Repository Instructions

Effect Error Style

• Prefer native Effect errors over plain JavaScript Error subclasses in Effect code.
• Use Data.TaggedError for internal typed errors that stay inside the Effect error channel.
• Use Schema.TaggedError when an error is part of a public schema or wire contract.
• Avoid throw for expected domain failures; return Effect.fail(...) with a typed error instead.

Export Style

• Prefer local declarations first and grouped exports at the end of each authored source file.
• Use export { ... } for runtime values and export type { ... } for type-only exports.
• Put default exports at the end too, using a named local value before export default name;.
• Do not hand-edit generated files just to satisfy this style rule.

Conventions

• Daily usage rows are keyed (deviceId, date, source, model) and upserted; sync must stay idempotent.
• Cost is frozen at first upload unless the usage (token counts) changes: the usage_days upsert keeps the stored cost_usd when all token columns match (ccusage re-prices unmarked Codex history with the device's current speed tier), takes the incoming cost when any token count differs, and prices a row stored at 0. Fixing already-stored cost means editing D1 directly; re-ingesting payloads with the same tokens is a no-op for cost.
• Only a parser change may lower a re-sent day. Each raw-ingested usage_days row records counted_by, the server PARSER_VERSION that produced it. A re-sent (device, day, source) is kept when every stored row has the current counted_by and the uploading owner, and the stored total is higher than the re-count. Kept means never lowered or pruned, though models that are new or grew are still taken. A lower re-count from the same parser means the device lost logs (deleted or pruned, a narrower source root, a full re-sync past Claude Code's 30-day cleanup), not a correction. The decision is per whole day, so a model relabeled without a loss is still replaced and pruned. Any other re-send replaces the day as before. So a correction that has to lower or drop stored usage ships as a PARSER_VERSION bump, also when it comes from a ccusage or CLI-side counting fix. A bump is required for any server change to the rows a report yields (parsing, model labels, plausibility limits). A bump re-counts every device's trailing window once, losses included. The structured /usage/sync path writes NULL, which is always replaced. Known gaps: a timezone change moves usage near midnight into the next day while the earlier day keeps its higher count, and two concurrent ingests from one device can race the check.
• date columns are opaque YYYY-MM-DD strings (ccusage local-time buckets); never parse them into Date objects for bucketing. Do day arithmetic with shiftDayKey/utcDayKey from @tokenmaxxing/api-contract (API windows live in apps/api/src/date-keys.ts).
• Published CLIs bundle a frozen contract. packages/api-contract/fixtures/ (recorded CLI requests + the CLI endpoint OpenAPI slice) is a compatibility contract: review diffs there as breaking-change reviews, and add a legacy/ fixture before changing a shape a released CLI sends. Responses go the other way: apps/api/src/http/cli-compat.test.ts decodes every replayed response and typed CLI error with frozen copies of the released CLI decoders, so fields may be added but never dropped, renamed or retyped, and error _tags never change. fixtures/full-api.openapi.json snapshots the whole API as a review aid; only the CLI slice is frozen.
• CLI tokens (tmx_ prefix) never expire; revocation (revokedAt) is the only kill switch.
• D1 caps a statement at 100 bound parameters, and the sqlite test harness (apps/api/src/testing/sqlite-d1.ts) enforces it. Never bind an unbounded list: chunk it, or pass it as one JSON parameter through json_each.
• Keep bun run test output clean. A test that exercises a logging path captures and asserts the log (makeTestLogger in apps/api/src/testing/logger.ts; vi.spyOn(console, ...) outside Effect) instead of letting it print. makeTestApp already captures into app.logs. Captured Effect logs are replayed when a test fails.

Reference Repos

• .repos/ holds upstream source as git submodules, for reading how a dependency actually behaves. It is read-only: never edit, import from, or build it (lint, format and tests already ignore it).
• .repos/effect-smol and .repos/alchemy-effect are pinned to the release tags matching the catalog versions in the root package.json (effect@X and vX tags). When bumping effect or alchemy in the catalog, move the submodule to the matching tag in the same PR: git -C .repos/effect-smol fetch --tags && git -C .repos/effect-smol checkout effect@X, then stage .repos/effect-smol. Don't use git submodule update --remote.
• .repos/opencode is not a dependency; it tracks upstream dev as a reference for OpenCode's data and CLI patterns.
• Fetch them with git submodule update --init .repos/<name>. If a checkout is missing, stale or disagrees with the installed version, node_modules is the source of truth.