fix(models): make pro tier include dynamically merged cloud models by dd373156 · Pull Request #1 · dwgx/WindsurfAPI

dd373156 · 2026-04-17T07:29:21Z

Problem

MODEL_TIER_ACCESS.pro currently references ALL_MODEL_KEYS, a snapshot of Object.keys(MODELS) taken at module load time. When mergeCloudModels() adds new models to the MODELS dict after startup (e.g. the five claude-opus-4-7-* entries from GetCascadeModelConfigs), they never appear in pro's frozen array.

Downstream this bites:

auth.js:getAvailableModelsForAccount() filters getTierModels(tier), so cloud-added models are missing from every account's availableModels.
handlers/chat.js preflight returns 403 model_not_entitled for every request targeting a cloud-added model, even on pro accounts that legitimately have access.

Repro

On a fresh pro account after mergeCloudModels has run:

# Catalog sees the models (listModels enumerates live MODELS each call)
curl -sS http://localhost:3003/v1/models | jq '.data[].id' | grep opus-4-7
# → claude-opus-4-7-low, -medium, -high, -xhigh, -none

# Account's availableModels does NOT (goes through getTierModels → ALL_MODEL_KEYS snapshot)
curl -sS http://localhost:3003/auth/accounts \
  | jq '.accounts[0].availableModels | map(select(contains("opus-4-7")))'
# → []   ← the mismatch

# And chat preflight 403s every request
curl -sS http://localhost:3003/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"claude-opus-4-7-high","messages":[{"role":"user","content":"hi"}]}'
# → {"error":{"message":"模型 claude-opus-4-7-high 在当前账号池中不可用（未订阅或已被封禁）","type":"model_not_entitled"}}

Fix

Replace the frozen-array reference with a getter so every access re-enumerates the live MODELS dictionary:

   export const MODEL_TIER_ACCESS = {
-    pro: ALL_MODEL_KEYS,
+    get pro() { return Object.keys(MODELS); },
     free: FREE_TIER_MODELS,
     unknown: FREE_TIER_MODELS,
     expired: [],
   };

One-line minimal diff. ALL_MODEL_KEYS is left declared (now unused) to keep the patch as small as possible for review; feel free to drop the dead variable in a follow-up cleanup if desired.

Verified

Applied on my instance — claude-opus-4-7-high and the other four opus-4-7-* entries now route correctly. Every future mergeCloudModels addition will flow through without further code changes.

Thanks for the excellent project!

ALL_MODEL_KEYS is a module-load-time snapshot. Models added later via mergeCloudModels (e.g. the opus-4-7 family) never make it into the snapshot, so getTierModels('pro') returns a stale list and the chat preflight returns 403 model_not_entitled for every cloud-added model. Replace the frozen-array reference with a getter so every call re-enumerates the live MODELS dictionary.

- #3 Firebase login: OAuth accounts now get clear guidance to use token method - #5 Long prompt timeout: cold stall detection adapts to input length (30s-90s) - PR #1: pro tier now dynamically includes cloud-merged models (getter) - README: rewritten in casual style, added FAQ, setup.sh one-click init - Dashboard: added OAuth hint on login panel - .gitignore: added codeium_ext/ codeium.zip *.db bugsy/ windsurf-grpc/

Dashboard 侧栏新增"关于 → 致谢"面板，列出外部贡献者： - dd373156 — PR #1 修 Pro 层模型合并 - colin1112a — PR #13 一次性审 15 个 bug 卡片含 GitHub 头像（github.com/:user.png）、PR 链接、合并日期、改动说明。CONTRIBUTORS 数组手动维护，后续有 PR 直接加一项就能渲染。底部放"提 issue / 提 PR"按钮鼓励继续贡献。 README.md / README.en.md 也加了对应的 Contributors 段放在 MIT 前。

PR #1 是 owner 还不熟 GitHub 流程时误开 PR 被自动 merge 的，不是主动审核后认可的贡献。从致谢面板和 README 移除。

fix(models): make pro tier include dynamically merged cloud models

- #3 Firebase login: OAuth accounts now get clear guidance to use token method - #5 Long prompt timeout: cold stall detection adapts to input length (30s-90s) - PR #1: pro tier now dynamically includes cloud-merged models (getter) - README: rewritten in casual style, added FAQ, setup.sh one-click init - Dashboard: added OAuth hint on login panel - .gitignore: added codeium_ext/ codeium.zip *.db bugsy/ windsurf-grpc/

