revert(core): revert Protocol enum & model-identity decoupling (#5089) - #5745
Conversation
|
Thanks for the PR, @wenshao. Template looks good ✓ — all required sections present, bilingual, test plan is concrete. On direction: this is reverting #5089, which was merged in error per the PR body. That's a valid motivation, but it's also backing out a shipped, foundational refactor — which means reopening at least six previously-closed issues (#5090, #4877, #5080, #4814, #4813, #4722) that #5089 claimed to resolve. Touching the model-identity / provider-selection core like this feels maintainer-territory enough that I'm flagging it for human judgment on the tradeoff rather than rubber-stamping either way. On approach: the two-commit structure is clean — (1) revert + re-adapt dependents, (2) add v5→v4 downgrade migration. The migration is the right call: anyone who ran v0.19.0 already has Scope-wise the diff is net -882 lines across 78 files, which is consistent with a revert. But I spotted two small pieces of drive-by formatting churn that don't belong in a revert:
Both are tiny, but a revert should be mechanically minimal. Worth splitting those out (or just dropping them) so the diff stays a pure structural revert. Moving on to code review. 🔍 中文说明感谢 @wenshao 的 PR。 模板完整 ✓ — 所有章节齐全，双语，测试计划具体。 方向：本 PR 回退被描述为"错误合入"的 #5089。动机合理，但同时也回退了一个已经发布的基础性重构 —— 意味着重新打开至少 6 个此前已被关闭的 issue（#5090、#4877、#5080、#4814、#4813、#4722）。涉及模型标识 / provider 选择的核心结构，个人判断这已经进入需要维护者拍板的范围，因此标记为需要人工判断，而不是直接通过或拒绝。 方案：两次提交的结构干净 ——（1）回退 + 适配依赖项，（2）新增 v5→v4 降级迁移。迁移是正确的选择：任何跑过 v0.19.0 的用户已经持有 范围上 diff 净 -882 行、涉及 78 个文件，符合回退的预期。但发现两处不属于回退的顺手格式改动：
都很小，但回退应当机械地最小化。建议把这些拆出去（或直接丢弃），让 diff 保持纯粹的结构性回退。 进入代码审查 🔍 — Qwen Code · qwen3.7-max |
qqqys
left a comment
qqqys
left a comment
There was a problem hiding this comment.
Reviewed the revert/downgrade migration path. Focused migration/model-provider tests passed; only noted a minor formatting issue in BaseTextInput.tsx.
wenshao
left a comment
wenshao
left a comment
There was a problem hiding this comment.
Summary: 1 Critical finding + 4 low-confidence Suggestions (terminal only).
Critical: VS Code findOpenaiModels drops V5 { protocol, models } wrapper handling. The extension writes settings.json directly without running the CLI migration — if a user's file is still at $version: 5 when the extension writes config, existing OpenAI models are silently lost.
Build ✅ | 1091 tests ✅ | TypeCheck: 0 new findings.
