Commit Graph
521 Commits
Author SHA1 Message Date
ZyyandCodex c0ad2ae656 fix(openai): restrict capacity cooldown to explicit overload
Count only confirmed capacity rejections toward account cooldown, excluding
generic server failures and unrecognized upstream errors. Classify streamed
invalid_prompt failures as invalid requests, matching official Codex.

Cover HTTP and WebSocket failures around the cooldown threshold while
preserving retry boundaries, upstream error payloads, and Smart scoring.
Update the capacity cooldown contract in the API and architecture docs.

Validation: Rustfmt, strict workspace Clippy, and all backend suites with
isolated PostgreSQL 18 / Redis 8 (--test-threads=4).

Reference: openai/codex@7498521d28,
codex-rs/codex-api/src/sse/responses.rs.

Refs #177

Co-authored-by: Codex <noreply@openai.com>
2026-09-20 14:01:29 +08:00
ZyyandCodex 798d4f518e fix(update): run updates in background and retain release notes (#188)
Detach online updates from HTTP request lifetimes and recover their persisted status in the admin UI. Retain current release notes when no upgrade is available while preserving upgrade restrictions.

Fixes #185

Co-authored-by: Codex <noreply@openai.com>
2026-09-20 13:29:48 +08:00
Akide LiuandClaude Fable 5 96cf46cca8 fix(openai): expand string input into a user message item (#179)
The public Responses API accepts `input` as a plain string, but the Codex
backend only accepts an item list and rejects such requests with
400 "Input must be a list". Expand a string `input` into a single user
message item in the downstream body compatibility layer so that HTTP/SSE
and WebSocket share the same rule; arrays and other shapes pass through
untouched.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-09-20 00:58:58 +08:00
ZyyandCodex 890fb3299b fix(openai): stop client retries for exhausted account pools
Return 429 usage_limit_reached when all eligible accounts have exhausted their quota. Preserve provider fallback, continuation recovery, and other unavailability causes.

Cover HTTP and WebSocket error contracts and account selection boundaries. Verified with 2516 backend integration tests, strict Clippy, and the official Codex CLI using gpt-5.6-luna and a quota-exhausted account.

Co-authored-by: Codex <noreply@openai.com>
2026-09-19 23:28:37 +08:00
helenandCodex a18f0a2e11 fix(openai): normalize system messages for Codex (#172)
Normalize explicit system messages to developer at the shared Codex request boundary while preserving message content and other fields. Cover HTTP and WebSocket request bodies and document the compatibility behavior.

Co-authored-by: Codex <noreply@openai.com>
2026-09-19 22:43:49 +08:00
ZyyandCodex a1ba6c0b61 fix(update): enforce release channel upgrade rules
Refs #169

Co-authored-by: Codex <noreply@openai.com>
2026-09-19 22:38:23 +08:00
ZyyandCodex 748c77da6a feat(ui): add key configuration and reorderable table columns
Add session-scoped Codex configuration and CCSwitch import for key holders. Persist table-column ordering with drag animations and refine key, quota, and usage layouts.

Validated with 395 gateway-api integration tests, Rustfmt, and strict Clippy for gateway-api and gateway-admin. Frontend validation is left to the maintainer as requested.

Co-authored-by: Codex <noreply@openai.com>
2026-09-19 19:44:14 +08:00
d0180b74b3 fix(openai): refresh expired quota windows after reset grace (#170)
Refresh non-exhausted accounts after the quota reset grace period while preserving observed usage and exhausted-window attribution. Keep periodic checks for exhausted accounts to detect early upstream resets, and cover refresh timing, throttling, and display regressions.

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-19 18:13:33 +08:00
4669a29b46 feat(usage): show key names and long-context billing badges (#171)
Show current client key names in usage and error records, and persist the selected long-context pricing band in billing snapshots.

Display a long-context badge without assuming a fixed multiplier. Keep missing historical flags false while preserving reconstructed billing details, with regression coverage for OpenAI and xAI.

Validated with 36 billing tests, strict Clippy, Rustfmt, frontend lint/typecheck/build, and browser checks for both usage views.

