New Features

• Added Request policies in System Settings: Request checks, Sessions and retries, and Channel health, with policy decision records on Usage Logs.
• Added Responses WebSocket relay, with a per-Channel toggle. Advanced Custom, Sub2API, and New API channels can use it when enabled; Advanced Custom only uses converter-free /v1/responses.
• Added a Model Mapping workbench, Quick options, and Pass-through controls in the Channel editor. Advanced Custom channels can enable Pass-through per path.
• Added vLLM and SGLang channels. Ollama channels can enable OpenAI-compatible chat per Channel.
• Added filtered task history and cleanup in System Info.
• Show subscription deductions on Usage Logs, and flag Response model mismatch when the upstream model differs.

Fixes and Improvements

• Interrupted Responses streams no longer settle at zero and refund the whole pre-consume Quota; estimated usage is billed unless upstream reported an explicit failure (#7427).
• Count Images by payload for Quota, and map image output token details into image-output billing (#7410).
• Hailuo / MiniMax Task Plugins bill only the dimensions each model family supports.
• Reserve model rate-limit slots and judge success by the response outcome, not HTTP status alone.
• Hide Task Plugin keys from the public no-Channel 503, while keeping plugin context for operators (#7414).
• Reject Gemini :countTokens as an unknown path instead of treating it as generateContent (#7388).
• Request stream usage on every cross-protocol conversion to OpenAI chat (#7389).
• Preserve Ollama tool calls from the final stream frame (#7376).
• Restore saved compatible Passkey domains; fix DeepSeek Channel balance fetch (#6814); accept mixed-case audio extensions (#7321).
• Handle legacy database constraints and scoped authorization policies on upgrade.

---

What's Changed

• test(web): stabilize flaky frontend suite timeouts and motion visibility races by @RedwindA in #7367
• fix: unify model provider detection and add Wan icon by @seefs001 in #7373
• fix(dashboard): align weekly default range with available preset by @jimmyleocn in #7355
• fix(web): prevent combobox dropdowns opening on dialog autofocus by @RedwindA in #7365
• fix(audio): case-insensitive file extension in GetAudioDuration by @chiliec in #7321
• fix(pricing): keep model status bar spacing uniform by @jimmyleocn in #7284
• Add Responses WebSocket relay support by @KoSukeWork in #5062
• feat: vllm channel && sglang channel by @seefs001 in #7332
• fix(pricing): 修复排序菜单打开时导航栏横向位移 by @jimmyleocn in #7145
• fix(ollama): preserve tool calls from final stream frame | 保留流式末尾帧的 tool calls by @zcxads666 in #7376
• feat(channel): 修复获取DeepSeek余额错误 by @sszj987-crypto in #6814
• feat(rate-limit): optimize in-memory limiter allocation and cleanup by @sszj987-crypto in #6807
• feat(ollama): add per-channel OpenAI-compatible chat switch by @seefs001 in #7382
• fix(gemini): accept case-insensitive thinkingLevel by @seefs001 in #7387
• fix(gemini): reject :countTokens as an unknown route by @seefs001 in #7388
• fix(relay): request stream usage on every cross-protocol conversion by @seefs001 in #7389
• feat: per-route pass-through for advanced custom channels by @seefs001 in #7386
• fix(middleware): 隐藏无可用渠道错误中的任务插件标识 by @blackdm666 in #7414
• docs(github): tighten issue and PR templates for pass-through by @seefs001 in #7438
• refactor: show subscription deductions and improve cost display by @seefs001 in #7272
• feat(log): flag upstream response model mismatches by @seefs001 in #7418
• fix(relay): 将 Images 的 output_tokens_details 映射到 img_o by @sxhstc in #7410
• fix(responses): estimate usage for streams cut before terminal usage by @Calcium-Ion in #7427
• doc: fix issue template by @seefs001 in #7441

New Contributors

• @chiliec made their first contribution in #7321
• @KoSukeWork made their first contribution in #5062
• @sszj987-crypto made their first contribution in #6814
• @blackdm666 made their first contribution in #7414
• @sxhstc made their first contribution in #7410

Full Changelog: v1.0.0-rc.37...v1.0.0-rc.38

New Features

• Added Convert to expression in Pricing Configuration: review a legacy-to-Expression draft with effective prices, then save to apply it. Legacy pricing will be deprecated soon.
• Extended Expression editors and usage details: Pricing by condition, Per-request, Time-based, and image usage / image cache. Confirmation is required before replacing an Expression.
• Added built-in Expression defaults for gpt-image-2, gpt-image-2.5-sunburst, and gpt-image-2.5-flare, including separate text and image cache prices. Existing administrator overrides stay in place.
• Expanded Task Plugins capabilities: Streaming, Usage prices and Expression per Task Plugin, and localized Changelog (English fallback).
• Unified Channel create and edit: shared provider selection, tabbed setup, inline model discovery, Model Mapping, and parameter overrides.
• Let Task Plugins take part in Channel provider and model selection, including a gateway category and preferring available plugins for legacy task providers.
• Added System updates reminders for administrators.
• Added multiple Passkey domains in Security & Access, with impact preview and confirmation before removing a domain still in use.

Fixes and Improvements

• Validate image quantities before reserving Quota, and keep image usage consistent through reservation, retries, and settlement.
• Stop Pricing Configuration writes from resetting or failing to persist on some long-lived deployments.
• Allow editing multi-key selection strategy without dropping existing keys; omit the strategy for single-key Channels.
• Restore the add split button in Advanced Custom (#7289).
• Correct Wan model protocols and usage accounting; preserve provider fields in Sora video queries; fix VolcEngine Channel model-list fetch (#7203).
• Unify server error notifications so each failed operation is reported once, while preserving backend details.
• Deduplicate /api/status requests (#7189).
• Keep Model pricing content in one scroll area.