— qwen3.7-max via Qwen Code /review
|
+1 on reverting #5089 — I just hit the exact crash this revert guards against, so it's real, not theoretical. What happened: v0.19.0 (which carries #5089) auto-migrated my The painful part is the migration is one-way: once v0.19.0 touches the file, every older binary reading it crashes, with no way back short of this revert. Plain version skew between the binary that wrote the config and the one reading it is enough to trigger it (in my case an older standalone install shadowing a newer npm one on So this revert plus the v5→v4 downgrade migration is the right call. One thing worth a line in the release note: the downgrade only kicks in on builds that include this PR, so users pinned to a pre-#5089 binary still have to upgrade past it first. |
Code ReviewIndependent proposal (before reading the diff): To revert #5089's structural changes while preserving post-#5089 features, I would: (1) restore Comparison with the PR: The approach matches my proposal exactly. Every structural element is reverted cleanly, the migration is well-designed, and the dependent PRs are adapted without losing their features. Findings:
Test ResultsReal-Scenario Test: v5→v4 Auto-MigrationCreated After exit, settings.json was rewritten to {
"$version": 4,
"modelProviders": {
"openai": [
{
"id": "gpt-4o",
"name": "GPT-4o",
"baseUrl": "https://api.openai.com/v1",
"envKey": "OPENAI_API_KEY"
}
]
}
}中文说明代码审查独立方案（读 diff 之前）： 要回退 #5089 的结构性改动同时保留之后的功能，我会：（1）把 与 PR 的对比： 方案完全吻合。每个结构要素都干净地回退了，迁移设计良好，依赖的 PR 在不丢失功能的前提下适配完成。 发现：
测试结果真实场景测试：v5→v4 自动迁移在 启动正常，无 — Qwen Code · qwen3.7-max |
|
This PR does exactly what it says and does it well. The structural revert is mechanically clean — The convergence architecture ( Two cosmetic changes (desktop-pet table formatting, qwen-serve trailing whitespace) are noise in a revert PR, but not worth blocking a time-sensitive revert over. The maintainer (@qqqys) has already signaled intent to merge this quickly. 1092 unit tests pass. Build and typecheck clean. Migration verified end-to-end in tmux. LGTM. ✅ 中文说明这个 PR 做到了它所承诺的一切。结构性回退机械地干净 —— 汇聚架构（ 两处装饰性改动（desktop-pet 表格格式化、qwen-serve 尾随空白）是回退 PR 中的噪音，但不值得用来阻断一个时间敏感的回退。维护者（@qqqys）已经表态要快速合入。 1092 个单元测试通过。构建和类型检查全部通过。迁移在 tmux 中端到端验证完成。LGTM。✅ — Qwen Code · qwen3.7-max |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
LGTM, looks ready to ship. ✅
wenshao
left a comment
wenshao
left a comment
There was a problem hiding this comment.
Re-review of the incremental changes since the last review (554dec7..da43b37). All findings from the previous review are resolved and verified against the new code:
modelRegistry.registerAuthTypeModelsnow guardsArray.isArray(skip-and-warn instead of an opaquemodels is not iterable) — placed inside the method, so it also covers the reload path. ✅needsMigrationJSDoc corrected to match the=== SETTINGS_VERSIONlogic (and the higher-version fall-through). ✅- The v5→v4 downgrade test now also asserts the persisted
.tmppayload ($version: 4+ unwrapped arrays), not just the in-memory result. ✅ findOpenaiModels(VS Code extension) restored read-side tolerance for the wrapped{ protocol, models }shape, so a not-yet-downgraded file no longer drops existing OpenAI models. ✅- Integration suites'
CURRENT_SETTINGS_VERSIONrealigned to 4. ✅
No new issues. Verified locally: full typecheck ✅, full build ✅, and the affected suites all pass (modelRegistry 56, settings + migration 272, settingsWriter 17). The new tests genuinely exercise each fix.
Downgraded from Approve to Comment: self-PR; CI still running. — claude-opus-4-8 via Qwen Code /qreview
…M#5089) Reverts the structural changes from QwenLM#5089 back to the pre-QwenLM#5089 shape: AuthType stays a fixed enum (not `string`), the Protocol enum is removed, modelProviders is `Record<authType, ModelConfig[]>` again (not `{ protocol, models }`), and createContentGenerator dispatches on authType. The v4->v5 settings migration is removed and SETTINGS_VERSION reverts to 4. Features merged on top of QwenLM#5089 are kept and re-adapted to the old enum+array structure (not reverted): - QwenLM#5632 fastOnly/voiceOnly model flags (test fixtures reshaped to arrays) - QwenLM#5638 workspace provider defaults (readProviderModels already tolerates both shapes; test fixtures reshaped to arrays) - QwenLM#5729 active-runtime-model listing (pre-QwenLM#5089 getAllConfiguredModels already enumerates Object.values(AuthType), so the runtime model is included natively) - QwenLM#5728 ACP set_config_option deterministic provider fixture (reshaped to array; the flake fix is preserved) KNOWN DOWNGRADE CAVEAT: settings already migrated to $version:5 (shipped in v0.19.0) retain the v5 `{ protocol, models }` modelProviders shape, which the reverted ModelRegistry consumes as an array. Such settings will throw on load until re-configured. A v5->v4 downgrade guard/migration is a separate follow-up if backward compatibility for migrated users is needed.
