impuestospy/DECISIONS.md

DECISIONS

One entry per decision that is not already obvious from the specs. Newest phase last.

Markers used in the code:

• // SPEC-GAP: the specs did not settle this and a choice was made here.
• // TODO-TAX-VERIFY: a tax rule that RULES.md does not state. Never invented, always flagged.

---

Blocking gaps in the source material

RULES.md is missing

docs/ ships SPEC.md, FLOWS.md, COPY.md and CONTRACTS.md. RULES.md, which the prompt names as authoritative for every tax rule, is not present. Phase 1 is entirely RULES.md and phases 4 and 5 depend on it.

Consequence for phase 0: packages/rules exists and exports only RULES_VERSION. Nothing in this phase computes a tax number, a check digit or a deadline, so nothing was invented. The seed deliberately stops short of profiles for the same reason: CONTRACTS.md section 4 asks for RUC base 4123456 "with computed DV" and a deadlineDigit, both of which are RULES.md algorithms. Seeding accounts only keeps phase 0 honest.

Needed before phase 1 starts.

boneyard and canvas-ui are not the packages the prompt means

Both names resolve on npm to unrelated projects: boneyard@0.1.4 is a 2015 Backbone "architectural toolkit", canvas-ui@0.2.3 is a Mesosphere Bootstrap theme. Neither does skeleton loading or canvas effects. Neither is needed before phase 7.

Plan unless corrected: keep the behaviour the prompt specifies (skeletons on every content load and never a spinner; canvas effects in exactly two places, degrading gracefully) behind a single <Skeleton name> component and a single effect component, so swapping in the real library later is a one file change.

Please confirm the intended packages before phase 7.

Neither named skill is installed

ponytail and ui-ux-pro-max-skill are not available in this environment. Their stated intent was applied by hand: nothing speculative, no unused configuration, and FLOWS.md section 1 as the design constraint.

---

Phase 0

Migrations: better-auth generates its own four tables, we own the rest

src/db/migrator.ts runs two ordered steps: better-auth's getMigrations() creates and updates user, session, account and verification, then the Kysely migrator applies src/db/migrations. Delegating the auth tables keeps them in step with the installed better-auth version and emits correct DDL for both dialects, so no hand written dialect SQL was needed for them. Upgrading better-auth in a way that adds a column means adding a migration that calls the same generator again.

The whole schema ships in migration 001_core, not phase by phase

Every table in SPEC.md section 5 is created now. The schema is fully specified and stable; splitting it across phases would produce a pile of migration files and no benefit before release. Later phases add modules on top, not tables.

Timestamps and money are portable by construction

Timestamps are ISO-8601 text and dates are YYYY-MM-DD text in both dialects: they sort chronologically as strings, so no dialect specific date type or comparison is needed anywhere. Money is bigint, because guaranies pass int4 at about Gs. 2.100.000.000, and apps/api/src/db/postgres.ts registers an int8 parser that returns a number and throws outside the safe integer range.

/api is proxied by a route handler, not a Next rewrite

SPEC-GAP against SPEC.md section 2, which specifies Next rewrites. Next bakes rewrite destinations into the build manifest, so API_INTERNAL_URL would become a build time value and one image could not serve both compose and k8s. apps/web/app/api/[...path]/route.ts forwards at request time instead. The single origin model is unchanged: the browser only ever sees the web origin, cookies stay first party, and there is still no CORS anywhere.

Workspace packages ship TypeScript source with no build step

packages/* have no dist. The web app lists them in transpilePackages and tsup bundles them into the API. Their relative imports are extensionless, because Turbopack does not rewrite a .js specifier onto a .ts source file.

The es catalog is split from the strings COPY.md does not define

catalogs/es.ts is generated from COPY.md and is verbatim; copy-parity.test.ts re-derives it from docs/COPY.md on every run and fails on any drift, in either direction. Strings the product needs that COPY.md does not list (the seven error envelope messages, three sign in labels, the language switcher) live in catalogs/es.extra.ts and follow the tone rules in COPY.md section 0. es is the merge of the two.

decl.approve is both a message and a namespace

COPY.md defines decl.approve (the button) alongside decl.approve.confirmTitle. A nested message tree cannot hold both, and next-intl walks a nested tree. unflatten moves such a message to a reserved _ child and resolveKey maps the key for callers, so components still address messages by their COPY.md key. apps/web/src/i18n/t.ts is the wrapper; it is computed from the catalog, so a future collision is handled without another change.

Locale negotiation on /

next-intl's default detection is left on: a browser asking for English lands on /en, anything else falls back to es. The en catalog exists for expats and international users (COPY.md section 0-EN), which is exactly the population whose browser is in English. An explicit choice through the switcher always wins and is in the URL.

SQLite refuses JOBS_INLINE=false

Implemented literally as SPEC.md section 15 instructs, which is narrower than the mode matrix in the same section: that table allows "SQLite, 1 dedicated worker" under Split small. Two pollers cannot be made safe against a single writer, so the boot check wins and SQLite stays single process. Worth reconciling in SPEC.md.

Deferred to the phase that needs them, deliberately

• RateLimiter: SPEC.md section 6 specifies a token bucket, but the first endpoint with a stated limit is GET /lookup/ruc/:number in phase 2. better-auth's own rate limiting covers the auth routes until then.
• DTO schemas in packages/contracts: enums, the error envelope, the client and ProfileDto exist because phase 0 uses them. The rest arrive with their endpoints.
• Storage in /readyz: the check covers the database and pending migrations. The storage driver probe is added in phase 3 with the driver.

Smaller choices

• TypeScript 5.9, not 7.x: typescript-eslint@8 declares typescript <6.1.0.
• better-sqlite3 is kept out of onlyBuiltDependencies: it ships prebuilt binaries, so letting pnpm run the implicit node-gyp rebuild would compile it for nothing and force a toolchain into the image.
• The language switcher is a native <select>: one dependency fewer than a popover, and the better mobile and keyboard experience for a two item choice.
• apps/web/proxy.ts, not middleware.ts: Next 16 deprecates the middleware convention.
• The dark mode palette and the dark variant are wired now; the toggle itself is phase 7 polish. System preference works today.
• audit_log is append only by construction: the module exposes no update or delete. A database trigger was written and then removed, because it would have been the only piece of dialect specific SQL outside the two files SPEC.md section 5 allows.

Legal pages

/legal/privacidad and /legal/terminos do not exist yet (they belong to phase 2's marketing routes). Per COPY.md section 13 they will ship with clearly marked placeholder content and no generated legal text. TODO: human written before launch.