feat: Add requested model to logs, enhance background task detection, and introduce AI SDK compatibility utilities.

- thread-stream test fixtures (intentionally malformed) were being picked
  up by Turbopack during production build, causing 111 compile errors
- IgnorePlugin excludes /test/ within thread-stream context
- thread-stream added to serverExternalPackages to prevent bundling
- /app removed: it is a stale npm-package prebuild artifact, not source code

.agents/workflows/generate-release.md

@@ -4,16 +4,36 @@ description: Create a new release, bump version up to 1.x.10 threshold, update c
 
 # Generate Release Workflow
 
-Bump version, finalize CHANGELOG, commit, tag, push, publish to npm, and create GitHub release.
+Bump version, finalize CHANGELOG, commit, open a **PR to main** and wait for user confirmation before tagging, publishing, and deploying.
 
 > **VERSION RULE: Always use PATCH bumps (2.x.y → 2.x.y+1)**
 > NEVER use `npm version minor` or `npm version major`.
 > Always use: `npm version patch --no-git-tag-version`
 > The threshold rule: when `y` reaches 10, bump to `2.(x+1).0` — e.g. `2.1.10` → `2.2.0`.
 
-## Steps
+---
 
-### 1. Determine new version
+## ⚠️ Two-Phase Flow
+
+```
+Phase 1 (automated): bump → docs → i18n → commit → push → open PR
+  ↕  🛑 STOP: Notify user, wait for PR confirmation
+Phase 2 (post-merge): tag → publish → GitHub release → Docker → deploy
+```
+
+**NEVER push directly to main or create tags before the user confirms the PR.**
+
+---
+
+## Phase 1: Pre-Merge
+
+### 1. Create release branch
+
+```bash
+git checkout -b release/v2.x.y
+```
+
+### 2. Determine new version
 
 Check current version in `package.json` and increment the **patch** number only:
 
@@ -27,11 +47,6 @@ Version format: `2.x.y` — examples:
 - `2.1.9` → `2.1.10` (patch)
 - `2.1.10` → `2.2.0` (minor threshold — do manually with `sed`)
 
-```bash
-# ALWAYS use patch:
-npm version patch --no-git-tag-version
-```
-
 > **⚠️ ATOMIC COMMIT RULE — Version bump MUST happen before committing feature files.**
 >
 > **CORRECT order:**
@@ -53,7 +68,7 @@ npm version patch --no-git-tag-version
 > This ensures that `git show v2.x.y` always contains both code changes and the version bump together.
 > The GitHub release tag will point to a commit that includes ALL changes for that version.
 
-### 2. Regenerate lock file (REQUIRED after version bump)
+### 3. Regenerate lock file (REQUIRED after version bump)
 
 **Mandatory** — skipping causes `@swc/helpers` lock mismatch and CI failures:
 
@@ -61,7 +76,7 @@ npm version patch --no-git-tag-version
 npm install
 ```
 
-### 3. Finalize CHANGELOG.md
+### 4. Finalize CHANGELOG.md
 
 Replace `[Unreleased]` header with the new version and date.
 Keep an empty `## [Unreleased]` section above it.
@@ -74,7 +89,7 @@ Keep an empty `## [Unreleased]` section above it.
 ## [2.x.y] — YYYY-MM-DD
 ```
 
-### 4. Update openapi.yaml version ⚠️ MANDATORY
+### 5. Update openapi.yaml version ⚠️ MANDATORY
 
 > **CI will fail** if `docs/openapi.yaml` version ≠ `package.json` version (`check:docs-sync` enforces this).
 
@@ -84,33 +99,97 @@ Keep an empty `## [Unreleased]` section above it.
 VERSION=$(node -p "require('./package.json').version") && sed -i "s/  version: .*/  version: $VERSION/" docs/openapi.yaml && echo "✓ openapi.yaml → $VERSION"
 ```
 
-### 5. Stage, commit, and tag
+### 6. Update README.md and i18n docs
+
+Run `/update-docs` workflow steps to:
+
+- Update feature table rows in `README.md`
+- Sync changes to all 29 language `docs/i18n/*/README.md` files
+- Update `docs/FEATURES.md` if Settings section changed
+
+### 7. Run tests
+
+// turbo
+
+```bash
+npm test
+```
+
+All tests must pass before creating the PR.
+
+### 8. Stage, commit, and push
 
 // turbo-all
 
 ```bash
-git add package.json package-lock.json CHANGELOG.md docs/openapi.yaml
+git add -A
 git commit -m "chore(release): v2.x.y — summary of changes"
+git push origin release/v2.x.y
+```
+
+### 9. Open PR to main
+
+```bash
+gh pr create \
+  --repo diegosouzapw/OmniRoute \
+  --base main \
+  --head release/v2.x.y \
+  --title "chore(release): v2.x.y — summary" \
+  --body "## 🚀 Release v2.x.y
+
+### Changes
+...
+
+### Tests
+- X/X tests pass
+
+### ⚠️ After merging: run Phase 2 steps to tag, publish, and deploy."
+```
+
+### 10. 🛑 STOP — Notify User & Await PR Confirmation
+
+**This is a mandatory stop point.** Use `notify_user` with `BlockedOnUser: true`:
+
+Inform the user:
+
+- PR URL
+- Summary of changes
+- Test results
+- List of files changed
+
+**DO NOT proceed to Phase 2 until the user confirms the PR looks good and merges it.**
+
+---
+
+## Phase 2: Post-Merge (only after user confirms)
+
+> Run these steps only AFTER the user has merged the PR.
+
+### 11. Pull main and create tag
+
+```bash
+git checkout main
+git pull origin main
 git tag -a v2.x.y -m "Release v2.x.y"
 ```
 
-### 6. Push to GitHub
+### 12. Push tag to GitHub
 
 ```bash
-git push origin main --tags
+git push origin --tags
 ```
 
-### 7. Create GitHub release
+### 13. Create GitHub release
 
 ```bash
 gh release create v2.x.y --title "v2.x.y — summary" --notes "..."
 ```
 
-### 8. 🐳 Trigger Docker Hub build (MANDATORY — keep npm and Docker in sync)
+### 14. 🐳 Trigger Docker Hub build (MANDATORY — keep npm and Docker in sync)
 
 > **CRITICAL**: Docker Hub and npm MUST always publish the same version.
 > The Docker image is built automatically via GitHub Actions when a new tag is pushed.
-> After pushing the tag in step 5-6, **verify the workflow runs**:
+> After pushing the tag in step 11-12, **verify the workflow runs**:
 
 ```bash
 # Verify the Docker workflow triggered
@@ -129,7 +208,7 @@ If the Docker build was not triggered automatically, trigger it manually:
 gh workflow run docker-publish.yml --repo diegosouzapw/OmniRoute --ref v2.x.y
 ```
 
-### 9. Deploy to BOTH VPS environments (MANDATORY)
+### 15. Deploy to BOTH VPS environments (MANDATORY)
 
 > Always deploy to **both** environments after every release.
 > See `/deploy-vps` workflow for detailed steps.
@@ -151,18 +230,27 @@ curl -s -o /dev/null -w "LOCAL:  HTTP %{http_code}\n" http://192.168.0.15:20128/
 curl -s -o /dev/null -w "AKAMAI: HTTP %{http_code}\n" http://69.164.221.35:20128/
 ```
 
+### 16. Clean up release branch
+
+```bash
+git branch -d release/v2.x.y
+```
+
+---
+
 ## Notes
 
 - Always run `/update-docs` BEFORE this workflow (ensures CHANGELOG and README are current)
 - The `prepublishOnly` script runs `npm run build:cli` automatically during `npm publish`
 - After npm publish, verify with `npm info omniroute version`
 - Lock file sync errors are caused by skipping `npm install` after version bump
+- Use `gh auth switch -u diegosouzapw` if git push fails with wrong account
 
 ## Known CI Pitfalls
 
 | CI failure                                                                | Cause                                                    | Fix                                                                    |
 | ------------------------------------------------------------------------- | -------------------------------------------------------- | ---------------------------------------------------------------------- |
-| `[docs-sync] FAIL - OpenAPI version differs from package.json`            | Skipped step 4 — `docs/openapi.yaml` version not updated | Run step 4 (`sed -i ...`) and commit                                   |
+| `[docs-sync] FAIL - OpenAPI version differs from package.json`            | Skipped step 5 — `docs/openapi.yaml` version not updated | Run step 5 (`sed -i ...`) and commit                                   |
 | `[docs-sync] FAIL - CHANGELOG.md first section must be "## [Unreleased]"` | `## [Unreleased]` missing or not at top of CHANGELOG     | Add `## [Unreleased]\n\n---\n` before the first versioned `## [x.y.z]` |
 | Electron Linux `.deb` build fails (`FpmTarget` error)                     | `fpm` Ruby gem not installed on `ubuntu-latest` runner   | Already fixed in `electron-release.yml` (`gem install fpm` step)       |
 | Docker Hub `502 error writing layer blob`                                 | Transient Docker Hub network error during ARM64 push     | Re-run the Docker publish workflow; no code change needed              |

.github/workflows/docker-publish.yml

@@ -21,18 +21,18 @@ jobs:
       IMAGE_NAME: diegosouzapw/omniroute
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           ref: ${{ github.event_name == 'workflow_dispatch' && format('refs/tags/v{0}', inputs.version) || '' }}
 
       - name: Set up QEMU (for multi-arch builds)
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v4
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
 
       - name: Login to Docker Hub
-        uses: docker/login-action@v3
+        uses: docker/login-action@v4
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -50,7 +50,7 @@ jobs:
           echo "Publishing Docker image: $IMAGE_NAME:$VERSION"
 
       - name: Build and push multi-arch image
-        uses: docker/build-push-action@v6
+        uses: docker/build-push-action@v7
         with:
           context: .
           target: runner-base
@@ -70,7 +70,7 @@ jobs:
           docker buildx imagetools inspect "${{ env.IMAGE_NAME }}:${{ steps.version.outputs.version }}"
 
       - name: Update Docker Hub description
-        uses: peter-evans/dockerhub-description@v4
+        uses: peter-evans/dockerhub-description@v5
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}

.gitignore

@@ -89,6 +89,7 @@ docs/*
 !docs/MCP-SERVER.md
 !docs/CLI-TOOLS.md
 
+
 # open-sse tests
 open-sse/test/*
 
@@ -130,3 +131,6 @@ vscode-extension/
 *.sqlite-shm
 *.sqlite-wal
 *.sqlite-journal
+
+# Compiled npm-package build artifact (not source, should not be in git)
+/app