…vert After reverting QwenLM#5089, settings already migrated to $version:5 (shipped in v0.19.0) carry a modelProviders `{ protocol, models }` shape that the reverted v4 readers consume as arrays, throwing "models is not iterable" on load. This adds the inverse migration so those configs auto-converge to v4 on load (the user-facing "automatically migrate $version:5 to 4"). - V5ToV4Migration: unwraps each modelProviders `{ protocol, models }` back to its `models` array, drops the now-implicit protocol (warning only when the explicit protocol differs from the key-derived one), and resets $version to 4. - DOWNGRADE_MIGRATIONS keeps the downgrade out of the ascending forward ALL_MIGRATIONS chain (preserving its invariants); runMigrations and needsMigration consider both via a combined convergence set. - needsMigration now gates on `=== SETTINGS_VERSION` instead of `>=`, so a newer-but-handled version (v5) is reported as needing migration while a genuinely unknown newer version (v6+) is still left untouched. Covered by unit tests for the migration, the framework wiring, and an end-to-end loadSettings downgrade-on-load test.
The integration suites hard-coded CURRENT_SETTINGS_VERSION = 5 (introduced by QwenLM#5676), which mismatched the reverted SETTINGS_VERSION = 4 and failed the migration assertions ($version now writes 4, not 5). Revert the constant to 4 in both settings-migration and qwen-config-dir integration tests. Verified: QWEN_SANDBOX=false vitest run --root ./integration-tests cli/settings-migration.test.ts cli/qwen-config-dir.test.ts → 21 passed.
Addresses /qreview feedback on the revert:
- vscode findOpenaiModels: restore read-side tolerance for the V5
{ protocol, models } shape. The extension reads/writes settings.json
without running the CLI v5->v4 migration, so a not-yet-downgraded
$version:5 file would otherwise return [] and silently drop existing
OpenAI models on the next write. (Critical)
- modelRegistry.registerAuthTypeModels: guard against a non-array provider
value (skip + warn) instead of throwing an opaque "models is not
iterable" — covers hand-edited or unmigrated files the downgrade misses.
- needsMigration JSDoc: update the stale ">= SETTINGS_VERSION" wording to
match the "=== SETTINGS_VERSION, else fall through" logic the downgrade
path depends on.
- settings.test.ts: also assert the v5->v4 downgrade is persisted to disk
(.tmp write-back), not just the in-memory merged result.
Adds tests for the registry guard and the vscode V5 read tolerance.
da43b37 to
29cf90e
Compare
✅ Maintainer local verification — real
|
| Check | Command | Result |
|---|---|---|
| Build | npm run build |
✅ exit 0, no errors |
| Typecheck | npm run typecheck (acp-bridge, audio-capture, cli, core, sdk, webui) |
✅ all pass |
| VSCode types | npm --prefix packages/vscode-ide-companion run check-types |
✅ pass |
| Migration suites | vitest run --root packages/cli src/config/migration/ |
✅ 137 / 6 files (incl. new v5-to-v4.test.ts = 17) |
| CLI config | vitest run --root packages/cli src/config/ |
✅ 764 / 23 files |
| Core models/providers | vitest run --root packages/core src/models/ src/providers/ src/core/contentGenerator.test.ts |
✅ 328 / 18 files |
2. The new downgrade tests are non-vacuous (mutation test)
Neutered the core transform in v5-to-v4.ts — providers[key] = value instead of unwrapping value.models — and re-ran the suites: 5 tests FAIL in v5-to-v4.test.ts + index.test.ts (e.g. "downgrades v5 modelProviders back to v4 arrays", "treats a wrapped config with missing models as an empty array"). Restoring the line → 17/17 green again. The tests genuinely guard the unwrap, not just that a migration ran.