Co-authored-by: 陈燕山 <chenyanshan@yanshan-mac.local>
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-19 18:13:14 +08:00
e63c5fa28d chore(deps): align TLS with Codex Desktop 26.915 (#175)
Update rustls and AWS-LC to the versions bundled with Codex Core 0.155,
refresh the captured ClientHello baseline, and remove the rustls audit waiver.

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-19 17:39:48 +08:00
1acd4a5f59 feat(settings): unify OpenAI and xAI client identities (#174)
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-19 17:39:29 +08:00
ZyyandCodex 4c7e4e08f6 feat: refine admin controls and overload protection
Refine settings navigation, selective pricing sync, model deletion, and admin interactions. Consolidate unreleased migrations and keep Fast policy scoped to groups. Increase confirmed capacity-rejection feedback without changing ordinary failure recovery.

Co-authored-by: Codex <noreply@openai.com>
2026-09-19 06:41:03 +08:00
16ae8c7afa feat(auth): add administrator password changes (#166)
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-19 06:07:17 +08:00
956f627f68 feat(keys): add manual budget reset and compact row actions (#165)
Add an admin-only reset for daily, weekly, or all key budget usage with atomic auditing and serialized settlement boundaries.

Keep budget limits, original expiration times, and billing history while moving secondary row actions into the existing popover pattern.

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-19 04:46:22 +08:00
f3076e2338 feat(openai): fallback auxiliary account endpoints to official upstream on 404 (#160)
* feat(openai): fallback auxiliary account endpoints to official upstream on 404

When a custom upstream base URL is configured (such as a Codex Relay for chat and inference), non-inference auxiliary account endpoints (usage, profile statistics, reset credits, subscriptions) may return 404 Not Found if the relay only proxies /backend-api/codex/*. Automatically fallback to the official ChatGPT backend endpoint to retrieve account metadata and quota statistics, while keeping chat and model traffic routed through the configured relay.

* fix(openai): preserve account fallback failures and query budgets

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-19 02:44:52 +08:00
ZyyandCodex 7bcc207055 fix(openai): stop cancelled artifact reads during shutdown
Return a terminal I/O error for cancelled or expired Desktop artifact reads so standard read helpers cannot retry forever and block runtime shutdown.

Cover cancellation before reading and during an in-flight range request.

Co-authored-by: Codex <noreply@openai.com>
2026-09-19 00:14:25 +08:00
fba8941068 feat(pricing): add configurable model prices and manual catalog sync (#159)
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-18 20:22:15 +08:00
Zyy e6587b6ca2 feat(openai): add configurable upstream client identities (#156)
Add six Desktop and CLI presets, global defaults and per-key overrides, custom identity parameters, and official release tracking. Keep the resolved identity stable throughout each request and its retries.
2026-09-18 16:39:29 +08:00
ZyyandCodex 4d1ed7917d fix(api): make final overload failures retryable for clients (#152)
Project capacity errors at the HTTP, SSE, and WebSocket delivery boundary while preserving upstream facts and existing gateway recovery policies.

Co-authored-by: Codex <noreply@openai.com>
2026-09-18 16:38:08 +08:00
77b0c29354 feat(proxies): add dual-stack diagnostics without blocking binding (#132)
Probe IPv4 and IPv6 exit endpoints concurrently and accept either successful address family. Persist both exit addresses with migration 0015 while preserving existing migration checksums and the legacy exitIp field.

Remove stale frontend IP preference payloads and display exit addresses without redundant family prefixes. Treat probe results as diagnostics and allow users to select and bind saved proxies regardless of probe status, while preserving missing-proxy checks, configuration consistency, import guards, and no-direct-fallback behavior.

Co-authored-by: zhangwentian <2han9wen71an@gmail.com>
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 19:58:22 +08:00
487b2e87e2 feat: add global and group controls to disable Fast (#146)
Freeze global and bound-group Fast restrictions in each request snapshot, then apply the effective policy before OpenAI transport headers, observations, and billing. Preserve omitted update values and existing data with an additive migration.

Keep the current split settings layout, add a service-tier card, and require WebSocket transport from the first turn in the continuation regression test.

Validation: frontend lint, typecheck and build; Rust formatting; migration checksums; 15 focused tests. Live priority/default/priority requests preserved context on one WebSocket, but the upstream returned default for all three, so actual priority-tier execution remains unverified. Cloud CI was still running when the maintainer requested immediate merge.

