feat(providers): cc-switch 式供应商自动故障转移（同厂商同模型） - #385
Open
coder-hhx wants to merge 10 commits into
Open
Conversation
…viders
Adds automatic provider failover for chat requests, configured per vendor
tab in provider settings and modeled after cc-switch's proxy-side design:
- Failover switches providers, never models: the failed request is re-sent
to the next provider in the queue with the same model id the conversation
was using; queued providers without that model active are skipped.
- Per-target circuit breaker (consecutive-failure threshold, cooldown with
half-open probe) keyed by provider::model, plus an error classifier that
only fails over on provider-fault-class errors (5xx/network/auth/quota),
never on client-request-class errors (context overflow etc.).
- Stream wrapper buffers events until first committed content so consumers
never see output from a discarded attempt; sticky winner keeps follow-up
rounds on the provider that answered, and onSwitched pins the conversation
selection to it.
- Settings: modelFailover is now persisted (it previously round-tripped
through neither the SQLite nor the localStorage channel, so the toggle
and queue silently reset on every settings reopen); queue entries are
provider ids with legacy {customProviderId, model} entries migrated by
collapsing to their provider id.
- UI: failover card under each vendor tab manages an ordered provider
queue with breaker knobs; i18n zh/en updated, {vendor} placeholder now
replaced at every occurrence.
Covered by unit tests for the breaker/stream wrapper, plan builder, and
settings normalization/migration; verified end-to-end in the running app.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
StackCairn
marked this pull request as draft
August 5, 2026 13:11
Contributor
|
PR governance checks passed. Awaiting human review. |
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
coder-hhx
marked this pull request as ready for review
August 5, 2026 13:27
Contributor
|
没啥希望通过, 仓库所有者表示希望更简洁的provider, 如有此需求推荐安装一个cc-switch |
Contributor
|
pr可以留着, 之后有插件体系了可以放插件里 |
…s drawer The failover card no longer renders under each vendor's provider list; it now lives in the drawer behind the gear button in the providers header, scoped to the active vendor tab. The drawer is renamed from "Custom Settings" to "Advanced Settings" (自定义设置 → 高级设置) in zh/en on both the GUI and WebUI ends. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ouoiouo
pushed a commit
to ouoiouo/LiveAgent
that referenced
this pull request
Aug 8, 2026
…pstream PR Stack-Cairn#385) - 修复 modelFailover 设置无法持久化的问题（接入 localStorage 通道） - cc-switch 语义供应商级故障转移：只换供应商不换模型，未激活跳过 - provider::model 维度熔断器（失败阈值→熔断→冷却→半开探测） - 流包装器在首个可见内容前缓冲，丢弃尝试对用户不可见 - 成功应答供应商粘性保持并写回对话选择
ouoiouo
pushed a commit
to ouoiouo/LiveAgent
that referenced
this pull request
Aug 8, 2026
The 3-column grid was cramped inside the 440px advanced-settings drawer; one field per row reads better there. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The hint ("未选择时，标题生成会使用当前对话使用的模型") described the
title-model picker but sat under the drawer heading; it now renders
between the 标题生成模型 label and the picker on both ends.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Member
|
审核结论：当前不建议直接合入，存在 1 个功能阻断。 阻断：WebUI 的
|
5 tasks
…ync it localStorage made the failover config per-webview state, so the GUI and WebUI could silently diverge. Move it to the desktop config DB and the gateway sync protocol so both ends share one copy: - Rust: new model_failover_settings table (same single-row payload_json shape as memory_settings), settings_load_all returns it, settings_save_model_failover persists it, and the gateway settings snapshot includes it so remote WebUI clients receive and can update it. - GUI storage.ts: loads from SQLite (falling back to the pre-SQLite localStorage copy for migration) and saves through the new command; the localStorage channel no longer carries modelFailover. - sync.ts (both ends): modelFailover joins the sync payload type, the diffed field list, the full-snapshot builder, and the incoming-payload apply, so edits from either end propagate like memory/mcp settings. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Regression tests for the three scenarios raised in review: - a modelFailover edit shows up in buildGatewaySettingsSyncUpdatePayload (and an unchanged config produces no entry), - the config round-trips build → apply across ends, with a partial payload leaving the receiver's config untouched, - a config applied from a synced payload feeds buildModelFailoverPlan and yields real fallback candidates. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Collaborator