What's Changed

• fix(frontend): deduplicate /api/status requests (#7157) by @CreatorEdition in #7189
• fix: 修正火山方舟(VolcEngine)渠道获取模型列表端点路径 404 by @captainold in #7203
• fix: restore add split button in advanced custom routes by @seefs001 in #7289

New Contributors

• @captainold made their first contribution in #7203

Full Changelog: v1.0.0-rc.36...v1.0.0-rc.37

New Features

• Improved Task Plugins and the Marketplace: Plugin metadata, Plugin website, Plugin icon (including optional icon.svg / icon.png), integrity checks, Capabilities, and clearer install/update flows.
• Improved Channel setup for Task Plugins so a new Channel picks up the matching Plugin icon and Plugin metadata.
• Added Site currency in Pricing Configuration. Display and conversion use the site exchange rate; stored prices remain in USD.
• Improved Model management, pricing display, Catalog visibility, and Model square visibility. Channel-configured models appear without creating catalog metadata first. The Model square can show Pricing by condition, usage prices, and visibility reasons.
• Added a searchable Group filter on Usage Logs. Suggestions come from the current view; historical Group names can still be typed in.
• Refined the Usage Logs mobile layout so filters and quick actions stay available.
• Refined Quota display on API Keys and users: Remaining quota and used amounts, progress bars, Group ratio badges, and activity timestamps. Full API Keys are resolved only for explicit copy or chat actions.
• Added confirmed batch deletion for Redemption Codes, plus an optional Save as a file step (TXT / Markdown) after creation.

Fixes and Improvements

• Show desktop API Keys Remaining quota and used amounts side by side, with more room and consistent cell spacing. Mobile stacked labels are unchanged.
• Unified table cells and Quota details across list pages.

---

Full Changelog: v1.0.0-rc.35...v1.0.0-rc.36

New Features

• Added Alibaba Bailian wan3.0-video and wan3.0-video-prime all-in-one video models, with DashScope media, resolution/ratio, duration rules, and billing ratios.
• Supported wan3.0 smart duration (duration / seconds = -1): submit -1 upstream, pre-consume 30 seconds, then settle from actual usage. Other Wan models reject this marker instead of silently defaulting to 5 seconds.
• Added kimi-k3 to the Moonshot model list.

Fixes and Improvements

• Preserved Kimi K3 mid-conversation tool-loading messages, so dynamic tools are no longer dropped during relay and tool_choice=required no longer fails for missing tools (#7235).
• Made wan3.0 convert and billing follow the redirected upstream model, reject unknown sizes, restore the image-only guard for text-to-video models, and forward ratio/audio in native passthrough (#7240).
• Made the SunoAPI built-in honor model redirects and fixed lyrics/music rendering. Bumped sunoapi to 1.0.2.
• Replaced the model-card health strip’s last-three-bucket rates with an hourly series, so empty hours stay gray and the 24-slot bar can show a full day.
• Portaled Combobox and Select popups into drawer content so they stay clickable inside modal drawers.
• Switched the pricing card grid (and its skeleton) to three columns at the xl breakpoint.
• Routed common and relaykit JSON helpers through a host-injectable codec so the JSON engine can be swapped in one place.

---

Full Changelog: v1.0.0-rc.34...v1.0.0-rc.35

New Features

• Reworked model and vendor management: catalog metadata validation, vendor management, batch delete with channel/pricing cleanup, model pricing snapshot editing with optimistic concurrency, and upstream ratio sync with price cells. The models page is organized around vendors and model connections, with dedicated pricing and vendor-management dialogs.
• Moved account security settings to a dedicated page, with system access-token status, rotation, revocation, and access history. Generate and revoke require scoped, single-use verification.
• Added a dedicated audit log page (audit.read). Successful and failed API token and quota operations are recorded with role snapshots and safe metadata, correlated with top-up records. Model, vendor, and pricing mutations are also audited.
• Unified login and sensitive-operation verification: enrolled TOTP and Passkey are alternative factors. Every primary login transport is gated before a session is issued; WebAuthn requires user verification; login challenges are consumed atomically with session creation.
• Required a scoped, single-use proof for account deletion, rechecked the session inside the deletion transaction, and revoked all sessions afterward. Bindings, password changes, Passkey/2FA enrollment, and channel-key access are likewise bound to the initiating session and action.

Security

• Completed the account-security surface: shared verification UI, session-bound proofs, last usable login method preservation, and audit of security operation outcomes. OAuth callback secrets are redacted from logs.
• Applied Unicode-aware password length limits (8–128). New passwords default to Argon2id; existing bcrypt hashes remain valid. Profiles return has_password from the existing SELECT without extra queries.
• Migrated Telegram login, binding, and security verification to the unified OAuth flow (authorization code + PKCE + ID tokens). Legacy widget / bind endpoints have been removed. A custom OAuth provider must not use the reserved telegram slug.