3. Real-binary A/B — the v5→v4 downgrade migration (the marquee fix)
The same $version: 5 settings file (modelProviders.openai = { protocol, models: [...] } — the shape v0.19.0 shipped) was fed to two binaries:
- PRE-FIX = a pure
git revertof refactor(core): extract Protocol enum and decouple model identity from auth type #5089 — rebuilt with the downgrade migration and themodelRegistryhardening removed (mutants confirmed present indistbefore the run). - FIXED = PR head as-is.
| PRE-FIX (pure revert) | FIXED (this PR) | |
|---|---|---|
| Boot | ❌ TypeError: models is not iterable |
✅ boots normally |
| Crash site | ModelRegistry.registerAuthTypeModels → new Config → loadCliConfig |
— (config initializes fully) |
| Real TUI (tmux) | never renders, exit 1 |
renders; /model lists [openai] GPT-4o (gpt-4o) |
settings.json on disk |
unchanged — stuck at $version: 5 |
auto-rewritten to $version: 4, providers → arrays |
PRE-FIX — crash on the main CLI boot path (real binary, captured in tmux):
An unexpected critical error occurred:
TypeError: models is not iterable
at ModelRegistry.registerAuthTypeModels (.../models/modelRegistry.js)
at new ModelRegistry (.../models/modelRegistry.js)
at new ModelsConfig (.../models/modelsConfig.js)
at new Config (.../config/config.js)
at loadCliConfig (.../cli/dist/src/config/config.js)
PRE_FIX_EXIT=1
Because the throw is in loadCliConfig → new Config, a plain revert of #5089 without this PR would brick every launch for any user whose config v0.19.0 had already migrated to v5.
FIXED — the file is auto-migrated on load:
BEFORE $version: 5 modelProviders.openai = { protocol, models: [ { id: gpt-4o } ] }
AFTER $version: 4 modelProviders.openai = [ { id: gpt-4o } ]
…and the migrated model is live and selectable in the real TUI (/model):
│ Select Model │
│ › 1. [openai] GPT-4o (gpt-4o) │
(Headless, the FIXED binary gets all the way to a real model API call — failing only with 401 on the dummy key, i.e. past the exact new Config step PRE-FIX dies in.)
Multi-provider (mirrors a real openai + anthropic config) — all providers, models and order preserved:
BEFORE $ver=5 openai=obj anthropic=obj
AFTER $ver=4 openai=[gpt-4o, qwen3.7-max] anthropic=[claude-opus-4-6]
Not run (disclosed)
integration-tests/— needs the full build+run rig; not executed (consistent with the author's note). The reshapedacp-integration.test.tsfixture (object → array) matches the revertedreadProviderModels, which now reads arrays directly — confirmed by reading the diff, not by execution.
Verdict
The revert builds and typechecks clean, the preserved features' suites pass, the new downgrade tests are non-vacuous, and the real binary confirms the exact failure this PR exists to prevent (models is not iterable on the main boot path) is fixed: a v5 config auto-migrates to v4 and its models stay usable. LGTM from a runtime-verification standpoint.