Co-authored-by: Yanshan Chen <16632048+chenyanshan@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 18:55:31 +08:00
2han9wen71anandZyy 5d941008fc fix(providers): let disabled accounts source their own model catalogs (#143)
* fix(providers): let disabled accounts source their own model catalogs

The admin per-account model query (connection test model dropdown)
resolves the account's plan catalog through refresh_account_catalog,
whose candidates come from catalog_candidates_by_scope over the
scheduling account listing. Disabled accounts never appear there, so a
disabled account either failed with NoEligibleCredential (no enabled
account in its plan scope) or silently sourced the catalog from other
accounts instead of itself, and the panel showed no test models.

Keep the pinned account reachable: when the target account is missing
from the eligibility-filtered candidates, append it (for xAI, load its
full credential) and let the preferred-first ordering try it before any
same-scope peers. A fully filtered listing is now treated as empty
candidates instead of an early error so the pinned account can still
answer alone. Client-facing catalog discovery, background sync and
scheduling listings continue to exclude disabled accounts.

* fix(xai): keep catalog credential revision aligned

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
2026-09-17 16:07:58 +08:00
b44a790777 fix(openai): coalesce SOCKS proxy negotiation writes (#142)
* fix(openai): coalesce SOCKS proxy negotiation writes

Buffer proxy handshake writes until the existing protocol flush points to support SOCKS services that reject partial method greetings. Preserve the dependency handshake and return the unbuffered stream before TLS/WebSocket traffic.

Refs #139

Co-authored-by: Codex <noreply@openai.com>

* test(openai): make SOCKS negotiation coverage deterministic

Replace TCP peek boundary assumptions with an in-memory write transcript and fragmented responses. Cover authenticated and unauthenticated negotiation, retain the existing account proxy integration coverage, and verify the handshake buffer is drained.

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Zyy <39424122+zyycn@users.noreply.github.com>
2026-09-17 15:54:09 +08:00
06e94ef69b fix(settings): make Responses decompression limit configurable (#141)
Persist the output limit in runtime settings and use the authenticated snapshot for each compressed Responses request. Expose the setting in MiB and preserve bounded decoding and the independent zstd window limit.

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 15:29:14 +08:00
0f2e0b18a9 feat(openai): recover quota-exhausted continuations through client replay (#140)
Project replay-safe quota rejections to the existing client history recovery
contract while preserving upstream failure facts and account isolation.

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 15:07:31 +08:00
71e26f44b7 fix(usage): bill sent models and show response model mismatches (#136)
* feat(usage): show upstream response model mismatches

Persist observed response models separately from routing and billing facts.
Surface mismatches with the existing model arrow and native title tooltips.
Keep response observations isolated across retries and WebSocket reuse.

Co-authored-by: Codex <noreply@openai.com>

* fix(usage): calculate costs from the outbound model

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 14:55:54 +08:00
dcc61915e9 fix(openai): default missing downstream store to false (#134)
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 14:54:52 +08:00
47d08560a7 fix(api): allow ChatGPT remote Desktop sessions without app versions (#133)
* fix(api): allow ChatGPT remote Desktop sessions without app versions

Recognize the official Android and iOS remote User-Agent suffixes when
Desktop application version metadata is absent. Preserve explicit version
validation and Client Key authentication, cover the streaming request path,
and document the version enforcement boundary.

Refs #129

Co-authored-by: Codex <noreply@openai.com>

* fix(api): match ChatGPT remote client name family

Accept nonempty platforms in codex_chatgpt_*_remote while preserving suffix
token validation and explicit Desktop version enforcement. Cover future
platforms and malformed names in parser and streaming route tests.

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 14:54:17 +08:00
ZyyandCodex 590c4a163b fix(store): release rejected proxy import locks before returning
Explicitly unlock rejected import reservations before returning their error so immediate proxy operations do not race connection teardown. Cover repeated rejection followed by writes with PostgreSQL integration tests.