CHANGELOG.md

@@ -2,6 +2,523 @@
 
 ## [Unreleased]
 
+> **Coming next** — see [3.0.0-rc branch](https://github.com/diegosouzapw/OmniRoute/tree/3.0.0-rc).
+
+---
+
+## [3.0.0-rc.10] — 2026-03-23
+
+### 🔧 Bug Fixes
+
+- **#509 / #508** — Electron build regression: downgraded Next.js from `16.1.x` to `16.0.10` to eliminate Turbopack module-hashing instability that caused blank screens in the Electron desktop bundle.
+- **Unit test fixes** — Corrected two stale test assertions (`nanobanana-image-handler` aspect ratio/resolution, `thinking-budget` Gemini `thinkingConfig` field mapping) that had drifted after recent implementation changes.
+- **#541** — Responded to user feedback about installation complexity; no code changes required.
+
+---
+
+## [3.0.0-rc.9] — 2026-03-23
+
+### ✨ New Features
+
+- **T29** — Vertex AI SA JSON Executor: implemented using the `jose` library to handle JWT/Service Account auth, along with configurable regions in the UI and automatic partner model URL building.
+- **T42** — Image generation aspect ratio mapping: created `sizeMapper` logic for generic OpenAI formats (`size`), added native `imagen3` handling, and updated NanoBanana endpoints to utilize mapped aspect ratios automatically.
+- **T38** — Centralized model specifications: `modelSpecs.ts` created for limits and parameters per model.
+
+### 🔧 Improvements
+
+- **T40** — OpenCode CLI tools integration: native `opencode-zen` and `opencode-go` integration completed in earlier PR.
+
+---
+
+## [3.0.0-rc.8] — 2026-03-23
+
+### 🔧 Bug Fixes & Improvements (Fallback, Quota & Budget)
+
+- **T24** — `503` cooldown await fix + `406` mapping: mapped `406 Not Acceptable` to `503 Service Unavailable` with proper cooldown intervals.
+- **T25** — Provider validation fallback: graceful fallback to standard validation models when a specific `validationModelId` is not present.
+- **T36** — `403` vs `429` provider handling refinement: extracted into `errorClassifier.ts` to properly segregate hard permissions failures (`403`) from rate limits (`429`).
+- **T39** — Endpoint Fallback for `fetchAvailableModels`: implemented a tri-tier mechanism (`/models` -> `/v1/models` -> local generic catalog) + `list_models_catalog` MCP tool updates to reflect `source` and `warning`.
+- **T33** — Thinking level to budget conversion: translates qualitative thinking levels into precise budget allocations.
+- **T41** — Background task auto redirect: routes heavy background evaluation tasks to flash/efficient models automatically.
+- **T23** — Intelligent quota reset fallback: accurately extracts `x-ratelimit-reset` / `retry-after` header values or maps static cooldowns.
+
+---
+
+## [3.0.0-rc.7] — 2026-03-23 _(What's New vs v2.9.5 — will be released as v3.0.0)_
+
+> **Upgrade from v2.9.5:** 16 issues resolved · 2 community PRs merged · 2 new providers · 7 new API endpoints · 3 new features · DB migration 008+009 · 832 tests passing · 15 sub2api gap improvements (T01–T15 complete).
+
+### 🆕 New Providers
+
+| Provider         | Alias          | Tier | Notes                                                          |
+| ---------------- | -------------- | ---- | -------------------------------------------------------------- |
+| **OpenCode Zen** | `opencode-zen` | Free | 3 models via `opencode.ai/zen/v1` (PR #530 by @kang-heewon)    |
+| **OpenCode Go**  | `opencode-go`  | Paid | 4 models via `opencode.ai/zen/go/v1` (PR #530 by @kang-heewon) |
+
+Both providers use the new `OpencodeExecutor` with multi-format routing (`/chat/completions`, `/messages`, `/responses`, `/models/{model}:generateContent`).
+
+---
+
+### ✨ New Features
+
+#### 🔑 Registered Keys Provisioning API (#464)
+
+Auto-generate and issue OmniRoute API keys programmatically with per-provider and per-account quota enforcement.
+
+| Endpoint                              | Method    | Description                                      |
+| ------------------------------------- | --------- | ------------------------------------------------ |
+| `/api/v1/registered-keys`             | `POST`    | Issue a new key — raw key returned **once only** |
+| `/api/v1/registered-keys`             | `GET`     | List registered keys (masked)                    |
+| `/api/v1/registered-keys/{id}`        | `GET`     | Get key metadata                                 |
+| `/api/v1/registered-keys/{id}`        | `DELETE`  | Revoke a key                                     |
+| `/api/v1/registered-keys/{id}/revoke` | `POST`    | Revoke (for clients without DELETE support)      |
+| `/api/v1/quotas/check`                | `GET`     | Pre-validate quota before issuing                |
+| `/api/v1/providers/{id}/limits`       | `GET/PUT` | Configure per-provider issuance limits           |
+| `/api/v1/accounts/{id}/limits`        | `GET/PUT` | Configure per-account issuance limits            |
+| `/api/v1/issues/report`               | `POST`    | Report quota events to GitHub Issues             |
+
+**DB — Migration 008:** Three new tables: `registered_keys`, `provider_key_limits`, `account_key_limits`.
+**Security:** Keys stored as SHA-256 hashes. Raw key shown once on creation, never retrievable again.
+**Quota types:** `maxActiveKeys`, `dailyIssueLimit`, `hourlyIssueLimit` per provider and per account.
+**Idempotency:** `idempotency_key` field prevents duplicate issuance. Returns `409 IDEMPOTENCY_CONFLICT` if key was already used.
+**Budget per key:** `dailyBudget` / `hourlyBudget` — limits how many requests a key can route per window.
+**GitHub reporting:** Optional. Set `GITHUB_ISSUES_REPO` + `GITHUB_ISSUES_TOKEN` to auto-create GitHub issues on quota exceeded or issuance failures.
+
+#### 🎨 Provider Icons — @lobehub/icons (#529)
+
+All provider icons in the dashboard now use `@lobehub/icons` React components (130+ providers with SVG).
+Fallback chain: **Lobehub SVG → existing `/providers/{id}.png` → generic icon**. Uses a proper React `ErrorBoundary` pattern.
+
+#### 🔄 Model Auto-Sync Scheduler (#488)
+
+OmniRoute now automatically refreshes model lists for connected providers every **24 hours**.
+
+- Runs on server startup via the existing `/api/sync/initialize` hook
+- Configurable via `MODEL_SYNC_INTERVAL_HOURS` environment variable
+- Covers 16 major providers
+- Records last sync time in the settings database
+
+---
+
+### 🔧 Bug Fixes
+
+#### OAuth & Auth
+
+- **#537 — Gemini CLI OAuth:** Clear actionable error when `GEMINI_OAUTH_CLIENT_SECRET` is missing in Docker/self-hosted deployments. Previously showed cryptic `client_secret is missing` from Google. Now provides specific `docker-compose.yml` and `~/.omniroute/.env` instructions.
+
+#### Providers & Routing
+
+- **#536 — LongCat AI:** Fixed `baseUrl` (`api.longcat.chat/openai`) and `authHeader` (`Authorization: Bearer`).
+- **#535 — Pinned model override:** `body.model` is now correctly set to `pinnedModel` when context-cache protection is active.
+- **#532 — OpenCode Go key validation:** Now uses the `zen/v1` test endpoint (`testKeyBaseUrl`) — same key works for both tiers.
+
+#### CLI & Tools
+
+- **#527 — Claude Code + Codex loop:** `tool_result` blocks are now converted to text instead of dropped, stopping infinite tool-result loops.
+- **#524 — OpenCode config save:** Added `saveOpenCodeConfig()` handler (XDG_CONFIG_HOME aware, writes TOML).
+- **#521 — Login stuck:** Login no longer freezes after skipping password setup — redirects correctly to onboarding.
+- **#522 — API Manager:** Removed misleading "Copy masked key" button (replaced with a lock icon tooltip).
+- **#532 — OpenCode Go config:** Guide settings handler now handles `opencode` toolId.
+
+#### Developer Experience
+
+- **#489 — Antigravity:** Missing `googleProjectId` returns a structured 422 error with reconnect guidance instead of a cryptic crash.
+- **#510 — Windows paths:** MSYS2/Git-Bash paths (`/c/Program Files/...`) are now normalized to `C:\\Program Files\\...` automatically.
+- **#492 — CLI startup:** `omniroute` CLI now detects `mise`/`nvm`-managed Node when `app/server.js` is missing and shows targeted fix instructions.
+
+---
+
+### 📖 Documentation Updates
+
+- **#513** — Docker password reset: `INITIAL_PASSWORD` env var workaround documented
+- **#520** — pnpm: `pnpm approve-builds better-sqlite3` step documented
+
+---
+
+### ✅ Issues Resolved in v3.0.0
+
+`#464` `#488` `#489` `#492` `#510` `#513` `#520` `#521` `#522` `#524` `#527` `#529` `#532` `#535` `#536` `#537`
+
+---
+
+### 🔀 Community PRs Merged
+
+| PR       | Author       | Summary                                                                |
+| -------- | ------------ | ---------------------------------------------------------------------- |
+| **#530** | @kang-heewon | OpenCode Zen + Go providers with `OpencodeExecutor` and improved tests |
+
+---
+
+## [3.0.0-rc.7] - 2026-03-23
+
+### 🔧 Improvements (sub2api Gap Analysis — T05, T08, T09, T13, T14)
+
+- **T05** — Rate-limit DB persistence: `setConnectionRateLimitUntil()`, `isConnectionRateLimited()`, `getRateLimitedConnections()` in `providers.ts`. The existing `rate_limited_until` column is now exposed as a dedicated API — OAuth token refresh must NOT touch this field to prevent rate-limit loops.
+- **T08** — Per-API-key session limit: `max_sessions INTEGER DEFAULT 0` added to `api_keys` via auto-migration. `sessionManager.ts` gains `registerKeySession()`, `unregisterKeySession()`, `checkSessionLimit()`, and `getActiveSessionCountForKey()`. Callers in `chatCore.js` can enforce the limit and decrement on `req.close`.
+- **T09** — Codex vs Spark rate-limit scopes: `getCodexModelScope()` and `getCodexRateLimitKey()` in `codex.ts`. Standard models (`gpt-5.x-codex`, `codex-mini`) get scope `"codex"`; spark models (`codex-spark*`) get scope `"spark"`. Rate-limit keys should be `${accountId}:${scope}` so exhausting one pool doesn't block the other.
+- **T13** — Stale quota display fix: `getEffectiveQuotaUsage(used, resetAt)` returns `0` when the reset window has passed; `formatResetCountdown(resetAt)` returns a human-readable countdown string (e.g. `"2h 35m"`). Both exported from `providers.ts` + `localDb.ts` for dashboard consumption.
+- **T14** — Proxy fast-fail: new `src/lib/proxyHealth.ts` with `isProxyReachable(proxyUrl, timeoutMs=2000)` (TCP check, ≤2s instead of 30s timeout), `getCachedProxyHealth()`, `invalidateProxyHealth()`, and `getAllProxyHealthStatuses()`. Results cached 30s by default; configurable via `PROXY_FAST_FAIL_TIMEOUT_MS` / `PROXY_HEALTH_CACHE_TTL_MS`.
+
+### 🧪 Tests
+
+- Test suite: **832 tests, 0 failures**
+
+---
+
+## [3.0.0-rc.6] - 2026-03-23
+
+### 🔧 Bug Fixes & Improvements (sub2api Gap Analysis — T01–T15)
+
+- **T01** — `requested_model` column in `call_logs` (migration 009): track which model the client originally requested vs the actual routed model. Enables fallback rate analytics.
+- **T02** — Strip empty text blocks from nested `tool_result.content`: prevents Anthropic 400 errors (`text content blocks must be non-empty`) when Claude Code chains tool results.
+- **T03** — Parse `x-codex-5h-*` / `x-codex-7d-*` headers: `parseCodexQuotaHeaders()` + `getCodexResetTime()` extract Codex quota windows for precise cooldown scheduling instead of generic 5-min fallback.
+- **T04** — `X-Session-Id` header for external sticky routing: `extractExternalSessionId()` in `sessionManager.ts` reads `x-session-id` / `x-omniroute-session` headers with `ext:` prefix to avoid collision with internal SHA-256 session IDs. Nginx-compatible (hyphenated header).
+- **T06** — Account deactivated → permanent block: `isAccountDeactivated()` in `accountFallback.ts` detects 401 deactivation signals and applies a 1-year cooldown to prevent retrying permanently dead accounts.
+- **T07** — X-Forwarded-For IP validation: new `src/lib/ipUtils.ts` with `extractClientIp()` and `getClientIpFromRequest()` — skips `unknown`/non-IP entries in `X-Forwarded-For` chains (Nginx/proxy-forwarded requests).
+- **T10** — Credits exhausted → distinct fallback: `isCreditsExhausted()` in `accountFallback.ts` returns 1h cooldown with `creditsExhausted` flag, distinct from generic 429 rate limiting.
+- **T11** — `max` reasoning effort → 131072 budget tokens: `EFFORT_BUDGETS` and `THINKING_LEVEL_MAP` updated; reverse mapping now returns `"max"` for full-budget responses. Unit test updated.
+- **T12** — MiniMax M2.7 pricing entries added: `minimax-m2.7`, `MiniMax-M2.7`, `minimax-m2.7-highspeed` added to pricing table (sub2api PR #1120). M2.5/GLM-4.7/GLM-5/Kimi pricing already existed.
+- **T15** — Array content normalization: `normalizeContentToString()` helper in `openai-to-claude.ts` correctly collapses array-formatted system/tool messages to string before sending to Anthropic.
+
+### 🧪 Tests
+
+- Test suite: **832 tests, 0 failures** (unchanged from rc.5)
+
+---
+
+## [3.0.0-rc.5] - 2026-03-22
+
+### ✨ New Features
+
+- **#464** — Registered Keys Provisioning API: auto-issue API keys with per-provider & per-account quota enforcement
+  - `POST /api/v1/registered-keys` — issue keys with idempotency support
+  - `GET /api/v1/registered-keys` — list (masked) registered keys
+  - `GET /api/v1/registered-keys/{id}` — get key metadata
+  - `DELETE /api/v1/registered-keys/{id}` / `POST ../{id}/revoke` — revoke keys
+  - `GET /api/v1/quotas/check` — pre-validate before issuing
+  - `PUT /api/v1/providers/{id}/limits` — set provider issuance limits
+  - `PUT /api/v1/accounts/{id}/limits` — set account issuance limits
+  - `POST /api/v1/issues/report` — optional GitHub issue reporting
+  - DB migration 008: `registered_keys`, `provider_key_limits`, `account_key_limits` tables
+
+---
+
+## [3.0.0-rc.4] - 2026-03-22
+
+### ✨ New Features
+
+- **#530 (PR)** — OpenCode Zen and OpenCode Go providers added (by @kang-heewon)
+  - New `OpencodeExecutor` with multi-format routing (`/chat/completions`, `/messages`, `/responses`)
+  - 7 models across both tiers
+
+---
+
+## [3.0.0-rc.3] - 2026-03-22
+
+### ✨ New Features
+
+- **#529** — Provider icons now use [@lobehub/icons](https://github.com/lobehub/lobe-icons) with graceful PNG fallback and a `ProviderIcon` component (130+ providers supported)
+- **#488** — Auto-update model lists every 24h via `modelSyncScheduler` (configurable via `MODEL_SYNC_INTERVAL_HOURS`)
+
+### 🔧 Bug Fixes
+
+- **#537** — Gemini CLI OAuth: now shows clear actionable error when `GEMINI_OAUTH_CLIENT_SECRET` is missing in Docker/self-hosted deployments
+
+---
+
+## [3.0.0-rc.2] - 2026-03-22
+
+### 🔧 Bug Fixes
+
+- **#536** — LongCat AI key validation: fixed baseUrl (`api.longcat.chat/openai`) and authHeader (`Authorization: Bearer`)
+- **#535** — Pinned model override: `body.model` is now set to `pinnedModel` when context-cache protection detects a pinned model
+- **#524** — OpenCode config now saved correctly: added `saveOpenCodeConfig()` handler (XDG_CONFIG_HOME aware, writes TOML)
+
+---
+
+## [3.0.0-rc.1] - 2026-03-22
+
+### 🔧 Bug Fixes
+
+- **#521** — Login no longer gets stuck after skipping password setup (redirects to onboarding)
+- **#522** — API Manager: Removed misleading "Copy masked key" button (replaced with lock icon tooltip)
+- **#527** — Claude Code + Codex superpowers loop: `tool_result` blocks now converted to text instead of dropped
+- **#532** — OpenCode GO API key validation now uses the correct `zen/v1` endpoint (`testKeyBaseUrl`)
+- **#489** — Antigravity: missing `googleProjectId` returns structured 422 error with reconnect guidance
+- **#510** — Windows: MSYS2/Git-Bash paths (`/c/Program Files/...`) are now normalized to `C:\\Program Files\\...`
+- **#492** — `omniroute` CLI now detects `mise`/`nvm` when `app/server.js` is missing and shows targeted fix
+
+### 📖 Documentation
+
+- **#513** — Docker password reset: `INITIAL_PASSWORD` env var workaround documented
+- **#520** — pnpm: `pnpm approve-builds better-sqlite3` documented
+
+### ✅ Closed Issues
+
+#489, #492, #510, #513, #520, #521, #522, #525, #527, #532
+
+---
+
+## [2.9.5] — 2026-03-22
+
+> Sprint: New OpenCode providers, embedding credentials fix, CLI masked key bug, CACHE_TAG_PATTERN fix.
+
+### 🐛 Bug Fixes
+
+- **CLI tools save masked API key to config files** — `claude-settings`, `cline-settings`, and `openclaw-settings` POST routes now accept a `keyId` param and resolve the real API key from DB before writing to disk. `ClaudeToolCard` updated to send `keyId` instead of the masked display string. Fixes #523, #526.
+- **Custom embedding providers: `No credentials` error** — `/v1/embeddings` now tracks `credentialsProviderId` separately from the routing prefix, so credentials are fetched from the matching provider node ID rather than the public prefix string. Fixes a regression where `google/gemini-embedding-001` and similar custom-provider models would always fail with a credentials error. Fixes #532-related. (PR #528 by @jacob2826)
+- **Context cache protection regex misses `\n` prefix** — `CACHE_TAG_PATTERN` in `comboAgentMiddleware.ts` updated to match both literal `\n` (backslash-n) and actual newline U+000A that `combo.ts` streaming injects around the `<omniModel>` tag after fix #515. Fixes #531.
+
+### ✨ New Providers
+
+- **OpenCode Zen** — Free tier gateway at `opencode.ai/zen/v1` with 3 models: `minimax-m2.5-free`, `big-pickle`, `gpt-5-nano`
+- **OpenCode Go** — Subscription service at `opencode.ai/zen/go/v1` with 4 models: `glm-5`, `kimi-k2.5`, `minimax-m2.7` (Claude format), `minimax-m2.5` (Claude format)
+- Both providers use the new `OpencodeExecutor` which routes dynamically to `/chat/completions`, `/messages`, `/responses`, or `/models/{model}:generateContent` based on the requested model. (PR #530 by @kang-heewon)
+
+---
+
+## [2.9.4] — 2026-03-21
+
+> Sprint: Bug fixes — preserve Codex prompt cache key, fix tagContent JSON escaping, sync expired token status to DB.
+
+### 🐛 Bug Fixes
+
+- **fix(translator)**: Preserve `prompt_cache_key` in Responses API → Chat Completions translation (#517)
+  — The field is a cache-affinity signal used by Codex; stripping it was preventing prompt cache hits.
+  Fixed in `openai-responses.ts` and `responsesApiHelper.ts`.
+
+- **fix(combo)**: Escape `\n` in `tagContent` so injected JSON string is valid (#515)
+  — Template literal newlines (U+000A) are not allowed unescaped inside JSON string values.
+  Replaced with `\\n` literal sequences in `open-sse/services/combo.ts`.
+
+- **fix(usage)**: Sync expired token status back to DB on live auth failure (#491)
+  — When the Limits & Quotas live check returns 401/403, the connection `testStatus` is now updated
+  to `"expired"` in the database so the Providers page reflects the same degraded state.
+  Fixed in `src/app/api/usage/[connectionId]/route.ts`.
+
+---
+
+## [2.9.3] — 2026-03-21
+
+> Sprint: Add 5 new free AI providers — LongCat, Pollinations, Cloudflare AI, Scaleway, AI/ML API.
+
+### ✨ New Providers
+
+- **feat(providers/longcat)**: Add LongCat AI (`lc/`) — 50M tokens/day free (Flash-Lite) + 500K/day (Chat/Thinking) during public beta. OpenAI-compatible, standard Bearer auth.
+- **feat(providers/pollinations)**: Add Pollinations AI (`pol/`) — no API key required. Proxies GPT-5, Claude, Gemini, DeepSeek V3, Llama 4 (1 req/15s free). Custom executor handles optional auth.
+- **feat(providers/cloudflare-ai)**: Add Cloudflare Workers AI (`cf/`) — 10K Neurons/day free (~150 LLM responses or 500s Whisper audio). 50+ models on global edge. Custom executor builds dynamic URL with `accountId` from credentials.
+- **feat(providers/scaleway)**: Add Scaleway Generative APIs (`scw/`) — 1M free tokens for new accounts. EU/GDPR compliant (Paris). Qwen3 235B, Llama 3.1 70B, Mistral Small 3.2.
+- **feat(providers/aimlapi)**: Add AI/ML API (`aiml/`) — $0.025/day free credit, 200+ models (GPT-4o, Claude, Gemini, Llama) via single aggregator endpoint.
+
+### 🔄 Provider Updates
+
+- **feat(providers/together)**: Add `hasFree: true` + 3 permanently free model IDs: `Llama-3.3-70B-Instruct-Turbo-Free`, `Llama-Vision-Free`, `DeepSeek-R1-Distill-Llama-70B-Free`
+- **feat(providers/gemini)**: Add `hasFree: true` + `freeNote` (1,500 req/day, no credit card needed, aistudio.google.com)
+- **chore(providers/gemini)**: Rename display name to `Gemini (Google AI Studio)` for clarity
+
+### ⚙️ Infrastructure
+
+- **feat(executors/pollinations)**: New `PollinationsExecutor` — omits `Authorization` header when no API key provided
+- **feat(executors/cloudflare-ai)**: New `CloudflareAIExecutor` — dynamic URL construction requires `accountId` in provider credentials
+- **feat(executors)**: Register `pollinations`, `pol`, `cloudflare-ai`, `cf` executor mappings
+
+### 📝 Documentation
+
+- **docs(readme)**: Expanded free combo stack to 11 providers ($0 forever)
+- **docs(readme)**: Added 4 new free provider sections (LongCat, Pollinations, Cloudflare AI, Scaleway) with model tables
+- **docs(readme)**: Updated pricing table with 4 new free tier rows
+- **docs(i18n/pt-BR)**: Updated pricing table + added LongCat/Pollinations/Cloudflare AI/Scaleway sections in Portuguese
+- **docs(new-features/ai)**: 10 task spec files + master implementation plan in `docs/new-features/ai/`
+
+### 🧪 Tests
+
+- Test suite: **821 tests, 0 failures** (unchanged)
+
+---
+
+## [2.9.2] — 2026-03-21
+
+> Sprint: Fix media transcription (Deepgram/HuggingFace Content-Type, language detection) and TTS error display.
+
+### 🐛 Bug Fixes
+
+- **fix(transcription)**: Deepgram and HuggingFace audio transcription now correctly map `video/mp4` → `audio/mp4` and other media MIME types via new `resolveAudioContentType()` helper. Previously, uploading `.mp4` files consistently returned "No speech detected" because Deepgram was receiving `Content-Type: video/mp4`.
+- **fix(transcription)**: Added `detect_language=true` to Deepgram requests — auto-detects audio language (Portuguese, Spanish, etc.) instead of defaulting to English. Fixes non-English transcriptions returning empty or garbage results.
+- **fix(transcription)**: Added `punctuate=true` to Deepgram requests for higher-quality transcription output with correct punctuation.
+- **fix(tts)**: `[object Object]` error display in Text-to-Speech responses fixed in both `audioSpeech.ts` and `audioTranscription.ts`. The `upstreamErrorResponse()` function now correctly extracts nested string messages from providers like ElevenLabs that return `{ error: { message: "...", status_code: 401 } }` instead of a flat error string.
+
+### 🧪 Tests
+
+- Test suite: **821 tests, 0 failures** (unchanged)
+
+### Triaged Issues
+
+- **#508** — Tool call format regression: requested proxy logs and provider chain info (`needs-info`)
+- **#510** — Windows CLI healthcheck path: requested shell/Node version info (`needs-info`)
+- **#485** — Kiro MCP tool calls: closed as external Kiro issue (not OmniRoute)
+- **#442** — Baseten /models endpoint: closed (documented manual workaround)
+- **#464** — Key provisioning API: acknowledged as roadmap item
+
+---
+
+## [2.9.1] — 2026-03-21
+
+> Sprint: Fix SSE omniModel data loss, merge per-protocol model compatibility.
+
+### Bug Fixes
+
+- **#511** — Critical: `<omniModel>` tag was sent after `finish_reason:stop` in SSE streams, causing data loss. Tag is now injected into the first non-empty content chunk, guaranteeing delivery before SDKs close the connection.
+
+### Merged PRs
+
+- **PR #512** (@zhangqiang8vip): Per-protocol model compatibility — `normalizeToolCallId` and `preserveOpenAIDeveloperRole` can now be configured per client protocol (OpenAI, Claude, Responses API). New `compatByProtocol` field in model config with Zod validation.
+
+### Triaged Issues
+
+- **#510** — Windows CLI healthcheck_failed: requested PATH/version info
+- **#509** — Turbopack Electron regression: upstream Next.js bug, documented workarounds
+- **#508** — macOS black screen: suggested `--disable-gpu` workaround
+
+---
+
+## [2.9.0] — 2026-03-20
+
+> Sprint: Cross-platform machineId fix, per-API-key rate limits, streaming context cache, Alibaba DashScope, search analytics, ZWS v5, and 8 issues closed.
+
+### ✨ New Features
+
+- **feat(search)**: Search Analytics tab in `/dashboard/analytics` — provider breakdown, cache hit rate, cost tracking. New API: `GET /api/v1/search/analytics` (#feat/search-provider-routing)
+- **feat(provider)**: Alibaba Cloud DashScope added with custom endpoint path validation — configurable `chatPath` and `modelsPath` per node (#feat/custom-endpoint-paths)
+- **feat(api)**: Per-API-key request-count limits — `max_requests_per_day` and `max_requests_per_minute` columns with in-memory sliding-window enforcement returning HTTP 429 (#452)
+- **feat(dev)**: ZWS v5 — HMR leak fix (485 DB connections → 1), memory 2.4GB → 195MB, `globalThis` singletons, Edge Runtime warning fix (@zhangqiang8vip)
+
+### 🐛 Bug Fixes
+
+- **fix(#506)**: Cross-platform `machineId` — `getMachineIdRaw()` rewritten with try/catch waterfall (Windows REG.exe → macOS ioreg → Linux file read → hostname → `os.hostname()`). Eliminates `process.platform` branching that Next.js bundler dead-code-eliminated, fixing `'head' is not recognized` on Windows. Also fixes #466.
+- **fix(#493)**: Custom provider model naming — removed incorrect prefix stripping in `DefaultExecutor.transformRequest()` that mangled org-scoped model IDs like `zai-org/GLM-5-FP8`.
+- **fix(#490)**: Streaming + context cache protection — `TransformStream` intercepts SSE to inject `<omniModel>` tag before `[DONE]` marker, enabling context cache protection for streaming responses.
+- **fix(#458)**: Combo schema validation — `system_message`, `tool_filter_regex`, `context_cache_protection` fields now pass Zod validation on save.
+- **fix(#487)**: KIRO MITM card cleanup — removed ZWS_README, generified `AntigravityToolCard` to use dynamic tool metadata.
+
+### 🧪 Tests
+
+- Added Anthropic-format tools filter unit tests (PR #397) — 8 regression tests for `tool.name` without `.function` wrapper
+- Test suite: **821 tests, 0 failures** (up from 813)
+
+### 📋 Issues Closed (8)
+
+- **#506** — Windows machineId `head` not recognized (fixed)
+- **#493** — Custom provider model naming (fixed)
+- **#490** — Streaming context cache (fixed)
+- **#452** — Per-API-key request limits (implemented)
+- **#466** — Windows login failure (same root cause as #506)
+- **#504** — MITM inactive (expected behavior)
+- **#462** — Gemini CLI PSA (resolved)
+- **#434** — Electron app crash (duplicate of #402)
+
+## [2.8.9] — 2026-03-20
+
+> Sprint: Merge community PRs, fix KIRO MITM card, dependency updates.
+
+### Merged PRs
+
+- **PR #498** (@Sajid11194): Fix Windows machine ID crash (`undefined\REG.exe`). Replaces `node-machine-id` with native OS registry queries. **Closes #486.**
+- **PR #497** (@zhangqiang8vip): Fix dev-mode HMR resource leaks — 485 leaked DB connections → 1, memory 2.4GB → 195MB. `globalThis` singletons, Edge Runtime warning fix, Windows test stability. (+1168/-338 across 22 files)
+- **PRs #499-503** (Dependabot): GitHub Actions updates — `docker/build-push-action@7`, `actions/checkout@6`, `peter-evans/dockerhub-description@5`, `docker/setup-qemu-action@4`, `docker/login-action@4`.
+
+### Bug Fixes
+
+- **#505** — KIRO MITM card now displays tool-specific instructions (`api.anthropic.com`) instead of Antigravity-specific text.
+- **#504** — Responded with UX clarification (MITM "Inactive" is expected behavior when proxy is not running).
+
+---
+
+## [2.8.8] — 2026-03-20
+
+> Sprint: Fix OAuth batch test crash, add "Test All" button to individual provider pages.
+
+### Bug Fixes
+
+- **OAuth batch test crash** (ERR_CONNECTION_REFUSED): Replaced sequential for-loop with 5-connection concurrency limit + 30s per-connection timeout via `Promise.race()` + `Promise.allSettled()`. Prevents server crash when testing large OAuth provider groups (~30+ connections).
+
+### Features
+
+- **"Test All" button on provider pages**: Individual provider pages (e.g., `/providers/codex`) now show a "Test All" button in the Connections header when there are 2+ connections. Uses `POST /api/providers/test-batch` with `{mode: "provider", providerId}`. Results displayed in a modal with pass/fail summary and per-connection diagnosis.
+
+---
+
+## [2.8.7] — 2026-03-20
+
+> Sprint: Merge PR #495 (Bottleneck 429 drop), fix #496 (custom embedding providers), triage features.
+
+### Bug Fixes
+
+- **Bottleneck 429 infinite wait** (PR #495 by @xandr0s): On 429, `limiter.stop({ dropWaitingJobs: true })` immediately fails all queued requests so upstream callers can trigger fallback. Limiter is deleted from Map so next request creates a fresh instance.
+- **Custom embedding models unresolvable** (#496): `POST /v1/embeddings` now resolves custom embedding models from ALL provider_nodes (not just localhost). Enables models like `google/gemini-embedding-001` added via dashboard.
+
+### Issues Responded
+
+- **#452** — Per-API-key request-count limits (acknowledged, on roadmap)
+- **#464** — Auto-issue API keys with provider/account limits (needs more detail)
+- **#488** — Auto-update model lists (acknowledged, on roadmap)
+- **#496** — Custom embedding provider resolution (fixed)
+
+---
+
+## [2.8.6] — 2026-03-20
+
+> Sprint: Merge PR #494 (MiniMax role fix), fix KIRO MITM dashboard, triage 8 issues.
+
+### Features
+
+- **MiniMax developer→system role fix** (PR #494 by @zhangqiang8vip): Per-model `preserveDeveloperRole` toggle. Adds "Compatibility" UI in providers page. Fixes 422 "role param error" for MiniMax and similar gateways.
+- **roleNormalizer**: `normalizeDeveloperRole()` now accepts `preserveDeveloperRole` parameter with tri-state behavior (undefined=keep, true=keep, false=convert).
+- **DB**: New `getModelPreserveOpenAIDeveloperRole()` and `mergeModelCompatOverride()` in `models.ts`.
+
+### Bug Fixes
+
+- **KIRO MITM dashboard** (#481/#487): `CLIToolsPageClient` now routes any `configType: "mitm"` tool to `AntigravityToolCard` (MITM Start/Stop controls). Previously only Antigravity was hardcoded.
+- **AntigravityToolCard generic**: Uses `tool.image`, `tool.description`, `tool.id` instead of hardcoded Antigravity values. Guards against missing `defaultModels`.
+
+### Cleanup
+
+- Removed `ZWS_README_V2.md` (development-only docs from PR #494).
+
+### Issues Triaged (8)
+
+- **#487** — Closed (KIRO MITM fixed in this release)
+- **#486** — needs-info (Windows REG.exe PATH issue)
+- **#489** — needs-info (Antigravity projectId missing, OAuth reconnect needed)
+- **#492** — needs-info (missing app/server.js on mise-managed Node)
+- **#490** — Acknowledged (streaming + context cache blocking, fix planned)
+- **#491** — Acknowledged (Codex auth state inconsistency)
+- **#493** — Acknowledged (Modal provider model name prefix, workaround provided)
+- **#488** — Feature request backlog (auto-update model lists)
+
+---
+
+## [2.8.5] — 2026-03-19
+
+> Sprint: Fix zombie SSE streams, context cache first-turn, KIRO MITM, and triage 5 external issues.
+
+### Bug Fixes
+
+- **Zombie SSE Streams** (#473): Reduce `STREAM_IDLE_TIMEOUT_MS` from 300s → 120s for faster combo fallback when providers hang mid-stream. Configurable via env var.
+- **Context Cache Tag** (#474): Fix `injectModelTag()` to handle first-turn requests (no assistant messages) — context cache protection now works from the very first response.
+- **KIRO MITM** (#481): Change KIRO `configType` from `guide` → `mitm` so the dashboard renders MITM Start/Stop controls.
+- **E2E Test** (CI): Fix `providers-bailian-coding-plan.spec.ts` — dismiss pre-existing modal overlay before clicking Add API Key button.
+
+### Closed Issues
+
+- #473 — Zombie SSE streams bypass combo fallback
+- #474 — Context cache `<omniModel>` tag missing on first turn
+- #481 — MITM for KIRO not activatable from dashboard
+- #468 — Gemini CLI remote server (superseded by #462 deprecation)
+- #438 — Claude unable to write files (external CLI issue)
+- #439 — AppImage doesn't work (documented libfuse2 workaround)
+- #402 — ARM64 DMG "damaged" (documented xattr -cr workaround)
+- #460 — CLI not runnable on Windows (documented PATH fix)
+
 ---
 
 ## [2.8.4] — 2026-03-19

README.md

@@ -11,7 +11,9 @@ _Your universal API proxy — one endpoint, 44+ providers, zero downtime. Now wi
 <div align="center">
 
 [![npm version](https://img.shields.io/npm/v/omniroute?color=cb3837&logo=npm)](https://www.npmjs.com/package/omniroute)
+[![npm downloads](https://img.shields.io/npm/dm/omniroute?color=cb3837&logo=npm&label=npm%20downloads)](https://www.npmjs.com/package/omniroute)
 [![Docker Hub](https://img.shields.io/docker/v/diegosouzapw/omniroute?label=Docker%20Hub&logo=docker&color=2496ED)](https://hub.docker.com/r/diegosouzapw/omniroute)
+[![Docker Pulls](https://img.shields.io/docker/pulls/diegosouzapw/omniroute?logo=docker&color=2496ED&label=docker%20pulls)](https://hub.docker.com/r/diegosouzapw/omniroute)
 [![License](https://img.shields.io/github/license/diegosouzapw/OmniRoute)](https://github.com/diegosouzapw/OmniRoute/blob/main/LICENSE)
 [![Website](https://img.shields.io/badge/Website-omniroute.online-blue?logo=google-chrome&logoColor=white)](https://omniroute.online)
 [![WhatsApp](https://img.shields.io/badge/WhatsApp-Community-25D366?logo=whatsapp&logoColor=white)](https://chat.whatsapp.com/JI7cDQ1GyaiDHhVBpLxf8b?mode=gi_t)
@@ -24,6 +26,25 @@ _Your universal API proxy — one endpoint, 44+ providers, zero downtime. Now wi
 
 ---
 
+## 🆕 What's New in v3.0.0
+
+> **Upgrading from v2.9.5?** — See the [full CHANGELOG](CHANGELOG.md#300--2026-03-22-release-candidate--not-yet-merged-to-main) for all changes.
+
+| Area                         | Change                                                                                                                                                                    |
+| ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🔑 **Registered Keys API**   | Auto-provision API keys via `POST /api/v1/registered-keys` with per-provider/account quota enforcement, idempotency, SHA-256 storage, and optional GitHub issue reporting |
+| 🎨 **Provider Icons**        | 130+ provider logos via `@lobehub/icons` (SVG) with PNG → generic fallback chain                                                                                          |
+| 🔄 **Model Auto-Sync**       | 24h scheduler refreshes model lists for 16 providers on startup — configurable via `MODEL_SYNC_INTERVAL_HOURS`                                                            |
+| 🌐 **OpenCode Zen/Go**       | Two new providers from @kang-heewon via PR #530: free tier + subscription tier via `OpencodeExecutor`                                                                     |
+| 🐛 **Gemini CLI OAuth**      | Actionable error when `GEMINI_OAUTH_CLIENT_SECRET` is missing in Docker (was cryptic Google error)                                                                        |
+| 🐛 **OpenCode config**       | `saveOpenCodeConfig()` now correctly writes TOML to `XDG_CONFIG_HOME`                                                                                                     |
+| 🐛 **Pinned model override** | `body.model` correctly set to `pinnedModel` on context-cache protection                                                                                                   |
+| 🐛 **Codex/Claude loop**     | `tool_result` blocks now converted to text to stop infinite loops                                                                                                         |
+| 🐛 **Login redirect**        | Login no longer freezes after skipping password setup                                                                                                                     |
+| 🐛 **Windows paths**         | MSYS2/Git-Bash paths (`/c/...`) normalized to `C:\...` automatically                                                                                                      |
+
+---
+
 ## 🖼️ Main Dashboard
 
 <div align="center">
@@ -716,7 +737,7 @@ Outcome: deep fallback depth for deadline-critical workloads
 
 **Point any IDE/CLI to:** `http://localhost:20128/v1` · API Key: `any-string` · Done.
 
-> **Optional extra coverage (also free):** Groq API key (30 RPM free), NVIDIA NIM (40 RPM free, 70+ models), Cerebras (1M tok/day).
+> **Optional extra coverage (also free):** Groq API key (30 RPM free), NVIDIA NIM (40 RPM free, 70+ models), Cerebras (1M tok/day), LongCat API key (50M tokens/day!), Cloudflare Workers AI (10K Neurons/day, 50+ models).
 
 ## ⚡ Quick Start
 
@@ -921,18 +942,28 @@ When minimized, OmniRoute lives in your system tray with quick actions:
 | **🆓 FREE**         | iFlow                       | **$0**                    | Unlimited        | 5 models unlimited                |
 |                     | Qwen                        | **$0**                    | Unlimited        | 4 models unlimited                |
 |                     | Kiro                        | **$0**                    | Unlimited        | Claude Sonnet/Haiku (AWS Builder) |
+|                     | LongCat Flash-Lite 🆕       | **$0** (50M tok/day 🔥)   | 1 RPS            | Largest free quota on Earth       |
+|                     | Pollinations AI 🆕          | **$0** (no key needed)    | 1 req/15s        | GPT-5, Claude, DeepSeek, Llama 4  |
+|                     | Cloudflare Workers AI 🆕    | **$0** (10K Neurons/day)  | ~150 resp/day    | 50+ models, global edge           |
+|                     | Scaleway AI 🆕              | **$0** (1M tokens total)  | Rate limited     | EU/GDPR, Qwen3 235B, Llama 70B    |
 
 > 🆕 **New models added (Mar 2026):** Grok-4 Fast family at $0.20/$0.50/M (benchmarked at 1143ms — 30% faster than Gemini 2.5 Flash), GLM-5 via Z.AI with 128K output, MiniMax M2.5 reasoning, DeepSeek V3.2 updated pricing, Kimi K2.5 via Moonshot direct API.
 
