Placing product frontend code

A product's UI belongs in products/<name>/frontend/, not frontend/src/scenes/<name>/.

About 18 products still have UI in both trees, so "there's already a folder in scenes/" is not evidence that a new file belongs there. Check before you add.

# Where does one directory stand?
python3 .agents/skills/placing-product-frontend-code/scripts/scene_product_split.py data-warehouse

# Every scene dir with a product counterpart, most migrated first
python3 .agents/skills/placing-product-frontend-code/scripts/scene_product_split.py

The script is advisory and read-only — no baseline, no exit code to satisfy. It counts hand-written .ts/.tsx on each side, skipping generated/ (orval writes those, so they are not migration progress).

Deciding where a file goes

Situation	Where it goes
products/<name>/frontend/ holds as much or more than scenes/<name>/	products/<name>/frontend/ — the scenes copy is a remnant
The move is under way but early	products/<name>/frontend/, and move the files the new code touches if that is cheap
products/<name>/ exists, no scenes/<name>/ yet	products/<name>/frontend/ — never create the scenes directory
No product directory at all, and the feature is product-shaped	Bootstrap the product: bin/hogli product:bootstrap <name>, see products/README.md
App-level scene — settings, onboarding, billing, max, error-tracking shells	frontend/src/scenes/ is correct; these have nowhere else to go

The last row is the reason this is a skill and not a lint rule: deciding whether a new directory is a product or an app-level scene is a judgment call, and a check script that guesses gets it wrong on onboarding and settings.

Migrating an existing scene wholesale is welcome and is the point of the convention. Expect that PR to report every merge-queue target once, because scenes register in products/<name>/manifest.tsx — a one-time cost per migration.

Why the directory is a real boundary

.github/scripts/trunk-impacted-targets.js assigns merge-queue lanes by path.

• A change anywhere under frontend/ reports fe:core plus every fe:product:* target, so it serializes against every other frontend PR in the queue.
• A change confined to products/<name>/frontend/ reports one target.

The obvious fix — narrow lanes with a frontend dependency graph instead of paths — was measured and does not work. A static import graph over the frontend (8430 files, 43924 edges) puts 2226 modules in a single strongly connected component, 27% of the graph, spanning 31 products. Every member of an SCC has identical reverse-reachability by definition, so reverse-reachability from any file under frontend/src reaches 73 of 79 products — a leaf tab component five levels into the replay player gives the byte-identical answer to types.ts. Cutting all 498 lib/** → scenes|products back-edges moved it from 6918 to 6909 modules: hundreds of redundant cycles, not one bad edge.

bin/find-affected-stories already concedes the same thing by listing frontend/src/lib/ as a full-run invalidator.

So path is the only signal that discriminates, and moving files is the only thing that narrows a lane. Note what this does and does not buy: merge-queue parallelism, not decoupling — the import graph stays as tangled as it was.

Related

• products/README.md — product layout and setup
• products/architecture.md — DTOs, facades, isolated testing
• frontend/src/AGENTS.md — the rest of the frontend guide