Co-authored-by: Codex <noreply@openai.com>
2026-09-17 05:36:00 +08:00
ZyyandCodex 223021630e fix(openai): unify API key and OAuth passthrough
Share the simulated client profile and existing request/response paths across authentication modes. Forward native capabilities and unknown models upstream while preserving credential isolation and routing permissions.

Co-authored-by: Codex <noreply@openai.com>
2026-09-17 05:15:31 +08:00
ZyyandCodex bcfea792e8 fix(config): tolerate unused startup settings
Ignore unused YAML fields across startup configuration and report their paths without values. Report missing required fields while retaining existing type and value validation.

Validation: configuration and bootstrap tests, strict Clippy, and architecture checks passed.

Co-authored-by: Codex <noreply@openai.com>
2026-09-17 04:08:17 +08:00
ZyyandCodex 1e7ea8d46f refactor(store): consolidate unreleased account migrations
Default account auto-freeze to off in its initial migration and renumber request location to 0011. Update checksum entries and migration tests.

These migrations were not included in v3.8.0. Existing PR freeze rules remain unchanged.

Validation: 18 gateway-store tests with isolated PostgreSQL, upgrade from the v3.8.0 schema, migration checksums and numbering, and Rustfmt.

Co-authored-by: Codex <noreply@openai.com>
2026-09-17 03:47:53 +08:00
bd61e0391a feat(settings): configure global and per-proxy request location (#123)
* feat(proxies): configure per-proxy request location

* feat(settings): manage global request timezone and location

Move request location from YAML into runtime settings with an enable switch. Preserve client fields when disabled and keep per-proxy overrides first. Share location inputs and simplify proxy and settings controls.

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 03:30:54 +08:00
499b43043e feat(accounts): add opt-in capacity cooldown and recovery (#126)
Add opt-in account capacity cooldown, recovery probes, and atomic adaptive concurrency updates. Include overload and slow_down in Smart scheduling feedback, expose rate-limit reasons, and keep probe results fenced by cooldown generation.

Exclude diagnostic probes and local connection failures from capacity evidence. Default auto-freeze to disabled and preserve explicitly saved settings with a forward migration.

Closes #120

Co-authored-by: zhangwentian <2han9wen71an@gmail.com>
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 03:23:38 +08:00
1e6d69a349 fix(openai): align downstream requests with Codex protocol (#117)
Centralize downstream SDK, browser and proxy header compatibility in the
OpenAI provider. Normalize the session header alias and isolate known
non-Codex body parameters while preserving ordinary extensions.

Align account installation metadata with Codex Core and remove unverified
identity header classifications. Cover HTTP and WebSocket behavior and
document the request boundaries.

Validation: 1,105 scoped and architecture tests, Rustfmt and strict scoped
Clippy passed. Four real-account HTTP/WebSocket cases passed on gpt-5.5,
including connection reuse. Earlier gpt-reserve overload failures were
recorded separately without changing production retry policy.

Co-authored-by: Clarence <1297529513@qq.com>
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-17 00:30:52 +08:00
2han9wen71anandCodex dac57a81c6 fix(providers): keep disabled accounts reachable for admin connection tests (#122)
Admin connection tests pin an account and bypass local availability, but
the scheduling store excludes disabled accounts before selection. Load
the pinned diagnostic account explicitly for OpenAI and xAI while keeping
normal scheduling and disabled-account state-write guards unchanged.

Apply the shared authentication and transport filters after restoring
OpenAI candidates. Align the memory stores with PostgreSQL filtering and
cover disabled-account diagnostics plus API key transport restrictions.

Validation: 652 OpenAI, 403 xAI, 350 Core, and 42 architecture/bootstrap
tests passed; Rustfmt and strict workspace Clippy passed.

Co-authored-by: Codex <noreply@openai.com>
2026-09-16 18:53:52 +08:00
d162f9b7f6 fix(accounts): run bulk imports as background tasks (#119)
* fix(openai): sync account plans with quota updates

Co-authored-by: Codex <noreply@openai.com>

* fix(accounts): run bulk imports as background tasks

Co-authored-by: Codex <noreply@openai.com>

* test(store): preserve settings in account plan refresh fixture

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-16 18:24:34 +08:00
866f71e3fb fix(openai): sync account plans with quota updates (#112)
* fix(openai): sync account plans with quota updates

Co-authored-by: Codex <noreply@openai.com>

* test(store): preserve settings in account plan refresh fixture

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-16 18:22:01 +08:00
4fb10f33f1 fix(update): reuse the served asset directory (#111)
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-16 18:05:14 +08:00
c9318c54cd fix(proxy): support IPv6 egress probes (#110)
Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-16 18:05:08 +08:00
8d9aaf0bbd feat(accounts): support upstream API key accounts (#107)
Add Responses-compatible upstream accounts with isolated credentials, transport,
model catalogs, atomic settings updates, and account management controls.