 **💡 $0 Combo Stack — The Complete Free Setup:**
 
 ```
-Gemini CLI (180K/mo free)
-  → iFlow (unlimited: kimi-k2-thinking, qwen3-coder-plus, deepseek-r1)
-  → Kiro (Claude Sonnet 4.5 + Haiku — unlimited, via AWS Builder ID)
-  → Qwen (4 models — unlimited)
-  → Groq (14.4K req/day — ultra-fast)
-  → NVIDIA NIM (70+ models — 40 RPM forever)
+# 🆓 Ultimate Free Stack 2026 — 11 Providers, $0 Forever
+Kiro (kr/)             → Claude Sonnet/Haiku UNLIMITED
+iFlow (if/)            → kimi-k2-thinking, qwen3-coder-plus, deepseek-r1 UNLIMITED
+LongCat Lite (lc/)     → LongCat-Flash-Lite — 50M tokens/day 🔥
+Pollinations (pol/)    → GPT-5, Claude, DeepSeek, Llama 4 — no key needed
+Qwen (qw/)             → qwen3-coder-plus, qwen3-coder-flash, qwen3-coder-next UNLIMITED
+Gemini (gemini/)       → Gemini 2.5 Flash — 1,500 req/day free API key
+Cloudflare AI (cf/)    → Llama 70B, Gemma 3, Mistral — 10K Neurons/day
+Scaleway (scw/)        → Qwen3 235B, Llama 70B — 1M free tokens (EU)
+Groq (groq/)           → Llama/Gemma ultra-fast — 14.4K req/day
+NVIDIA NIM (nvidia/)   → 70+ open models — 40 RPM forever
+Cerebras (cerebras/)   → Llama/Qwen world-fastest — 1M tok/day
 ```
 
 **Zero cost. Never stops coding.** Configure this as one OmniRoute combo and all fallbacks happen automatically — no manual switching ever.
@@ -1003,19 +1034,66 @@ Available free: `llama-3.3-70b`, `llama-3.1-8b`, `deepseek-r1-distill-llama-70b`
 
 Available free: `llama-3.3-70b-versatile`, `gemma2-9b-it`, `mixtral-8x7b`, `whisper-large-v3`
 
-> **💡 The Ultimate Free Stack:**
+### 🔴 LONGCAT AI (Free API Key — longcat.chat) 🆕
+
+| Model                         | Prefix | Daily Free Quota  | Notes                   |
+| ----------------------------- | ------ | ----------------- | ----------------------- |
+| `LongCat-Flash-Lite`          | `lc/`  | **50M tokens** 💥 | Largest free quota ever |
+| `LongCat-Flash-Chat`          | `lc/`  | 500K tokens       | Multi-turn chat         |
+| `LongCat-Flash-Thinking`      | `lc/`  | 500K tokens       | Reasoning / CoT         |
+| `LongCat-Flash-Thinking-2601` | `lc/`  | 500K tokens       | Jan 2026 version        |
+| `LongCat-Flash-Omni-2603`     | `lc/`  | 500K tokens       | Multimodal              |
+
+> 100% free while in public beta. Sign up at [longcat.chat](https://longcat.chat) with email or phone. Resets daily 00:00 UTC.
+
+### 🟢 POLLINATIONS AI (No API Key Required) 🆕
+
+| Model      | Prefix | Rate Limit | Provider Behind    |
+| ---------- | ------ | ---------- | ------------------ |
+| `openai`   | `pol/` | 1 req/15s  | GPT-5              |
+| `claude`   | `pol/` | 1 req/15s  | Anthropic Claude   |
+| `gemini`   | `pol/` | 1 req/15s  | Google Gemini      |
+| `deepseek` | `pol/` | 1 req/15s  | DeepSeek V3        |
+| `llama`    | `pol/` | 1 req/15s  | Meta Llama 4 Scout |
+| `mistral`  | `pol/` | 1 req/15s  | Mistral AI         |
+
+> ✨ **Zero friction:** No signup, no API key. Add the Pollinations provider with an empty key field and it works immediately.
+
+### 🟠 CLOUDFLARE WORKERS AI (Free API Key — cloudflare.com) 🆕
+
+| Tier | Daily Neurons | Equivalent Usage                        | Notes                   |
+| ---- | ------------- | --------------------------------------- | ----------------------- |
+| Free | **10,000**    | ~150 LLM resp / 500s audio / 15K embeds | Global edge, 50+ models |
+
+Popular free models: `@cf/meta/llama-3.3-70b-instruct`, `@cf/google/gemma-3-12b-it`, `@cf/openai/whisper-large-v3-turbo` (free audio!), `@cf/qwen/qwen2.5-coder-15b-instruct`
+
+> Requires API Token + Account ID from [dash.cloudflare.com](https://dash.cloudflare.com). Store Account ID in provider settings.
+
+### 🟣 SCALEWAY AI (1M Free Tokens — scaleway.com) 🆕
+
+| Tier | Free Quota    | Location     | Notes                               |
+| ---- | ------------- | ------------ | ----------------------------------- |
+| Free | **1M tokens** | 🇫🇷 Paris, EU | No credit card needed within limits |
+
+Available free: `qwen3-235b-a22b-instruct-2507` (Qwen3 235B!), `llama-3.1-70b-instruct`, `mistral-small-3.2-24b-instruct-2506`, `deepseek-v3-0324`
+
+> EU/GDPR compliant. Get API key at [console.scaleway.com](https://console.scaleway.com).
+
+> **💡 The Ultimate Free Stack (11 Providers, $0 Forever):**
 >
 > ```
-> Kiro (Claude, unlimited)
->  → iFlow (5 models, unlimited)
->  → Qwen (4 models, unlimited)
->  → Gemini CLI (180K/mo)
->  → Cerebras (1M tok/day)
->  → Groq (14.4K req/day)
->  → NVIDIA NIM (40 RPM, 70+ models)
+> Kiro (kr/)             → Claude Sonnet/Haiku UNLIMITED
+> iFlow (if/)            → kimi-k2-thinking, qwen3-coder-plus, deepseek-r1 UNLIMITED
+> LongCat Lite (lc/)     → LongCat-Flash-Lite — 50M tokens/day 🔥
+> Pollinations (pol/)    → GPT-5, Claude, DeepSeek, Llama 4 — no key needed
+> Qwen (qw/)             → qwen3-coder models UNLIMITED
+> Gemini (gemini/)       → Gemini 2.5 Flash — 1,500 req/day free
+> Cloudflare AI (cf/)    → 50+ models — 10K Neurons/day
+> Scaleway (scw/)        → Qwen3 235B, Llama 70B — 1M free tokens (EU)
+> Groq (groq/)           → Llama/Gemma — 14.4K req/day ultra-fast
+> NVIDIA NIM (nvidia/)   → 70+ open models — 40 RPM forever
+> Cerebras (cerebras/)   → Llama/Qwen world-fastest — 1M tok/day
 > ```
->
-> Configure this as an OmniRoute combo and you'll never pay for AI again.
 
 ## 🎙️ Free Transcription Combo
 
@@ -1105,17 +1183,17 @@ OmniRoute v2.0 is built as an operational platform, not just a relay proxy.
 
 ### 🎵 Multi-Modal APIs
 
-| Feature                    | What It Does                                                                                                 |
-| -------------------------- | ------------------------------------------------------------------------------------------------------------ |
-| 🖼️ **Image Generation**    | `/v1/images/generations` with cloud and local backends                                                       |
-| 📐 **Embeddings**          | `/v1/embeddings` for search and RAG pipelines                                                                |
-| 🎤 **Audio Transcription** | `/v1/audio/transcriptions` (Whisper and additional providers)                                                |
-| 🔊 **Text-to-Speech**      | `/v1/audio/speech` (multiple engines/providers)                                                              |
-| 🎬 **Video Generation**    | `/v1/videos/generations` (ComfyUI + SD WebUI workflows)                                                      |
-| 🎵 **Music Generation**    | `/v1/music/generations` (ComfyUI workflows)                                                                  |
-| 🛡️ **Moderations**         | `/v1/moderations` safety checks                                                                              |
-| 🔀 **Reranking**           | `/v1/rerank` for relevance scoring                                                                           |
-| 🔍 **Web Search** 🆕       | `/v1/search` — 5 providers (Serper, Brave, Perplexity, Exa, Tavily), 6,500+ free/month, auto-failover, cache |
+| Feature                    | What It Does                                                                                                                                                               |
+| -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Image Generation**    | `/v1/images/generations` with cloud and local backends                                                                                                                     |
+| 📐 **Embeddings**          | `/v1/embeddings` for search and RAG pipelines                                                                                                                              |
+| 🎤 **Audio Transcription** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Text-to-Speech**      | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise) with correct error messages          |
+| 🎬 **Video Generation**    | `/v1/videos/generations` (ComfyUI + SD WebUI workflows)                                                                                                                    |
+| 🎵 **Music Generation**    | `/v1/music/generations` (ComfyUI workflows)                                                                                                                                |
+| 🛡️ **Moderations**         | `/v1/moderations` safety checks                                                                                                                                            |
+| 🔀 **Reranking**           | `/v1/rerank` for relevance scoring                                                                                                                                         |
+| 🔍 **Web Search** 🆕       | `/v1/search` — 5 providers (Serper, Brave, Perplexity, Exa, Tavily), 6,500+ free/month, auto-failover, cache                                                               |
 
 ### 🛡️ Resilience, Security & Governance
 

ZWS_README_V4.md

@@ -0,0 +1,374 @@
+# ZWS_README_V4 — 启动性能优化：HMR 泄漏修复与 Turbopack 迁移
+
+## 一、如何发现问题
+
+### 现象
+
+- `npm run dev` 后，首次打开浏览器白屏等待 **5-22 秒**不等。
+- 运行一段时间后 Node 进程内存飙升至 **2.4 GB**，触发 Next.js 内存阈值保护强制重启。
+- 重启后 `Ready in 82.6s`（正常冷启动仅 3.4s），之后每个页面首次编译需 **7-28 秒**。
+- 日志中大量重复输出，单次会话内：
+  - `[DB] SQLite database ready` 出现 **485 次**
+  - `[HealthCheck] Starting proactive token health-check` 出现 **586 次**
+  - `[CREDENTIALS] No external credentials file found` 出现 **432 次**
+
+### 排查过程
+
+1. **Terminal 日志分析**：统计关键日志出现次数，发现 DB 连接和 HealthCheck 定时器被反复创建。
+2. **代码审计**：追踪到所有受影响模块使用 `let initialized = false` 作为单例守卫——这在 Next.js dev 模式的 Webpack HMR 下会被重置。
+3. **对比**：`apiBridgeServer.ts` 使用了 `globalThis.__omnirouteApiBridgeStarted`，在日志中无重复初始化，验证了 `globalThis` 方案的有效性。
+4. **内存快照**：通过 `Get-Process node` 观察到两个 node 进程分别占用 1.7GB 和 1.0GB。
+5. **编译时间分析**：日志中 `compile:` 字段显示 Webpack 编译每个路由需 2-26 秒，对比 Turbopack 应在 0.5-3 秒。
+
+---
+
+## 二、根因分析
+
+### 根因 1（P0）：模块级单例在 HMR 中丢失
+
+Next.js dev 模式下，Webpack HMR 会重新执行被修改（或依赖链变化）的模块。模块级 `let` 变量在每次重新执行时被重置为初始值。
+
+```typescript
+// 修复前 — 每次 HMR 重新执行时 _db 重置为 null
+let _db: SqliteDatabase | null = null;
+
+export function getDbInstance() {
+  if (_db) return _db; // HMR 后这里永远 false
+  // ... 重新打开一个新的 DB 连接（旧连接泄漏）
+}
+```
+
+**受影响的模块与泄漏类型：**
+
+| 模块                    | 泄漏资源               | 累计次数 | 后果                    |
+| ----------------------- | ---------------------- | -------- | ----------------------- |
+| `db/core.ts`            | SQLite 连接            | 485      | 文件句柄泄漏 + 内存占用 |
+| `tokenHealthCheck.ts`   | `setInterval` 定时器   | 586      | CPU 空转 + DB 查询风暴  |
+| `localHealthCheck.ts`   | `setTimeout` 定时器链  | ~400     | 重复 HTTP 请求 + CPU    |
+| `consoleInterceptor.ts` | console 方法包装       | ~400     | 日志 double-write       |
+| `gracefulShutdown.ts`   | SIGTERM/SIGINT handler | ~400     | 信号处理器堆叠          |
+
+**级联效应**：泄漏的资源持续消耗内存和 CPU → 触发 Next.js 内存阈值保护 → 进程重启 → Webpack 从零重建模块图 → **Ready in 82.6s**。
+
+### 根因 2（P0）：强制使用 Webpack 而非 Turbopack
+
+`scripts/run-next.mjs` 中硬编码了 `--webpack` 标志：
+
+```javascript
+if (mode === "dev") {
+  args.splice(2, 0, "--webpack");
+}
+```
+
+Next.js 16 默认使用 Turbopack（Rust 编写的增量打包器），dev 编译速度是 Webpack 的 5-10 倍。强制回退到 Webpack 导致：
+
+| 指标                    | Webpack | Turbopack（预期） |
+| ----------------------- | ------- | ----------------- |
+| 首页编译                | 3.7s    | ~0.5s             |
+| Provider 详情页首次编译 | 22s     | ~2-3s             |
+| API route 首次编译      | 2-7s    | ~0.3-1s           |
+| 内存重启后 Ready        | 82.6s   | 不会触发          |
+
+### 根因 3（P1）：`node:crypto` 被拉入客户端 bundle
+
+`src/lib/db/proxies.ts` 使用了 `import { randomUUID } from "node:crypto"`。通过 `localDb.ts` 的 re-export 链，这个 Node.js 原生模块被间接拉入客户端组件的 bundle，导致 Webpack 报错：
+
+```
+UnhandledSchemeError: Reading from "node:crypto" is not handled by plugins
+Import trace: node:crypto → ./src/lib/db/proxies.ts → ./src/lib/localDb.ts → page.tsx
+```
+
+Webpack 无法处理 `node:` URI scheme 前缀。`crypto`（不带 `node:` 前缀）已在 `next.config.mjs` 的 `serverExternalPackages` 中声明为服务端外部包。
+
+### 根因 4（P1）：Edge Runtime 编译警告刷屏
+
+Next.js 16 会同时为 **Node.js** 和 **Edge** 两种运行时编译 `instrumentation.ts`。虽然 `register()` 函数内有 `process.env.NEXT_RUNTIME === "nodejs"` 的运行时守卫，但 Turbopack 在打包 Edge 版本时仍会**静态追踪**所有动态 `import()` 的依赖链：
+
+```
+instrumentation.ts
+  → import("@/lib/db/secrets")
+    → @/lib/db/core.ts  →  fs, path, better-sqlite3
+    → @/lib/dataPaths.ts →  path, os
+    → @/lib/db/migrationRunner.ts → fs, path, url
+```
+
+对每个 Node.js 原生模块，Turbopack 都输出一条 "not supported in Edge Runtime" 警告。每次有新请求触发热编译时，这组 **10+ 条警告重复刷一遍**，严重污染终端输出，干扰开发调试。
+
+### 根因 5（P2）：启动 import 完全串行
+
+`instrumentation.ts` 中 9 个 `await import()` 完全串行执行，每个都可能触发 Webpack 编译其依赖树：
+
+```typescript
+await ensureSecrets();                              // 串行 1
+const { initConsoleInterceptor } = await import(...); // 串行 2
+const { initGracefulShutdown } = await import(...);   // 串行 3
+const { initApiBridgeServer } = await import(...);    // 串行 4
+const { startBackgroundRefresh } = await import(...); // 串行 5
+const { getSettings } = await import(...);            // 串行 6
+const { setCustomAliases } = await import(...);       // 串行 7
+const { setDefaultFastServiceTierEnabled } = await import(...); // 串行 8
+const { initAuditLog, cleanupExpiredLogs } = await import(...); // 串行 9
+```
+
+其中 4-6 互不依赖，7-8 互不依赖，完全可以并行。
+
+---
+
+## 三、修复方案
+
+### 修复 1：globalThis 单例守卫（core.ts, tokenHealthCheck.ts, localHealthCheck.ts, consoleInterceptor.ts, gracefulShutdown.ts）
+
+**原理**：`globalThis` 对象在 Node.js 进程生命周期内全局唯一，不受 Webpack 模块重新执行的影响。
+
+```typescript
+// 修复后 — globalThis 在 HMR 后依然保留
+declare global {
+  var __omnirouteDb: import("better-sqlite3").Database | undefined;
+}
+
+function getDb() {
+  return globalThis.__omnirouteDb ?? null;
+}
+function setDb(db) {
+  /* ... */
+}
+
+export function getDbInstance() {
+  const existing = getDb();
+  if (existing) return existing; // HMR 后命中缓存
+  // ...
+}
+```
+
+**每个模块的具体改动：**
+
+| 模块                    | globalThis key                      | 守卫内容                                                    |
+| ----------------------- | ----------------------------------- | ----------------------------------------------------------- |
+| `db/core.ts`            | `__omnirouteDb`                     | SQLite 连接实例                                             |
+| `tokenHealthCheck.ts`   | `__omnirouteTokenHC`                | `{ initialized, interval }`                                 |
+| `localHealthCheck.ts`   | `__omnirouteLocalHC`                | `{ initialized, sweepTimer, healthCache, sweepInProgress }` |
+| `consoleInterceptor.ts` | `__omnirouteConsoleInterceptorInit` | `boolean`                                                   |
+| `gracefulShutdown.ts`   | `__omnirouteShutdownInit`           | `boolean`                                                   |
+
+**优点**：
+
+- 零依赖，无需额外库。
+- 与 `apiBridgeServer.ts` 已有模式一致。
+- 对生产环境零影响（非 HMR 场景下行为完全相同）。
+
+**缺点/注意**：
+
+- `globalThis` 键名需全局唯一，使用 `__omniroute` 前缀避免冲突。
+- 需要 `declare global` 类型声明以保持 TypeScript 类型安全。
+- 生产构建中 `globalThis` 存储略冗余（但仅是一个对象引用，几乎零开销）。
+
+### 修复 2：支持通过环境变量切换 Turbopack（run-next.mjs）
+
+```javascript
+// 修复后 — 默认仍用 webpack（保持原有行为），设置环境变量可启用 Turbopack
+if (mode === "dev" && process.env.OMNIROUTE_USE_TURBOPACK !== "1") {
+  args.splice(2, 0, "--webpack");
+}
+```
+
+**默认行为不变**：dev 模式仍使用 Webpack，与修复前完全一致。设置 `OMNIROUTE_USE_TURBOPACK=1` 可切换到 Turbopack 以获得更快的 dev 编译速度。
+
+**优点**：
+
+- 零风险：不改变任何人的现有体验。
+- 需要时设置 `OMNIROUTE_USE_TURBOPACK=1` 即可获得 5-10 倍编译加速。
+- `next.config.mjs` 中已有 `turbopack.resolveAlias` 配置，说明项目已在准备 Turbopack 迁移。
+
+**缺点/注意**：
+
+- Turbopack 对某些 Webpack 特定配置（如自定义 externals 函数）的支持方式不同，启用前需测试兼容性。
+- 默认走 Webpack 意味着不主动启用 Turbopack 的用户无法享受编译加速。
+
+### 修复 3：`node:crypto` → `crypto`（proxies.ts, errorResponse.ts）
+
+```typescript
+// 修复前
+import { randomUUID } from "node:crypto";
+
+// 修复后
+import { randomUUID } from "crypto";
+```
+
+**优点**：
+
+- `crypto`（无 `node:` 前缀）已在 `next.config.mjs` 的 `serverExternalPackages` 列表中，Webpack/Turbopack 会正确将其标记为外部包。
+- 消除 `UnhandledSchemeError` 构建失败。
+- Node.js 中 `crypto` 和 `node:crypto` 解析到同一模块。
+
+**缺点**：
+
+- 无。`crypto` 是 Node.js 内建模块，两种写法功能完全等价。
+
+### 修复 4：分离 Edge/Node.js Instrumentation（instrumentation.ts → instrumentation-node.ts）
+
+**问题**：`instrumentation.ts` 中所有 Node.js 逻辑（`ensureSecrets`、DB 初始化、审计日志等）虽然只在 `NEXT_RUNTIME === "nodejs"` 时执行，但 Turbopack 编译 Edge 版本时仍静态追踪其 import 链，对每个 `fs`/`path`/`os`/`better-sqlite3` 等原生模块输出警告。
+
+**方案**：将所有 Node.js 专属逻辑提取到 `src/instrumentation-node.ts`，主文件通过**计算的 import 路径**引入，阻止 Turbopack 静态解析：
+
+```typescript
+// src/instrumentation.ts — 精简后仅 ~20 行
+export async function register() {
+  if (process.env.NEXT_RUNTIME === "nodejs") {
+    // 拼接路径阻止 Turbopack 在 Edge 编译时静态解析模块依赖
+    const nodeMod = "./instrumentation-" + "node";
+    const { registerNodejs } = await import(nodeMod);
+    await registerNodejs();
+  }
+}
+```
+
+```typescript
+// src/instrumentation-node.ts — 包含全部 Node.js 启动逻辑
+export async function registerNodejs(): Promise<void> {
+  await ensureSecrets();
+  // initConsoleInterceptor, initGracefulShutdown, initApiBridgeServer, ...
+  // （原 instrumentation.ts 的完整 Node.js 逻辑）
+}
+```
+
+**关键技术**：`"./instrumentation-" + "node"` 是运行时拼接的字符串，Turbopack 无法在编译期确定其值，因此**不会追踪**该 import 的依赖树。Node.js 运行时则正常解析该路径并执行。
+
+**优点**：
+
+- Edge 编译时完全跳过 Node.js 模块追踪，**10+ 条重复警告全部消除**。
+- Node.js 运行时行为与修复前完全一致。
+- 启动时间从 **13.9s → 1.25s**（Turbopack 不再在 Edge 编译中处理 Node.js 模块图）。
+
+**缺点/注意**：
+
+- 新增一个文件 `instrumentation-node.ts`，需同步维护。
+- 计算 import 路径是有意为之的 bundler 逃逸技巧，需加注释说明原因防止后续重构时被"优化"回静态字符串。
+
+### 修复 5：并行化 instrumentation.ts 中的启动 import
+
+```typescript
+// 修复后 — 4 个独立模块并行导入
+const [
+  { initGracefulShutdown },
+  { initApiBridgeServer },
+  { startBackgroundRefresh },
+  { getSettings },
+] = await Promise.all([
+  import("@/lib/gracefulShutdown"),
+  import("@/lib/apiBridgeServer"),
+  import("@/domain/quotaCache"),
+  import("@/lib/db/settings"),
+]);
+
+// 2 个 open-sse 模块也并行导入
+const [{ setCustomAliases }, { setDefaultFastServiceTierEnabled }] = await Promise.all([
+  import("@omniroute/open-sse/services/modelDeprecation.ts"),
+  import("@omniroute/open-sse/executors/codex.ts"),
+]);
+```
+
+**优点**：
+
+- `consoleInterceptor` 仍保持第一个（必须在任何日志前初始化）。
+- 后续 4 个无依赖模块并行加载，节省 3 次串行等待。
+- open-sse 的 2 个模块也并行加载。
+
+**缺点**：
+
+- 并行 import 的错误堆栈略复杂（Promise.all 中某一个失败会 reject 整个组）。
+- 这里的 compliance 模块仍保持独立 try/catch 串行，因为它有自己的错误处理逻辑。
+
+---
+
+## 四、预期效果
+
+| 指标                          | 修复前                    | 修复后（预期）           |
+| ----------------------------- | ------------------------- | ------------------------ |
+| DB 连接创建次数               | 485 次/会话               | 1 次                     |
+| HealthCheck 定时器            | 586 个泄漏                | 1 个                     |
+| 信号处理器注册                | ~400 次重复               | 1 次                     |
+| Console 拦截层数              | ~400 层嵌套               | 1 层                     |
+| 内存使用峰值                  | 2.4 GB → OOM 重启         | 预期 < 500 MB            |
+| 冷启动 Ready                  | 3.4s                      | ~3s（略快）              |
+| 内存重启 Ready                | 82.6s                     | 不再触发内存重启         |
+| Login 页首次编译              | 3.7s                      | ~0.5s (需启用 Turbopack) |
+| Provider 详情页首次编译       | 22s                       | ~2-3s (需启用 Turbopack) |
+| `node:crypto` 构建错误        | 反复出现                  | 消除                     |
+| Edge Runtime 编译警告         | 每次热编译刷出 10+ 条     | **0 条**                 |
+| instrumentation 启动耗时      | 13.9s（含 Edge 模块追踪） | **1.25s**                |
+| instrumentation import 并行度 | 9 次串行 import           | 3 批并行 import          |
+
+---
+
+## 五、涉及文件清单
+
+| 区域                | 文件                            | 改动类型                                                           |
+| ------------------- | ------------------------------- | ------------------------------------------------------------------ |
+| DB 单例             | `src/lib/db/core.ts`            | `let _db` → `globalThis.__omnirouteDb`                             |
+| Token 健康检查      | `src/lib/tokenHealthCheck.ts`   | `let initialized` → `globalThis.__omnirouteTokenHC`                |
+| 本地节点健康检查    | `src/lib/localHealthCheck.ts`   | `let initialized` → `globalThis.__omnirouteLocalHC`                |
+| Console 拦截        | `src/lib/consoleInterceptor.ts` | `let initialized` → `globalThis.__omnirouteConsoleInterceptorInit` |
+| 优雅关停            | `src/lib/gracefulShutdown.ts`   | 新增 `globalThis.__omnirouteShutdownInit` 守卫                     |
+| Dev 启动脚本        | `scripts/run-next.mjs`          | 新增 `OMNIROUTE_USE_TURBOPACK=1` 开关                              |
+| Proxy 注册表        | `src/lib/db/proxies.ts`         | `node:crypto` → `crypto`                                           |
+| API 错误响应        | `src/lib/api/errorResponse.ts`  | `node:crypto` → `crypto`                                           |
+| 启动钩子（主入口）  | `src/instrumentation.ts`        | 精简为 ~20 行，计算 import 路径阻止 Edge 追踪                      |
+| 启动钩子（Node.js） | `src/instrumentation-node.ts`   | 新文件，承载全部 Node.js 启动逻辑 + `Promise.all` 并行             |
+
+---
+
+## 六、回退方案
+
+- **启用 Turbopack**：设置 `OMNIROUTE_USE_TURBOPACK=1` 环境变量；不设置则默认使用 Webpack（原有行为不变）。
+- **globalThis 方案异常**：所有 globalThis key 都以 `__omniroute` 为前缀，可通过 `delete globalThis.__omnirouteDb` 等方式手动重置。
+- **Edge 警告回退**：若 `instrumentation-node.ts` 拆分导致问题，可将其内容合并回 `instrumentation.ts`，恢复为直接 `import()` 调用（警告会重新出现但不影响功能）。
+- **生产环境**：以上修复对生产构建无负面影响——生产环境不存在 HMR，globalThis 单例仅在首次调用时初始化一次。计算 import 路径在 `next build` 时由 Node.js 正常解析，不影响打包产物。
+
+---
+
+## 七、单元测试与备份恢复（pre-commit 验证通过）
+
+为保证提交前必须通过验证（不再使用 `--no-verify`），对以下失败用例与生产逻辑做了修复与加固。
+
+### 问题与根因
+
+| 失败项                                | 根因                                                                                                                                    |
+| ------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
+| bootstrap-env 4 个用例                | Windows 上 DATA_DIR 解析用 `APPDATA`/`homedir()`，测试只设了 `HOME`，脚本读不到测试用的 `.env`。                                        |
+| domain-persistence costRules 2 个用例 | `core` 在首次 import 时缓存 `DATA_DIR`；测试每测一个 tmpDir 并在 afterEach 删目录，导致后续 describe 使用的 DB 路径已被删，读写得到 0。 |
+| fixes-p1 restoreDbBackup              | 测试在 DB 仍打开时写 stale 侧文件；`restoreDbBackup` 内 pre-restore 备份未 await 就关库，Windows 上句柄未及时释放，unlink 报 EBUSY。    |
+| fixes-p1 resetStorage 及后续用例      | 上一测留下 DB 打开，下一测 `resetStorage()` 删目录时文件仍被占用，EBUSY。                                                               |
+
+### 修复 6：bootstrap-env 测试（tests/unit/bootstrap-env.test.mjs）
+
+在每个用例的 `withTempEnv` 回调开头增加 `process.env.DATA_DIR = dataDir`，使脚本在任意平台（含 Windows）都使用测试临时目录，而不是依赖 `HOME`/`APPDATA`。
+
+### 修复 7：domain-persistence 测试（tests/unit/domain-persistence.test.mjs）
+
+- **单例 tmpDir**：全文件共用一个 `fileTmpDir`，在模块加载时创建并设置 `process.env.DATA_DIR`，与 `core` 首次加载时缓存的路径一致。
+- **每测清 DB 不清目录**：`beforeEach` 中 `resetDbInstance()` 后删除 `storage.sqlite` 及其 `-wal`/`-shm`/`-journal`，保证每测干净 DB，不在 afterEach 删目录，避免路径失效。
+- **收尾**：`after()` 中恢复 `DATA_DIR` 并删除 `fileTmpDir`。
+- **costRules 断言**：改为小容差精确校验（`assertAlmostEqual`），继续验证 `4.5` / `4.0` 这类业务关键值，避免把真实累计错误放过去。
+
+### 修复 8：fixes-p1 测试（tests/unit/fixes-p1.test.mjs）
+
+- **restoreDbBackup 用例**：在写入 stale 侧文件前调用 `core.resetDbInstance()`，避免 DB 仍打开时写 `-wal`/`-shm` 触发 Windows 锁错误。
+- **Windows 跳过**：该用例在 Windows 上仍使用 `test(..., { skip: isWindows })`。原因不是业务逻辑不支持 Windows，而是 better-sqlite3 关闭后底层句柄释放存在时序抖动，这条真实 sidecar 集成测试容易退化成不稳定的文件锁测试；Linux/macOS 上照常运行。
+- **核心兜底测试**：新增平台无关的 `unlinkFileWithRetry` 单测，直接模拟 `EBUSY` / `EPERM` 后重试并最终成功，确保 Windows 相关的重试删除逻辑被稳定覆盖，而不是完全依赖 flaky 的真实文件锁时序。
+- **resetStorage**：改为 async，对 `rmSync(TEST_DATA_DIR)` 做最多 10 次、间隔 100ms 的 EBUSY/EPERM 重试，避免下一测因上一测句柄未释放而失败。
+
+### 修复 9：备份恢复逻辑（src/lib/db/backup.ts）
+
+- **pre-restore 备份改为同步等待**：在 `restoreDbBackup` 内用内联逻辑做 pre-restore 备份并 `await` 完成，再调用 `resetDbInstance()`，避免异步 backup 未结束就关库导致后续 unlink 失败。
+- **节流语义保持一致**：pre-restore 备份成功后补回 `_lastBackupAt = Date.now()`，避免恢复后紧接着又触发一轮额外自动备份。
+- **关库后短延迟**：`resetDbInstance()` 后 `await new Promise(r => setTimeout(r, 500))`，再执行 unlink，给 Windows 等平台释放句柄留时间。
+- **unlink 重试**：将主库及 `-wal`/`-shm`/`-journal` 的删除提取为 `unlinkFileWithRetry`，统一做最多 10 次、间隔 100ms 的 EBUSY/EPERM 重试，提高恢复流程在锁释放较慢环境下的成功率，也便于单测直接覆盖重试逻辑。
+
+### 涉及文件（本节）
+
+| 区域     | 文件                                     | 改动类型                                                                                                        |
+| -------- | ---------------------------------------- | --------------------------------------------------------------------------------------------------------------- |
+| 单元测试 | `tests/unit/bootstrap-env.test.mjs`      | 各用例内设置 `process.env.DATA_DIR = dataDir`                                                                   |
+| 单元测试 | `tests/unit/domain-persistence.test.mjs` | 单例 tmpDir、beforeEach 清 DB 文件、after 删目录；costRules 改为小容差精确断言                                  |
+| 单元测试 | `tests/unit/fixes-p1.test.mjs`           | restoreDbBackup 前 resetDbInstance、Windows skip 说明、resetStorage 重试、`unlinkFileWithRetry` 核心单测        |
+| 备份恢复 | `src/lib/db/backup.ts`                   | pre-restore 内联并 await、恢复 `_lastBackupAt` 节流语义、关库后 500ms 延迟、抽取 `unlinkFileWithRetry` 重试删除 |