Fixes and Improvements

• Fixed OpenAI Chat capability detection: match gpt-<n> by major version (n >= 5) so later generations such as gpt-6-astra use max_completion_tokens and drop unsupported temperature / top_p / logprobs. gpt-4.1, gpt-4o, gpt-oss, gpt-image, and gpt-realtime are unchanged (#7211).
• Upgraded the SQLite GORM driver to distinguish unique indexes from constraints, and normalized equivalent MySQL decimal defaults and PostgreSQL CHAR metadata. Unchanged databases no longer emit redundant DDL on restart.
• Bulk-copied RawMessage fields to reduce request deep-copy cost (#7221).
• Collapsed the completed setup guide by default so it no longer occupies the dashboard.
• Upgraded the ClickHouse driver to v2.46.0.

---

Full Changelog: v1.0.0-rc.33...v1.0.0-rc.34

New Features

• Added gpt-6-astra to the OpenAI model list with built-in expression pricing for standard and long-context requests, including input, output, cache-read, and cache-write usage. Existing administrator pricing overrides take precedence.
• Added explicit trailing @key:value model modifiers for non-pass-through relay requests:
  • @thinking:on, @thinking:off, @thinking:adaptive, or @thinking:<budget>
  • @effort:none|minimal|low|medium|high|xhigh|max
  • @temperature:<number> and @topp:<number>
• Model modifiers override corresponding request-body fields and remain effective across channel model mapping. Modifiers on the mapped model take precedence over those on the requested model; repeated keys use the rightmost value.
• Added administrator-visible conversion diagnostics for modifier conflicts, duplicates, and provider-specific adjustments.
• Added re:-prefixed Go regular expressions to Models that skip thinking suffix processing, allowing real model names containing @key:value to bypass modifier parsing and validation. Unknown modifier keys and invalid values otherwise return HTTP 400.

Billing and Routing

• Added canonical billing names such as base@effort:high@thinking:on and base@thinking:on, with fallback to the base model when no matching variant price is configured.
• Equivalent modifier variants share canonical pricing names. Numeric thinking-budget values, temperature, and topp do not create separate canonical pricing variants.
• Updated pricing settings and pricing synchronization to expose the effective built-in billing defaults used at runtime.
• Improved channel selection, model mapping, and token model-limit matching for models with modifiers or supported legacy aliases.
• Preserved existing legacy alias pricing and Gemini thinking-budget wildcard entries.

Fixes and Improvements

• Fixed implicit reasoning-effort remapping: max and xhigh now remain distinct through OpenAI Chat and Responses conversion paths.
• Preserved native reasoning controls when no model modifier or protocol conversion requires adaptation, avoiding unnecessary normalization or rejection of provider-native values.
• Restricted generic legacy suffix parsing to recognized GPT/o-series, Claude, and Gemini model families, protecting real model names such as qwen-max and kimi-k2-thinking.
• Applied thinking-suffix exemptions consistently across DeepSeek, Volcengine, and xAI adaptors.
• Improved Claude reasoning conversion by adapting unsupported combinations where possible and recording diagnostic warnings.
• In global or channel-level pass-through mode, model modifiers are forwarded unchanged without local parsing or validation.
• Improved HTTP 503 messages for models claimed by task plugins: errors now identify the claiming plugin and explain when it has no enabled channel in the selected group. Plugin model ownership remains unchanged (#7185).
• Fixed release builds selecting an incorrect version by resolving the version from the triggering Git tag.

---

Full Changelog: v1.0.0-rc.31...v1.0.0-rc.33

New Features

• Added explicit trailing model modifiers for non-pass-through relay requests:
  • @thinking:on, @thinking:off, @thinking:adaptive, or @thinking:<budget>
  • @effort:none|minimal|low|medium|high|xhigh|max
  • @temperature:<number> and @topp:<number>
• Model modifiers override the corresponding request-body fields and remain effective across channel model mapping. Modifiers configured on the mapped model take precedence over those on the requested model.
• Added conversion diagnostics for overridden, duplicated, or adjusted reasoning and sampling settings. These warnings are recorded with consumption logs for troubleshooting.
• Unknown modifier keys and invalid values now return HTTP 400.
• Added re:-prefixed Go regular expressions to Models that skip thinking suffix processing, allowing real model IDs containing @key:value segments to bypass modifier parsing and validation.

Billing and Routing

• Added canonical billing identities for model variants. Pricing can now target forms such as base@effort:high@thinking:on or base@thinking:on, with automatic fallback to the base model.
• Canonical billing is independent of modifier order, duplicate modifiers, and numeric thinking-budget values. temperature and topp never affect the selected billing identity.
• Updated channel selection, model listing, model mapping, and token model-limit checks to recognize the underlying base model when modifiers or supported legacy aliases are used.
• Preserved existing wildcard pricing behavior for legacy Gemini thinking-budget model names.

Relay Improvements

• Limited legacy -thinking, -nothinking, thinking-budget, and effort suffix parsing to recognized GPT/o-series, Claude, and Gemini model families, preventing ordinary model IDs such as qwen-max from being misinterpreted.
• Improved Claude reasoning conversion by safely adapting unsupported thinking modes, effort levels, budgets, and sampling combinations while recording conversion warnings.
• Native Claude and Gemini requests without host-specific model modifiers continue to preserve their provider-native reasoning controls.
• Global and channel-level pass-through requests remain byte-identical: model modifiers are not parsed or validated and are forwarded verbatim.

Bug Fixes

• Improved task-plugin HTTP 503 errors to identify the plugin that owns the requested model when it has no enabled channel in the selected group. The error now recommends disabling or overriding the claiming plugin instead of returning only a generic “no available channel” message (#7185).
• Fixed release builds resolving an incorrect version when the workflow is triggered from a tag.