Dashboard 侧栏新增"关于 → 致谢"面板，列出外部贡献者： - dd373156 — PR #1 修 Pro 层模型合并 - colin1112a — PR #13 一次性审 15 个 bug 卡片含 GitHub 头像（github.com/:user.png）、PR 链接、合并日期、改动说明。CONTRIBUTORS 数组手动维护，后续有 PR 直接加一项就能渲染。底部放"提 issue / 提 PR"按钮鼓励继续贡献。 README.md / README.en.md 也加了对应的 Contributors 段放在 MIT 前。

PR #1 是 owner 还不熟 GitHub 流程时误开 PR 被自动 merge 的，不是主动审核后认可的贡献。从致谢面板和 README 移除。

fix(models): make pro tier include dynamically merged cloud models

- #3 Firebase login: OAuth accounts now get clear guidance to use token method - #5 Long prompt timeout: cold stall detection adapts to input length (30s-90s) - PR #1: pro tier now dynamically includes cloud-merged models (getter) - README: rewritten in casual style, added FAQ, setup.sh one-click init - Dashboard: added OAuth hint on login panel - .gitignore: added codeium_ext/ codeium.zip *.db bugsy/ windsurf-grpc/

Dashboard 侧栏新增"关于 → 致谢"面板，列出外部贡献者： - dd373156 — PR #1 修 Pro 层模型合并 - colin1112a — PR #13 一次性审 15 个 bug 卡片含 GitHub 头像（github.com/:user.png）、PR 链接、合并日期、改动说明。CONTRIBUTORS 数组手动维护，后续有 PR 直接加一项就能渲染。底部放"提 issue / 提 PR"按钮鼓励继续贡献。 README.md / README.en.md 也加了对应的 Contributors 段放在 MIT 前。

PR #1 是 owner 还不熟 GitHub 流程时误开 PR 被自动 merge 的，不是主动审核后认可的贡献。从致谢面板和 README 移除。

…/v1/responses + /v1/messages spec gaps 7 fixes from a project-wide gpt-5.5 audit: #59 sub-bug 3 — tool-boundary text split (P1 真因 = parser bug) - ToolCallStreamParser.feed() 之前 `{text, toolCalls}` 两个数组返回，丢失 text/tool 相对顺序 - 改成同时返回 `items: [{type:'text',text}|{type:'tool_call',toolCall}]` 保留顺序 - chat.js 流式消费方按 items 顺序 emit，不再先发全部 tool 再发文本 - 老 `text/toolCalls` 字段保留向后兼容 #66 — 300秒限速误报 - rateLimitCooldownMs 解析具体的 retry-after N seconds/minutes/hours，不再一刀切 5min - markRateLimited 改成 max-extend 而不是覆盖，并发 429 不会把 cooldown 不断后推 - preflight checkMessageRateLimit 没拿到 retryAfterMs 时不再本地标 cooldown，本次 skip 即可 - windsurf-api.js 透传上游 retryAfterMs #63 follow-up + P1 #2 — /v1/responses 规范缺口 - 非 function tools (web_search_preview 等) 静默 drop → 改成 400 直接拒 - function-call-only 响应不再带空 message item 在 output P1 #1 — /v1/messages 丢失 thinking + tool_choice - anthropicToOpenAI 透传 body.thinking - Anthropic tool_choice (auto/any/tool/none) 映射到 OpenAI 形状 P1 #3 — 全账号 RPM 满返回 429 不是 503 - isAllTemporarilyUnavailable 聚合 rate_limit / model_rate_limit / rpm_full / strict_reuse_busy - 非流式路径在 queue 超时后返回 429 + Retry-After，503 只在真没账号时返 - 流式 SSE 头已发，body error type 至少标对 rate_limit_exceeded P1 #4 — preflight skip 还在吃本地 RPM headroom - account 加 _lastReservationAt + getApiKey 返回 reservationTimestamp - refundReservation(apiKey, ts) 把最近一条 _rpmHistory 退回去 - preflight !hasCapacity 路径自动调用，避免被跳过的账号占住本地配额 P2 cache-hit chunk 不一致 - cache HIT 流式分支拆成 finish_reason chunk + 单独 usage chunk，跟 live-stream 路径同 shape 测试： - 12 个新单测 + 修了既有的 4 个，全套 158 个测试通过 - 新文件: test/rate-limit.test.js, test/messages.test.js, test/chat-cache-hit.test.js - 改动: test/tool-emulation.test.js (items 顺序), test/responses.test.js (unsupported tool 400 + empty msg) 来源审计报告: tmp/audit-report-2026-04-26.md (gpt-5.5 high reasoning 出的 296 行 P0/P1/P2 全列)