ZWS_README_V5.md

@@ -0,0 +1,332 @@
+# ZWS_README_V5 — 按协议配置模型兼容性 + 前端性能优化
+
+V4 内容（HMR 泄漏修复、Edge 警告消除、测试稳定性）已完成；V5 在 V4 基础上实现**按协议维度配置模型兼容性**，新增前端查找性能优化与类型安全改进。
+
+---
+
+## 一、如何发现问题
+
+### 现象
+
+- 同一模型被 **OpenAI Chat Completions**、**OpenAI Responses API**、**Anthropic Messages** 三种客户端请求形态调用时，V2 的兼容性开关（工具 ID 9 位、不保留 developer 角色）是**全局生效**的——无法为不同协议设置不同的兼容策略。
+- 例如：用户希望 OpenAI Responses API 请求时不保留 developer 角色（MiniMax 422 修复），但 OpenAI Chat Completions 请求时保留。V2 下只能二选一。
+- 前端兼容性弹层未标明当前配置对应哪种协议，容易误导。
+- 前端组件中 `Array.find()` 在每次渲染时对 customModels 和 modelCompatOverrides 做 O(n) 线性扫描，模型数量多时存在不必要的性能开销。
+- `ModelCompatPatch` 类型定义与运行时逻辑不一致：`preserveOpenAIDeveloperRole` 字段需要支持 `null`（表示取消设置/恢复默认），但类型仅允许 `boolean`。
+
+### 排查过程
+
+1. **需求分析**：梳理 `detectFormat(body)` 返回的三种协议键（`openai`、`openai-responses`、`claude`），确认每种协议对 developer 角色和 tool call ID 的需求不同。
+2. **数据模型设计**：在现有 `normalizeToolCallId` / `preserveOpenAIDeveloperRole` 顶层字段基础上，设计 `compatByProtocol` 嵌套结构，按协议键细分。
+3. **构建问题**：客户端 `"use client"` 组件直接从 `@/lib/localDb` 引入常量时，间接拉入了 `node:crypto`（经由 `db/proxies.ts`），触发 Webpack `UnhandledSchemeError`。需将常量拆到 `shared/` 层。
+4. **前端性能**：通过 React DevTools 和代码审计发现 `effectiveNormalizeForProtocol` 等函数每次调用都对数组做 `find()`，在渲染列表时存在 O(n²) 的隐患。
+
+---
+
+## 二、根因分析
+
+### 根因 1（P0）：兼容选项无协议维度
+
+V2 的 `normalizeToolCallId` / `preserveOpenAIDeveloperRole` 存储在模型级别的顶层字段，无法区分请求来源协议。`chatCore.ts` 中的 getter 函数只接收 `(providerId, modelId)` 两个参数，不感知当前请求的 `sourceFormat`。
+
+**影响**：跨协议场景下用户只能设置一个全局值，无法精确控制。
+
+### 根因 2（P1）：客户端构建拉入 Node.js 模块
+
+`page.tsx`（"use client"）→ `@/lib/localDb` → `db/proxies.ts` → `import { randomUUID } from "node:crypto"`
+
+Webpack 无法处理 `node:` URI scheme，报 `UnhandledSchemeError`。虽然 V4 已将 `node:crypto` → `crypto` 修复了 `proxies.ts`，但 `localDb.ts` 的 barrel export 链仍然存在风险——客户端组件不应引入任何可能传递到 Node.js 模块的路径。
+
+### 根因 3（P2）：前端查找性能
+
+`effectiveNormalizeForProtocol`、`effectivePreserveForProtocol`、`anyNormalizeCompatBadge`、`anyNoPreserveCompatBadge` 四个函数每次调用都使用 `Array.find()` 在 `customModels` 和 `modelCompatOverrides` 数组中查找目标模型。在模型列表渲染时，每个模型行会调用多次这些函数，导致 O(n × m) 的查找开销（n = 模型数，m = 每行调用次数）。
+
+### 根因 4（P2）：类型定义与运行时不一致
+
+```typescript
+// V3 暂存区版本（有问题）
+export type ModelCompatPatch = Partial<
+  Pick<
+    ModelCompatOverride,
+    "normalizeToolCallId" | "preserveOpenAIDeveloperRole" | "compatByProtocol"
+  >
+>;
+```
+
+`ModelCompatOverride.preserveOpenAIDeveloperRole` 类型为 `boolean | undefined`，但 `mergeModelCompatOverride()` 内部有 `=== null` 判断（用于取消设置/恢复默认），类型层面无法覆盖。
+
+---
+
+## 三、修复方案
+
+### 修复 1：`compatByProtocol` 存储与读取（models.ts）
+
+**新增数据结构**：
+
+```typescript
+type CompatByProtocolMap = Partial<Record<ModelCompatProtocolKey, ModelCompatPerProtocol>>;
+
+export type ModelCompatOverride = {
+  id: string;
+  normalizeToolCallId?: boolean;
+  preserveOpenAIDeveloperRole?: boolean;
+  compatByProtocol?: CompatByProtocolMap; // 新增
+};
+```
+
+**读取优先级链**（适用于 `getModelNormalizeToolCallId` 和 `getModelPreserveOpenAIDeveloperRole`）：
+
+```
+compatByProtocol[sourceFormat].field  →  顶层 field  →  默认值
+```
+
+1. 若 `sourceFormat` 属于已知协议键（`openai` / `openai-responses` / `claude`），且 `compatByProtocol[sourceFormat]` 中存在目标字段，使用该值。
+2. 否则回退到顶层字段。
+3. 顶层字段也不存在时使用默认值（normalizeToolCallId=false，preserveOpenAIDeveloperRole=undefined）。
+
+**深度合并逻辑** `deepMergeCompatByProtocol()`：
+
+- 对每个协议键，逐字段合并而非覆盖。
+- `normalizeToolCallId=false` 时删除该字段（不存储 false，减少冗余）。
+- 合并后若整个协议条目为空对象，删除该协议条目。
+- 协议键通过 `isCompatProtocolKey()` 白名单校验，拒绝未知键。
+
+**Getter 签名扩展**（向后兼容，第三参数可选）：
+
+```typescript
+export function getModelNormalizeToolCallId(
+  providerId: string,
+  modelId: string,
+  sourceFormat?: string | null
+): boolean;
+
+export function getModelPreserveOpenAIDeveloperRole(
+  providerId: string,
+  modelId: string,
+  sourceFormat?: string | null
+): boolean | undefined;
+```
+
+**优点**：
+
+- 完全向后兼容：无 `sourceFormat` 参数时行为与 V2 完全一致。
+- 协议键白名单校验防止存储污染。
+- 深度合并保留未变更协议的配置。
+
+**缺点/注意**：
+
+- JSON 存储体积略增（每个模型最多增加 3 个协议条目）。
+- 新增 ~80 行 TypeScript 代码。
+
+### 修复 2：请求管线传入 sourceFormat（chatCore.ts）
+
+```typescript
+const normalizeToolCallId = getModelNormalizeToolCallId(
+  provider || "",
+  model || "",
+  sourceFormat // 新增第三参
+);
+const preserveDeveloperRole = getModelPreserveOpenAIDeveloperRole(
+  provider || "",
+  model || "",
+  sourceFormat // 新增第三参
+);
+```
+
+`sourceFormat` 由已有的 `detectFormat(body)` 返回，无需新增检测逻辑。
+
+**优点**：
+
+- 改动仅 2 行，精准传参。
+- 不影响其他 handler（embeddings、imageGeneration 等不涉及 developer 角色和 tool call ID）。
+
+### 修复 3：API 路由支持 compatByProtocol（route.ts）
+
+**PUT 请求体扩展**：
+
+- 解构 `compatByProtocol` 并传入 `updateCustomModel()`。
+- `compatOnly` 判断扩展：仅含 `provider` + `modelId` + 兼容字段时，走 `mergeModelCompatOverride()` 路径。
+- 使用 `ModelCompatPatch` 类型替代行内类型定义，统一类型来源。
+
+**Zod 校验 schema**：
+
+```typescript
+const modelCompatPerProtocolSchema = z.object({
+  normalizeToolCallId: z.boolean().optional(),
+  preserveOpenAIDeveloperRole: z.boolean().optional(),
+}).strict();  // strict: 拒绝额外字段
+
+compatByProtocol: z
+  .record(z.enum(["openai", "openai-responses", "claude"]), modelCompatPerProtocolSchema)
+  .optional(),
+```
+
+**优点**：
+
+- `.strict()` 防止客户端注入额外字段。
+- `z.enum()` 限定协议键，与后端白名单一致。
+- 仅传 `compatByProtocol` 即可更新，前端无需拼装完整模型对象。
+
+### 修复 4：客户端安全常量拆分（modelCompat.ts）
+
+**新增** `src/shared/constants/modelCompat.ts`：
+
+```typescript
+export const MODEL_COMPAT_PROTOCOL_KEYS = ["openai", "openai-responses", "claude"] as const;
+export type ModelCompatProtocolKey = (typeof MODEL_COMPAT_PROTOCOL_KEYS)[number];
+```
+
+- 不依赖 Node.js / DB 代码，客户端组件可安全引入。
+- `models.ts` 从此模块引入并再导出。
+- `localDb.ts` 新增 `ModelCompatPatch` 类型导出（供 route.ts 使用），不导出协议常量。
+- `page.tsx` 改为从 `@/shared/constants/modelCompat` 引入。
+
+**优点**：
+
+- 彻底切断客户端 → localDb → db → proxies → node:crypto 的依赖链。
+- 协议键定义单一来源（Single Source of Truth）。
+
+### 修复 5：前端协议选择器与按协议解析（page.tsx）
+
+**ModelCompatPopover 重构**：
+
+- 新增协议下拉选择器（`<select>`），可选 OpenAI Chat / OpenAI Responses / Anthropic Messages。
+- 两个开关（工具 ID 9 位、不保留 developer）**针对选中协议**生效。
+- 选择 Claude 协议时隐藏 developer 角色开关（developer 仅对 OpenAI 系有意义）。
+- 保存时以 `{ compatByProtocol: { [protocol]: payload } }` 形式提交，后端按协议合并。
+- 深色模式适配：下拉框使用 `bg-white dark:bg-zinc-800`、`text-zinc-900 dark:text-zinc-100`。
+
+**Props 接口重构**：
+
+旧接口（4 个独立值/回调）：
+
+```typescript
+(normalizeToolCallId, preserveDeveloperRole, onNormalizeChange, onPreserveChange);
+```
+
+新接口（3 个函数式 props）：
+
+```typescript
+effectiveModelNormalize: (protocol: string) => boolean
+effectiveModelPreserveDeveloper: (protocol: string) => boolean
+onCompatPatch: (protocol: string, payload: {...}) => void
+```
+
+所有消费方（`ModelRow`、`PassthroughModelRow`、`CustomModelsSection`、`CompatibleModelsSection`）已同步更新。
+
+**角标显示逻辑**：
+
+- `anyNormalizeCompatBadge()`：任意协议或顶层存在 `normalizeToolCallId=true` 即显示「ID×9」角标。
+- `anyNoPreserveCompatBadge()`：任意协议或顶层存在 `preserveOpenAIDeveloperRole=false` 即显示「不保留」角标。
+
+**CustomModelsSection 增强**：
+
+- 新增 `modelCompatOverrides` 状态，从 API 响应中获取。
+- 新增 `saveCustomCompat()` 函数，支持仅传 `compatByProtocol` 的独立保存。
+
+### 修复 6：前端 Map 查找性能优化（page.tsx）
+
+**问题**：`effectiveNormalizeForProtocol` 等函数对 `customModels` 和 `modelCompatOverrides` 用 `Array.find()` 做 O(n) 查找，在列表渲染时每个模型行多次调用。
+
+**方案**：使用 `useMemo` + `Map` 将数组预建为 O(1) 查找表。
+
+```typescript
+type CompatModelMap = Map<string, CompatModelRow>;
+
+function buildCompatMap(rows: CompatModelRow[]): CompatModelMap {
+  const m = new Map<string, CompatModelRow>();
+  for (const r of rows) if (r.id) m.set(r.id, r);
+  return m;
+}
+
+// 在组件内
+const customMap = useMemo(() => buildCompatMap(modelMeta.customModels), [modelMeta.customModels]);
+const overrideMap = useMemo(
+  () => buildCompatMap(modelMeta.modelCompatOverrides),
+  [modelMeta.modelCompatOverrides]
+);
+```
+
+所有查找函数签名从 `(modelId, protocol, customModels[], overrides[])` 改为 `(modelId, protocol, customMap, overrideMap)`，内部使用 `Map.get()` 替代 `Array.find()`。
+
+**优点**：
+
+- 查找从 O(n) 降为 O(1)。
+- `useMemo` 依赖项正确，仅在数据变化时重建 Map。
+- `CustomModelsSection` 内部也独立构建 Map，不依赖父组件。
+
+### 修复 7：ModelCompatPatch 类型修正（models.ts）
+
+```typescript
+// 修复后 — 显式允许 null
+export type ModelCompatPatch = {
+  normalizeToolCallId?: boolean;
+  preserveOpenAIDeveloperRole?: boolean | null; // null = 取消设置/恢复默认
+  compatByProtocol?: CompatByProtocolMap;
+};
+```
+
+与 `mergeModelCompatOverride()` 内的 `=== null` 判断逻辑一致，类型安全。
+
+### 修复 8：CompatByProtocolMap 类型收紧（page.tsx）
+
+客户端 `CompatByProtocolMap` 从 `Record<string, ...>` 改为 `Record<ModelCompatProtocolKey, ...>`，增强类型安全，防止传入未知协议键。
+
+### 修复 9：i18n 文案新增
+
+| 键名                            | 中文                                          | 英文                                                           |
+| ------------------------------- | --------------------------------------------- | -------------------------------------------------------------- |
+| `compatProtocolLabel`           | 客户端请求协议                                | Client request protocol                                        |
+| `compatProtocolHint`            | 以下选项在 OmniRoute 识别到该请求形态时生效。 | These options apply when OmniRoute detects this request shape. |
+| `compatProtocolOpenAI`          | OpenAI Chat Completions                       | OpenAI Chat Completions                                        |
+| `compatProtocolOpenAIResponses` | OpenAI Responses API                          | OpenAI Responses API                                           |
+| `compatProtocolClaude`          | Anthropic Messages                            | Anthropic Messages                                             |
+
+---
+
+## 四、使用方式
+
+1. 点击模型行的 **「兼容性」** 按钮。
+2. 在弹层内先选择 **「客户端请求协议」**（OpenAI Chat / OpenAI Responses / Anthropic Messages）。
+3. 勾选该协议下的「工具 ID 9 位」或「不保留 developer 角色」。
+4. 保存后，仅在该协议形态的请求下生效。
+5. 未配置某协议时，该协议下行为回退到顶层兼容字段（若存在），再回退到默认值（保留 developer、不规范化 tool id）。
+6. 角标「ID×9」「不保留」在任意协议存在对应配置时显示。
+
+---
+
+## 五、预期效果
+
+| 指标                      | 修复前                | 修复后                                     |
+| ------------------------- | --------------------- | ------------------------------------------ |
+| 兼容性配置维度            | 全局（模型级）        | 按协议（OpenAI Chat / Responses / Claude） |
+| developer 角色精确控制    | 不支持                | 支持（如：仅 Responses API 不保留）        |
+| 前端兼容性查找性能        | O(n) Array.find       | O(1) Map.get（useMemo 缓存）               |
+| ModelCompatPatch 类型安全 | null 值无类型覆盖     | 显式 `boolean \| null`                     |
+| 客户端构建风险            | 可能引入 Node.js 模块 | 已隔离（shared/constants 层）              |
+| API 验证                  | 无 compatByProtocol   | Zod strict schema 校验                     |
+| 深色模式                  | 协议选择器不可读      | bg/text 适配 dark 主题                     |
+
+---
+
+## 六、涉及文件清单
+
+| 区域       | 文件                                                        | 改动类型                                                                                         |
+| ---------- | ----------------------------------------------------------- | ------------------------------------------------------------------------------------------------ |
+| 协议常量   | `src/shared/constants/modelCompat.ts`                       | **新建**，客户端安全的协议键与类型                                                               |
+| 存储与读写 | `src/lib/db/models.ts`                                      | `compatByProtocol` 数据结构、深度合并、getter 第三参 `sourceFormat`、`ModelCompatPatch` 类型修正 |
+| 再导出层   | `src/lib/localDb.ts`                                        | 新增 `ModelCompatPatch` 类型导出                                                                 |
+| API 路由   | `src/app/api/provider-models/route.ts`                      | PUT 支持 `compatByProtocol`，使用 `ModelCompatPatch` 类型                                        |
+| 输入校验   | `src/shared/validation/schemas.ts`                          | `modelCompatPerProtocolSchema`（strict）+ `compatByProtocol` 记录校验                            |
+| 请求管线   | `open-sse/handlers/chatCore.ts`                             | `getModelNormalizeToolCallId` / `getModelPreserveOpenAIDeveloperRole` 传入 `sourceFormat`        |
+| 前端 UI    | `src/app/(dashboard)/dashboard/providers/[id]/page.tsx`     | 协议选择器、按协议解析/保存、角标逻辑、Map 性能优化、类型收紧                                    |
+| i18n       | `src/i18n/messages/en.json`，`src/i18n/messages/zh-CN.json` | 5 条新文案                                                                                       |
+
+---
+
+## 七、回退方案
+
+- **禁用按协议配置**：删除 `compatByProtocol` 字段后，getter 自动回退到顶层字段，行为与 V2 一致。
+- **前端 Map 优化回退**：将 `Map.get()` 改回 `Array.find()` 即可，纯性能优化无功能耦合。
+- **客户端常量回退**：将 `MODEL_COMPAT_PROTOCOL_KEYS` 定义移回 `models.ts` 并从 `localDb.ts` 导出（需同时确保 `node:crypto` 问题不再存在）。
+- **生产环境**：以上修复对生产构建无负面影响。`compatByProtocol` 为可选字段，未配置时默认行为不变。API Zod 校验确保不会接受畸形数据。

bin/omniroute.mjs

@@ -189,8 +189,27 @@ const serverJs = join(APP_DIR, "server.js");
 
 if (!existsSync(serverJs)) {
   console.error("\x1b[31m✖ Server not found at:\x1b[0m", serverJs);
-  console.error("  This usually means the package was not built correctly.");
-  console.error("  Try reinstalling: npm install -g omniroute");
+  console.error("  The package may not have been built correctly.");
+  console.error("");
+  // (#492) Detect common non-standard Node managers that cause this issue
+  const nodeExec = process.execPath || "";
+  const isMise = nodeExec.includes("mise") || nodeExec.includes(".local/share/mise");
+  const isNvm = nodeExec.includes(".nvm") || nodeExec.includes("nvm");
+  if (isMise) {
+    console.error(
+      "  \x1b[33m⚠ mise detected:\x1b[0m If you installed via `npm install -g omniroute`,"
+    );
+    console.error("    try: \x1b[36mnpx omniroute@latest\x1b[0m  (downloads a fresh copy)");
+    console.error("    or:  \x1b[36mmise exec -- npx omniroute\x1b[0m");
+  } else if (isNvm) {
+    console.error(
+      "  \x1b[33m⚠ nvm detected:\x1b[0m Try reinstalling after loading the correct Node version:"
+    );
+    console.error("    \x1b[36mnvm use --lts && npm install -g omniroute\x1b[0m");
+  } else {
+    console.error("  Try: \x1b[36mnpm install -g omniroute\x1b[0m  (reinstall)");
+    console.error("  Or:  \x1b[36mnpx omniroute@latest\x1b[0m");
+  }
   process.exit(1);
 }
 

docs/API_REFERENCE.md

@@ -38,15 +38,20 @@ Content-Type: application/json
 
 ### Custom Headers
 
-| Header                   | Direction | Description                       |
-| ------------------------ | --------- | --------------------------------- |
-| `X-OmniRoute-No-Cache`   | Request   | Set to `true` to bypass cache     |
-| `X-OmniRoute-Progress`   | Request   | Set to `true` for progress events |
-| `Idempotency-Key`        | Request   | Dedup key (5s window)             |
-| `X-Request-Id`           | Request   | Alternative dedup key             |
-| `X-OmniRoute-Cache`      | Response  | `HIT` or `MISS` (non-streaming)   |
-| `X-OmniRoute-Idempotent` | Response  | `true` if deduplicated            |
-| `X-OmniRoute-Progress`   | Response  | `enabled` if progress tracking on |
+| Header                   | Direction | Description                                      |
+| ------------------------ | --------- | ------------------------------------------------ |
+| `X-OmniRoute-No-Cache`   | Request   | Set to `true` to bypass cache                    |
+| `X-OmniRoute-Progress`   | Request   | Set to `true` for progress events                |
+| `X-Session-Id`           | Request   | Sticky session key for external session affinity |
+| `x_session_id`           | Request   | Underscore variant also accepted (direct HTTP)   |
+| `Idempotency-Key`        | Request   | Dedup key (5s window)                            |
+| `X-Request-Id`           | Request   | Alternative dedup key                            |
+| `X-OmniRoute-Cache`      | Response  | `HIT` or `MISS` (non-streaming)                  |
+| `X-OmniRoute-Idempotent` | Response  | `true` if deduplicated                           |
+| `X-OmniRoute-Progress`   | Response  | `enabled` if progress tracking on                |
+| `X-OmniRoute-Session-Id` | Response  | Effective session ID used by OmniRoute           |
+
+> Nginx note: if you rely on underscore headers (for example `x_session_id`), enable `underscores_in_headers on;`.
 