---

Full Changelog: v1.0.0-rc.31...v1.0.0-rc.32

New Features

• Expanded request, response, and streaming conversion across OpenAI Chat Completions, OpenAI Responses, Claude Messages, and Gemini. Hosted tools, citations, grounding metadata, reasoning intent, and authoritative usage are now preserved more faithfully across supported relay paths (#7137, #7170).
• Added the Anthropic-compatible POST /v1/messages/count_tokens endpoint and enabled Claude channels to serve Gemini-format clients through protocol conversion.
• Released RelayKit v0.2.0 with conversion diagnostics and configurable tool-loss policies.
• Added MiniMax-H3 (Hailuo-03) /v2 video generation to the Hailuo task plugin, including multimodal image, video, and audio references, 768P/2K output, and validated 4–15 second durations (#7168, #7171).
• Added MiniMax-H3 usage facts for input_images and input_video_seconds. Existing ModelBillingExpr values are not migrated automatically and must reference these fields to include input-media charges.

Reliability Improvements

• Hardened asynchronous task polling with host-level HTTP classification. Upstream 404/410 responses now fail and refund immediately, while transient, authentication, hook, and unrecognized-response failures use a bounded consecutive-failure counter.
• Added TASK_POLL_MAX_FAILURES, defaulting to 20. A successful poll resets the counter; reaching the limit fails and refunds the task. Set it to 0 or lower to disable this limit. The existing TASK_TIMEOUT_MINUTES deadline remains unchanged.
• Reduced unnecessary anonymous cold-start authentication requests and added ETag/304 revalidation for public notice, home, about, and legal-content APIs (#7166).
• Fixed unchanged system-task state writes being incorrectly treated as task-lock loss (#7135).

Billing, Privacy, and Bug Fixes

• Improved billing integrity across converted and streamed requests by preserving authoritative upstream usage and fixing Claude cache accounting, Gemini modality usage, tool-call surcharges, and interrupted-stream settlement (#7137, #7170).
• Fixed MiniMax-H3 input-media accounting so bounded estimates can be settled against valid upstream image and duration usage (#7171).
• Hardened usage-log privacy by separating user, administrator, root, and audit metadata and stripping sensitive channel and rejection details from user-visible log responses.
• Corrected Video API links across localized documentation (#7116).

New Contributors

• @To3akaRin made their first contribution in #7116.
• @NebulaMao made their first contribution in #7168.
• @CreatorEdition made their first contribution in #7166.

---

Full Changelog: v1.0.0-rc.30...v1.0.0-rc.31

Bug Fixes

• Fixed usage-log statistics showing a zero quota total despite matching consumption records containing non-zero quota. This was a reporting-only issue; consumption logs, actual billing, and account balances were not affected (#7106, #7108).
• Added a safe and idempotent PostgreSQL migration for known legacy unique constraints on tokens.key, preserving existing token data and key uniqueness while preventing startup migration conflicts.

New Contributors

• @jimmyleocn made their first contribution in #7108.

---

Full Changelog: v1.0.0-rc.29...v1.0.0-rc.30

Bug Fixes

• Fixed database initialization failures on existing MySQL and PostgreSQL deployments caused by incompatible GORM core and database driver versions in v1.0.0-rc.28 (#7098, #7099).
• Added a safe and idempotent PostgreSQL migration for legacy prefill_groups.name uniqueness constraints, preserving existing data and the intended soft-delete behavior.

---

Full Changelog: v1.0.0-rc.28...v1.0.0-rc.29

⚠️ IMPORTANT WARNING / 重要警告 ⚠️

自 v1.0.0-rc.27 起引入的实验性插件系统仍处于测试阶段，并涉及较大的架构调整。如遇问题，请及时反馈。从 v1.0.0-rc.26 或更早版本升级的用户，必须重新配置所有视频模型价格；已在 v1.0.0-rc.27 完成配置的用户无需重复操作。不推荐在生产环境中使用本版本。

The experimental plugin system introduced in v1.0.0-rc.27 remains under testing and includes major architectural changes. Please report any issues promptly. Users upgrading from v1.0.0-rc.26 or earlier must reconfigure all video model prices; users who already completed this step on v1.0.0-rc.27 do not need to repeat it. This release is not recommended for production use.

New Features

• Added optional RSA-OAEP encryption for password-login request bodies. Enable it with PASSWORD_LOGIN_ENCRYPTION_ENABLED=true; this is an additional safeguard and does not replace HTTPS (#6743).
• Added AQBot to the default list of third-party chat options (#7079).
• Added support for channel model-mapping aliases and model-name case variants in task-plugin routing, including consistent upstream mapping and billing resolution.

Improvements

• Built-in task plugins now update with new-api releases; Marketplace actions no longer create unintended persistent overrides.
• Improved concurrent-write reliability for the default SQLite configuration by enabling WAL, a 30-second busy timeout, and immediate transactions (#7030).
• Improved PostgreSQL transaction-pooler compatibility by disabling prepared statements, and fixed JSON-column handling under the simple protocol.

Bug Fixes

• Prevented unbounded memory retention and potential OOM when an upstream does not return response headers (#6949). RELAY_RESPONSE_HEADER_TIMEOUT now defaults to 1800 seconds; set it to 0 to restore the previous unbounded behavior. Streaming after response headers arrive is unaffected.
• Fixed the frontend display of the "subscription-only" preference when no valid subscription exists (#7086).
• Fixed setup-status detection after a page reload (#6968).
• Changed relay request-validation failures to return HTTP 400 (#6774).
• Fixed Ali image models to honor the requested response format (#7048).

