Port upstream 0.66.0: OpenAI per-day usage history, core and bridge (stacked on #645)

[Finesssee]

Summary

Part (a) of the OpenAI per-day usage chart (upstream 0.66.0): the OpenAI Admin usage fetch now also returns a per-UTC-day history (cost, requests, input/cached/output/total tokens, line items, models). The history travels as ProviderFetchResult.open_ai_api_usage, reaches the frontend as the openAiApiUsage bridge field, and is typed in types/bridge.ts. No UI in this PR; the chart is part (b).

• rust/src/core/usage_snapshot.rs: OpenAiApiUsageHistory, OpenAiApiDailyUsage, OpenAiApiLineItemCost, OpenAiApiModelUsage; ProviderFetchResult::with_open_ai_api_usage.
• rust/src/providers/openaiapi/history.rs (new): per-day bucketing keyed by start_time. result_from_admin_usage now derives its totals, top models and top line items from these days, so there is one parsing path.
• apps/desktop-tauri/src-tauri/src/commands/bridge/openai_usage.rs (new): camelCase DTOs (openAiApiUsage, epoch seconds) and From conversions; ProviderUsageSnapshot.open_ai_api_usage (omitted when absent).
• apps/desktop-tauri/src/types/bridge.ts: OpenAiApiUsageSnapshot and friends, ProviderUsageSnapshot.openAiApiUsage.
• The balance fallback (billing-api) has no history, as designed.

Approved design

Design note: design-openai-per-day-chart.md (approved; open questions resolved by the note's recommendations). Summary:

• Upstream builds card.openAIAPIUsage = {historyDays, projectID|null, daily[]} beside the cost summary. Each UTC-day bucket holds startTime/endTime, costUSD, requests, inputTokens (input + input_audio), cachedInputTokens (separate, a subset of input), outputTokens (output + output_audio), totalTokens, lineItems[] (desc by cost, then name) and models[] (desc by tokens, then name).
• Windows: provider id openaiapi, Admin API path only. Reuses the two Admin calls already made (/v1/organization/costs grouped by line_item, /v1/organization/usage/completions grouped by model, bucket_width=1d); no new endpoint, credential or dependency.
• Model: OpenAiApiUsageHistory in usage_snapshot.rs, attached to ProviderFetchResult; bridge DTO openAiApiUsage (camelCase, epoch seconds) mirrored in types/bridge.ts. Upstream bounds apply (366 days, 10000 line-item + model entries).
• Later UI (part b, not here): collapsed "Daily usage" section in the tray/pop-out MenuCard and a Settings > Providers ChartsSection tab, reusing components/charts/BarChart, Cost | Tokens toggle, one bar per UTC day, keyboard-navigable day selection, detail panel per day, empty state via DetailChartEmpty, 13 new OpenAIChart* locale keys across the 8 locales.
• Open questions taken as recommended: window fixed at 30 days until an OPENAI_HISTORY_DAYS setting exists (the payload carries historyDays); USD only, no currency conversion; project id masking under hide-personal-info is a UI concern for part (b); the openaiapi id stays out of PROVIDER_CHART_DATA_IDS.

Upstream reference

• Release 0.66.0 (OpenAI per-day usage chart). All reads pinned to tag v0.66.0.
• Sources/CodexBarCore/Resources/Plugins/openai.js (daily map, bucket(), card.openAIAPIUsage).
• Sources/CodexBarCore/Providers/OpenAI/OpenAIAPIProviderDescriptor.swift (mapPluginCard bounds and validation).
• Sources/CodexBarCore/Providers/OpenAI/OpenAIAPIUsageSnapshot.swift (bucket shape, sort orders).
• Tests/CodexBarTests/OpenAIAPIUsageFetcherTests.swift (parses admin costs and completions usage into daily summaries; its wire pages are the fixture for the new tests).

Ported / Deferred

Ported: per-day bucketing with upstream semantics (buckets keyed by start_time, first end_time wins, audio tokens folded into input/output, cached kept separate, days after now dropped, newest historyDays kept, upstream sort orders, blank names fall back to API / Responses and Chat Completions, negative or beyond-2^53 counts are parse failures like upstream integer()).

One deliberate deviation: upstream fails the whole fetch when a card bound is broken (a bucket whose end is not after its start, or more than 10000 line-item + model rows). Here the history is dropped with a tracing::warn! and the spend summary still shows, because the chart is auxiliary. With the fixed 30-day window the 366-day bound cannot be reached.

Deferred (per the design): all UI (part b), locale keys, OPENAI_HISTORY_DAYS, currency conversion, proof-harness variant for seeding an OpenAI API snapshot (CODEXBAR_SEED_USAGE_JSON currently seeds only Codex). The design note's risk about payload size (up to 366 days through provider_cache and events) is not an issue at 30 days; revisit if the window becomes configurable.

Behavior change to existing summary: totals, top models and top line items now come from the per-day buckets, so a bucket that starts after now no longer counts (upstream filter(start <= now)), and equal-cost line items tie-break by name.

Validation

Toolchain cargo +1.98.0, E-core wrappers, slot-5 target dir.