🇨🇳 中文版
✅ 维护者本地验证 —— 真实 qwen 二进制（tmux TUI + headless）
从 PR head 构建真实 qwen 二进制并实际驱动运行，不止跑单测。
环境 —— 在 PR head da43b37a0（base origin/main d350dd8）建独立 worktree，全新 npm ci；macOS（darwin），Node v22.22.2。
1. 构建 · 类型检查 · 定向套件 —— 全绿
| 检查项 | 命令 | 结果 |
|---|---|---|
| 构建 | npm run build |
✅ exit 0，无报错 |
| 类型检查 | npm run typecheck（acp-bridge、audio-capture、cli、core、sdk、webui） |
✅ 全部通过 |
| VSCode 类型 | npm --prefix packages/vscode-ide-companion run check-types |
✅ 通过 |
| 迁移套件 | vitest run --root packages/cli src/config/migration/ |
✅ 137 / 6 文件（含新增 v5-to-v4.test.ts = 17） |
| CLI config | vitest run --root packages/cli src/config/ |
✅ 764 / 23 文件 |
| Core models/providers | vitest run --root packages/core src/models/ src/providers/ src/core/contentGenerator.test.ts |
✅ 328 / 18 文件 |
2. 新增的降级测试非空过（变异测试）
把 v5-to-v4.ts 的核心转换中和 —— 用 providers[key] = value 取代解包 value.models —— 重跑套件：v5-to-v4.test.ts + index.test.ts 中 5 个测试 FAIL（如 "downgrades v5 modelProviders back to v4 arrays"、"treats a wrapped config with missing models as an empty array"）。恢复该行后 → 重新 17/17 全绿。说明测试真正守卫了 unwrap 转换，而非只校验"有迁移跑过"。
3. 真实二进制 A/B —— v5→v4 降级迁移（核心修复）
把同一份 $version: 5 设置文件（modelProviders.openai = { protocol, models: [...] }，即 v0.19.0 发布时写入的形状）分别喂给两个二进制：
- PRE-FIX = 对 refactor(core): extract Protocol enum and decouple model identity from auth type #5089 的纯
git revert—— 重新构建时把降级迁移和modelRegistry加固都移除（运行前已 grep 确认 mutant 在dist中）。 - FIXED = PR head 原样。
| PRE-FIX（纯回退） | FIXED（本 PR） | |
|---|---|---|
| 启动 | ❌ TypeError: models is not iterable |
✅ 正常启动 |
| 崩溃点 | ModelRegistry.registerAuthTypeModels → new Config → loadCliConfig |
—（config 完整初始化） |
| 真实 TUI（tmux） | 界面从未渲染，exit 1 |
正常渲染；/model 列出 [openai] GPT-4o (gpt-4o) |
磁盘上的 settings.json |
不变 —— 卡在 $version: 5 |
自动改写为 $version: 4，provider → 数组 |
PRE-FIX —— 崩在主 CLI 启动路径（真实二进制，tmux 捕获）：
An unexpected critical error occurred:
TypeError: models is not iterable
at ModelRegistry.registerAuthTypeModels (.../models/modelRegistry.js)
at new ModelRegistry (.../models/modelRegistry.js)
at new ModelsConfig (.../models/modelsConfig.js)
at new Config (.../config/config.js)
at loadCliConfig (.../cli/dist/src/config/config.js)
PRE_FIX_EXIT=1
抛错位置在 loadCliConfig → new Config，因此对 #5089 做纯回退而不带本 PR，会让任何在 v0.19.0 已迁移到 v5 的用户每次启动都崩。
FIXED —— 文件在加载时被自动迁移：
BEFORE $version: 5 modelProviders.openai = { protocol, models: [ { id: gpt-4o } ] }
AFTER $version: 4 modelProviders.openai = [ { id: gpt-4o } ]
迁移后的模型在真实 TUI 里可见、可选（/model）：
│ Select Model │
│ › 1. [openai] GPT-4o (gpt-4o) │
（headless 下 FIXED 一路走到真实模型 API 调用 —— 只因 dummy key 返回 401，即已越过 PRE-FIX 崩溃的那个 new Config 步骤。）
多 provider（对应真实的 openai + anthropic 配置）—— 所有 provider、模型与顺序都保留：
BEFORE $ver=5 openai=obj anthropic=obj
AFTER $ver=4 openai=[gpt-4o, qwen3.7-max] anthropic=[claude-opus-4-6]
未执行（如实说明）
integration-tests/—— 需要完整 build+run harness，未执行（与作者说明一致）。被改回数组的acp-integration.test.tsfixture（object → array）与回退后的readProviderModels一致 —— 后者现在直接读数组，已通过阅读 diff 确认，但未实际运行。
结论
回退构建与类型检查干净，被保留功能的套件通过，新增降级测试非空过，真实二进制确认本 PR 要防的那个故障（主启动路径上的 models is not iterable）已修复：v5 配置自动迁移到 v4，模型仍可用。从运行时验证角度 LGTM。
… paths Addresses /review suggestions on the revert: - contentGenerator: import PROVIDER_SOURCED_FIELDS from constants.js (where it is actually defined) instead of modelsConfig.js, breaking the runtime import cycle contentGenerator -> modelsConfig -> contentGenerator. constants.js only references contentGenerator at the type level, which is erased at runtime, so no cycle remains. - contentGenerator.test: add coverage for the two authType error paths the revert restored (missing authType -> "must have an authType"; unknown authType -> "Unsupported authType"), which QwenLM#5089's protocol-based tests had replaced. Neither was covered before. The acpAgent z.nativeEnum(AuthType).parse(methodId) suggestion is left as-is: that line is byte-identical to pre-QwenLM#5089, so it is pre-existing behavior the revert faithfully restores rather than a regression of this PR.