---

What's Changed

• fix(relay): 请求参数校验错误返回 HTTP 400 by @ax2 in #6774
• fix(web): recheck setup status after page reload by @seefs001 in #6968
• feat(chat): 将 AQBot 加入第三方聊天默认列表 by @Licoy in #7079
• fix(ali): honor image response format (#5513) by @PuppetKL in #7048
• fix(subscription): 无有效订阅时前端如实显示「仅用订阅」偏好 (#6222) by @CR-Yun in #7086
• fix(sqlite): enable WAL + working busy timeout + _txlock=immediate to stop concurrent write lockouts by @LinineTy in #7030
• fix(relay): bound the wait for upstream response headers (fixes unbounded heap growth → OOM) by @txgo in #6949

New Contributors

• @ax2 made their first contribution in #6774
• @PuppetKL made their first contribution in #7048
• @CR-Yun made their first contribution in #7086
• @LinineTy made their first contribution in #7030
• @txgo made their first contribution in #6949

Full Changelog: v1.0.0-rc.27...v1.0.0-rc.28

⚠️ IMPORTANT WARNING / 重要警告 ⚠️

最新版已增加插件系统测试版本，系统架构大改，有bug请立即报告。对于切换到这个版本的用户，需要重新配置所有的视频模型价格！不推荐生产环境使用。

The latest version introduces an experimental plugin system with major architectural changes. Please report any bugs immediately. Users switching to this version MUST reconfigure all video model prices! Not recommended for production environments.

New Features

• Replaced built-in task adaptors with a new sandboxed JavaScript plugin system (#7076)
• Added support for GLM channel /v1/responses endpoint (#7050)
• Enabled passthrough of Claude Messages and OpenAI Responses for Ollama (#7051)

Bug Fixes

• Fixed an issue in billing where a constantly true time rule expression caused multipliers to apply all day (#6934)
• Restored the ability for administrators to unbind built-in providers in the web interface (#6987)
• Fixed docker development build context by adding relaykit go.mod (#7072)

Other

• Updated PR templates and removed the PR Check workflow (#7053)

---

What's Changed

• feat: glm chanel /v1/responses by @seefs001 in #7050
• feat(ollama): passthrough Claude Messages and OpenAI Responses by @seefs001 in #7051
• docs: update PR template and remove PR Check workflow by @seefs001 in #7053
• fix(web): restore admin unbinding for built-in providers by @zcxads666 in #6987
• fix(billing): 修复时间规则区间恒真表达式导致倍率全天生效 by @zcxads666 in #6934
• fix(docker): add relaykit go.mod to dev build context by @VladKabiak in #7072
• feat(task): replace built-in task adaptors with a sandboxed JS plugin system by @Calcium-Ion in #7076

New Contributors

• @VladKabiak made their first contribution in #7072

Full Changelog: v1.0.0-rc.26...v1.0.0-rc.27

⚠️ Important Warnings / 重要警告

• Database Migration Risk / 数据库迁移风险: This release changes quota fields to support values larger than 4,294,967,295. Upon startup, the system will check database table structures. Startup may fail if the column types are incorrect. In such cases, manual modification is required (instructions will be in the error message; consult AI if unsure how to proceed).
  本次更新修复了额度上限问题，支持大于 42 亿的数值。启动时系统会检测数据库表结构，如果不匹配可能会导致启动失败。遇到此问题需根据错误提示手动修改数据库表结构（如不熟悉数据库操作，可寻求 AI 辅助）。

• Deprecation of 32-bit Systems / 废弃 32 位系统支持: Support for 32-bit operating systems has been temporarily deprecated to accommodate these changes.
  为配合上述更改，我们已“临时”废弃对 32 位操作系统架构的支持。

New Features

• Added support for thinking_token_budget in vLLM (#7027)

Bug Fixes

• Prevented browser auto-fill from incorrectly populating credentials in usage log filters (#6966)

Improvements

• Removed 32-bit integer limitations to allow quota values to exceed 4294 million (#7025)

---

What's Changed

• refactor(web): prevent credential autofill in usage log filters by @seefs001 in #6966
• refactor: deprecate int32 by @seefs001 in #7025
• feat: vllm thinking_token_budget by @seefs001 in #7027

Full Changelog: v1.0.0-rc.25...v1.0.0-rc.26

Highlights

Channel testing and gateway administration gain safer workflows, user/group-aware overrides, and request-field passthrough controls. This release also improves provider compatibility, Responses API accounting, and recharge/quota reliability.

New Features

• Channel testing has been expanded with safer test workflows, including an auto-disable-only mode for failing channels (#6917).
• Parameter overrides can now use requester user and group context for more flexible routing and policy rules (#6534).
• Gateway channels now let administrators choose which request fields are passed through to upstream providers (#6847).

Bug Fixes

• Ali provider requests now avoid sending unset top_p values and correctly handle mapped image models (#6772).
• Qwen TTS models are now classified correctly so they route through the proper text-to-speech path (#6711).
• Responses API handling now preserves penalties and prompt-cache settings while settling cached-token usage accurately (#6892).
• Ollama and Claude conversions now preserve reasoning/tool-call context while avoiding invalid empty tool payloads (#6862).
• Gemini-style model listing requests to /v1/models now return correctly (#6199).
• Recharge, refund, and quota updates are now settled more consistently, including concurrent and uncreditable order cases (#6845).
• Web forms now keep fetched model selections in sync, and the mobile sidebar behaves correctly (#6841).
• Usage logs now record reasoning effort consistently (#6641).
• Request length validation is now enforced consistently on the backend (#5548).