…rate-limit, /v1/responses + /v1/messages spec gaps 7 fixes from a project-wide gpt-5.5 audit: dwgx#59 sub-bug 3 — tool-boundary text split (P1 真因 = parser bug) - ToolCallStreamParser.feed() 之前 `{text, toolCalls}` 两个数组返回，丢失 text/tool 相对顺序 - 改成同时返回 `items: [{type:'text',text}|{type:'tool_call',toolCall}]` 保留顺序 - chat.js 流式消费方按 items 顺序 emit，不再先发全部 tool 再发文本 - 老 `text/toolCalls` 字段保留向后兼容 dwgx#66 — 300秒限速误报 - rateLimitCooldownMs 解析具体的 retry-after N seconds/minutes/hours，不再一刀切 5min - markRateLimited 改成 max-extend 而不是覆盖，并发 429 不会把 cooldown 不断后推 - preflight checkMessageRateLimit 没拿到 retryAfterMs 时不再本地标 cooldown，本次 skip 即可 - windsurf-api.js 透传上游 retryAfterMs dwgx#63 follow-up + P1 dwgx#2 — /v1/responses 规范缺口 - 非 function tools (web_search_preview 等) 静默 drop → 改成 400 直接拒 - function-call-only 响应不再带空 message item 在 output P1 dwgx#1 — /v1/messages 丢失 thinking + tool_choice - anthropicToOpenAI 透传 body.thinking - Anthropic tool_choice (auto/any/tool/none) 映射到 OpenAI 形状 P1 dwgx#3 — 全账号 RPM 满返回 429 不是 503 - isAllTemporarilyUnavailable 聚合 rate_limit / model_rate_limit / rpm_full / strict_reuse_busy - 非流式路径在 queue 超时后返回 429 + Retry-After，503 只在真没账号时返 - 流式 SSE 头已发，body error type 至少标对 rate_limit_exceeded P1 dwgx#4 — preflight skip 还在吃本地 RPM headroom - account 加 _lastReservationAt + getApiKey 返回 reservationTimestamp - refundReservation(apiKey, ts) 把最近一条 _rpmHistory 退回去 - preflight !hasCapacity 路径自动调用，避免被跳过的账号占住本地配额 P2 cache-hit chunk 不一致 - cache HIT 流式分支拆成 finish_reason chunk + 单独 usage chunk，跟 live-stream 路径同 shape 测试： - 12 个新单测 + 修了既有的 4 个，全套 158 个测试通过 - 新文件: test/rate-limit.test.js, test/messages.test.js, test/chat-cache-hit.test.js - 改动: test/tool-emulation.test.js (items 顺序), test/responses.test.js (unsupported tool 400 + empty msg) 来源审计报告: tmp/audit-report-2026-04-26.md (gpt-5.5 high reasoning 出的 296 行 P0/P1/P2 全列)

…/v1/responses + /v1/messages spec gaps 7 fixes from a project-wide gpt-5.5 audit: #59 sub-bug 3 — tool-boundary text split (P1 真因 = parser bug) - ToolCallStreamParser.feed() 之前 `{text, toolCalls}` 两个数组返回，丢失 text/tool 相对顺序 - 改成同时返回 `items: [{type:'text',text}|{type:'tool_call',toolCall}]` 保留顺序 - chat.js 流式消费方按 items 顺序 emit，不再先发全部 tool 再发文本 - 老 `text/toolCalls` 字段保留向后兼容 #66 — 300秒限速误报 - rateLimitCooldownMs 解析具体的 retry-after N seconds/minutes/hours，不再一刀切 5min - markRateLimited 改成 max-extend 而不是覆盖，并发 429 不会把 cooldown 不断后推 - preflight checkMessageRateLimit 没拿到 retryAfterMs 时不再本地标 cooldown，本次 skip 即可 - windsurf-api.js 透传上游 retryAfterMs #63 follow-up + P1 #2 — /v1/responses 规范缺口 - 非 function tools (web_search_preview 等) 静默 drop → 改成 400 直接拒 - function-call-only 响应不再带空 message item 在 output P1 #1 — /v1/messages 丢失 thinking + tool_choice - anthropicToOpenAI 透传 body.thinking - Anthropic tool_choice (auto/any/tool/none) 映射到 OpenAI 形状 P1 #3 — 全账号 RPM 满返回 429 不是 503 - isAllTemporarilyUnavailable 聚合 rate_limit / model_rate_limit / rpm_full / strict_reuse_busy - 非流式路径在 queue 超时后返回 429 + Retry-After，503 只在真没账号时返 - 流式 SSE 头已发，body error type 至少标对 rate_limit_exceeded P1 #4 — preflight skip 还在吃本地 RPM headroom - account 加 _lastReservationAt + getApiKey 返回 reservationTimestamp - refundReservation(apiKey, ts) 把最近一条 _rpmHistory 退回去 - preflight !hasCapacity 路径自动调用，避免被跳过的账号占住本地配额 P2 cache-hit chunk 不一致 - cache HIT 流式分支拆成 finish_reason chunk + 单独 usage chunk，跟 live-stream 路径同 shape 测试： - 12 个新单测 + 修了既有的 4 个，全套 158 个测试通过 - 新文件: test/rate-limit.test.js, test/messages.test.js, test/chat-cache-hit.test.js - 改动: test/tool-emulation.test.js (items 顺序), test/responses.test.js (unsupported tool 400 + empty msg) 来源审计报告: tmp/audit-report-2026-04-26.md (gpt-5.5 high reasoning 出的 296 行 P0/P1/P2 全列)