 ---
 

docs/USER_GUIDE.md

@@ -578,6 +578,22 @@ Configure via **Dashboard → Settings → Routing**.
 | **Least Used**                 | Routes to the account with the oldest `lastUsedAt` timestamp, distributing traffic evenly        |
 | **Cost Optimized**             | Routes to the account with the lowest priority value, optimizing for lowest-cost providers       |
 
+#### External Sticky Session Header
+
+For external session affinity (for example, Claude Code/Codex agents behind reverse proxies), send:
+
+```http
+X-Session-Id: your-session-key
+```
+
+OmniRoute also accepts `x_session_id` and returns the effective session key in `X-OmniRoute-Session-Id`.
+
+If you use Nginx and send underscore-form headers, enable:
+
+```nginx
+underscores_in_headers on;
+```
+
 #### Wildcard Model Aliases
 
 Create wildcard patterns to remap model names:

docs/i18n/ar/README.md

@@ -932,8 +932,8 @@ npm run electron:build:linux   # Linux (.AppImage)
 | ميزة | ماذا يفعل || -------------------------- | ------------------------------------------------------------- |
 | 🖼️ **إنشاء الصور** | `/v1/images/generations` مع الواجهات الخلفية السحابية والمحلية |
 | 📐 **المضامين** | `/v1/embeddings` للبحث وخطوط أنابيب RAG |
-| 🎤 **نسخ صوتي** | `/v1/audio/transcriptions` (مقدمو خدمات الهمس والإضافيون) |
-| 🔊 **تحويل النص إلى كلام** | `/v1/audio/speech` (محركات/موفرو متعددون) |
+| 🎤 **نسخ صوتي** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **تحويل النص إلى كلام** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise) |
 | 🎬 **توليد الفيديو** | `/v1/videos/generations` (سير عمل ComfyUI + SD WebUI) |
 | 🎵 **جيل الموسيقى** | `/v1/music/generations` (سير عمل ComfyUI) |
 | 🛡️ **اعتدالات** | فحوصات السلامة `/v1/moderations` |

docs/i18n/bg/README.md

@@ -933,8 +933,8 @@ OmniRoute v2.0 е създаден като операционна платфо
 | Характеристика | Какво прави || -------------------------- | ------------------------------------------------------------ |
 | 🖼️ **Генериране на изображения** | `/v1/images/generations` с облак и локален бекенд |
 | 📐 **Вграждания** | `/v1/embeddings` за търсене и RAG тръбопроводи |
-| 🎤 **Аудио транскрипция** | `/v1/audio/transcriptions` (Whisper и допълнителни доставчици) |
-| 🔊 **Текст към говор** | `/v1/audio/speech` (множество машини/доставчици) |
+| 🎤 **Аудио транскрипция** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Текст към говор** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise) |
 | 🎬 **Видео генериране** | `/v1/videos/generations` (работни процеси ComfyUI + SD WebUI) |
 | 🎵 **Музикално поколение** | `/v1/music/generations` (работни процеси на ComfyUI) |
 | 🛡️ **Модерации** | `/v1/moderations` проверки за безопасност |

docs/i18n/cs/README.md

@@ -2,7 +2,7 @@
 
 ### Nikdy nepřestávejte s kódováním. Chytré směrování k **BEZPLATNÝM a levným modelům AI** s automatickým přepínáním mezi záložními systémy.
 
-*Váš univerzální API proxy – jeden endpoint, více než 44 poskytovatelů, nulové výpadky. Nyní s orchestrací agentů **MCP a A2A** .*
+_Váš univerzální API proxy – jeden endpoint, více než 44 poskytovatelů, nulové výpadky. Nyní s orchestrací agentů **MCP a A2A** ._
 
 **Dokončení chatu • Vkládání • Generování obrázků • Video • Hudba • Audio • Změna pořadí • **Vyhledávání na webu** • MCP server • A2A protokol • 100% TypeScript**
 
@@ -30,26 +30,23 @@
 <summary><b>Kliknutím zobrazíte snímky obrazovky z řídicího panelu</b></summary>
 </details>
 
-Strana | Snímek obrazovky
---- | ---
-**Poskytovatelé** | ![Poskytovatelé](docs/screenshots/01-providers.png)
-**Kombinace** | ![Kombinace](docs/screenshots/02-combos.png)
-**Analytika** | ![Analytika](docs/screenshots/03-analytics.png)
-**Zdraví** | ![Zdraví](docs/screenshots/04-health.png)
-**Překladatel** | ![Překladatel](docs/screenshots/05-translator.png)
-**Nastavení** | ![Nastavení](docs/screenshots/06-settings.png)
-**Nástroje CLI** | ![Nástroje CLI](docs/screenshots/07-cli-tools.png)
-**Protokoly používání** | ![Používání](docs/screenshots/08-usage.png)
-**Koncové body** | ![Koncové body](docs/screenshots/09-endpoint.png)
-
-
-
+| Strana                  | Snímek obrazovky                                    |
+| ----------------------- | --------------------------------------------------- |
+| **Poskytovatelé**       | ![Poskytovatelé](docs/screenshots/01-providers.png) |
+| **Kombinace**           | ![Kombinace](docs/screenshots/02-combos.png)        |
+| **Analytika**           | ![Analytika](docs/screenshots/03-analytics.png)     |
+| **Zdraví**              | ![Zdraví](docs/screenshots/04-health.png)           |
+| **Překladatel**         | ![Překladatel](docs/screenshots/05-translator.png)  |
+| **Nastavení**           | ![Nastavení](docs/screenshots/06-settings.png)      |
+| **Nástroje CLI**        | ![Nástroje CLI](docs/screenshots/07-cli-tools.png)  |
+| **Protokoly používání** | ![Používání](docs/screenshots/08-usage.png)         |
+| **Koncové body**        | ![Koncové body](docs/screenshots/09-endpoint.png)   |
 
 ---
 
 ### 🤖 Bezplatný poskytovatel umělé inteligence pro vaše oblíbené programátory
 
-*Připojte libovolný nástroj IDE nebo CLI s umělou inteligencí přes OmniRoute — bezplatnou API bránu pro neomezené kódování.*
+_Připojte libovolný nástroj IDE nebo CLI s umělou inteligencí přes OmniRoute — bezplatnou API bránu pro neomezené kódování._
 
   <table>
     <tr>
@@ -68,7 +65,6 @@ Strana | Snímek obrazovky
     </tr>
   </table>
 
-
 <sub>📡 Všichni agenti se připojují přes <code>http://localhost:20128/v1</code> nebo <code>http://cloud.omniroute.online/v1</code> — jedna konfigurace, neomezené modely a kvóty</sub>
 
 ---
@@ -161,9 +157,6 @@ Vývojáři platí za Claude Pro, Codex Pro nebo GitHub Copilot 20–200 dolarů
 - **Vlastní kombinace** — Přizpůsobitelné záložní řetězce se 6 strategiemi vyvažování (fill-first, round robin, P2C, náhodné, nejméně používané, nákladově optimalizované)
 - **Codex Business Quotas** — Sledování kvót pracovního prostoru firmy/týmu přímo v dashboardu
 
-
-
-
 <details>
 <summary><b>🔌 2. „Potřebuji použít více poskytovatelů, ale každý má jiné API“</b></summary>
 </details>
@@ -180,9 +173,6 @@ OpenAI používá jeden formát, Claude (Anthropic) jiný a Gemini ještě třet
 - **Strukturovaný výstup pro Gemini** — `json_schema` → automatická konverze `responseMimeType` / `responseSchema`
 - **Výchozí hodnota `stream` je `false`** – Odpovídá specifikaci OpenAI, čímž se zabrání neočekávanému SSE v Python/Rust/Go SDK.
 
-
-
-
 <details>
 <summary><b>🌐 3. „Můj poskytovatel AI blokuje můj region/zemi“</b></summary>
 </details>
@@ -199,9 +189,6 @@ Poskytovatelé jako OpenAI/Codex blokují přístup z určitých geografických
 - **TLS Fingerprint Spoofing** — Otisk prstu TLS podobný prohlížeči pomocí `wreq-js` pro obcházení detekce botů
 - **🔏 Porovnávání otisků prstů v CLI** — Změní pořadí záhlaví a polí v těle serveru tak, aby odpovídala nativním binárním podpisům v CLI, čímž drasticky snižuje riziko nahlašování účtu. IP adresa proxy je zachována — získáte současně stealth **i** maskování IP adresy.
 
-
-
-
 <details>
 <summary><b>🆓 4. „Chci používat umělou inteligenci pro kódování, ale nemám peníze“</b></summary>
 </details>
@@ -216,9 +203,6 @@ Ne každý si může dovolit zaplatit 20–200 dolarů měsíčně za předplatn
 - **NVIDIA NIM Free Access** — ~40 RPM developerský přístup k více než 70 modelům na build.nvidia.com (přechod z kreditů na čisté limity rychlosti)
 - **Strategie optimalizace nákladů** – Strategie směrování, která automaticky vybere nejlevnějšího dostupného poskytovatele
 
-
-
-
 <details>
 <summary><b>🔒 5. „Potřebuji chránit svou bránu umělé inteligence před neoprávněným přístupem“</b></summary>
 </details>
@@ -236,9 +220,6 @@ Při zpřístupnění brány umělé inteligence síti (LAN, VPS, Docker) může
 - **Ochrana proti vkládání výzev** – Sanitizace proti škodlivým vzorcům výzev
 - **Šifrování AES-256-GCM** – přihlašovací údaje jsou v klidovém stavu šifrovány
 
-
-
-
 <details>
 <summary><b>🛑 6. „Můj poskytovatel selhal a já ztratil/a programovací tok“</b></summary>
 </details>
@@ -254,9 +235,6 @@ Poskytovatelé umělé inteligence se mohou stát nestabilními, vracet chyby 5x
 - **Kombinovaný jistič** – Automaticky deaktivuje selhávajícího poskytovatele v rámci kombinovaného řetězce
 - **Dashboard stavu** — Monitorování provozuschopnosti, stavy jističů, uzamčení, statistiky mezipaměti, latence p50/p95/p99
 
-
-
-
 <details>
 <summary><b>🔧 7. „Konfigurace každého nástroje umělé inteligence je zdlouhavá a opakující se“</b></summary>
 </details>
@@ -270,9 +248,6 @@ Vývojáři používají Cursor, Claude Code, Codex CLI, OpenClaw, Gemini CLI, K
 - **Průvodce zaváděním** – 4krokové nastavení pro začínající uživatele
 - **Jeden koncový bod, všechny modely** – jednou nakonfigurujte `http://localhost:20128/v1` a získejte přístup k více než 44 poskytovatelům
 
-
-
-
 <details>
 <summary><b>🔑 8. „Správa OAuth tokenů od více poskytovatelů je peklo“</b></summary>
 </details>
@@ -288,9 +263,6 @@ Claude Code, Codex, Gemini CLI, Copilot – všechny používají OAuth 2.0 s to
 - **OAuth Behind Nginx** — Používá `window.location.origin` pro kompatibilitu s reverzní proxy
 - **Průvodce vzdáleným OAuth** – Podrobný návod k přihlašovacím údajům Google Cloud na VPS/Dockeru
 
-
-
-
 <details>
 <summary><b>📊 9. „Nevím, kolik utrácím ani kde“</b></summary>
 </details>
@@ -305,9 +277,6 @@ Vývojáři používají více placených poskytovatelů, ale nemají jednotný
 - **Statistiky použití pro každý klíč API** — Počet požadavků a časové razítko posledního použití pro každý klíč
 - **Analytický panel** – Statistické karty, graf využití modelu, tabulka poskytovatelů s mírou úspěšnosti a latencí
 
-
-
-
 <details>
 <summary><b>🐛 10. „Nedokážu diagnostikovat chyby a problémy ve volání umělé inteligence.“</b></summary>
 </details>
@@ -324,9 +293,6 @@ Když volání selže, vývojář neví, zda se jednalo o limit rychlosti, vypr
 - **Souborové protokolování s rotací** – Konzolový interceptor zachycuje vše do protokolu JSON s rotací na základě velikosti
 - **Zpráva o systémových informacích** — příkaz `npm run system-info` vygeneruje `system-info.txt` s kompletním popisem vašeho prostředí (verze uzlu, verze OmniRoute, operační systém, nástroje CLI, stav Dockeru/PM2). Přiložte jej při hlášení problémů pro okamžité třídění.
 
-
-
-
 <details>
 <summary><b>🏗️ 11. „Nasazení a údržba brány je složitá“</b></summary>
 </details>
@@ -343,9 +309,6 @@ Instalace, konfigurace a údržba AI proxy v různých prostředích (lokální,
 - **Cloud Sync** – Konfigurace synchronizace mezi zařízeními pomocí Cloudflare Workers
 - **Zálohy databází** — Automatické zálohování, obnovení, export a import všech nastavení
 
-
-
-
 <details>
 <summary><b>🌍 12. „Rozhraní je pouze v angličtině a můj tým nemluví anglicky“</b></summary>
 </details>
@@ -359,9 +322,6 @@ Týmy v neanglicky mluvících zemích, zejména v Latinské Americe, Asii a Evr
 - **Vícejazyčné soubory README** — 30 kompletních překladů dokumentace
 - **Výběr jazyka** — Ikona glóbu v záhlaví pro přepínání v reálném čase
 
-
-
-
 <details>
 <summary><b>🔄 13. „Potřebuji víc než jen chat – potřebuji vložené soubory, obrázky, zvuk.“</b></summary>
 </details>
@@ -380,9 +340,6 @@ Umělá inteligence není jen dokončování chatu. Vývojáři potřebují gene
 - **Změna pořadí** — `/v1/rerank` — Změna pořadí relevance dokumentu
 - **Responses API** — Plná podpora `/v1/responses` pro Codex
 
-
-
-
 <details>
 <summary><b>🧪 14. „Nemám způsob, jak testovat a porovnávat kvalitu napříč modely.“</b></summary>
 </details>
@@ -397,9 +354,6 @@ Vývojáři chtějí vědět, který model je pro jejich případ použití nejl
 - **Tester chatu** — Kompletní okružní cesta s vizuálním vykreslováním odpovědí
 - **Živý monitor** — Stream všech požadavků procházejících proxy serverem v reálném čase
 
-
-
-
 <details>
 <summary><b>📈 15. „Potřebuji škálovat bez ztráty výkonu“</b></summary>
 </details>
@@ -415,9 +369,6 @@ S rostoucím objemem požadavků generují stejné otázky bez ukládání do me
 - **Mezipaměť pro ověření klíčů API** — třívrstvá mezipaměť pro výkon produkčního prostředí
 - **Dashboard s telemetrií** – latence p50/p95/p99, statistiky mezipaměti, dostupnost
 
-
-
-
 <details>
 <summary><b>🤖 16. „Chci mít chování modelů globálně pod kontrolou“</b></summary>
 </details>
@@ -434,9 +385,6 @@ Vývojáři, kteří chtějí všechny odpovědi v určitém jazyce, se specific
 - **Přepínání poskytovatele** – Povolení/zakázání všech připojení pro poskytovatele jedním kliknutím
 - **Blokovaní poskytovatelé** – Vyloučení konkrétních poskytovatelů ze seznamu `/v1/models`
 
-
-
-
 <details>
 <summary><b>🧰 17. „Potřebuji nástroje MCP jako prvotřídní produktové funkce.“</b></summary>
 </details>
@@ -449,9 +397,6 @@ Mnoho bran umělé inteligence odhaluje MCP pouze jako skrytý implementační d
 - Vyhrazená stránka pro správu MCP s procesy, nástroji, rozsahy a auditem
 - Vestavěný rychlý start pro `omniroute --mcp` a onboarding klienta
 
-
-
-
 <details>
 <summary><b>🧠 18. „Potřebuji orchestraci A2A se synchronizací a cestami úloh streamu.“</b></summary>
 </details>
@@ -464,9 +409,6 @@ Pracovní postupy agentů vyžadují jak přímé odpovědi, tak dlouhodobé str
 - Streamování SSE s šířením stavu terminálu
 - Rozhraní API životního cyklu úloh pro `tasks/get` a `tasks/cancel`
 
-
-
-
 <details>
 <summary><b>🛰️ 19. „Potřebuji skutečný stav procesu MCP, ne odhadovaný stav.“</b></summary>
 </details>
@@ -479,9 +421,6 @@ Provozní týmy potřebují vědět, zda je MCP skutečně aktivní, nejen zda j
 - API stavu MCP kombinující prezenční signál a nedávnou aktivitu
 - Karty stavu uživatelského rozhraní pro zobrazení aktuálnosti procesů/provozuschopnosti/prezenčního signálu
 
-
-
-
 <details>
 <summary><b>📋 20. „Potřebuji auditovatelné provedení nástroje MCP“</b></summary>
 </details>
@@ -494,9 +433,6 @@ Když nástroje mění konfiguraci nebo spouštějí operační akce, týmy pot
 - Filtruje podle nástroje, úspěchu/neúspěchu, klíče API a stránkování
 - Tabulka auditu dashboardu + koncové body statistik pro automatizaci
 
-
-
-
 <details>
 <summary><b>🔐 21. „Potřebuji omezená oprávnění MCP pro každou integraci.“</b></summary>
 </details>
@@ -509,9 +445,6 @@ Různí klienti by měli mít přístup ke kategoriím nástrojů s nejnižším
 - Vynucení rozsahu a viditelnost v uživatelském rozhraní správy MCP
 - Bezpečná výchozí poloha pro provozní nástroje
 
-
-
-
 <details>
 <summary><b>⚙️ 22. „Potřebuji provozní kontroly bez nutnosti přesouvání“</b></summary>
 </details>
@@ -524,9 +457,6 @@ Týmy potřebují rychlé změny v běhovém prostředí během incidentů nebo
 - Používejte profily odolnosti z předdefinovaných balíčků zásad
 - Resetujte stav jističe ze stejného ovládacího panelu
 
-
-
-
 <details>
 <summary><b>🔄 23. „Potřebuji živý přehled o životním cyklu úkolů A2A a jejich zrušení.“</b></summary>
 </details>
@@ -539,9 +469,6 @@ Bez přehledu o životním cyklu je obtížné třídit incidenty úkolů.
 - Podrobný přehled metadat úloh, událostí a artefaktů
 - Koncový bod zrušení úlohy a akce uživatelského rozhraní s potvrzením
 
-
-
-
 <details>
 <summary><b>🌊 24. „Potřebuji metriky aktivního streamu pro A2A zátěž“</b></summary>
 </details>
@@ -554,9 +481,6 @@ Streamovací pracovní postupy vyžadují provozní přehled o souběžnosti a
 - Časové razítko posledního úkolu a počty pro jednotlivé stavy
 - Karty A2A dashboardu pro monitorování provozu v reálném čase
 
-
-
-
 <details>
 <summary><b>🪪 25. „Potřebuji standardní vyhledávání agentů pro klienty“</b></summary>
 </details>
@@ -569,9 +493,6 @@ Externí klienti a orchestratoři potřebují pro onboarding strojově čitelná
 - Schopnosti a dovednosti zobrazené v uživatelském rozhraní pro správu
 - API pro stav A2A zahrnuje metadata pro zjišťování pro automatizaci
 
-
-
-
 <details>
 <summary><b>🧭 26. „Potřebuji v uživatelském rozhraní produktu zjistitelnost protokolu.“</b></summary>
 </details>
@@ -584,9 +505,6 @@ Pokud uživatelé nemohou objevit protokolové povrchy, kvalita přijetí a podp
 - Přepínání stavu inline služby (Online/Offline) pro MCP a A2A
 - Odkazy z přehledu na vyhrazené karty pro správu
 
-
-
-
 <details>
 <summary><b>🧪 27. „Potřebuji komplexní ověření protokolu se skutečnými klienty.“</b></summary>
 </details>
@@ -599,9 +517,6 @@ Simulované testy nestačí k ověření kompatibility protokolu před vydáním
 - Klientské testy A2A pro toky zjišťování, odesílání, streamování, načítání a zrušení
 - Křížová kontrola tvrzení oproti API pro audit MCP a úkoly A2A
 
-
-
-
 <details>
 <summary><b>📡 28. „Potřebuji jednotnou pozorovatelnost napříč všemi rozhraními“</b></summary>
 </details>
@@ -614,9 +529,6 @@ Rozdělení pozorovatelnosti podle protokolu vytváří slepá místa a delší
 - Stav + audit + telemetrie požadavků napříč vrstvami OpenAI, MCP a A2A
 - Provozní API pro stav a automatizaci
 
-
-
-
 <details>
 <summary><b>💼 29. „Potřebuji jeden runtime pro proxy + nástroje + orchestraci agentů“</b></summary>
 </details>
@@ -629,9 +541,6 @@ Spouštění mnoha samostatných služeb zvyšuje provozní náklady a počet po
 - Sdílené ověřování, odolnost, úložiště dat a pozorovatelnost
 - Konzistentní model politik napříč všemi interakčními plochami
 
-
-
-
 <details>
 <summary><b>🚀 30. „Potřebuji agentské pracovní postupy bez slepení kódu.“</b></summary>
 </details>
@@ -644,9 +553,6 @@ Týmy ztrácejí rychlost při spojování více ad-hoc služeb a skriptů.
 - Vestavěná uživatelská rozhraní pro správu protokolů a cesty pro ověřování kouře
 - Základy připravené pro produkční prostředí (zabezpečení, protokolování, odolnost, zálohování)
 
-
-
-
 ### Příklady herních plánů (integrované případy užití)
 
 **Příručka A: Maximalizace placeného předplatného + levné zálohování**
@@ -701,13 +607,13 @@ Outcome: deep fallback depth for deadline-critical workloads
 
 > Nastavte si kódování s umělou inteligencí během několika minut za **0 $/měsíc** . Propojte tyto bezplatné účty a využijte vestavěnou kombinaci **Free Stack** .
 
-Krok | Akce | Poskytovatelé odemčeni
---- | --- | ---
-1 | Připojení **Kiro** (AWS Builder ID OAuth) | Claude Sonnet 4.5, Haiku 4.5 – **neomezeně**
-2 | Připojení k **iFlow** (Google OAuth) | kimi-k2-myšlení, qwen3-coder-plus, deepseek-r1... — **neomezeně**
-3 | Připojení **Qwen** (kód zařízení) | qwen3-coder-plus, qwen3-coder-flash... — **neomezeně**
-4 | Připojení **rozhraní příkazového řádku Gemini** (Google OAuth) | gemini-3-flash, gemini-2.5-pro — **180 000 GBP/měsíc zdarma**
-5 | `/dashboard/combos` → Šablona **Free Stack (0 $)** | Automatické zařazení všech bezplatných poskytovatelů do routingu
+| Krok | Akce                                                           | Poskytovatelé odemčeni                                            |
+| ---- | -------------------------------------------------------------- | ----------------------------------------------------------------- |
+| 1    | Připojení **Kiro** (AWS Builder ID OAuth)                      | Claude Sonnet 4.5, Haiku 4.5 – **neomezeně**                      |
+| 2    | Připojení k **iFlow** (Google OAuth)                           | kimi-k2-myšlení, qwen3-coder-plus, deepseek-r1... — **neomezeně** |
+| 3    | Připojení **Qwen** (kód zařízení)                              | qwen3-coder-plus, qwen3-coder-flash... — **neomezeně**            |
+| 4    | Připojení **rozhraní příkazového řádku Gemini** (Google OAuth) | gemini-3-flash, gemini-2.5-pro — **180 000 GBP/měsíc zdarma**     |
+| 5    | `/dashboard/combos` → Šablona **Free Stack (0 $)**             | Automatické zařazení všech bezplatných poskytovatelů do routingu  |
 
 **V libovolném IDE/CLI naveďte:** `http://localhost:20128/v1` · Klíč API: `any-string` · Hotovo.
 
@@ -732,13 +638,13 @@ omniroute
 
 Dashboard se otevírá na `http://localhost:20128` a základní URL API je `http://localhost:20128/v1` .
 
-Příkaz | Popis
---- | ---
-`omniroute` | Spuštění serveru ( `PORT=20128` , API a dashboard na stejném portu)
-`omniroute --port 3000` | Nastavte kanonický/API port na 3000
-`omniroute --mcp` | Spuštění MCP serveru (transport stdio)
-`omniroute --no-open` | Neotevírat prohlížeč automaticky
-`omniroute --help` | Zobrazit nápovědu
+| Příkaz                  | Popis                                                               |
+| ----------------------- | ------------------------------------------------------------------- |
+| `omniroute`             | Spuštění serveru ( `PORT=20128` , API a dashboard na stejném portu) |
+| `omniroute --port 3000` | Nastavte kanonický/API port na 3000                                 |
+| `omniroute --mcp`       | Spuštění MCP serveru (transport stdio)                              |
+| `omniroute --no-open`   | Neotevírat prohlížeč automaticky                                    |
+| `omniroute --help`      | Zobrazit nápovědu                                                   |
 
 Volitelný režim s rozděleným portem:
 