4 participants
What this PR does
This reverts the structural changes introduced by #5089 back to the pre-#5089 shape, while preserving the features that were merged on top of it, and adds a downgrade migration so configs already migrated by the released version keep loading.
Reverted back to the pre-#5089 structure:
AuthTypeis a fixed enum again instead ofstring; theProtocolenum is removed;modelProvidersisRecord<authType, ModelConfig[]>(arrays) again instead of{ protocol, models };createContentGeneratordispatches onauthType; the v4→v5 settings migration is removed andSETTINGS_VERSIONreturns to 4.Preserved and re-adapted to the enum + array structure (NOT reverted): #5632 fastOnly/voiceOnly model flags, #5638 workspace provider defaults, #5729 active-runtime-model listing, and #5728 ACP
set_config_optiondeterministic provider fixture. Their production code was already format-agnostic or tolerant (e.g.readProviderModelshandles both shapes); only test fixtures that hard-coded the v5{ protocol, models }shape were reshaped back to arrays.Added a v5→v4 settings downgrade migration: settings already migrated to
$version: 5(shipped in v0.19.0) carry amodelProvidersshape the reverted v4 readers consume as arrays, which would throwmodels is not iterableon load. The new migration unwraps each{ protocol, models }back to itsmodelsarray, drops the now key-derived protocol, and resets$versionto 4 automatically on load.Why it's needed
#5089 was merged in error and needs to be backed out. A plain
git revertof the squash commit does not build, because several PRs merged after it (#5632, #5638, #5729, #5728) were written against itsProtocol/ProviderConfigstructures. This PR backs out only the structural change and re-adapts those dependents so the tree still builds and their features keep working, and it ships the downgrade migration so users who already ran v0.19.0 are not left with an unreadable config.Reviewer Test Plan
How to verify
npm run build— full monorepo build succeeds.npm run typecheck— all packages typecheck (alsonpm --prefix packages/vscode-ide-companion run check-types).npx vitest run --root packages/core src/models/ src/providers/ src/core/contentGenerator.test.tsandnpx vitest run --root packages/cli src/config/.~/.qwen/settings.jsonat"$version": 5withmodelProviders: { "openai": { "protocol": "openai", "models": [ { "id": "gpt-4o" } ] } }, start the CLI, and confirm the file is rewritten to"$version": 4withmodelProviders: { "openai": [ { "id": "gpt-4o" } ] }and the app loads withoutmodels is not iterable.Evidence (Before & After)
N/A — non-user-visible refactor/revert. Verified by full-workspace typecheck, full
npm run build, and the affected unit suites (core models/providers/config + cli config, plus a new end-to-endloadSettingsdowngrade test). The integration testintegration-tests/cli/acp-integration.test.tshad its fixture reshaped to the v4 array form but was not executed locally (requires the full build+run harness).Tested on
Risk & Scope
modelProviderssettings format reverts from{ protocol, models }to plain arrays. The included v5→v4 migration auto-converts existing$version: 5configs on load, so no manual user action is required. (A pathological untagged file mixing v1-era keys with a v5-shapemodelProvidersis not handled — that state cannot be produced by any shipped version.)Linked Issues
Reverts #5089.