• cargo fmt --all: clean.
• cargo clippy --workspace --all-targets -- -D warnings: pass (both manifests).
• cargo test -p codexbar openaiapi: 40 passed, 0 failed (11 new tests: 10 in history_tests.rs, 1 wire-page end-to-end in tests.rs; plus a no-history assertion on the balance-fallback test).
• cargo test -p codexbar (full, shared core touched): 2194 passed, 0 failed, 1 ignored.
• cargo test -p codexbar-desktop-tauri (full): 465 passed, 1 failed. The failure is commands::tests::bootstrap_payload_exposes_every_provider_variant (catalog has 79 entries, 78 active providers); this PR does not touch the provider catalog, and the test reads this machine's settings, so it looks environment-dependent. Not confirmed against the base branch.
  • New bridge tests in openai_usage_tests.rs (4) pass: camelCase epoch-second payload, omitted when absent, round trip, empty history.
• pnpm exec tsc --noEmit: clean. pnpm run lint: only pre-existing warnings. vitest run src/components/MenuCard.test.tsx: 31 passed.

File sizes: no file crosses 1000 lines. mod.rs 783 -> 748, usage_snapshot.rs 819 -> 881; bridge.rs, commands/tests.rs and types/bridge.ts were already over 1000 (bridge.rs and tests.rs only gain the one-line open_ai_api_usage: None field in literals; bridge.ts gains ~45 lines of type declarations per the design).

Affected areas

• Rust backend (rust/src/core, rust/src/providers/openaiapi)
• Tauri shell bridge (commands/bridge)
• Frontend bridge types only (types/bridge.ts)
• Settings, tray, float bar, or any rendered surface
• New dependencies or tooling

UI proof

Not applicable: no rendered surface changes in this PR. CUA proof belongs to part (b).

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 1f65ff44-074e-4b7c-b30e-e185909534aa

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Finesssee]

Reviewed by Codex gpt-6-luna (xhigh); verified and validated by Claude

Thermo-nuclear review of #699. Part (a) is complete against the approved design note (Admin API history, typed bridge, balance-fallback behavior, backend and bridge tests). Chart UI and locale keys belong to part (b).

• P2 rust/src/core/usage_snapshot.rs:671: populated history was serialized with ProviderFetchResult although documented as transient. Fixed: #[serde(skip)] plus a regression test.
• P2 rust/src/providers/openaiapi/history.rs:186: individually valid counts could aggregate past JavaScript's exact integer range and reach the bridge as rounded numbers. Fixed: checked aggregation, unsafe chart history omitted while spend summaries are kept, plus a regression test.

[Finesssee]

Fixes for both findings landed in commit "Address thermo review". No items left open.

Commands run: cargo +1.98.0 fmt --all; cargo +1.98.0 clippy --all-targets -D warnings on rust and apps/desktop-tauri/src-tauri (clean); cargo +1.98.0 test --lib filtered to openai/open_ai/usage_snapshot (88 passed). No frontend change, so vitest not run.

[Finesssee]

CI failure diagnosed and fixed at the new head f328c66 (was 5d99354):

Diagnosis. The failing pr-check (2026-09-30T10:02Z, 10m9s, CircleCI) could not be reproduced in a full local CI-slice run of the recorded head 5d99354 — every step green with the CI-pinned toolchain (rust 1.98.0 from scripts/circleci-pinned-rust.txt): fmt ✓, clippy --workspace (both crates) 0 errors ✓, cargo test codexbar 2196/0 fail/1 ign ✓ (openai focused 75/75), vitest 402/0 ✓, lint 0 errors ✓, check:anti-slop ✓, test:anti-slop ✓, build ✓, interaction-guard tests ✓, circleci-pr.tests.ps1 ✓. The only local failure is the documented #684 bootstrap env baseline, which passes on clean CI hosts. The recorded head's tests are hermetic (fixed clocks), and the run's start time rules out the 00:00-01:00 UTC window… but the flaky cost_scanner fixtures (CI builds 977/978/979: codex_source_recovery_keeps_appended_duplicate_unpriced_after_cache_reload + both paginated fork tests) are the only known CI-only test failures on this codebase in that period, and this branch's base predates the fix (#721, fix/cost-scanner-midnight-fixtures at 400f805, delivered on main on 2026-10-01 — after this CI run). Any CI run of this branch could flake the same way while it stays unfixed in the branch's history.

Fix. Cherry-picked 400f805 ("Keep Codex cost fixtures on the local day") onto this branch as f328c66 — test-fixture-only change (rust/src/cost_scanner/tests.rs + tests/paginated.rs), no production code, merge-tree verified conflict-free against the recorded head. This is the same flake fix already delivered for main; it does not alter the PR's behavior.

Validation at f328c66 (with the pinned 1.98.0 toolchain):

• cargo fmt --all --check — clean.
• cargo test -p codexbar — 2197 passed / 0 failed / 1 ignored (cost_scanner focused 90/90).
• cargo test -p codexbar-desktop-tauri — 465 passed / 1 failed = the documented Isolate bootstrap payload test from real settings #684 env baseline (dev machine only; hermetic fix Make the bootstrap catalog test hermetic (#684) #711 not on this base).
• cargo clippy both crates, --all-targets -- -D warnings — 0 errors.

CI re-check. Hosted pr-check at f328c66: pass (13m54s, workflow 33fc20fe). Fast-forward push 5d99354..f328c66; ls-remote verified.

Stack note: #701 (chart UI) is stacked on this branch and already carries its own validation; the cherry-pick is behind #701's head (813458d contains 5d99354), so #701 should merge the new head in — its CI will then include the flake fix too.

[Finesssee]

Shipped in v0.70.0: this PR's head is included in main via #735 (merge commit 9d0a37a). Closing as integrated.