@@ -847,10 +753,10 @@ docker compose --profile base up -d
 docker compose --profile cli up -d
 ```
 
-Obraz | Štítek | Velikost | Popis
---- | --- | --- | ---
-`diegosouzapw/omniroute` | `latest` | ~250 MB | Nejnovější stabilní verze
-`diegosouzapw/omniroute` | `1.0.3` | ~250 MB | Aktuální verze
+| Obraz                    | Štítek   | Velikost | Popis                     |
+| ------------------------ | -------- | -------- | ------------------------- |
+| `diegosouzapw/omniroute` | `latest` | ~250 MB  | Nejnovější stabilní verze |
+| `diegosouzapw/omniroute` | `1.0.3`  | ~250 MB  | Aktuální verze            |
 
 ---
 
@@ -893,41 +799,47 @@ Po minimalizaci se OmniRoute nachází v systémové liště a nabízí rychlé
 
 ## 💰 Přehled cen
 
-Úroveň | Poskytovatel | Náklady | Obnovení kvóty | Nejlepší pro
---- | --- | --- | --- | ---
-**💳 PŘEDPLATNÉ** | Claude Code (profesionál) | 20 dolarů měsíčně | 5 hodin + týdně | Již přihlášen/a k odběru
- | Kodex (Plus/Pro) | 20–200 USD/měsíc | 5 hodin + týdně | Uživatelé OpenAI
- | Rozhraní příkazového řádku Gemini | **UVOLNIT** | 180 tisíc měsíčně + 1 tisíc denně | Každý!
- | GitHub Copilot | 10–19 USD/měsíc | Měsíční | Uživatelé GitHubu
-**🔑 KLÍČ API** | NVIDIA NIM | **ZDARMA** (vývoj navždy) | ~40 ot./min | 70+ otevřených modelů
- | Mozky | **ZDARMA** (1 milion tok/den) | 60 000 otáček za minutu / 30 ot./min | Nejrychlejší na světě
- | Groq | **ZDARMA** (30 ot./min.) | 14,4 tisíc otáček za minutu | Ultrarychlá lama/gema
- | DeepSeek V3.2 | 0,27/1,10 USD za 1 milion | Žádný | Nejlepší zdůvodnění ceny a kvality
- | xAI Grok-4 Rychlý | **0,20/0,50 USD za 1 milion** 🆕 | Žádný | Nejrychlejší + volání nástroje, ultranízké
- | xAI Grok-4 (standardní) | 0,20/1,50 USD za 1 milion 🆕 | Žádný | Vlajková loď Reasoning od xAI
- | Mistral | Zkušební verze zdarma + placené | Omezená sazba | Evropská umělá inteligence
- | OpenRouter | Platba za použití | Žádný | Více než 100 modelů agregováno.
-**💰 LEVNÉ** | GLM-5 (přes Z.AI) 🆕 | 0,5 USD/1 milion | Denně v 10:00 | Výstup 128 tisíc obrazových bodů, nejnovější vlajková loď
- | GLM-4.7 | 0,6 USD/1 milion | Denně v 10:00 | Záloha rozpočtu
- | MiniMax M2.5 🆕 | Vstup 0,3 USD/1 milion | 5hodinové válcování | Úvaha + agentní úkoly
- | MiniMax M2.1 | 0,2 USD/1 milion | 5hodinové válcování | Nejlevnější varianta
- | Kimi K2.5 (Moonshot API) 🆕 | Platba za použití | Žádný | Přímý přístup k Moonshot API
- | Kimi K2 | 9 dolarů měsíčně bez závazků | 10 milionů tokenů/měsíc | Předvídatelné náklady
-**🆓 ZDARMA** | iFlow | **0 dolarů** | Neomezený | 5 modelů neomezeně
- | Qwen | **0 dolarů** | Neomezený | 4 modely neomezeně
- | Kiro | **0 dolarů** | Neomezený | Claude Sonnet/Haiku (tvorce AWS)
+| Úroveň                            | Poskytovatel                     | Náklady                              | Obnovení kvóty                             | Nejlepší pro                                              |
+| --------------------------------- | -------------------------------- | ------------------------------------ | ------------------------------------------ | --------------------------------------------------------- |
+| **💳 PŘEDPLATNÉ**                 | Claude Code (profesionál)        | 20 dolarů měsíčně                    | 5 hodin + týdně                            | Již přihlášen/a k odběru                                  |
+| Kodex (Plus/Pro)                  | 20–200 USD/měsíc                 | 5 hodin + týdně                      | Uživatelé OpenAI                           |
+| Rozhraní příkazového řádku Gemini | **UVOLNIT**                      | 180 tisíc měsíčně + 1 tisíc denně    | Každý!                                     |
+| GitHub Copilot                    | 10–19 USD/měsíc                  | Měsíční                              | Uživatelé GitHubu                          |
+| **🔑 KLÍČ API**                   | NVIDIA NIM                       | **ZDARMA** (vývoj navždy)            | ~40 ot./min                                | 70+ otevřených modelů                                     |
+| Mozky                             | **ZDARMA** (1 milion tok/den)    | 60 000 otáček za minutu / 30 ot./min | Nejrychlejší na světě                      |
+| Groq                              | **ZDARMA** (30 ot./min.)         | 14,4 tisíc otáček za minutu          | Ultrarychlá lama/gema                      |
+| DeepSeek V3.2                     | 0,27/1,10 USD za 1 milion        | Žádný                                | Nejlepší zdůvodnění ceny a kvality         |
+| xAI Grok-4 Rychlý                 | **0,20/0,50 USD za 1 milion** 🆕 | Žádný                                | Nejrychlejší + volání nástroje, ultranízké |
+| xAI Grok-4 (standardní)           | 0,20/1,50 USD za 1 milion 🆕     | Žádný                                | Vlajková loď Reasoning od xAI              |
+| Mistral                           | Zkušební verze zdarma + placené  | Omezená sazba                        | Evropská umělá inteligence                 |
+| OpenRouter                        | Platba za použití                | Žádný                                | Více než 100 modelů agregováno.            |
+| **💰 LEVNÉ**                      | GLM-5 (přes Z.AI) 🆕             | 0,5 USD/1 milion                     | Denně v 10:00                              | Výstup 128 tisíc obrazových bodů, nejnovější vlajková loď |
+| GLM-4.7                           | 0,6 USD/1 milion                 | Denně v 10:00                        | Záloha rozpočtu                            |
+| MiniMax M2.5 🆕                   | Vstup 0,3 USD/1 milion           | 5hodinové válcování                  | Úvaha + agentní úkoly                      |
+| MiniMax M2.1                      | 0,2 USD/1 milion                 | 5hodinové válcování                  | Nejlevnější varianta                       |
+| Kimi K2.5 (Moonshot API) 🆕       | Platba za použití                | Žádný                                | Přímý přístup k Moonshot API               |
+| Kimi K2                           | 9 dolarů měsíčně bez závazků     | 10 milionů tokenů/měsíc              | Předvídatelné náklady                      |
+| **🆓 ZDARMA**                     | iFlow                            | **0 dolarů**                         | Neomezený                                  | 5 modelů neomezeně                                        |
+| Qwen                              | **0 dolarů**                     | Neomezený                            | 4 modely neomezeně                         |
+| Kiro                              | **0 dolarů**                     | Neomezený                            | Claude Sonnet/Haiku (tvorce AWS)           |
 
 > 🆕 **Přidány nové modely (březen 2026):** řada Grok-4 Fast za 0,20 USD/0,50 USD/M (benchmarkováno na 1143 ms – o 30 % rychlejší než Gemini 2.5 Flash), GLM-5 přes Z.AI s výstupem 128K, uvažování MiniMax M2.5, aktualizované ceny DeepSeek V3.2, Kimi K2.5 přes Moonshot Direct API.
 
 **💡 Kombinovaný balík za 0 $ — Kompletní bezplatná instalace:**
 
 ```
-Gemini CLI (180K/mo free)
-  → iFlow (unlimited: kimi-k2-thinking, qwen3-coder-plus, deepseek-r1)
-  → Kiro (Claude Sonnet 4.5 + Haiku — unlimited, via AWS Builder ID)
-  → Qwen (4 models — unlimited)
-  → Groq (14.4K req/day — ultra-fast)
-  → NVIDIA NIM (70+ models — 40 RPM forever)
+# 🆓 Ultimate Free Stack 2026 — 11 Providers, $0 Forever
+Kiro (kr/)             → Claude Sonnet/Haiku UNLIMITED
+iFlow (if/)            → kimi-k2-thinking, qwen3-coder-plus, deepseek-r1 UNLIMITED
+LongCat Lite (lc/)     → LongCat-Flash-Lite — 50M tokens/day 🔥
+Pollinations (pol/)    → GPT-5, Claude, DeepSeek, Llama 4 — no key needed
+Qwen (qw/)             → qwen3-coder-plus, qwen3-coder-flash, qwen3-coder-next UNLIMITED
+Gemini (gemini/)       → Gemini 2.5 Flash — 1,500 req/day free API key
+Cloudflare AI (cf/)    → Llama 70B, Gemma 3, Mistral — 10K Neurons/day
+Scaleway (scw/)        → Qwen3 235B, Llama 70B — 1M free tokens (EU)
+Groq (groq/)           → Llama/Gemma ultra-fast — 14.4K req/day
+NVIDIA NIM (nvidia/)   → 70+ open models — 40 RPM forever
+Cerebras (cerebras/)   → Llama/Qwen world-fastest — 1M tok/day
 ```
 
 **Nulové náklady. Nikdy nepřestávejte s kódováním.** Nakonfigurujte si to jako jednu kombinaci OmniRoute a všechny záložní režimy se provede automaticky – žádné ruční přepínání.
@@ -942,59 +854,59 @@ Gemini CLI (180K/mo free)
 
 ### 🔵 CLAUDE MODELS (přes Kiro — AWS Builder ID)
 
-Model | Předpona | Omezit | Limit rychlosti
---- | --- | --- | ---
-`claude-sonnet-4.5` | `kr/` | **Neomezený** | Žádný hlášený denní limit
-`claude-haiku-4.5` | `kr/` | **Neomezený** | Žádný hlášený denní limit
-`claude-opus-4.6` | `kr/` | **Neomezený** | Nejnovější opus od Kira
+| Model               | Předpona | Omezit        | Limit rychlosti           |
+| ------------------- | -------- | ------------- | ------------------------- |
+| `claude-sonnet-4.5` | `kr/`    | **Neomezený** | Žádný hlášený denní limit |
+| `claude-haiku-4.5`  | `kr/`    | **Neomezený** | Žádný hlášený denní limit |
+| `claude-opus-4.6`   | `kr/`    | **Neomezený** | Nejnovější opus od Kira   |
 
 ### 🟢 MODELY IFLOW (Bezplatné OAuth — bez nutnosti platit kreditní kartou)
 
-Model | Předpona | Omezit | Limit rychlosti
---- | --- | --- | ---
-`kimi-k2-thinking` | `if/` | **Neomezený** | Žádný hlášený strop
-`qwen3-coder-plus` | `if/` | **Neomezený** | Žádný hlášený strop
-`deepseek-r1` | `if/` | **Neomezený** | Žádný hlášený strop
-`minimax-m2.1` | `if/` | **Neomezený** | Žádný hlášený strop
-`kimi-k2` | `if/` | **Neomezený** | Žádný hlášený strop
+| Model              | Předpona | Omezit        | Limit rychlosti     |
+| ------------------ | -------- | ------------- | ------------------- |
+| `kimi-k2-thinking` | `if/`    | **Neomezený** | Žádný hlášený strop |
+| `qwen3-coder-plus` | `if/`    | **Neomezený** | Žádný hlášený strop |
+| `deepseek-r1`      | `if/`    | **Neomezený** | Žádný hlášený strop |
+| `minimax-m2.1`     | `if/`    | **Neomezený** | Žádný hlášený strop |
+| `kimi-k2`          | `if/`    | **Neomezený** | Žádný hlášený strop |
 
 ### 🟡 MODELY QWEN (Ověření kódu zařízení)
 
-Model | Předpona | Omezit | Limit rychlosti
---- | --- | --- | ---
-`qwen3-coder-plus` | `qw/` | **Neomezený** | Žádný hlášený strop
-`qwen3-coder-flash` | `qw/` | **Neomezený** | Žádný hlášený strop
-`qwen3-coder-next` | `qw/` | **Neomezený** | Žádný hlášený strop
-`vision-model` | `qw/` | **Neomezený** | Multimodální (obrázky)
+| Model               | Předpona | Omezit        | Limit rychlosti        |
+| ------------------- | -------- | ------------- | ---------------------- |
+| `qwen3-coder-plus`  | `qw/`    | **Neomezený** | Žádný hlášený strop    |
+| `qwen3-coder-flash` | `qw/`    | **Neomezený** | Žádný hlášený strop    |
+| `qwen3-coder-next`  | `qw/`    | **Neomezený** | Žádný hlášený strop    |
+| `vision-model`      | `qw/`    | **Neomezený** | Multimodální (obrázky) |
 
 ### 🟣 Rozhraní GEMINI CLI (Google OAuth)
 
-Model | Předpona | Omezit | Limit rychlosti
---- | --- | --- | ---
-`gemini-3-flash-preview` | `gc/` | **180 tisíc tok/měsíc** + 1 tisíc/den | Měsíční reset
-`gemini-2.5-pro` | `gc/` | 180 tisíc měsíčně (sdílený bazén) | Vysoká kvalita
+| Model                    | Předpona | Omezit                                | Limit rychlosti |
+| ------------------------ | -------- | ------------------------------------- | --------------- |
+| `gemini-3-flash-preview` | `gc/`    | **180 tisíc tok/měsíc** + 1 tisíc/den | Měsíční reset   |
+| `gemini-2.5-pro`         | `gc/`    | 180 tisíc měsíčně (sdílený bazén)     | Vysoká kvalita  |
 
 ### ⚫ NVIDIA NIM (Bezplatný klíč API — build.nvidia.com)
 
-Úroveň | Denní limit | Limit rychlosti | Poznámky
---- | --- | --- | ---
-Zdarma (vývojář) | Žádný limit tokenů | **~40 ot./min** | Více než 70 modelů; přechod na čisté limity sazeb v polovině roku 2025
+| Úroveň           | Denní limit        | Limit rychlosti | Poznámky                                                               |
+| ---------------- | ------------------ | --------------- | ---------------------------------------------------------------------- |
+| Zdarma (vývojář) | Žádný limit tokenů | **~40 ot./min** | Více než 70 modelů; přechod na čisté limity sazeb v polovině roku 2025 |
 
 Oblíbené bezplatné modely: `moonshotai/kimi-k2.5` (Kimi K2.5), `z-ai/glm4.7` (GLM 4.7), `deepseek-ai/deepseek-v3.2` (DeepSeek V3.2), `nvidia/llama-3.3-70b-instruct` , `deepseek/deepseek-r1`
 
 ### ⚪ CEREBRAS (Bezplatný klíč API — inference.cerebras.ai)
 
-Úroveň | Denní limit | Limit rychlosti | Poznámky
---- | --- | --- | ---
-Uvolnit | **1 milion tokenů/den** | 60 000 otáček za minutu / 30 ot./min | Nejrychlejší inference LLM na světě; denně se resetuje
+| Úroveň  | Denní limit             | Limit rychlosti                      | Poznámky                                               |
+| ------- | ----------------------- | ------------------------------------ | ------------------------------------------------------ |
+| Uvolnit | **1 milion tokenů/den** | 60 000 otáček za minutu / 30 ot./min | Nejrychlejší inference LLM na světě; denně se resetuje |
 
 Dostupné zdarma: `llama-3.3-70b` , `llama-3.1-8b` , `deepseek-r1-distill-llama-70b`
 
 ### 🔴 GROQ (Bezplatný API klíč — console.groq.com)
 
-Úroveň | Denní limit | Limit rychlosti | Poznámky
---- | --- | --- | ---
-Uvolnit | **14,4 tisíc otáček za minutu** | 30 ot./min na model | Žádná kreditní karta; limit 429, neúčtováno
+| Úroveň  | Denní limit                     | Limit rychlosti     | Poznámky                                    |
+| ------- | ------------------------------- | ------------------- | ------------------------------------------- |
+| Uvolnit | **14,4 tisíc otáček za minutu** | 30 ot./min na model | Žádná kreditní karta; limit 429, neúčtováno |
 
 K dispozici zdarma: `llama-3.3-70b-versatile` , `gemma2-9b-it` , `mixtral-8x7b` , `whisper-large-v3`
 
@@ -1016,11 +928,11 @@ K dispozici zdarma: `llama-3.3-70b-versatile` , `gemma2-9b-it` , `mixtral-8x7b`
 
 > Přepisujte libovolné audio/video za **0 $** – Deepgram leady za 200 $ zdarma, AssemblyAI za 50 $ jako záložní nástroj, Groq Whisper jako neomezená nouzová záloha.
 
-Poskytovatel | Bezplatné kredity | Nejlepší model | Limit rychlosti
---- | --- | --- | ---
-🟢 **Deepgram** | **200 dolarů zdarma** (registrace) | `nova-3` — nejvyšší přesnost, více než 30 jazyků | Žádný limit RPM pro kredity zdarma
-🔵 **AssemblyAI** | **50 dolarů zdarma** (registrace) | `universal-3-pro` — kapitoly, sentiment, osobní údaje | Žádný limit RPM pro kredity zdarma
-🔴 **Groq** | **Navždy zdarma** | `whisper-large-v3` — OpenAI Šepot | 30 ot./min (omezená rychlost)
+| Poskytovatel      | Bezplatné kredity                  | Nejlepší model                                        | Limit rychlosti                    |
+| ----------------- | ---------------------------------- | ----------------------------------------------------- | ---------------------------------- |
+| 🟢 **Deepgram**   | **200 dolarů zdarma** (registrace) | `nova-3` — nejvyšší přesnost, více než 30 jazyků      | Žádný limit RPM pro kredity zdarma |
+| 🔵 **AssemblyAI** | **50 dolarů zdarma** (registrace)  | `universal-3-pro` — kapitoly, sentiment, osobní údaje | Žádný limit RPM pro kredity zdarma |
+| 🔴 **Groq**       | **Navždy zdarma**                  | `whisper-large-v3` — OpenAI Šepot                     | 30 ot./min (omezená rychlost)      |
 
 **Navrhovaná kombinace v `/dashboard/combos` :**
 
@@ -1041,118 +953,118 @@ OmniRoute v2.0 je navržen jako operační platforma, nikoli pouze jako proxy pr
 
 ### 🆕 Nové — Vylepšení inspirovaná ClawRouterem (březen 2026)
 
-Funkce | Co to dělá
---- | ---
-⚡ **Grok-4 Rychlá rodina** | Modely xAI za 0,20 USD/0,50 USD/M – v benchmarku 1143 ms (o 30 % rychlejší než Gemini 2.5 Flash)
-🧠 **GLM-5 přes Z.AI** | 128 tisíc výstupních dat, 0,5 USD/1 milion USD – nejnovější vlajková loď rodiny GLM
-🔮 **MiniMax M2.5** | Úvaha + agentní úkoly za 0,30 USD/1 milion – významný upgrade oproti M2.1
-🎯 **Příznak volání nástroje pro každý model** | `toolCalling: true/false` v registru — AutoCombo přeskakuje modely, které nepodporují nástroje.
-🌍 **Detekce vícejazyčného záměru** | Klíčová slova PT/ZH/ES/AR v bodování AutoCombo – lepší výběr modelu pro neanglický obsah
-📊 **Záložní metody řízené benchmarkem** | Skutečná latence p95 z živých požadavků poskytuje kombinované skóre – AutoCombo se učí ze skutečných dat
-🔁 **Požádat o deduplikaci** | Okno pro deduplikaci na základě hashování obsahu – bezpečné pro více agentů, zabraňuje duplicitním platbám
-🔌 **Strategie pro zásuvné routery** | Rozšiřitelné rozhraní `RouterStrategy` – přidejte si vlastní logiku směrování jako pluginy
+| Funkce                                         | Co to dělá                                                                                                 |
+| ---------------------------------------------- | ---------------------------------------------------------------------------------------------------------- |
+| ⚡ **Grok-4 Rychlá rodina**                    | Modely xAI za 0,20 USD/0,50 USD/M – v benchmarku 1143 ms (o 30 % rychlejší než Gemini 2.5 Flash)           |
+| 🧠 **GLM-5 přes Z.AI**                         | 128 tisíc výstupních dat, 0,5 USD/1 milion USD – nejnovější vlajková loď rodiny GLM                        |
+| 🔮 **MiniMax M2.5**                            | Úvaha + agentní úkoly za 0,30 USD/1 milion – významný upgrade oproti M2.1                                  |
+| 🎯 **Příznak volání nástroje pro každý model** | `toolCalling: true/false` v registru — AutoCombo přeskakuje modely, které nepodporují nástroje.            |
+| 🌍 **Detekce vícejazyčného záměru**            | Klíčová slova PT/ZH/ES/AR v bodování AutoCombo – lepší výběr modelu pro neanglický obsah                   |
+| 📊 **Záložní metody řízené benchmarkem**       | Skutečná latence p95 z živých požadavků poskytuje kombinované skóre – AutoCombo se učí ze skutečných dat   |
+| 🔁 **Požádat o deduplikaci**                   | Okno pro deduplikaci na základě hashování obsahu – bezpečné pro více agentů, zabraňuje duplicitním platbám |
+| 🔌 **Strategie pro zásuvné routery**           | Rozšiřitelné rozhraní `RouterStrategy` – přidejte si vlastní logiku směrování jako pluginy                 |
 
 ### 🚀 Předchozí verze v2.0.9+ — Hřiště, otisky prstů v CLI a ACP
 
-Funkce | Co to dělá
---- | ---
-🎮 **Modelové hřiště** | Stránka řídicího panelu pro přímé testování libovolného modelu – selektory poskytovatele/modelu/koncového bodu, editor Monaco, streamování, přerušení, načasování
-🔏 **Porovnávání otisků prstů v CLI** | Řazení hlaviček/těl serveru podle poskytovatele tak, aby odpovídalo nativním podpisům CLI – přepínání pro jednotlivé poskytovatele v Nastavení &gt; Zabezpečení. **Vaše IP adresa proxy serveru je zachována.**
-🤝 **Podpora ACP (Agent Client Protocol)** | Vyhledávání agentů CLI (Codex, Claude, Goose, Gemini CLI, OpenClaw + 9 dalších), generátor procesů, koncový bod `/api/acp/agents`
-🤖 **Řídicí panel agentů ACP** | Ladění › Stránka Agenti — mřížka 14 agentů se stavem instalace, verzí a formulářem pro vlastní agenta pro libovolný nástroj CLI. Uživatelé **OpenCode** získají tlačítko „Stáhnout opencode.json“, které automaticky vygeneruje konfiguraci připravenou k použití se všemi dostupnými modely.
-🔧 **Směrování `apiFormat` pro vlastní model** | Vlastní modely s `apiFormat: "responses"` nyní správně směrují do překladače Responses API.
-🏢 **Izolace pracovního prostoru Codexu** | Více pracovních prostorů Codexu na jeden e-mail – OAuth správně odděluje připojení podle ID pracovního prostoru
-🔄 **Automatická aktualizace elektronů** | Desktopová aplikace kontroluje aktualizace + automaticky se instaluje po restartu
+| Funkce                                         | Co to dělá                                                                                                                                                                                                                                                                                    |
+| ---------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🎮 **Modelové hřiště**                         | Stránka řídicího panelu pro přímé testování libovolného modelu – selektory poskytovatele/modelu/koncového bodu, editor Monaco, streamování, přerušení, načasování                                                                                                                             |
+| 🔏 **Porovnávání otisků prstů v CLI**          | Řazení hlaviček/těl serveru podle poskytovatele tak, aby odpovídalo nativním podpisům CLI – přepínání pro jednotlivé poskytovatele v Nastavení &gt; Zabezpečení. **Vaše IP adresa proxy serveru je zachována.**                                                                               |
+| 🤝 **Podpora ACP (Agent Client Protocol)**     | Vyhledávání agentů CLI (Codex, Claude, Goose, Gemini CLI, OpenClaw + 9 dalších), generátor procesů, koncový bod `/api/acp/agents`                                                                                                                                                             |
+| 🤖 **Řídicí panel agentů ACP**                 | Ladění › Stránka Agenti — mřížka 14 agentů se stavem instalace, verzí a formulářem pro vlastní agenta pro libovolný nástroj CLI. Uživatelé **OpenCode** získají tlačítko „Stáhnout opencode.json“, které automaticky vygeneruje konfiguraci připravenou k použití se všemi dostupnými modely. |
+| 🔧 **Směrování `apiFormat` pro vlastní model** | Vlastní modely s `apiFormat: "responses"` nyní správně směrují do překladače Responses API.                                                                                                                                                                                                   |
+| 🏢 **Izolace pracovního prostoru Codexu**      | Více pracovních prostorů Codexu na jeden e-mail – OAuth správně odděluje připojení podle ID pracovního prostoru                                                                                                                                                                               |
+| 🔄 **Automatická aktualizace elektronů**       | Desktopová aplikace kontroluje aktualizace + automaticky se instaluje po restartu                                                                                                                                                                                                             |
 
 ### 🤖 Operace s agenty a protokoly (v2.0)
 
-Funkce | Co to dělá
---- | ---
-🔧 **MCP Server (16 nástrojů)** | Nástroje IDE/agent prostřednictvím 3 transportů: stdio, SSE ( `/api/mcp/sse` ), Streamovatelný HTTP ( `/api/mcp/stream` )
-🤝 **A2A server (JSON-RPC + SSE)** | Spouštění úloh mezi agenty se synchronizací a streamováním
-🧭 **Konsolidovaná stránka koncových bodů** | Stránka pro správu s kartami Endpoint Proxy, MCP, A2A a API Endpoints
-🎚️ **Přepínače pro povolení/zakázání služby** | Přepínače ZAP/VYP pro MCP a A2A s trvalým nastavením (výchozí: VYP)
-🛰️ **Srdeční tep za běhu MCP** | Skutečný stav procesu (pid, doba provozuschopnosti, stáří heartbeatu, transport, režim rozsahu)
-📋 **Auditní záznam MCP** | Filtrovatelné protokoly auditu s hodnocením úspěchu/neúspěchu a klíčovým přiřazením
-🔐 **Vynucování rozsahu MCP** | 9 podrobných oprávnění pro řízený přístup k nástrojům
-📡 **Správa životního cyklu úkolů A2A** | Seznam/filtrování úloh, kontrola událostí/artefaktů, zrušení spuštěných úloh
-📋 **Objevení karty agenta** | `/.well-known/agent.json` pro automatické vyhledávání klientů
-🧪 **Testovací postroj Protocol E2E** | Skutečné MCP SDK + toky klientů A2A v `test:protocols:e2e`
-⚙️ **Provozní kontroly** | Kombinace přepínačů, použití profilů odolnosti, resetování jističů z jednoho ovládacího panelu
+| Funkce                                        | Co to dělá                                                                                                                |
+| --------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- |
+| 🔧 **MCP Server (16 nástrojů)**               | Nástroje IDE/agent prostřednictvím 3 transportů: stdio, SSE ( `/api/mcp/sse` ), Streamovatelný HTTP ( `/api/mcp/stream` ) |
+| 🤝 **A2A server (JSON-RPC + SSE)**            | Spouštění úloh mezi agenty se synchronizací a streamováním                                                                |
+| 🧭 **Konsolidovaná stránka koncových bodů**   | Stránka pro správu s kartami Endpoint Proxy, MCP, A2A a API Endpoints                                                     |
+| 🎚️ **Přepínače pro povolení/zakázání služby** | Přepínače ZAP/VYP pro MCP a A2A s trvalým nastavením (výchozí: VYP)                                                       |
+| 🛰️ **Srdeční tep za běhu MCP**                | Skutečný stav procesu (pid, doba provozuschopnosti, stáří heartbeatu, transport, režim rozsahu)                           |
+| 📋 **Auditní záznam MCP**                     | Filtrovatelné protokoly auditu s hodnocením úspěchu/neúspěchu a klíčovým přiřazením                                       |
+| 🔐 **Vynucování rozsahu MCP**                 | 9 podrobných oprávnění pro řízený přístup k nástrojům                                                                     |
+| 📡 **Správa životního cyklu úkolů A2A**       | Seznam/filtrování úloh, kontrola událostí/artefaktů, zrušení spuštěných úloh                                              |
+| 📋 **Objevení karty agenta**                  | `/.well-known/agent.json` pro automatické vyhledávání klientů                                                             |
+| 🧪 **Testovací postroj Protocol E2E**         | Skutečné MCP SDK + toky klientů A2A v `test:protocols:e2e`                                                                |
+| ⚙️ **Provozní kontroly**                      | Kombinace přepínačů, použití profilů odolnosti, resetování jističů z jednoho ovládacího panelu                            |
 
 ### 🧠 Směrování a inteligence
 
-Funkce | Co to dělá
---- | ---
-🎯 **Inteligentní čtyřúrovňový záložní systém** | Automatická trasa: Předplatné → API klíč → Levné → Zdarma
-📊 **Sledování kvót v reálném čase** | Počet tokenů v reálném čase + odpočet resetování pro každého poskytovatele
-🔄 **Překlad formátu** | OpenAI ↔ Claude ↔ Gemini ↔ Odpovědi s konverzemi bezpečnými pro schéma
-👥 **Podpora více účtů** | Více účtů na poskytovatele s inteligentním výběrem
-🔄 **Automatická aktualizace tokenů** | Tokeny OAuth se automaticky obnovují při opakovaném pokusu.
-🎨 **Vlastní kombinace** | 6 vyvažovacích strategií + řízení záložního řetězce
-🌐 **Směrovač se zástupnými znaky** | dynamické směrování `provider/*`
-🧠 **Přemýšlení o rozpočtových kontrolách** | Limity pro průchozí, automatické, vlastní a adaptivní uvažování
-🔀 **Aliasy modelů** | Vestavěné + vlastní aliasování modelů a bezpečnost migrace
-⚡ **Degradace pozadí** | Směrujte úlohy na pozadí s nízkou prioritou na levnější modely
-🧪 **Chytré směrování s ohledem na úkoly** | Automatický výběr modelu podle typu obsahu (kódování/vize/analýza/sumarizace)
-💬 **Vstřikování do systému** | Globální kontroly chování uplatňované konzistentně
-📄 **Kompatibilita API pro odpovědi** | Plná podpora `/v1/responses` pro Codex a pokročilé agentické pracovní postupy
+| Funkce                                          | Co to dělá                                                                    |
+| ----------------------------------------------- | ----------------------------------------------------------------------------- |
+| 🎯 **Inteligentní čtyřúrovňový záložní systém** | Automatická trasa: Předplatné → API klíč → Levné → Zdarma                     |
+| 📊 **Sledování kvót v reálném čase**            | Počet tokenů v reálném čase + odpočet resetování pro každého poskytovatele    |
+| 🔄 **Překlad formátu**                          | OpenAI ↔ Claude ↔ Gemini ↔ Odpovědi s konverzemi bezpečnými pro schéma        |
+| 👥 **Podpora více účtů**                        | Více účtů na poskytovatele s inteligentním výběrem                            |
+| 🔄 **Automatická aktualizace tokenů**           | Tokeny OAuth se automaticky obnovují při opakovaném pokusu.                   |
+| 🎨 **Vlastní kombinace**                        | 6 vyvažovacích strategií + řízení záložního řetězce                           |
+| 🌐 **Směrovač se zástupnými znaky**             | dynamické směrování `provider/*`                                              |
+| 🧠 **Přemýšlení o rozpočtových kontrolách**     | Limity pro průchozí, automatické, vlastní a adaptivní uvažování               |
+| 🔀 **Aliasy modelů**                            | Vestavěné + vlastní aliasování modelů a bezpečnost migrace                    |
+| ⚡ **Degradace pozadí**                         | Směrujte úlohy na pozadí s nízkou prioritou na levnější modely                |
+| 🧪 **Chytré směrování s ohledem na úkoly**      | Automatický výběr modelu podle typu obsahu (kódování/vize/analýza/sumarizace) |
+| 💬 **Vstřikování do systému**                   | Globální kontroly chování uplatňované konzistentně                            |
+| 📄 **Kompatibilita API pro odpovědi**           | Plná podpora `/v1/responses` pro Codex a pokročilé agentické pracovní postupy |
 
 ### 🎵 Multimodální API
 
-Funkce | Co to dělá
---- | ---
-🖼️ **Generování obrázků** | `/v1/images/generations` s cloudovým a lokálním backendem
-📐 **Vložení** | `/v1/embeddings` pro vyhledávání a RAG pipelines
-🎤 **Přepis zvuku** | `/v1/audio/transcriptions` (Whisper a další poskytovatelé)
-🔊 **Převod textu na řeč** | `/v1/audio/speech` (více enginů/poskytovatelů)
-🎬 **Generování videa** | `/v1/videos/generations` (pracovní postupy ComfyUI + SD WebUI)
-🎵 **Hudební generace** | `/v1/music/generations` (pracovní postupy ComfyUI)
-🛡️ **Moderování** | Bezpečnostní kontroly `/v1/moderations`
-🔀 **Změna pořadí** | `/v1/rerank` pro hodnocení relevance
-🔍 **Vyhledávání na webu** 🆕 | `/v1/search` — 5 poskytovatelů (Serper, Brave, Perplexity, Exa, Tavily), více než 6 500 zdarma/měsíc, automatické přepnutí na záložní systém, mezipaměť
+| Funkce                        | Co to dělá                                                                                                                                              |
+| ----------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Generování obrázků**     | `/v1/images/generations` s cloudovým a lokálním backendem                                                                                               |
+| 📐 **Vložení**                | `/v1/embeddings` pro vyhledávání a RAG pipelines                                                                                                        |
+| 🎤 **Přepis zvuku**           | `/v1/audio/transcriptions` (Whisper a další poskytovatelé)                                                                                              |
+| 🔊 **Převod textu na řeč**    | `/v1/audio/speech` (více enginů/poskytovatelů)                                                                                                          |
+| 🎬 **Generování videa**       | `/v1/videos/generations` (pracovní postupy ComfyUI + SD WebUI)                                                                                          |
+| 🎵 **Hudební generace**       | `/v1/music/generations` (pracovní postupy ComfyUI)                                                                                                      |
+| 🛡️ **Moderování**             | Bezpečnostní kontroly `/v1/moderations`                                                                                                                 |
+| 🔀 **Změna pořadí**           | `/v1/rerank` pro hodnocení relevance                                                                                                                    |
+| 🔍 **Vyhledávání na webu** 🆕 | `/v1/search` — 5 poskytovatelů (Serper, Brave, Perplexity, Exa, Tavily), více než 6 500 zdarma/měsíc, automatické přepnutí na záložní systém, mezipaměť |
 
 ### 🛡️ Odolnost, bezpečnost a správa věcí veřejných
 
-Funkce | Co to dělá
---- | ---
-🔌 **Jističe** | Vypnutí/obnovení pro každý model s ovládáním prahových hodnot
-🎯 **Modely s ohledem na koncové body** | Vlastní modely deklarují podporované koncové body + formát API
-🛡️ **Stádo proti hromům** | Ochrana mutexu a semaforu při událostech opakování/rychlosti
-🧠 **Sémantická + podpisová mezipaměť** | Snížení nákladů/latence díky dvěma vrstvám mezipaměti
-⚡ **Žádost o idempotenci** | Okno ochrany proti duplikacím
-🔒 **Falšování otisků prstů pomocí TLS** | Otisk TLS podobný prohlížeči – **snižuje detekci botů a nahlašování účtů**
-🔏 **Porovnávání otisků prstů v CLI** | Shoduje se s nativními podpisy požadavků CLI – **snižuje riziko zablokování a zároveň zachovává IP adresu proxy**
-🌐 **Filtrování IP adres** | Ovládání seznamu povolených/blokovaných položek pro odhalená nasazení
-📊 **Upravitelné limity rychlosti** | Konfigurovatelné globální/na úrovni poskytovatele limity s perzistencí
-🔑 **Správa klíčů API a stanovení rozsahu** | Bezpečné vydávání/rotace klíčů a kontroly modelu/poskytovatele
-🛡️ **Chráněné `/models`** | Volitelné ověřování a skrytí poskytovatele pro katalog modelů
+| Funkce                                      | Co to dělá                                                                                                        |
+| ------------------------------------------- | ----------------------------------------------------------------------------------------------------------------- |
+| 🔌 **Jističe**                              | Vypnutí/obnovení pro každý model s ovládáním prahových hodnot                                                     |
+| 🎯 **Modely s ohledem na koncové body**     | Vlastní modely deklarují podporované koncové body + formát API                                                    |
+| 🛡️ **Stádo proti hromům**                   | Ochrana mutexu a semaforu při událostech opakování/rychlosti                                                      |
+| 🧠 **Sémantická + podpisová mezipaměť**     | Snížení nákladů/latence díky dvěma vrstvám mezipaměti                                                             |
+| ⚡ **Žádost o idempotenci**                 | Okno ochrany proti duplikacím                                                                                     |
+| 🔒 **Falšování otisků prstů pomocí TLS**    | Otisk TLS podobný prohlížeči – **snižuje detekci botů a nahlašování účtů**                                        |
+| 🔏 **Porovnávání otisků prstů v CLI**       | Shoduje se s nativními podpisy požadavků CLI – **snižuje riziko zablokování a zároveň zachovává IP adresu proxy** |
+| 🌐 **Filtrování IP adres**                  | Ovládání seznamu povolených/blokovaných položek pro odhalená nasazení                                             |
+| 📊 **Upravitelné limity rychlosti**         | Konfigurovatelné globální/na úrovni poskytovatele limity s perzistencí                                            |
+| 🔑 **Správa klíčů API a stanovení rozsahu** | Bezpečné vydávání/rotace klíčů a kontroly modelu/poskytovatele                                                    |
+| 🛡️ **Chráněné `/models`**                   | Volitelné ověřování a skrytí poskytovatele pro katalog modelů                                                     |
 
 ### 📊 Pozorovatelnost a analytika
 
-Funkce | Co to dělá
---- | ---
-📝 **Žádost + protokolování proxy** | Úplné protokolování požadavků/odpovědí a proxy
-📋 **Sjednocený panel protokolů** | Zobrazení požadavků, proxy, auditu a konzole na jedné stránce
-🔍 **Vyžádat si telemetrii** | Latence p50/p95/p99 a trasování požadavků
-🏥 **Panel zdraví** | Doba provozuschopnosti, stavy jističů, uzamčení, statistiky mezipaměti
-💰 **Sledování nákladů** | Kontrola rozpočtu a přehled o cenách pro jednotlivé modely
-📈 **Analytické vizualizace** | Přehledy využití modelů/poskytovatelů a zobrazení trendů
-🧪 **Rámec hodnocení** | Testování zlaté sady s konfigurovatelnými strategiemi shody
+| Funkce                              | Co to dělá                                                             |
+| ----------------------------------- | ---------------------------------------------------------------------- |
+| 📝 **Žádost + protokolování proxy** | Úplné protokolování požadavků/odpovědí a proxy                         |
+| 📋 **Sjednocený panel protokolů**   | Zobrazení požadavků, proxy, auditu a konzole na jedné stránce          |
+| 🔍 **Vyžádat si telemetrii**        | Latence p50/p95/p99 a trasování požadavků                              |
+| 🏥 **Panel zdraví**                 | Doba provozuschopnosti, stavy jističů, uzamčení, statistiky mezipaměti |
+| 💰 **Sledování nákladů**            | Kontrola rozpočtu a přehled o cenách pro jednotlivé modely             |
+| 📈 **Analytické vizualizace**       | Přehledy využití modelů/poskytovatelů a zobrazení trendů               |
+| 🧪 **Rámec hodnocení**              | Testování zlaté sady s konfigurovatelnými strategiemi shody            |
 
 ### ☁️ Nasazení a platforma
 
-Funkce | Co to dělá
---- | ---
-🌐 **Nasazení kdekoli** | Localhost, VPS, Docker, cloudová prostředí
-💾 **Synchronizace s cloudem** | Synchronizace konfigurace přes cloud worker
-🔄 **Zálohování/Obnovení** | Toky exportu/importu a obnovy po havárii
-🧙 **Průvodce nástupem** | Průvodce prvním spuštěním
-🔧 **Panel nástrojů CLI** | Nastavení oblíbených kódovacích nástrojů jedním kliknutím
-🎮 **Modelové hřiště** | Otestujte libovolného poskytovatele/model/koncový bod z řídicího panelu
-🔏 **Přepínač otisků prstů v příkazovém řádku** | Porovnávání otisků prstů podle poskytovatele v Nastavení &gt; Zabezpečení
-🌐 **i18n (30 jazyků)** | Plná jazyková podpora dashboardu a dokumentace s psaním zprava doleva
-📂 **Adresář vlastních dat** | Přepsání `DATA_DIR` pro umístění úložiště
+| Funkce                                          | Co to dělá                                                                |
+| ----------------------------------------------- | ------------------------------------------------------------------------- |
+| 🌐 **Nasazení kdekoli**                         | Localhost, VPS, Docker, cloudová prostředí                                |
+| 💾 **Synchronizace s cloudem**                  | Synchronizace konfigurace přes cloud worker                               |
+| 🔄 **Zálohování/Obnovení**                      | Toky exportu/importu a obnovy po havárii                                  |
+| 🧙 **Průvodce nástupem**                        | Průvodce prvním spuštěním                                                 |
+| 🔧 **Panel nástrojů CLI**                       | Nastavení oblíbených kódovacích nástrojů jedním kliknutím                 |
+| 🎮 **Modelové hřiště**                          | Otestujte libovolného poskytovatele/model/koncový bod z řídicího panelu   |
+| 🔏 **Přepínač otisků prstů v příkazovém řádku** | Porovnávání otisků prstů podle poskytovatele v Nastavení &gt; Zabezpečení |
+| 🌐 **i18n (30 jazyků)**                         | Plná jazyková podpora dashboardu a dokumentace s psaním zprava doleva     |
+| 📂 **Adresář vlastních dat**                    | Přepsání `DATA_DIR` pro umístění úložiště                                 |
 
 ### Hluboký pohled na funkce
 
@@ -1203,12 +1115,12 @@ Předinstalovaná sada „OmniRoute Golden Set“ obsahuje testovací případy
 
 ### Strategie hodnocení
 
-Strategie | Popis | Příklad
---- | --- | ---
-`exact` | Výstup se musí přesně shodovat | `"4"`
-`contains` | Výstup musí obsahovat podřetězec (bez rozlišení velkých a malých písmen) | `"Paris"`
-`regex` | Výstup musí odpovídat vzoru regulárních výrazů | `"1.*2.*3"`
-`custom` | Vlastní JS funkce vrací true/false | `(output) => output.length > 10`
+| Strategie  | Popis                                                                    | Příklad                          |
+| ---------- | ------------------------------------------------------------------------ | -------------------------------- |
+| `exact`    | Výstup se musí přesně shodovat                                           | `"4"`                            |
+| `contains` | Výstup musí obsahovat podřetězec (bez rozlišení velkých a malých písmen) | `"Paris"`                        |
+| `regex`    | Výstup musí odpovídat vzoru regulárních výrazů                           | `"1.*2.*3"`                      |
+| `custom`   | Vlastní JS funkce vrací true/false                                       | `(output) => output.length > 10` |
 
 ---
 
@@ -1240,9 +1152,6 @@ Užitečná API pro automatizaci:
 - `GET /api/mcp/audit`
 - `GET /api/mcp/audit/stats`
 
-
-
-
 <details>
 <summary><b>🤝 Nastavení A2A (Agent2Agent)</b></summary>
 </details>
@@ -1272,9 +1181,6 @@ Provozní uživatelské rozhraní:
 
 - `/dashboard/a2a` pro pozorovatelnost úloh/stavů/streamů a akce kouření
 
-
-
-
 <details>
 <summary><b>🧪 Komplexní validace protokolu</b></summary>
 </details>
@@ -1291,9 +1197,6 @@ Tím se ověřuje:
 - A2A objevování/odesílání/streamování/získávání/zrušení
 - Křížová kontrola dat v auditu MCP a API pro správu úloh A2A
 
-
-
-
 <details>
 <summary><b>💳 Poskytovatelé předplatného</b></summary>
 </details>
@@ -1369,9 +1272,6 @@ Models:
   gh/gemini-3-pro
 ```
 