中文说明
这个 PR 做了什么
回退 #5089 引入的结构性改动到 #5089 之前的形态，同时保留在它之后合入、叠加于其上的功能，并新增一个降级迁移，使已被发布版本迁移过的配置仍能正常加载。
回退到 #5089 之前的结构：
AuthType重新变回固定枚举而不是string；移除Protocol枚举；modelProviders重新变回Record<authType, ModelConfig[]>（数组）而不是{ protocol, models }；createContentGenerator按authType分发；移除 v4→v5 设置迁移，SETTINGS_VERSION回到 4。保留并适配回枚举 + 数组结构（未回退）：#5632 fastOnly/voiceOnly 模型标记、#5638 workspace provider defaults、#5729 运行时模型列表、#5728 ACP
set_config_option确定性 provider fixture。它们的生产代码本就与格式无关或双格式兼容（例如readProviderModels同时处理两种形状）；只有硬编码了 v5{ protocol, models }形状的测试 fixture 被改回数组。新增 v5→v4 设置降级迁移：已迁移到
$version: 5（随 v0.19.0 发布）的设置，其modelProviders形状会被回退后的 v4 读取逻辑当作数组消费，加载时抛models is not iterable。新迁移在加载时自动把每个{ protocol, models }解包回models数组、丢弃可由 key 推导的 protocol，并把$version重置为 4。为什么需要
#5089 是错误合入的，需要回退。直接对 squash commit 做
git revert编译不过，因为在它之后合入的多个 PR（#5632、#5638、#5729、#5728）是基于它的Protocol/ProviderConfig结构编写的。本 PR 只回退结构性改动并把这些依赖项适配回去，使代码仍能构建、功能仍然可用；同时附带降级迁移，让已经运行过 v0.19.0 的用户不会被卡在无法读取的配置上。Reviewer 测试计划
如何验证
npm run build—— 全仓库构建通过。npm run typecheck—— 所有包类型检查通过（外加npm --prefix packages/vscode-ide-companion run check-types）。npx vitest run --root packages/core src/models/ src/providers/ src/core/contentGenerator.test.ts与npx vitest run --root packages/cli src/config/。"$version": 5且modelProviders: { "openai": { "protocol": "openai", "models": [ { "id": "gpt-4o" } ] } }的~/.qwen/settings.json，启动 CLI，确认文件被改写为"$version": 4、modelProviders: { "openai": [ { "id": "gpt-4o" } ] }，且加载时不再抛models is not iterable。证据（前后对比）
N/A —— 非用户可见的重构/回退。通过全仓库类型检查、完整
npm run build，以及受影响的单元测试套件（core models/providers/config + cli config，外加一个新的端到端loadSettings降级测试）验证。集成测试integration-tests/cli/acp-integration.test.ts的 fixture 已改回 v4 数组形式，但未在本地执行（需要完整的构建 + 运行 harness）。测试环境
风险与范围
modelProviders设置格式从{ protocol, models }回退为纯数组。附带的 v5→v4 迁移会在加载时自动转换现有$version: 5配置，用户无需手动操作。（一种把 v1 时代的 key 与 v5 形状modelProviders混在一起、且无版本标记的病态文件不被处理 —— 任何已发布版本都不会产生该状态。）关联 Issue
Reverts #5089.