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-16 18:04:41 +08:00
ZyyandCodex cca89f2f96 feat(accounts): add per-account model restrictions (#105)
Add per-account model allowlists and denylists to frozen routing, account settings, and CPR import/export. Preserve existing scheduling and account state boundaries.

Closes #97

Co-authored-by: Codex <noreply@openai.com>
2026-09-16 02:29:56 +08:00
e25aad2231 feat(gateway): add bounded concurrency queues for keys and accounts (#104)
* feat(auth): add API key workspace

Add persistent API key sessions and enforce admin/key authorization boundaries. Expose key-scoped quotas, trends, health, and paginated usage records through server-projected data.

Share the application shell, system overview, and usage views between admin and key sessions, reorganize frontend domains around view ownership, and omit account, provider, and authentication details from the key workspace.

BREAKING CHANGE: Browser authentication now uses /api/auth/* and cpr_session; existing browser sessions must sign in again.

Validated frontend lint, type checking, production build, backend workspace tests, and live authenticated key views.

Co-authored-by: Codex <noreply@openai.com>

* feat(console): refine key workspace experience

Unify role-aware console routes and use mode/role in authentication contracts.
Reuse request state, stable pagination and performance projections across workspaces.
Refine quota cards, usage panels, empty states and key-aware theme previews.
Skip session restoration on login and retain form inputs until unmount.
Keep the full health timeline grid visible without samples.

Co-authored-by: Codex <noreply@openai.com>

* feat(key-usage): replace shared workspace with focused usage page

Preserve unified admin/key authentication, restore admin pages, and remove the previous client workspace implementation. Add a scoped usage page with metrics, trends, budgets, request logs, and section loading placeholders.

Co-authored-by: Codex <noreply@openai.com>

* feat(gateway): queue requests at key and account concurrency limits

Share one wait deadline across client admission and account selection.

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-16 02:00:26 +08:00
84382761fc feat(auth): add focused API key usage page (#102)
* feat(auth): add API key workspace

Add persistent API key sessions and enforce admin/key authorization boundaries. Expose key-scoped quotas, trends, health, and paginated usage records through server-projected data.

Share the application shell, system overview, and usage views between admin and key sessions, reorganize frontend domains around view ownership, and omit account, provider, and authentication details from the key workspace.

BREAKING CHANGE: Browser authentication now uses /api/auth/* and cpr_session; existing browser sessions must sign in again.

Validated frontend lint, type checking, production build, backend workspace tests, and live authenticated key views.

Co-authored-by: Codex <noreply@openai.com>

* feat(console): refine key workspace experience

Unify role-aware console routes and use mode/role in authentication contracts.
Reuse request state, stable pagination and performance projections across workspaces.
Refine quota cards, usage panels, empty states and key-aware theme previews.
Skip session restoration on login and retain form inputs until unmount.
Keep the full health timeline grid visible without samples.

Co-authored-by: Codex <noreply@openai.com>

* feat(key-usage): replace shared workspace with focused usage page

Preserve unified admin/key authentication, restore admin pages, and remove the previous client workspace implementation. Add a scoped usage page with metrics, trends, budgets, request logs, and section loading placeholders.

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-16 01:49:37 +08:00
0a09c1572f feat(accounts): add notes and refine account interactions (#100)
Persist optional account notes across creation, editing, imports, and credential refreshes. Show notes in compact account rows with overflow-only popovers.

Support vertically resizable textareas, bounded select menus with custom scrollbars, and quota updates without table loading.