@dwgx

Operational/navigation links updated to MYMDO/WindsurfAPI: - package.json: homepage, repository.url, bugs.url - install-ls.sh: OUR_RELEASE - update.sh: RELEASE_URL - docker-compose.yml: ghcr.io/mymdo/windsurf-api:latest (lowercased per GHCR) - SECURITY.md: 2x security advisory URLs - .github/ISSUE_TEMPLATE/config.yml: security advisory URL - .github/workflows/release.yml: comment - README.{md,en,ua,zh}.md: clone URLs, GitHub Pages catalog, Issues/PR CTAs - docs/index.html: nav GitHub, hero CTA, deploy clone, contributors CTA, footer (GitHub/Releases/Issues/Security/READMEs/CONTRIBUTING) - src/dashboard/index.html + index-sketch.html: Issue/PR CTA buttons, RELEASE_NOTES blob link KEPT at dwgx (intentional): - Historical PR references (PR dwgx#1, dwgx#13, dwgx#36, dwgx#43, dwgx#44, dwgx#45) — they exist only in dwgx/WindsurfAPI - @dwgx profile link in footer (attribution) - (c) 2026 dwgx (copyright attribution per MIT) - package.json author field (original creator) - bydwgx1337 brand strings in dashboard UI / server provider / version BRAND - contributors.json (login + historical narrative) - test fixtures and code comments referencing dwgx - docs/releases/RELEASE_NOTES_*.md (historical archives)

dwgx merged commit f9783eb into dwgx:master Apr 20, 2026

dwgx added a commit that referenced this pull request Apr 21, 2026

credits: 移除 dd373156

f501f96

PR #1 是 owner 还不熟 GitHub 流程时误开 PR 被自动 merge 的，不是主动审核后认可的贡献。从致谢面板和 README 移除。

dwgx added a commit that referenced this pull request Apr 21, 2026

Merge pull request #1 from dd373156/fix-opus-4-7-tier-access

5613416

fix(models): make pro tier include dynamically merged cloud models

dwgx added a commit that referenced this pull request Apr 21, 2026

credits: 移除 dd373156

5aff219

PR #1 是 owner 还不熟 GitHub 流程时误开 PR 被自动 merge 的，不是主动审核后认可的贡献。从致谢面板和 README 移除。

dwgx added a commit that referenced this pull request Apr 21, 2026

Merge pull request #1 from dd373156/fix-opus-4-7-tier-access

4ff1baf

fix(models): make pro tier include dynamically merged cloud models

dwgx added a commit that referenced this pull request Apr 21, 2026

credits: 移除 dd373156

60af392

PR #1 是 owner 还不熟 GitHub 流程时误开 PR 被自动 merge 的，不是主动审核后认可的贡献。从致谢面板和 README 移除。

lihengcn mentioned this pull request Apr 22, 2026

[Bug] claude code里使用时，中文对话时经常回复英文 #35

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(models): make pro tier include dynamically merged cloud models#1

fix(models): make pro tier include dynamically merged cloud models#1
dwgx merged 1 commit into
dwgx:masterfrom
dd373156:fix-opus-4-7-tier-access

dd373156 commented Apr 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

dd373156 commented Apr 17, 2026

Problem

Repro

Fix

Verified

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants