Port upstream 0.66.0: OpenAI Admin pagination, token totals, balance-fallback gate, retry

[Finesssee]

Summary

OpenAI Admin usage (rust/src/providers/openaiapi/mod.rs) now matches upstream v0.66.0 openai.js / native OpenAIAPIUsageFetcher behavior:

• /v1/organization/costs (group_by=line_item) and /v1/organization/usage/completions (group_by=model) use bucket_width=1d, limit <= 31 and UTC-day-aligned 31-day ranges, following has_more / next_page (100-page cap per range; a missing, blank or repeated cursor is a parse error). A busy org is no longer silently truncated.
• Token totals are input + input_audio + output + output_audio. Cached input is a subset of input and is no longer added on top (it was double counted in the Tokens row and model ranking).
• Legacy credit-balance fallback is skipped for a project-scoped Admin key (the endpoint is not project-filtered); unscoped keys and non-Admin keys keep the billing-api label and API balance login method.
• Each Admin GET gets one transient retry (408/429/500/502/503/504, timeout, refused connection); numeric Retry-After is honored up to 10 s, default 1 s. Auth failures and other transport errors are never retried. Request timeout is 20 s as upstream.
• Existing tests moved to a sibling tests.rs (mod.rs stays under 1000 lines).

Upstream reference

• Release 0.66.0 bullet: "OpenAI and Fireworks: use bundled provider plugins while preserving OpenAI usage charts and project labels" (commit a634f55c; native behavior from v0.65.0 OpenAIAPIUsageFetcher.swift).
• Tag-pinned v0.66.0: Sources/CodexBarCore/Resources/Plugins/openai.js (ranges, pages, queryURL, tokens), Sources/CodexBarCore/Providers/OpenAI/OpenAIAPIProviderDescriptor.swift (OpenAIAPIUsageCredential.allowsLegacyBalanceFallback), tests OpenAIAPIUsageFetcherTests, OpenAIAPIProjectScopeTests.

Ported / Deferred

Ported: pagination, UTC day ranges, token totals, balance-fallback gate, transient retry.

Deferred:

• OPENAI_HISTORY_DAYS (1-365): there is no Windows setting for it, so the window stays 30 days and the label stays Last 30 days. usage_ranges already takes the day count, so wiring a setting later is a one-line change. Documented in the module header.
• Per-day usage chart card (card.openAIAPIUsage): separate UI decision, no Windows chart counterpart.
• Deliberate deviation: the final range end is clamped to now (upstream ends at tomorrow's midnight) because the API rejects a future end_time.

Validation

All on the pinned toolchain cargo +1.98.0, E-core wrappers, slot-3.

• cargo +1.98.0 fmt --all: clean.
• cargo +1.98.0 clippy -p codexbar --all-targets -- -D warnings: pass.
• cargo +1.98.0 test -p codexbar openaiapi -- --test-threads=4: 29 passed, 0 failed (6 pre-existing tests kept, 23 new: pagination cursors, missing/repeated cursor, 100-page cap, long-history ranges, token totals, fallback gate matrix, retry policy, retry once then report status, no retry for auth).
• Workspace clippy was not run: the codexbar-desktop-tauri build script fails in slot-3 on a stale tauri permissions path from another slot's cache (environmental). This PR does not touch that crate.

Affected areas

• Provider (rust/src/providers/openaiapi)
• Tauri shell / frontend / settings / tray / float bar
• Docs

UI proof

Not applicable (Tokens/Requests row values change; no UI code touched).

[Finesssee]

Thermo-nuclear review

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

Findings (1):

• Medium, rust/src/providers/openaiapi/mod.rs:287: costs and completions pages are fetched sequentially although independent. Codex suggested tokio::try_join!.

Disposition: not applied. Running both paginated fetches concurrently doubles the request burst against the rate-limited Admin API, which this PR already guards with transient retry; the latency gain is marginal for a background refresh. No other defects found against the 0.66.0 openai-admin-parity spec. No code changes needed; no commit pushed.

[Finesssee]

Validated 2026-10-02 at 44dada5 (PR head):

• Reviewed against the 0.66.0.md PR 13 spec (OpenAI Admin pagination, token totals, balance-fallback gate, transient retry): the diff matches — /v1/organization/costs (group_by=line_item) and /v1/organization/usage/completions (group_by=model) paginate with bucket_width=1d, limit <= 31, UTC-day-aligned 31-day ranges, has_more/next_page cursor following with a 100-page cap and missing/repeated-cursor errors; token totals are input + input_audio + output + output_audio with cached input tracked separately (never added); balance fallback only when the key is not project-scoped Admin (project_id.is_none() || !is_admin), keeping the billing-api label; one transient retry for Admin GETs (408/429/500/502/503/504 + timeout) with Retry-After capped at 10 s, base 1 s, never on 401/403. The prior thermo review's Medium finding (sequential costs/completions fetch) was dispositioned as not applied with rationale — consistent with upstream behavior. No defects found; no changes needed.
• cargo fmt --all --check — clean.
• cargo test -p codexbar — 2183 passed / 0 failed / 1 ignored (openai focused 62/62, incl. pagination-cursor, project-scope, retry, and fallback gates).
• cargo test -p codexbar-desktop-tauri — 461 passed / 1 failed = the documented Isolate bootstrap payload test from real settings #684 bootstrap env baseline (79 vs 78; hermetic fix Make the bootstrap catalog test hermetic (#684) #711 not on this base; also fails on unmodified base).
• cargo clippy --all-targets -- -D warnings — only the three documented pre-existing main-history lint sites (alibabatokenplan/kiro/openai-subscription, rustc 1.96 drift, all outside this diff; QUEUE INTEGRATION FINDINGS).
• No push needed (head 44dada5 already on origin); no fixes required.

Chain note: base for #699 (per-day usage core) and #701 (chart UI). Proceeding to #699's CI diagnosis next.