-
-
-
 <details>
 <summary><b>🔑 Poskytovatelé klíčů API</b></summary>
 </details>
@@ -1381,7 +1281,7 @@ Models:
 1. Registrace: [build.nvidia.com](https://build.nvidia.com)
 2. Získejte zdarma klíč API (včetně 1000 inferenčních kreditů)
 3. Ovládací panel → Přidat poskytovatele → NVIDIA NIM:
-    - Klíč API: `nvapi-your-key`
+   - Klíč API: `nvapi-your-key`
 
 **Modely:** `nvidia/llama-3.3-70b-instruct` , `nvidia/mistral-7b-instruct` a více než 50 dalších
 
@@ -1413,9 +1313,6 @@ Models:
 
 **Modely:** Získejte přístup k více než 100 modelům od všech hlavních poskytovatelů prostřednictvím jediného klíče API.
 
-
-
-
 <details>
 <summary><b>💰 Levní poskytovatelé (záložní)</b></summary>
 </details>
@@ -1425,8 +1322,8 @@ Models:
 1. Registrace: [Zhipu AI](https://open.bigmodel.cn/)
 2. Získejte klíč API z kódovacího plánu
 3. Nástěnka → Přidat klíč API:
-    - Poskytovatel: `glm`
-    - Klíč API: `your-key`
+   - Poskytovatel: `glm`
+   - Klíč API: `your-key`
 
 **Použití:** `glm/glm-4.7`
 
@@ -1452,9 +1349,6 @@ Models:
 
 **Tip pro profesionály:** Fixních 9 $/měsíc za 10 milionů tokenů = efektivní náklady 0,90 $/1 milion!
 
-
-
-
 <details>
 <summary><b>🆓 BEZPLATNÍ poskytovatelé (nouzové zálohování)</b></summary>
 </details>
@@ -1498,9 +1392,6 @@ Models:
   kr/claude-haiku-4.5
 ```
 
-
-
-
 <details>
 <summary><b>🎨 Vytvořte kombinace</b></summary>
 </details>
@@ -1531,9 +1422,6 @@ Models:
 Cost: $0 forever!
 ```
 
-
-
-
 <details>
 <summary><b>🔧 Integrace s rozhraním příkazového řádku</b></summary>
 </details>
@@ -1637,9 +1525,6 @@ opencode
 
 > **Tip:** Do sekce `models` přidejte jakýkoli model dostupný ve vašem koncovém bodu OmniRoute `/v1/models` . Použijte formát `provider/model-id` z vašeho dashboardu OmniRoute.
 
-
-
-
 ---
 
 ## 🐛 Řešení problémů
@@ -1880,14 +1765,8 @@ Chcete-li získat přístup k kriterii pověření, můžete použít adresu **U
 
 > Toto řešení funguje na základě autorizačního kódu na adrese URL a nezávislého přesměrování přesměrování nebo jiného.
 
-
-
-
 ---
 
-
-
-
 ## 🛠️ Technologický stack
 
 <details>
@@ -1909,28 +1788,25 @@ Chcete-li získat přístup k kriterii pověření, můžete použít adresu **U
 - **Docker** : [hub.docker.com/r/diegosouzapw/omniroute](https://hub.docker.com/r/diegosouzapw/omniroute)
 - **Odolnost** : Jistič, exponenciální odstavení, ochrana proti hromům, falešné TLS, automatické kombinované samoopravování
 
-
-
-
 ---
 
 ## 📖 Dokumentace
 
-Dokument | Popis
---- | ---
-[Uživatelská příručka](docs/USER_GUIDE.md) | Poskytovatelé, kombinace, integrace CLI, nasazení
-[Referenční informace k API](docs/API_REFERENCE.md) | Všechny koncové body s příklady
-[MCP server](open-sse/mcp-server/README.md) | 16 nástrojů MCP, konfigurace IDE, klienti Python/TS/Go
-[Server A2A](src/lib/a2a/README.md) | Protokol JSON-RPC 2.0, dovednosti, streamování, správa úloh
-[Auto-Combo Engine](docs/auto-combo.md) | 6faktorové bodování, balíčky režimů, samoléčba
-[Odstraňování problémů](docs/TROUBLESHOOTING.md) | Běžné problémy a jejich řešení
-[Architektura](docs/ARCHITECTURE.md) | Architektura a interní prvky systému
-[Přispívání](CONTRIBUTING.md) | Nastavení a pokyny pro vývoj
-[Specifikace OpenAPI](docs/openapi.yaml) | Specifikace OpenAPI 3.0
-[Bezpečnostní zásady](SECURITY.md) | Hlášení zranitelností a bezpečnostní postupy
-[Nasazení virtuálního počítače](docs/VM_DEPLOYMENT_GUIDE.md) | Kompletní průvodce: Nastavení virtuálního počítače + nginx + Cloudflare
-[Galerie funkcí](docs/FEATURES.md) | Vizuální prohlídka řídicího panelu se snímky obrazovky
-[Kontrolní seznam vydání](docs/RELEASE_CHECKLIST.md) | Kroky ověření před vydáním
+| Dokument                                                     | Popis                                                                   |
+| ------------------------------------------------------------ | ----------------------------------------------------------------------- |
+| [Uživatelská příručka](docs/USER_GUIDE.md)                   | Poskytovatelé, kombinace, integrace CLI, nasazení                       |
+| [Referenční informace k API](docs/API_REFERENCE.md)          | Všechny koncové body s příklady                                         |
+| [MCP server](open-sse/mcp-server/README.md)                  | 16 nástrojů MCP, konfigurace IDE, klienti Python/TS/Go                  |
+| [Server A2A](src/lib/a2a/README.md)                          | Protokol JSON-RPC 2.0, dovednosti, streamování, správa úloh             |
+| [Auto-Combo Engine](docs/auto-combo.md)                      | 6faktorové bodování, balíčky režimů, samoléčba                          |
+| [Odstraňování problémů](docs/TROUBLESHOOTING.md)             | Běžné problémy a jejich řešení                                          |
+| [Architektura](docs/ARCHITECTURE.md)                         | Architektura a interní prvky systému                                    |
+| [Přispívání](CONTRIBUTING.md)                                | Nastavení a pokyny pro vývoj                                            |
+| [Specifikace OpenAPI](docs/openapi.yaml)                     | Specifikace OpenAPI 3.0                                                 |
+| [Bezpečnostní zásady](SECURITY.md)                           | Hlášení zranitelností a bezpečnostní postupy                            |
+| [Nasazení virtuálního počítače](docs/VM_DEPLOYMENT_GUIDE.md) | Kompletní průvodce: Nastavení virtuálního počítače + nginx + Cloudflare |
+| [Galerie funkcí](docs/FEATURES.md)                           | Vizuální prohlídka řídicího panelu se snímky obrazovky                  |
+| [Kontrolní seznam vydání](docs/RELEASE_CHECKLIST.md)         | Kroky ověření před vydáním                                              |
 
 ---
 
@@ -1938,14 +1814,14 @@ Dokument | Popis
 
 OmniRoute má **v plánu více než 210 funkcí** v několika fázích vývoje. Zde jsou klíčové oblasti:
 
-Kategorie | Plánované funkce | Hlavní body
---- | --- | ---
-🧠 **Směrování a inteligence** | 25+ | Směrování s nejnižší latencí, směrování založené na tagech, kontrola kvót před výstupem, výběr účtu P2C
-🔒 **Zabezpečení a dodržování předpisů** | 20+ | Zpevnění SSRF, maskování přihlašovacích údajů, limit rychlosti pro každý koncový bod, stanovení rozsahu klíčů pro správu
-📊 **Pozorovatelnost** | 15+ | Integrace OpenTelemetry, sledování kvót v reálném čase, sledování nákladů podle modelu
-🔄 **Integrace poskytovatelů** | 20+ | Dynamický registr modelů, doba zchlazení poskytovatelů, Codex pro více účtů, analýza kvót Copilota
-⚡ **Výkon** | 15+ | Dvojitá vrstva mezipaměti, mezipaměť výzev, mezipaměť odpovědí, udržování streamování, dávkové API
-🌐 **Ekosystém** | 10+ | WebSocket API, horké opětovné načítání konfigurace, distribuované úložiště konfigurace, komerční režim
+| Kategorie                                | Plánované funkce | Hlavní body                                                                                                              |
+| ---------------------------------------- | ---------------- | ------------------------------------------------------------------------------------------------------------------------ |
+| 🧠 **Směrování a inteligence**           | 25+              | Směrování s nejnižší latencí, směrování založené na tagech, kontrola kvót před výstupem, výběr účtu P2C                  |
+| 🔒 **Zabezpečení a dodržování předpisů** | 20+              | Zpevnění SSRF, maskování přihlašovacích údajů, limit rychlosti pro každý koncový bod, stanovení rozsahu klíčů pro správu |
+| 📊 **Pozorovatelnost**                   | 15+              | Integrace OpenTelemetry, sledování kvót v reálném čase, sledování nákladů podle modelu                                   |
+| 🔄 **Integrace poskytovatelů**           | 20+              | Dynamický registr modelů, doba zchlazení poskytovatelů, Codex pro více účtů, analýza kvót Copilota                       |
+| ⚡ **Výkon**                             | 15+              | Dvojitá vrstva mezipaměti, mezipaměť výzev, mezipaměť odpovědí, udržování streamování, dávkové API                       |
+| 🌐 **Ekosystém**                         | 10+              | WebSocket API, horké opětovné načítání konfigurace, distribuované úložiště konfigurace, komerční režim                   |
 
 ### 🔜 Již brzy
 

docs/i18n/da/README.md

@@ -934,8 +934,8 @@ OmniRoute v2.0 er bygget som en operationel platform, ikke kun en relæ-proxy.
 | Funktion | Hvad det gør || -------------------------- | -------------------------------------------------------------------- |
 | 🖼️ **Billedgenerering** | `/v1/images/generations` med cloud og lokale backends |
 | 📐 **Indlejringer** | `/v1/embeddings` til søgning og RAG-rørledninger |
-| 🎤 **Lydtransskription** | `/v1/audio/transcriptions` (Whisper og yderligere udbydere) |
-| 🔊 **Tekst-til-tale** | `/v1/audio/speech` (flere motorer/udbydere) |
+| 🎤 **Lydtransskription** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Tekst-til-tale** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise) |
 | 🎬 **Videogenerering** | `/v1/videos/generations` (ComfyUI + SD WebUI-arbejdsgange) |
 | 🎵 **Music Generation** | `/v1/music/generations` (ComfyUI-arbejdsgange) |
 | 🛡️ **Moderationer** | `/v1/moderations` sikkerhedstjek |

docs/i18n/de/README.md

@@ -939,8 +939,8 @@ OmniRoute v2.0 ist als Betriebsplattform konzipiert und nicht nur als Relay-Prox
 | Funktion | Was es tut || -------------------------- | ------------------------------------------------------------- |
 | 🖼️ **Bilderzeugung** | `/v1/images/generations` mit Cloud- und lokalen Backends |
 | 📐 **Einbettungen** | `/v1/embeddings` für Such- und RAG-Pipelines |
-| 🎤 **Audio-Transkription** | `/v1/audio/transcriptions` (Whisper und zusätzliche Anbieter) |
-| 🔊 **Text-to-Speech** | `/v1/audio/speech` (mehrere Engines/Anbieter) |
+| 🎤 **Audio-Transkription** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Text-to-Speech** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise) |
 | 🎬 **Videogenerierung** | `/v1/videos/generations` (ComfyUI + SD WebUI-Workflows) |
 | 🎵 **Musikgeneration** | `/v1/music/generations` (ComfyUI-Workflows) |
 | 🛡️ **Moderationen** | `/v1/moderations` Sicherheitsprüfungen |

docs/i18n/es/README.md

@@ -877,14 +877,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 APIs Multi-Modal
 