Author
|
@su-fen 感谢细致的审核。你指出的阻断在你审核的 head（
麻烦基于最新 head 复审。 🤖 Generated with Claude Code |
Member
|
审查结论：暂不建议直接合入，建议 Request changes。 发现两个需要先修复的行为问题：
验证情况：GitHub checks 全绿；PR 与最新 |
… budget withProviderFailover sized the attempt window as maxSwitches + 1 regardless of which candidate the plan started on. When the primary's breaker was already open, the plan began on the first fallback — a switch that had effectively happened — yet the budget still allowed maxSwitches more moves, so maxSwitches=1 could walk through two fallbacks (two real switches). Charge that initial skip: when attemptPlan[0] is not the primary, the window shrinks by one, so budget semantics match the user-visible number of provider changes. Covered by two new tests: the skip exhausting a budget of 1, and a budget of 2 still reaching the second fallback after the skip. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The failover plan built in useSendChatTurn was only handed to the agent branch, so with executionMode=text an enabled failover queue was dead config: the request failed on the primary provider without ever trying the queue. Wire the same per-turn plan into the text runtime: - textOnlyRuntime.streamAssistantMessage accepts an optional failover param and wraps each attempt in withProviderFailover, mirroring the agent runner's per-round candidate wiring (lazy fallback proxy/model preparation, sticky winner across recovery turns, breaker-open candidates skipped by the wrapper). - runTextConversationTurn threads the plan through, surfaces the "switching to X" status line during a switch, and forwards onSwitched so the winning fallback becomes the conversation's selection exactly like agent mode. - useSendChatTurn passes the already-built failoverParams to the text branch too. Regression tests drive streamAssistantMessage against a mocked streamSimpleByApi: uncommitted provider fault fails over to the queued provider (discarded attempt invisible, callbacks fired), committed content and client-class errors never switch, and the no-failover path is unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Linked issue
Closes #384
Summary
修复自动故障转移设置无法持久化的 bug，并将故障转移重构为 cc-switch 语义的供应商级设计：
modelFailover此前未接入任何持久化通道，退出设置页重进即丢失。现持久化到桌面配置库新表model_failover_settings（settings_load_all读取、settings_save_model_failover写入），并加入 Gateway settings-sync 协议（payload 类型 / 字段白名单 / build / apply 四处），GUI 与 WebUI 共享同一份配置：桌面改动推快照给 WebUI，WebUI 改动经SettingsUpdate通道落桌面 SQLite 并回推。localStorage 仅保留一次性旧数据迁移回退。provider::model维度的熔断器（连续失败阈值 → 熔断 → 冷却 → 半开探测）；错误分类器仅对供应商故障类错误（5xx/网络/鉴权/配额）转移，客户端请求类错误（如上下文超限）不转移；流包装器在首个可见内容前缓冲事件，丢弃的尝试对用户不可见；成功应答的供应商粘性保持并写回对话选择。{vendor}占位符多次出现时只替换首个的问题。Change scope
crates/agent-gui/src-tauri/src/commands/config/settings/（新表model_failover_settings+ 读写命令 + 网关快照）crates/agent-gui/src/lib/providers/runtime/providerFailover.ts（新增：熔断器 + 错误分类 + 流包装器）crates/agent-gui/src/lib/settings/{index,storage,sync}.ts（队列模型 + 迁移 + SQLite 持久化 + 同步协议）crates/agent-gui/src/pages/chat/runtime/{providerRuntimeConfig,useSendChatTurn}.ts（每轮 failover 计划）crates/agent-gui/src/lib/chat/runner/agentRunner.ts（候选装配 + withProviderFailover 接入）crates/agent-gui/src/pages/settings/ProvidersSection.tsx+ web 端孪生 + 双端 i18n / sync.tsScreenshots / preview
供应商配置右上角设置按钮 →「高级设置」抽屉——供应商级故障转移队列：
实测转移过程：将 P1 供应商（RightCode sale）的 API key 置为无效后发起对话，状态栏显示「第 1 轮：RightCode（sale）· claude-fable-5 不可用，正在切换到 RightCode · claude-fable-5…」，随后由 P2 供应商以同一模型完成应答，对话选择自动切到应答的供应商。
Verification
cargo check/cargo test --lib initialize_schema✅（新表 schema 测试通过）cd crates/agent-gui && ./node_modules/.bin/tsc --noEmit✅（web 端同样通过）biome check两端 0 error ✅node --test test/settings/model-failover-normalize.test.mjs test/providers/provider-failover-plan.test.mjs test/providers/provider-failover.test.mjs— 30 项通过（熔断器/流包装器 16、计划构建 6、归一化+迁移 8）modelFailover变更进buildGatewaySettingsSyncUpdatePayload()；双端 build→apply round-trip（部分 payload 不覆盖现有配置）；同步来的配置可被buildModelFailoverPlan()构造实际候选 — GUInormalization.test.mjs+ webweb-settings.test.mjs+provider-failover-plan.test.mjsnode scripts/check-mirror.mjs✅（115 文件）pnpm tauri dev启动桌面端，复现原 bug 路径（退出设置→重进），配置保留；旧数据自动迁移；实际断供触发转移成功（见截图说明）Pre-submit checklist
🤖 Generated with Claude Code