Improvements

• Billing logs now highlight the conditional multiplier that matched a request, making cost calculations easier to audit (#6561).
• Large server and model-list searches in the web UI are debounced, improving responsiveness on big installations (#6727).
• The playground now displays streamed output with smoother word fade-ins and a more robust editor experience (#6895).
• Custom channel route editing now has a more capable advanced editor (#6865).

---

What's Changed

• fix(ali): stop injecting top_p into requests that omit it by @ENCHIGO in #6674
• fix(channels): classify Qwen TTS models correctly by @seefs001 in #6711
• feat(channels): add auto-disable-only channel test mode by @seefs001 in #6728
• perf(web): debounce server and large-list searches by @seefs001 in #6727
• fix: record reasoning effort consistently in usage logs by @seefs001 in #6641
• feat(relay): expose user and group context to parameter override by @seefs001 in #6534
• fix(ollama): preserve reasoning and tool-call context by @seefs001 in #6605
• fix: backend length validation by @seefs001 in #5548
• feat(billing): highlight matched conditional multipliers in logs by @seefs001 in #6561
• fix(web): require confirmation before rotating access token by @seefs001 in #6749
• fix(relay): preserve presence/frequency penalty in Responses conversion by @UzkiS in #6654
• refactor(responses): remove compact model suffix handling by @seefs001 in #6770
• fix: mobile sidebar by @seefs001 in #6760
• fix(web): refresh Turnstile token after login attempt by @thiswod in #6764
• fix: 修复 Gemini 风格 /v1/models 列表请求 by @zhibisora in #6199
• fix(billing): 异步任务退款时同步减少 used_quota by @wans10 in #6795
• fix(web): keep fetched model selection in sync with form by @enwaiax in #6841
• fix(ali): 修复阿里图片模型映射后仍使用原始模型名判断协议的问题 by @wxwwt in #6772
• chore(deps): bump builder-util-runtime and electron-builder in /electron by @dependabot[bot] in #6467
• chore(deps): bump dompurify from 3.4.11 to 3.4.13 in /web by @dependabot[bot] in #6735
• chore(deps-dev): bump fast-uri from 3.1.4 to 3.1.5 in /electron by @dependabot[bot] in #6846
• chore(deps-dev): bump tar from 7.5.16 to 7.5.22 in /electron by @dependabot[bot] in #6468
• fix(topup): reject uncreditable orders before payment by @enwaiax in #6845
• chore(deps-dev): bump js-yaml from 4.3.0 to 4.3.1 in /electron by @dependabot[bot] in #6704
• chore(deps-dev): bump electron from 39.8.5 to 39.8.10 in /electron by @dependabot[bot] in #6705
• fix: prompt_cache_key openai chat -> openai responses by @seefs001 in #6861
• feat: add field passthrough controls for gateway channels by @seefs001 in #6847
• fix(relay): stop injecting empty tools into Claude requests that omit them by @zcxads666 in #6837
• fix(oauth): align custom binding response fields in frontend by @seefs001 in #6818
• test(web): standardize frontend tests on Vitest by @QuentinHsu in #6569
• fix(relaykit): preserve parameterless tools in Claude conversion by @seefs001 in #6862
• refactor: advanced custom channel route editor by @seefs001 in #6865
• feat: channel test by @seefs001 in #6917
• feat(web): fade in streamed response words and harden playground editor by @RedwindA in #6895
• fix: settle Responses cached token usage by @LiaoQi98 in #6892

New Contributors

• @ENCHIGO made their first contribution in #6674
• @UzkiS made their first contribution in #6654
• @thiswod made their first contribution in #6764
• @zhibisora made their first contribution in #6199
• @enwaiax made their first contribution in #6841
• @zcxads666 made their first contribution in #6837
• @LiaoQi98 made their first contribution in #6892

Full Changelog: v1.0.0-rc.24...v1.0.0-rc.25

Release: v1.0.0-rc.24

New Features

• Added user critical rate limiting middleware for access token and affiliate transfer routes to enhance security and stability.
• Refined the categorization of fetched models in channels to improve organization and retrieval.

Bug Fixes

• Resolved an issue with redemption code quotas experiencing precision loss, ensuring accurate credit allocation (#6685).
• Addressed an HTTP/2 transport issue where upstream stream resets could prevent transparent retries by correctly setting Request.GetBody in the relay (#6249).
• Ensured Claude and Gemini endpoints are tested with the native request format for improved compatibility.

Other

• Enhanced the release synchronization workflow with optional file syncing capabilities in CI.

---

What's Changed

• fix(relay): set Request.GetBody so the HTTP/2 transport can transparently retry after an upstream stream reset by @BlockchainHe in #6249
• feat(channels): refine fetched model categorization by @seefs001 in #6632
• fix: test Claude/Gemini endpoints with native request format by @seefs001 in #6698
• fix: 修复兑换码额度精度损失(#6680) by @lihu-001 in #6685

New Contributors

• @BlockchainHe made their first contribution in #6249
• @lihu-001 made their first contribution in #6685

Full Changelog: v1.0.0-rc.23...v1.0.0-rc.24

Highlights: This release expands provider and channel compatibility with New API channels, DeepSeek Responses API support, zstd request decompression, and automatic group routing. It also improves operational accuracy with stream visibility in logs, more reliable tiered-retry billing, and upstream Bedrock cancellation when clients disconnect.

New Features

• Added New API channel support with Gemini-to-OpenAI streaming conversion, Advanced Custom Responses Compact handling, and multipart image edit support (#6559).
• Added DeepSeek Responses API support (#6562).
• Added automatic group selection for request routing (#6590).
• Added per-channel HTTP transport controls so administrators can tune connection behavior per channel.
• Added support for zstd-compressed requests (#6545).
• Log owners can now see whether a request was streamed (#6558).

Bug Fixes

• Qwen thinking_budget is now preserved when requests are relayed (#5836).
• Tiered retry billing now settles usage against the final selected group, including group-switch retries (#6570).
• Bedrock requests are now cancelled when clients disconnect, reducing wasted upstream work (#6589).

Improvements

• Public header navigation links now use a consistent font size (#6557).

---

What's Changed

• refactor: extract protocol conversion layer into standalone relaykit module by @Calcium-Ion in #6369
• fix: 慢查询/错误 SQL 日志参数化 by @feitianbubu in #6493
• feat(oidc): 支持自定义 OIDC 登录显示名称 by @mrchi in #6012
• fix: preserve Qwen thinking_budget passthrough by @wxwwt in #5836
• feat: support zstd request decompression by @seefs001 in #6545
• feat(logs): expose stream status to log owners by @seefs001 in #6558
• fix(relay): preserve multipart image edits for New API channels by @seefs001 in #6559
• fix(oauth): stop treating a foreign window.opener as a bind flow by @neimaravila in #6425
• style: unify public header nav link font size to 14px (text-sm) by @feitianbubu in #6557
• feat: deepseek responses api by @seefs001 in #6562
• fix(billing): settle tiered retries with final group by @faithleysath in #6518
• fix(billing): harden tiered retry group-switch billing (follow-up to #6518) by @Calcium-Ion in #6570
• fix(aws): cancel Bedrock requests on client disconnect by @Calcium-Ion in #6589
• Feat/auto group by @Calcium-Ion in #6590

New Contributors

• @mrchi made their first contribution in #6012
• @neimaravila made their first contribution in #6425
• @faithleysath made their first contribution in #6518

Full Changelog: v1.0.0-rc.22...v1.0.0-rc.23

This release expands channel and model administration with upstream model discovery, new Gemini image models, configurable billing options, and Sub2API support. It also improves reliability across streaming, pricing, refunds, user management, and admin UI workflows.

New Features

• Channels can now discover available models from advanced custom and Codex-compatible upstreams (#6184).
• Added support for the latest Gemini image GA models (#6371).
• Added configurable tool pricing with Sub2API channel support and alpha-search billing.
• OpenAI Realtime GA models are now available with improved routing compatibility (#6032).

Bug Fixes

• Responses-to-Chat streaming no longer emits duplicate tool calls (#6225).
• Creating a model with an existing name no longer clears its saved pricing (#6365).
• The unset-price models view now lists only models available from the selected channels (#6126).
• Very large user quota values are now handled without overflow (#6134).
• User list sorting now works correctly with paginated results (#6194).
• MiniMax models are now correctly associated with the MiniMax vendor (#6164).
• Playground model group selectors now handle auto-group listings and long options correctly (#6163).
• Task refunds are no longer duplicated during concurrent status updates (#6074).
• Top-up amount fields can now be cleared completely while editing (#6473).
• Channel inline priority updates are more reliable (#6415).

Improvements

• Admin JSON editing now has a more consistent experience and avoids UI issues that could hide content (#6421).
• Users page search is debounced to reduce unnecessary requests and improve responsiveness (#6474).
• Proxy-backed channel requests are more compatible and stable (#6157).
• Upstream error logs now provide more useful fallback details when providers return unclear errors (#6431).

---

What's Changed

• fix: list only channel models in unset price models tab by @feitianbubu in #6126
• fix: purge authentication data on hard user deletion by @seefs001 in #6168
• fix(users): prevent large quota values from overflowing by @QuentinHsu in #6134
• fix: infer MiniMax vendor for MiniMax models by @fux-dev in #6164
• fix: add server-side sorting to user list, prevent client-side sort on paged data by @DawnMoon1542 in #6194
• fix(data-table): update table body immediately when toggling columns by @DawnMoon1542 in #6253
• fix: prevent duplicate tool calls in Responses-to-Chat streaming by @nikuleo in #6225
• feat(channel): support Codex upstream model discovery by @seefs001 in #6184
• feat: support upstream model fetch for advanced custom channels by @seefs001 in #5971
• refactor(auth): replace dashboard sessions with stateless tokens and session control by @Calcium-Ion in #6329
• fix: 中转 gpt-realtime 系列 GA 模型时不再发送 realtime beta 标识 by @feitianbubu in #6032
• fix(playground): prevent model group option stretching by @QuentinHsu in #6120
• perf(data-table): optimize action column width allocation by @QuentinHsu in #6135
• fix(web): show used quota for unlimited API keys by @RedwindA in #6224
• fix(channel): improve proxy client compatibility and cache lifecycle by @yyhhyyyyyy in #6157
• fix(playground): resolve auto group model listing by @QuentinHsu in #6163
• fix: prevent duplicate suno task refunds via cas status update by @feitianbubu in #6074
• chore(github): migrate issue templates to required forms by @seefs001 in #6452
• fix(github): restore issue form visibility by @seefs001 in #6454
• fix: log response body when parsed upstream error message is empty by @bigsongeth in #6431
• fix: prevent model create from wiping existing pricing for same name by @feitianbubu in #6365
• fix(web/channel): stabilize inline priority updates by @RedwindA in #6415
• perf(json-editor): unify admin JSON editing experience by @QuentinHsu in #6421
• feat: add gemini-3-pro-image and gemini-3.1-flash-image GA models by @feitianbubu in #6371
• feat: support Tencent TokenHub API key via OpenAI-compatible protocol by @feitianbubu in #6232
• fix: allow topup amount input to be fully cleared by @feitianbubu in #6473
• fix: debounce users page search input by @feitianbubu in #6474

New Contributors

• @fux-dev made their first contribution in #6164
• @DawnMoon1542 made their first contribution in #6194
• @nikuleo made their first contribution in #6225
• @bigsongeth made their first contribution in #6431

Full Changelog: v1.0.0-rc.21...v1.0.0-rc.22

Highlights

The headline change is GPT-5.6 cache write billing: OpenAI cache_write_tokens are now charged at the cache-creation rate, with accounting safeguards around cached and uncached prompt tokens. This release also improves pricing workflows, Playground controls, channel table usability, and billing/logging reliability.

New Features

• Added GPT-5.6 cache write billing support, charging OpenAI cache_write_tokens at the cache-creation rate while keeping cached and uncached prompt token accounting consistent.
• Pricing pages now support group-aware dynamic calculations and avoid showing sample-only special groups (#5906).
• Model pricing settings now include an Unset Price Models tab plus status and sync filters, making it easier to find and configure models without prices (#6124).
• The Playground now includes request parameter panels for chat and generation tuning, so users can adjust model behavior directly in the UI (#6044).
• Channel lists now support manual column resizing and keep resized tables filling the available width (#5948).
• Logs now show stream timing metrics and task detail views for clearer troubleshooting.
• Added enhanced text protocol conversion and advanced custom routing capabilities for more flexible channel routing (#5825).

Bug Fixes

• Browser translation can no longer corrupt React-rendered pages (#5963).
• Custom model names that differ only by case can now be added consistently, including from multi-select controls (#6061).
• Billing now handles quota pre-consumption and image-stream disconnect edge cases more accurately.
• Subscription reset handling is more reliable under concurrent updates (#6057).
• Codex response passthrough fields now sync correctly (#6018).
• Playground model selector lists now stay synchronized after model changes.
• Custom home iframes can trigger top navigation after a user action as intended (#5955).
• Restored paste behavior for default channel connection settings.

Improvements

• Upstream price sync tables in model pricing now load and render more efficiently (#6092).
• Refined table badges, pricing layouts, sync dialogs, and default theme styling for a cleaner admin UI.
• Font assets now resolve correctly across workspace layouts.

---

What's Changed

• fix(web): prevent browser translation from mutating React roots by @cat0825 in #5963
• feat(web): 支持渠道列表手动调整列宽 by @zuiho-kai in #5948
• fix(web): let channel table fill available width by @zuiho-kai in #6031
• chore: remove dead files resurrected by v1.0 launch commit by @feitianbubu in #6041
• fix: use GORM v2 row locking for subscription resets by @seefs001 in #6057
• feat(playground): add playground parameter settings panel by @QuentinHsu in #6044
• fix: allow adding model names that differ only by case in multi-select by @feitianbubu in #6061
• fix: remove sample special usable groups leaking into pricing page by @feitianbubu in #5906
• perf(model-pricing): optimize upstream price sync tables by @QuentinHsu in #6092
• fix(web): allow user-activated top navigation for custom home iframe by @Heykode in #5955
• chore(deps): bump golang.org/x/crypto from 0.51.0 to 0.52.0 by @dependabot[bot] in #6096
• fix: sync codex responses passthrough fields by @seefs001 in #6018
• feat: enhance text protocol conversion and advanced custom routing by @Calcium-Ion in #5825
• feat: add unset price models tab to model pricing settings by @feitianbubu in #6124

New Contributors

• @cat0825 made their first contribution in #5963

Full Changelog: v1.0.0-rc.20...v1.0.0-rc.21

Highlights

This release adds GPT-5.6 pricing support and stale instance cleanup tools for administrators. It also fixes several quota and pricing edge cases, including wallet reward transfers and decimal ratio input.

New Features

• Added token pricing support for GPT-5.6 models.
• Added system-info actions for cleaning up stale instances. (#5953)

Bug Fixes

• Wallet reward transfers now respect the configured quota unit instead of using an incorrect fixed unit. (#5808)
• Group ratio fields now allow decimal values such as 0.05 to be typed normally. (#5995)
• Tiered pricing now uses a default token estimate during pre-consumption, improving quota calculation for affected pricing rules.

---

What's Changed

• fix: 分组倍率 Ratio 输入框保留小数中间态,修复无法输入 0.05 by @feitianbubu in #5995
• feat(system-info): add stale instance cleanup actions by @seefs001 in #5953
• fix(wallet): honor configured quota units for reward transfers by @A-Words in #5808

New Contributors

• @A-Words made their first contribution in #5808

Full Changelog: v1.0.0-rc.19-i18nfix.2...v1.0.0-rc.20

New Features

• Admins can now reset subscription quotas and review quota saturation events for auditing, making quota management easier (#5952).
• Added Traditional Chinese (zh-TW) localization for the new UI (#5942).

Bug Fixes

• Fixed billing quota calculations for very large quotas so limits are applied correctly.

---

What's Changed

• Localized new ui to zh-TW by @MeowTechOpenSource in #5942
• feat(subscription): add admin quota reset actions by @seefs001 in #5952

New Contributors

• @MeowTechOpenSource made their first contribution in #5942

Full Changelog: v1.0.0-rc.18...v1.0.0-rc.19

v1.0.0-rc.18

Compare

Choose a tag to compare