is |
| 261 | reported as blocked. All tokens a builder spends are reported, including |
| 262 | for blocked apps and time spent getting toolchains working. |
| 263 | |
| 264 | ## Left to the builders |
| 265 | |
| 266 | - Each app's libraries, framework, server, templating and SQLite binding. |
| 267 | - How the JS talks to the server (HTML fragments or small JSON). |
| 268 | - Internal code layout, within the repo shape above. |
| 269 | |
| 270 | ## Calls made by the orchestrator, not by Paul |
| 271 | |
| 272 | - Real paths instead of `#/` routes, and per-visitor lists by anonymous |
| 273 | cookie. |
| 274 | - Fable for blind readability scoring, separate from Astra's review. |
| 275 | - One review-and-fix round per app. |
| 276 | - All apps in one repo under `apps/<slug>/`. |
| 277 | - App names `todo-<slug>`, and the report app named `bakeoff`. |
| 278 | - The shared test suite and load test are written once by the |
| 279 | orchestrator, not by each builder. |
| 280 | - The effort limit: 4 hours of wall-clock time per builder. |
| 281 | - Build time for apps that can't cross-compile on docs-t1 is measured |
| 282 | where they're built, with the machine named, and isn't ranked against |
| 283 | the others. |
| 284 | |
amadan