[roadmap] Stage 9: Endgame: the browser and every dependency #4858

A goal of the ROADMAP tab (app.t27.ai/queen, key m): the whole stack rewritten in .t27, generated to its target (trios CLAUDE.md, law L0). Part of #1219.

Stage 9: Endgame: the browser and every dependency

Stage 2 rewrites only the part of BrowserOS that is deployed as api.t27.ai. This goal covers everything the stack runs that is still in another language once stages 1-8 are done:

• the browser itself;
• every third-party library.

It is the top row of the comb and opens after stage 8.

• Repositories: gHashTag/BrowserOS (the whole repository, dev), upstream Chromium with the repositories its DEPS pins, and the third-party packages locked by each deployed service.
• Languages to rewrite: C++, C, Java, Objective-C, TypeScript, JavaScript, Go, Rust (the dependencies).
• Measured size today: 1,341.7 MB for Chromium src + V8 + Skia (below). That is seven times the 192.2 MB the stack count holds for everything else.
• Target: t27c gen-rust, gen-c.

What is measured (2026-09-27)

All sizes below are counted by the rules in trinity apps/website/scripts/roadmap-stack.mjs: vendored trees, build/, third_party/ and lockfiles are skipped. Config, Docker and Make files are left out of the totals.

BrowserOS dev at 0934096 (2026-09-20): 5.48 MB of source.

• Most of the agent is already stage 2.
  • 1,067 of the 1,354 files in packages/browseros-agent have the same path in trios/agent-server on fix/queen-worker-provider-and-prompt-size.
  • 1,007 of them are byte-identical.
• What BrowserOS itself owns of the browser is the 343 files under chromium_patches/.

The browser: Chromium 146.0.7680.31. packages/browseros/CHROMIUM_VERSION pins this version, and BrowserOS's BASE_COMMIT 4d3225104176d is that tag's commit. The engine was read from the GitHub mirrors (github.com/chromium/chromium, v8/v8 and google/skia) at the commits Chromium pins:

• Where Chromium's bytes are, by the rules:
  • chrome/ 406 MB;
  • components/ 187 MB;
  • content/ 98 MB;
  • ui/ 63 MB;
  • ash/ 62 MB;
  • ios/ 56 MB;
  • net/ 36 MB;
  • media/ 34 MB.
• Tests are a large share. First-party test paths are 481.2 MB of Chromium's 1,878.6 MB, and 74.9 MB of V8's 147.1 MB.
• Blink is not in the 1,341.7 MB. It is Chromium's own engine, but it lives under third_party/blink, which the rules skip.
  • 132.4 MB outside its tests (C++ 97.3, C 28.3);
  • 282.0 MB of web tests.
• Chromium src pins 260 other repositories as DEPS (gitlinks in its tree). Only V8 and Skia are measured.
  • Roughly 60 of them are Google-internal or branding data, going by their paths (internal, clank, signing_keys, google_chrome). They are not fetched by a public checkout.
• Not counted anywhere: assets with no language. The largest are PNG 552 MB, .xtb translations 379 MB and .tflite models 159 MB.

Third-party packages, counted from the lockfiles of what runs today. These are package counts. Their bytes are not measured.

What has to be decided before it is worked

1. The dependency count is not fixed. It is what stages 2-7 leave behind.
   • Porting api.t27.ai to .t27 → gen-rust removes its 2,099 npm packages and adds whatever crates the generated Rust needs.
   • So this goal's list is re-measured after each stage, not rewritten from today's list.
2. The seed. L0 names t27c as a hand-written Rust exception, so its 239 bootstrap crates are the seed's dependencies. Do they fall under the seed exception, or are they endgame work?
3. The browser engine. t27c has gen-rust, gen-c and gen-verilog. It has no C++ target and no interface target (see [roadmap] Stage 8: Interfaces: the open question #4550). The measured engine is 1,341.7 MB, most of it C++, so there are two ways forward:
   • a target for C++ interop, so the engine can be rewritten component by component;
   • a documented exception for upstream Chromium, which keeps only the 343 patched files as ours.
4. What is still unmeasured:
   • the other 258 DEPS repositories;
   • the bytes of the npm, crate and Go dependency set.

How this goal is worked

This issue is a GOAL, not a bee task. It has no Boundary on purpose, so the Queen does not dispatch it.

• The work is filed as one issue per file or module. Each has its own Boundary, Given/When/Then scenario and measurable requirements, in the shape of the existing Port tools/<file>.py ... to specs/port/... issues.
• The roadmap feeder (tools/queen/feed_roadmap.py, feat(queen): feed the swarm from the roadmap, with a brief that says what the game is #4855) indexes this stage as refused. It files nothing until stage 8 is done and the questions above are answered.
• The ROADMAP tab reads the 1,341.7 MB from goals.json measured. That is the endgame boss's HP.
• This goal closes when either condition holds:
  • the lockfiles of the deployed services list only packages generated from .t27, and the browser's languages are gone from the count;
  • what remains is an explicit, documented exception.