The rest of this suite maps the territory — the eight-step flow, the edge and agent architecture, the real GCP pipeline, the generative-AI touchpoints, the UI bets. This page sequences the crossing. Five phases, each with the dependencies that order them, exit criteria you can measure, and the decisions that gate them. The reader who finishes here should know exactly what to do next, and in what order. The phases and decisions are CONCEPT · proposed; everything they build on is grounded in the pages they cite. The audience includes external Game Builders — terms are glossed on first use across the suite.
The plan is not a feature list — it is a dependency graph with gates. P0 · Decide unblocks everything by recording the choices the build pages keep deferring. P1 · Foundations makes a spec scaffold to a previewable build with no hand edits. P2 · Reskin and P3 · Tune then take engineers out of the two highest-volume paths, in parallel. P4 moves the sandbox to the real edge and opens the external portal. The diagram places the phases on a timeline with their hard dependencies; the critical path P0→P1→P2 is drawn in flame.
Read the flame edges as the critical path: nothing in P2 can start until P1 ships scaffold v2 and template extraction, and nothing in P1 starts until P0 records the UI-treatment and import decisions it depends on. P3 needs only P1's simulation-harness foundations, so it overlaps P2. P4 needs both P2 and P3 plus the builder-auth decision from P0 — which is why that one P0 decision has a long edge all the way to the end.
Depends on: nothing — this is the unblocker. Everything downstream waits on it.
Four decisions are blocking the build, and all are framed elsewhere in the suite but not yet made. P0 is the cheapest, highest-leverage phase: it is conversation and a written record, not code.
ownerPrincipal) is already in the data model, but the auth choice gates the external portal in P4. architecture · two audiences, one surface.Depends on: P0's UI-treatment and import decisions (the surfaces these tools feed). Unblocks: P2 (needs scaffold v2 + extraction) and P3 (needs the sim-harness foundations laid here).
The plumbing phase. It closes the manual gaps traced in variant → live · gaps and makes the Scaffolder agent buildable on a real CLI rather than a hypothetical one.
GameTemplates from each engine DEFAULT_CONFIG at build time. Opens all 15 engine games to the Builder (not three) and kills template drift at the source. build-changes #2.sync-games.ts, patches the CI GAMES input, tsc --noEmits the variant, validates the registry entry, writes the action to audit_log, and exposes itself over MCP. build-changes #3.versions[] / baseGameRef on the grown spec record the PR and Worker, so provenance lives in the dashboard, not only in variant.meta.json. architecture · data model.sync-games.ts or hand-edit the CI GAMES arg, P1 is not done.Depends on: P1's scaffold v2 + template extraction. On the critical path — this is the first phase that delivers the zero-engineering promise to a real path.
The keystone phase. The single highest-leverage change in the whole plan — the shared client shell — finally makes the captured theme read instead of dead data in variant.meta.json.
assets.ts (generative-ai · map).Depends on: P1's simulation-harness foundations. Overlaps P2 — different people, no shared keystone, so the two run in parallel.
Where tuning earns its proof. The Math-Sim agent's tool turns "reviewers eyeball numbers" into "reviewers approve a number a simulation established."
Depends on: P2 + P3 (the agent workforce and proof loop must exist) and the builder-auth decision from P0. The longest phase, and the one the whole vision points at.
Where the edge target stops being a beachhead and the second audience arrives. Incremental, never a big bang.
trunk-raider-edge, edge-engine, edge-do, edge-game-sdk into the Workers-for-Platforms namespaces game by game. The migration is credible because the experiments already run the target runtime — see variant → live · today vs target.ownerPrincipal so a builder sees only their own games and no revenue or pipeline surfaces.A roadmap is only as honest as its open questions. Some choices the suite has already decided and built on; others are pending and genuinely block P0's exit. The pending rows below are not rhetorical — four of them carry a "decide by P0" hint precisely because P1 cannot start until they are recorded. Marketplace is marked an open direction: it is documented, not scheduled.
| Decision | Where it is recorded / framed | Status / decide by |
|---|---|---|
| Decided — built on across the suite | ||
| Workers for Platforms is the target runtime | Specified at architecture · edge delivery; today-vs-target contrast at variant → live. | decided |
| Builder Validation folded into Builder Studio | The validation surface is step 5 of the one flow — ui-variants · validation; not a separate product. | decided |
| Mocks carry honesty badges | The honesty rule on the index: every mock is GROUNDED or CONCEPT. | decided |
| Proposal cadence — one step per proposal | The seven-part proposal shape on the index; flow/01 and flow/04 are the exemplars. | decided |
| Pending — these block P0 exit | ||
| UI treatments per surface | Framed with recommendations and cheapest validation at ui-variants · deciding. | pending · decide by P0 |
| Builder auth model | Two options weighed at architecture · two audiences; restated as an open question at open questions. | pending · decide by P0 |
| Import — PDF info sheet up front? | Framed in flow/01 open questions; the import benchmark (0:25-to-spec) assumes a PDF was supplied. | pending · decide by P0 |
| Mechanic-prototype surfacing | Framed in flow/01 open questions — sixth card vs New-game variant vs advanced toggle. | pending · decide by P0 |
| Marketplace direction | Documented as an open direction at architecture · open questions (mechanics as licensable products) and the mind map's Marketplace branch. | open direction · no commitment |
The four "decide by P0" rows are the literal content of P0's exit criteria. They are not blockers on the architecture (the suite is careful to keep the rest un-gated on them), but they are blockers on starting to build: you cannot write scaffold v2's audited-MCP surface without knowing which validation treatment it reports into, and you cannot finish the data model's portal index without the auth choice.
The north star is "little to no engineering per game." That is only real if it is instrumented. This is the panel — and the honest starting point is that none of it is measurable today because there is no telemetry on the build pipeline. The benchmarks below come from the worked examples (a wall-clock trace), not from a live dashboard; the job of P1 is to make them live.
| Measure | Target / benchmark | Today |
|---|---|---|
| Time-to-playable per path | Reskin 2:05 · Tune 3:10 (math gate dominated) · Import 0:25-to-spec — from the worked examples. | Not measurable — no build telemetry exists. |
| Engineering hours per game | 0 for reskin and tune (the P2 and P3 exit bars). | Hours per reskin (hand-edited assets.ts + Tailwind + art); not tracked. |
| Builds started per week | Trending up as the Builder opens to all 15 games and external builders. | Not tracked. |
| Gate first-pass rate | Per gate — rising as agents and proof loops mature. | Not tracked; gates are not yet wired as runs. |
| Agent autonomy promotions | Per touchpoint, up the autonomy ladder (assistive → autonomous·gated → autonomous), earned on track record. | N/A — no agents wired; every touchpoint starts assistive. |
| Sandbox sessions per build | A signal of iteration depth before handoff. | Not tracked — Launch Lab is mocked, not wired. |
Instrument these in the dashboard from P1, so every later phase proves itself with numbers, not vibes — the P2 and P3 exit bars (hours = 0, a-day-with-proof) are only credible if the panel above is live to read them off.
The immediate actions — most of them P0, one of them the first piece of P1 you can start in parallel because it depends on no decision.
origin/feat/builder-studio-path-picker-mock and origin/feat/builder-studio-readiness-grounded a disposition once the UI treatments are chosen, so neither lingers as an ambiguous half-state.That is the whole crossing: the suite mapped the territory, this page sequenced it, and the cadence keeps it moving — one decision recorded, one proposal elevated, one phase exited at a time.