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-16 01:36:34 +08:00
2han9wen71an b1c24e276b fix(gateway-api): project bare WebSocket errors to response.failed (#99)
codex 的 WS 客户端(codex-api/src/endpoint/responses_websocket.rs)只消费
两类 error 帧:带非 2xx HTTP status 的包装错误,以及
websocket_connection_limit_reached / previous_response_not_found 内置可重
试错误码。OpenAI 上游的中流错误帧(如 server_is_overloaded)没有 status
字段,map_wrapped_websocket_error_event 返回 None,落回
process_responses_event 也没有 error 分支——整帧被静默忽略。

客户端收不到可识别的终态,只能空等到自身 300s stream_idle_timeout 超时
再重连,5 次重连累计 20-30 分钟"正在思考",最后才把失败呈现出来。CPA
等实现之所以秒级报错,正是因为它把错误负载重写成 codex 可消费的终态。

#90 已在 SSE 边界做了同样的投影,但客户端 WS 边界(push_websocket)仍原
样转发;使用 WS transport 的 codex 桌面端不受 #90 保护。本提交在该边界
对不可消费的错误帧投影成 response.failed(消息 JSON 自带 type 字段,经
process_responses_event 的 response.failed 分支映射为可重试错误并立即重
试);带非 2xx status 或特殊错误码的可消费形状原样透传,保持
deliverable_business_failures_are_not_rewritten... 既有契约。
2026-09-15 16:31:07 +08:00
2han9wen71an dfabf0f9dd fix(openai): recover quota after same-window usage drops (#95)
OpenAI 周窗口的 used_percent 可以在同一窗口内回落(滚动计量),而
reset_at 固定到窗口边界才变化。恢复规则要求「reset 前进 + 用量低于
10%」,两个条件在滑动周窗口上都无法成立:耗尽账号被钉在旧 reset
日期(最长一周)无法参与调度,尽管 /usage 已显式报告 allowed=true /
limit_reached=false 且 5h 窗口全新。

新增第二条解除路径:同一窗口(reset 未滚动)内连续两次观测都未触顶
即解除。候选记录持久化在 _quota_recovery,跨 worker 周期与重启累积;
连续两次的要求保持对「耗尽后立刻拉到的旧快照」竞态的防护(单次旧
快照撑不过两次观测)。原有「reset 前进 + 用量<10%」快路径不变。
2026-09-15 15:38:46 +08:00
ZyyandCodex cc2dfab19a fix(openai): preserve client location when unset
Accept omitted and null request locations without rewriting web search location or environment dates and timezones. Keep complete configured overrides validated and document the passthrough behavior.

Validated 626 OpenAI tests, 38 gateway tests including architecture checks, Rustfmt, and strict workspace Clippy.

Co-authored-by: Codex <noreply@openai.com>
2026-09-15 15:02:36 +08:00
1bbae67040 fix(gateway-api): translate upstream WebSocket error events into SSE response.failed (#90)
* fix(gateway-api): translate upstream WebSocket error events into SSE response.failed

WS 上游协议用 `error` 事件报告流中途失败(如 OpenAI 的 server_is_overloaded),
SSE 协议的对应形态是 `response.failed`。SSE 交付边界此前把 `error` 事件按
未知事件原样透传,而 codex 等 SSE 客户端会静默忽略未知事件类型:客户端只看到
无原因的流 EOF("stream closed before response.completed"),丢失了上游的
完整错误负载(message/code/type),也无法据此分类重试。

现在 SSE 编码器在交付边界把 `error` wire 事件翻译成 `response.failed`:
错误负载(type/code/message/param)原样保留,response id 沿用已交付的
response.created。WS 客户端交付不变(WS 协议原生支持 error 帧)。
原有的会话级失败兜底(finish_with_gateway_error 合成 response.failed)不变。

* fix(gateway-api): preserve upstream response failures

Co-authored-by: Codex <noreply@openai.com>

---------

Co-authored-by: Zyy <zyycn@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-15 10:49:29 +08:00
ZyyandCodex c7823a7821 feat(openai): make request location configurable
Group country, region, city, and timezone under openai.wire_profile.location while retaining the existing defaults. Share the configured location across web search and environment context, preserve epoch timestamps and release updates, and add validation, regression tests, and deployment documentation.

Co-authored-by: Codex <noreply@openai.com>
2026-09-15 10:49:29 +08:00