-| Característica                | Qué Hace                                               |
-| ----------------------------- | ------------------------------------------------------ |
-| 🖼️ **Generación de Imágenes** | `/v1/images/generations` — 4 proveedores, 9+ modelos   |
-| 📐 **Embeddings**             | `/v1/embeddings` — 6 proveedores, 9+ modelos           |
-| 🎤 **Transcripción de Audio** | `/v1/audio/transcriptions` — Compatible con Whisper    |
-| 🔊 **Texto a Voz**            | `/v1/audio/speech` — Síntesis de audio multi-proveedor |
-| 🛡️ **Moderaciones**           | `/v1/moderations` — Verificaciones de seguridad        |
-| 🔀 **Reranking**              | `/v1/rerank` — Reranking de relevancia de documentos   |
+| Característica                | Qué Hace                                                                                                                                                                   |
+| ----------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Generación de Imágenes** | `/v1/images/generations` — 4 proveedores, 9+ modelos                                                                                                                       |
+| 📐 **Embeddings**             | `/v1/embeddings` — 6 proveedores, 9+ modelos                                                                                                                               |
+| 🎤 **Transcripción de Audio** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Texto a Voz**            | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderaciones**           | `/v1/moderations` — Verificaciones de seguridad                                                                                                                            |
+| 🔀 **Reranking**              | `/v1/rerank` — Reranking de relevancia de documentos                                                                                                                       |
 
 ### 🛡️ Resiliencia y Seguridad
 

docs/i18n/fi/README.md

@@ -874,14 +874,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Multimodaaliset sovellusliittymät
 
-| Ominaisuus                | Mitä se tekee                                             |
-| ------------------------- | --------------------------------------------------------- |
-| 🖼️ **Kuvan luominen**     | `/v1/images/generations` — 4 toimittajaa, 9+ mallia       |
-| 📐 **Upotukset**          | `/v1/embeddings` — 6 toimittajaa, 9+ mallia               |
-| 🎤 **Äänitranskriptio**   | `/v1/audio/transcriptions` — Kuiskausyhteensopiva         |
-| 🔊 **Tekstistä puheeksi** | `/v1/audio/speech` — Usean palveluntarjoajan äänisynteesi |
-| 🛡️ **Moderaatiot**        | `/v1/moderations` — Sisällön turvallisuustarkistukset     |
-| 🔀 **Uudelleenjärjestys** | `/v1/rerank` — Asiakirjan osuvuuden uudelleensijoitus     |
+| Ominaisuus                | Mitä se tekee                                                                                                                                                              |
+| ------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Kuvan luominen**     | `/v1/images/generations` — 4 toimittajaa, 9+ mallia                                                                                                                        |
+| 📐 **Upotukset**          | `/v1/embeddings` — 6 toimittajaa, 9+ mallia                                                                                                                                |
+| 🎤 **Äänitranskriptio**   | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Tekstistä puheeksi** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderaatiot**        | `/v1/moderations` — Sisällön turvallisuustarkistukset                                                                                                                      |
+| 🔀 **Uudelleenjärjestys** | `/v1/rerank` — Asiakirjan osuvuuden uudelleensijoitus                                                                                                                      |
 
 ### 🛡️ Joustavuus ja turvallisuus
 

docs/i18n/fr/README.md

@@ -875,14 +875,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 APIs multi-modales
 
-| Fonctionnalité             | Ce qu'elle fait                                         |
-| -------------------------- | ------------------------------------------------------- |
-| 🖼️ **Génération d'images** | `/v1/images/generations` — 4 fournisseurs, 9+ modèles   |
-| 📐 **Embeddings**          | `/v1/embeddings` — 6 fournisseurs, 9+ modèles           |
-| 🎤 **Transcription audio** | `/v1/audio/transcriptions` — compatible Whisper         |
-| 🔊 **Texte vers parole**   | `/v1/audio/speech` — synthèse audio multi-fournisseur   |
-| 🛡️ **Modérations**         | `/v1/moderations` — vérifications de sécurité           |
-| 🔀 **Reranking**           | `/v1/rerank` — reclassement de pertinence des documents |
+| Fonctionnalité             | Ce qu'elle fait                                                                                                                                                            |
+| -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Génération d'images** | `/v1/images/generations` — 4 fournisseurs, 9+ modèles                                                                                                                      |
+| 📐 **Embeddings**          | `/v1/embeddings` — 6 fournisseurs, 9+ modèles                                                                                                                              |
+| 🎤 **Transcription audio** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Texte vers parole**   | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Modérations**         | `/v1/moderations` — vérifications de sécurité                                                                                                                              |
+| 🔀 **Reranking**           | `/v1/rerank` — reclassement de pertinence des documents                                                                                                                    |
 
 ### 🛡️ Résilience & Sécurité
 

docs/i18n/he/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 ממשקי API רב-מודאליים
 
-| תכונה               | מה זה עושה                                    |
-| ------------------- | --------------------------------------------- |
-| 🖼️ **יצירת תמונות** | `/v1/images/generations` — 4 ספקים, 9+ דגמים  |
-| 📐 **הטבעות**       | `/v1/embeddings` — 6 ספקים, 9+ דגמים          |
-| 🎤 **תמלול אודיו**  | `/v1/audio/transcriptions` — תואם לחישה       |
-| 🔊 **טקסט לדיבור**  | `/v1/audio/speech` — סינתזת אודיו מרובה ספקים |
-| 🛡️ **מנחים**        | `/v1/moderations` — בדיקות בטיחות תוכן        |
-| 🔀 **דירוג מחדש**   | `/v1/rerank` — דירוג מחדש של רלוונטיות המסמך  |
+| תכונה               | מה זה עושה                                                                                                                                                                 |
+| ------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **יצירת תמונות** | `/v1/images/generations` — 4 ספקים, 9+ דגמים                                                                                                                               |
+| 📐 **הטבעות**       | `/v1/embeddings` — 6 ספקים, 9+ דגמים                                                                                                                                       |
+| 🎤 **תמלול אודיו**  | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **טקסט לדיבור**  | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **מנחים**        | `/v1/moderations` — בדיקות בטיחות תוכן                                                                                                                                     |
+| 🔀 **דירוג מחדש**   | `/v1/rerank` — דירוג מחדש של רלוונטיות המסמך                                                                                                                               |
 
 ### 🛡️ חוסן וביטחון
 

docs/i18n/hu/README.md

@@ -874,14 +874,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Multimodális API-k
 
-| Funkció                | Mit csinál                                                   |
-| ---------------------- | ------------------------------------------------------------ |
-| 🖼️ **Képgenerálás**    | `/v1/images/generations` — 4 szolgáltató, 9+ modell          |
-| 📐 **Beágyazás**       | `/v1/embeddings` — 6 szolgáltató, 9+ modell                  |
-| 🎤 **Audio átírás**    | `/v1/audio/transcriptions` — Suttogás-kompatibilis           |
-| 🔊 **Szövegfelolvasó** | `/v1/audio/speech` — Hangszintézis több szolgáltatónál       |
-| 🛡️ **Moderálás**       | `/v1/moderations` — Tartalombiztonsági ellenőrzések          |
-| 🔀 **Átsorolás**       | `/v1/rerank` — A dokumentumok relevancia szerinti átsorolása |
+| Funkció                | Mit csinál                                                                                                                                                                 |
+| ---------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Képgenerálás**    | `/v1/images/generations` — 4 szolgáltató, 9+ modell                                                                                                                        |
+| 📐 **Beágyazás**       | `/v1/embeddings` — 6 szolgáltató, 9+ modell                                                                                                                                |
+| 🎤 **Audio átírás**    | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Szövegfelolvasó** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderálás**       | `/v1/moderations` — Tartalombiztonsági ellenőrzések                                                                                                                        |
+| 🔀 **Átsorolás**       | `/v1/rerank` — A dokumentumok relevancia szerinti átsorolása                                                                                                               |
 
 ### 🛡️ Rugalmasság és biztonság
 

docs/i18n/id/README.md

@@ -874,14 +874,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 API Multi-Modal
 
-| Fitur                      | Apa Fungsinya                                          |
-| -------------------------- | ------------------------------------------------------ |
-| 🖼️ **Pembuatan Gambar**    | `/v1/images/generations` — 4 penyedia, 9+ model        |
-| 📐 **Sematan**             | `/v1/embeddings` — 6 penyedia, 9+ model                |
-| 🎤 **Transkripsi Audio**   | `/v1/audio/transcriptions` — Kompatibel dengan bisikan |
-| 🔊 **Teks-ke-Ucapan**      | `/v1/audio/speech` — Sintesis audio multi-penyedia     |
-| 🛡️ **Moderasi**            | `/v1/moderations` — Pemeriksaan keamanan konten        |
-| 🔀 **Pemeringkatan Ulang** | `/v1/rerank` — Pemeringkatan ulang relevansi dokumen   |
+| Fitur                      | Apa Fungsinya                                                                                                                                                              |
+| -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Pembuatan Gambar**    | `/v1/images/generations` — 4 penyedia, 9+ model                                                                                                                            |
+| 📐 **Sematan**             | `/v1/embeddings` — 6 penyedia, 9+ model                                                                                                                                    |
+| 🎤 **Transkripsi Audio**   | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Teks-ke-Ucapan**      | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderasi**            | `/v1/moderations` — Pemeriksaan keamanan konten                                                                                                                            |
+| 🔀 **Pemeringkatan Ulang** | `/v1/rerank` — Pemeringkatan ulang relevansi dokumen                                                                                                                       |
 
 ### 🛡️ Ketahanan & Keamanan
 

docs/i18n/in/README.md

@@ -770,14 +770,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 मल्टी-मॉडल एपीआई
 
-| फ़ीचर                        | यह क्या करता है                                   |
-| ---------------------------- | ------------------------------------------------- |
-| 🖼️ **छवि निर्माण**           | `/v1/images/generations` - 4 प्रदाता, 9+ मॉडल     |
-| 📐 **एंबेडिंग**              | `/v1/embeddings` — 6 प्रदाता, 9+ मॉडल             |
-| 🎤 **ऑडियो ट्रांस्क्रिप्शन** | `/v1/audio/transcriptions` - कानाफूसी-संगत        |
-| 🔊 **टेक्स्ट-टू-स्पीच**      | `/v1/audio/speech` - बहु-प्रदाता ऑडियो संश्लेषण   |
-| 🛡️ **संयम**                  | `/v1/moderations` — सामग्री सुरक्षा जांच          |
-| 🔀 **पुनर्रैंकिंग**          | `/v1/rerank` — दस्तावेज़ प्रासंगिकता पुनर्रैंकिंग |
+| फ़ीचर                        | यह क्या करता है                                                                                                                                                            |
+| ---------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **छवि निर्माण**           | `/v1/images/generations` - 4 प्रदाता, 9+ मॉडल                                                                                                                              |
+| 📐 **एंबेडिंग**              | `/v1/embeddings` — 6 प्रदाता, 9+ मॉडल                                                                                                                                      |
+| 🎤 **ऑडियो ट्रांस्क्रिप्शन** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **टेक्स्ट-टू-स्पीच**      | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **संयम**                  | `/v1/moderations` — सामग्री सुरक्षा जांच                                                                                                                                   |
+| 🔀 **पुनर्रैंकिंग**          | `/v1/rerank` — दस्तावेज़ प्रासंगिकता पुनर्रैंकिंग                                                                                                                          |
 
 ### 🛡️ लचीलापन और सुरक्षा
 

docs/i18n/it/README.md

@@ -874,14 +874,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 API Multi-modali
 
-| Funzionalità                | Cosa Fa                                              |
-| --------------------------- | ---------------------------------------------------- |
-| 🖼️ **Generazione immagini** | `/v1/images/generations` — 4 provider, 9+ modelli    |
-| 📐 **Embeddings**           | `/v1/embeddings` — 6 provider, 9+ modelli            |
-| 🎤 **Trascrizione audio**   | `/v1/audio/transcriptions` — Compatibile Whisper     |
-| 🔊 **Testo a voce**         | `/v1/audio/speech` — Sintesi audio multi-provider    |
-| 🛡️ **Moderazioni**          | `/v1/moderations` — Controlli di sicurezza           |
-| 🔀 **Reranking**            | `/v1/rerank` — Riclassificazione rilevanza documenti |
+| Funzionalità                | Cosa Fa                                                                                                                                                                    |
+| --------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Generazione immagini** | `/v1/images/generations` — 4 provider, 9+ modelli                                                                                                                          |
+| 📐 **Embeddings**           | `/v1/embeddings` — 6 provider, 9+ modelli                                                                                                                                  |
+| 🎤 **Trascrizione audio**   | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Testo a voce**         | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderazioni**          | `/v1/moderations` — Controlli di sicurezza                                                                                                                                 |
+| 🔀 **Reranking**            | `/v1/rerank` — Riclassificazione rilevanza documenti                                                                                                                       |
 
 ### 🛡️ Resilienza & Sicurezza
 

docs/i18n/ja/README.md

@@ -874,14 +874,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 マルチモーダル API
 
-| 特集                    | 何をするのか                                                    |
-| ----------------------- | --------------------------------------------------------------- |
-| 🖼️ **画像生成**         | `/v1/images/generations` — 4 つのプロバイダー、9 つ以上のモデル |
-| 📐 **埋め込み**         | `/v1/embeddings` — 6 つのプロバイダー、9 つ以上のモデル         |
-| 🎤 **音声文字起こし**   | `/v1/audio/transcriptions` — ウィスパー互換                     |
-| 🔊 **テキスト読み上げ** | `/v1/audio/speech` — マルチプロバイダーのオーディオ合成         |
-| 🛡️ **モデレーション**   | `/v1/moderations` — コンテンツの安全性チェック                  |
-| 🔀 **再ランキング**     | `/v1/rerank` — ドキュメントの関連性の再ランキング               |
+| 特集                    | 何をするのか                                                                                                                                                               |
+| ----------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **画像生成**         | `/v1/images/generations` — 4 つのプロバイダー、9 つ以上のモデル                                                                                                            |
+| 📐 **埋め込み**         | `/v1/embeddings` — 6 つのプロバイダー、9 つ以上のモデル                                                                                                                    |
+| 🎤 **音声文字起こし**   | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **テキスト読み上げ** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **モデレーション**   | `/v1/moderations` — コンテンツの安全性チェック                                                                                                                             |
+| 🔀 **再ランキング**     | `/v1/rerank` — ドキュメントの関連性の再ランキング                                                                                                                          |
 
 ### 🛡️ 復元力とセキュリティ
 

docs/i18n/ko/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 다중 모드 API
 
-| 기능                    | 그것이 하는 일                                         |
-| ----------------------- | ------------------------------------------------------ |
-| 🖼️ **이미지 생성**      | `/v1/images/generations` — 4개 공급자, 9개 이상의 모델 |
-| 📐 **임베딩**           | `/v1/embeddings` — 6개 공급자, 9개 이상의 모델         |
-| 🎤 **오디오 전사**      | `/v1/audio/transcriptions` — 속삭임 호환               |
-| 🔊 **텍스트 음성 변환** | `/v1/audio/speech` — 다중 제공자 오디오 합성           |
-| 🛡️ **조정**             | `/v1/moderations` — 콘텐츠 안전 확인                   |
-| 🔀 **재순위**           | `/v1/rerank` — 문서 관련성 재순위                      |
+| 기능                    | 그것이 하는 일                                                                                                                                                             |
+| ----------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **이미지 생성**      | `/v1/images/generations` — 4개 공급자, 9개 이상의 모델                                                                                                                     |
+| 📐 **임베딩**           | `/v1/embeddings` — 6개 공급자, 9개 이상의 모델                                                                                                                             |
+| 🎤 **오디오 전사**      | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **텍스트 음성 변환** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **조정**             | `/v1/moderations` — 콘텐츠 안전 확인                                                                                                                                       |
+| 🔀 **재순위**           | `/v1/rerank` — 문서 관련성 재순위                                                                                                                                          |
 
 ### 🛡️ 복원력 및 보안
 

docs/i18n/ms/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 API Berbilang Modal
 
-| Ciri                     | Apa yang Dilakukan                                     |
-| ------------------------ | ------------------------------------------------------ |
-| 🖼️ **Penjanaan Imej**    | `/v1/images/generations` — 4 pembekal, 9+ model        |
-| 📐 **Pembenaman**        | `/v1/embeddings` — 6 pembekal, 9+ model                |
-| 🎤 **Transkripsi Audio** | `/v1/audio/transcriptions` — Serasi dengan bisikan     |
-| 🔊 **Teks-ke-Ucapan**    | `/v1/audio/speech` — Sintesis audio berbilang pembekal |
-| 🛡️ **Kesederhanaan**     | `/v1/moderations` — Pemeriksaan keselamatan kandungan  |
-| 🔀 **Penyusunan semula** | `/v1/rerank` — Penarafan semula perkaitan dokumen      |
+| Ciri                     | Apa yang Dilakukan                                                                                                                                                         |
+| ------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Penjanaan Imej**    | `/v1/images/generations` — 4 pembekal, 9+ model                                                                                                                            |
+| 📐 **Pembenaman**        | `/v1/embeddings` — 6 pembekal, 9+ model                                                                                                                                    |
+| 🎤 **Transkripsi Audio** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Teks-ke-Ucapan**    | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Kesederhanaan**     | `/v1/moderations` — Pemeriksaan keselamatan kandungan                                                                                                                      |
+| 🔀 **Penyusunan semula** | `/v1/rerank` — Penarafan semula perkaitan dokumen                                                                                                                          |
 
 ### 🛡️ Ketahanan & Keselamatan
 

docs/i18n/nl/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Multimodale API's
 
-| Kenmerk                  | Wat het doet                                              |
-| ------------------------ | --------------------------------------------------------- |
-| 🖼️ **Beeldgeneratie**    | `/v1/images/generations` — 4 providers, 9+ modellen       |
-| 📐 **Insluitingen**      | `/v1/embeddings` — 6 providers, 9+ modellen               |
-| 🎤 **Audiotranscriptie** | `/v1/audio/transcriptions` — Whisper-compatibel           |
-| 🔊 **Tekst-naar-spraak** | `/v1/audio/speech` — Audiosynthese van meerdere providers |
-| 🛡️ **Moderaties**        | `/v1/moderations` — Veiligheidscontroles van inhoud       |
-| 🔀 **Herschikking**      | `/v1/rerank` — Herschikking van documentrelevantie        |
+| Kenmerk                  | Wat het doet                                                                                                                                                               |
+| ------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Beeldgeneratie**    | `/v1/images/generations` — 4 providers, 9+ modellen                                                                                                                        |
+| 📐 **Insluitingen**      | `/v1/embeddings` — 6 providers, 9+ modellen                                                                                                                                |
+| 🎤 **Audiotranscriptie** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Tekst-naar-spraak** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderaties**        | `/v1/moderations` — Veiligheidscontroles van inhoud                                                                                                                        |
+| 🔀 **Herschikking**      | `/v1/rerank` — Herschikking van documentrelevantie                                                                                                                         |
 
 ### 🛡️ Veerkracht en veiligheid
 

docs/i18n/no/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Multi-Modal APIer
 
-| Funksjon                | Hva det gjør                                           |
-| ----------------------- | ------------------------------------------------------ |
-| 🖼️ **Bildegenerering**  | `/v1/images/generations` — 4 leverandører, 9+ modeller |
-| 📐 **Innbygging**       | `/v1/embeddings` — 6 leverandører, 9+ modeller         |
-| 🎤 **Lydtranskripsjon** | `/v1/audio/transcriptions` — Whisper-kompatibel        |
-| 🔊 **Tekst-til-tale**   | `/v1/audio/speech` — Multi-leverandør lydsyntese       |
-| 🛡️ **Moderasjoner**     | `/v1/moderations` — Innholdssikkerhetssjekker          |
-| 🔀 **Omrangering**      | `/v1/rerank` — Rerangering av dokumentrelevans         |
+| Funksjon                | Hva det gjør                                                                                                                                                               |
+| ----------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Bildegenerering**  | `/v1/images/generations` — 4 leverandører, 9+ modeller                                                                                                                     |
+| 📐 **Innbygging**       | `/v1/embeddings` — 6 leverandører, 9+ modeller                                                                                                                             |
+| 🎤 **Lydtranskripsjon** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Tekst-til-tale**   | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderasjoner**     | `/v1/moderations` — Innholdssikkerhetssjekker                                                                                                                              |
+| 🔀 **Omrangering**      | `/v1/rerank` — Rerangering av dokumentrelevans                                                                                                                             |
 
 ### 🛡️ Spenst og sikkerhet
 

docs/i18n/phi/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Mga Multi-Modal na API
 
-| Tampok                     | Ano ang Ginagawa Nito                                        |
-| -------------------------- | ------------------------------------------------------------ |
-| 🖼️ **Pagbuo ng Larawan**   | `/v1/images/generations` — 4 na provider, 9+ na modelo       |
-| 📐 **Mga Pag-embed**       | `/v1/embeddings` — 6 na provider, 9+ na modelo               |
-| 🎤 **Audio Transcription** | `/v1/audio/transcriptions` — Whisper-compatible              |
-| 🔊 **Text-to-Speech**      | `/v1/audio/speech` — Multi-provider audio synthesis          |
-| 🛡️ **Mga Pag-moderate**    | `/v1/moderations` — Mga pagsusuri sa kaligtasan ng nilalaman |
-| 🔀 **Reranking**           | `/v1/rerank` — Muling pagraranggo ng kaugnayan ng dokumento  |
+| Tampok                     | Ano ang Ginagawa Nito                                                                                                                                                      |
+| -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Pagbuo ng Larawan**   | `/v1/images/generations` — 4 na provider, 9+ na modelo                                                                                                                     |
+| 📐 **Mga Pag-embed**       | `/v1/embeddings` — 6 na provider, 9+ na modelo                                                                                                                             |
+| 🎤 **Audio Transcription** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Text-to-Speech**      | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Mga Pag-moderate**    | `/v1/moderations` — Mga pagsusuri sa kaligtasan ng nilalaman                                                                                                               |
+| 🔀 **Reranking**           | `/v1/rerank` — Muling pagraranggo ng kaugnayan ng dokumento                                                                                                                |
 
 ### 🛡️ Katatagan at Seguridad
 

docs/i18n/pl/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Wielomodalne interfejsy API
 
-| Funkcja                       | Co to robi                                             |
-| ----------------------------- | ------------------------------------------------------ |
-| 🖼️ **Generowanie obrazu**     | `/v1/images/generations` — 4 dostawców, ponad 9 modeli |
-| 📐 **Osadzenia**              | `/v1/embeddings` — 6 dostawców, ponad 9 modeli         |
-| 🎤 **Transkrypcja audio**     | `/v1/audio/transcriptions` — Kompatybilny z szeptem    |
-| 🔊 **Zamiana tekstu na mowę** | `/v1/audio/speech` — Synteza dźwięku wielu dostawców   |
-| 🛡️ **Moderacje**              | `/v1/moderations` — Kontrola bezpieczeństwa treści     |
-| 🔀 **Ponowna pozycja**        | `/v1/rerank` — Zmiana rankingu trafności dokumentu     |
+| Funkcja                       | Co to robi                                                                                                                                                                 |
+| ----------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Generowanie obrazu**     | `/v1/images/generations` — 4 dostawców, ponad 9 modeli                                                                                                                     |
+| 📐 **Osadzenia**              | `/v1/embeddings` — 6 dostawców, ponad 9 modeli                                                                                                                             |
+| 🎤 **Transkrypcja audio**     | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Zamiana tekstu na mowę** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderacje**              | `/v1/moderations` — Kontrola bezpieczeństwa treści                                                                                                                         |
+| 🔀 **Ponowna pozycja**        | `/v1/rerank` — Zmiana rankingu trafności dokumentu                                                                                                                         |
 
 ### 🛡️ Odporność i bezpieczeństwo
 

docs/i18n/pt-BR/README.md

@@ -819,24 +819,28 @@ Quando minimizado, o OmniRoute fica na bandeja do sistema com ações rápidas:
 
 ## 💰 Preços Resumidos
 
-| Tier              | Provedor          | Custo                        | Reset de Cota     | Melhor Para             |
-| ----------------- | ----------------- | ---------------------------- | ----------------- | ----------------------- |
-| **💳 ASSINATURA** | Claude Code (Pro) | $20/mês                      | 5h + semanal      | Já é assinante          |
-|                   | Codex (Plus/Pro)  | $20-200/mês                  | 5h + semanal      | Usuários OpenAI         |
-|                   | Gemini CLI        | **GRATUITO**                 | 180K/mês + 1K/dia | Todos!                  |
-|                   | GitHub Copilot    | $10-19/mês                   | Mensal            | Usuários GitHub         |
-| **🔑 API KEY**    | NVIDIA NIM        | **GRATUITO** (1000 créditos) | Único             | Testes gratuitos        |
-|                   | DeepSeek          | Por uso                      | Nenhum            | Melhor preço/qualidade  |
-|                   | Groq              | Tier gratuito + pago         | Limitado          | Inferência ultra-rápida |
-|                   | xAI (Grok)        | Por uso                      | Nenhum            | Modelos Grok            |
-|                   | Mistral           | Tier gratuito + pago         | Limitado          | IA Europeia             |
-|                   | OpenRouter        | Por uso                      | Nenhum            | 100+ modelos            |
-| **💰 BARATO**     | GLM-4.7           | $0.6/1M                      | Diário 10h        | Backup econômico        |
-|                   | MiniMax M2.1      | $0.2/1M                      | Rotativo 5h       | Opção mais barata       |
-|                   | Kimi K2           | $9/mês fixo                  | 10M tokens/mês    | Custo previsível        |
-| **🆓 GRATUITO**   | iFlow             | $0                           | Ilimitado         | 8 modelos gratuitos     |
-|                   | Qwen              | $0                           | Ilimitado         | 3 modelos gratuitos     |
-|                   | Kiro              | $0                           | Ilimitado         | Claude gratuito         |
+| Tier              | Provedor          | Custo                        | Reset de Cota     | Melhor Para                    |
+| ----------------- | ----------------- | ---------------------------- | ----------------- | ------------------------------ |
+| **💳 ASSINATURA** | Claude Code (Pro) | $20/mês                      | 5h + semanal      | Já é assinante                 |
+|                   | Codex (Plus/Pro)  | $20-200/mês                  | 5h + semanal      | Usuários OpenAI                |
+|                   | Gemini CLI        | **GRATUITO**                 | 180K/mês + 1K/dia | Todos!                         |
+|                   | GitHub Copilot    | $10-19/mês                   | Mensal            | Usuários GitHub                |
+| **🔑 API KEY**    | NVIDIA NIM        | **GRATUITO** (1000 créditos) | Único             | Testes gratuitos               |
+|                   | DeepSeek          | Por uso                      | Nenhum            | Melhor preço/qualidade         |
+|                   | Groq              | Tier gratuito + pago         | Limitado          | Inferência ultra-rápida        |
+|                   | xAI (Grok)        | Por uso                      | Nenhum            | Modelos Grok                   |
+|                   | Mistral           | Tier gratuito + pago         | Limitado          | IA Europeia                    |
+|                   | OpenRouter        | Por uso                      | Nenhum            | 100+ modelos                   |
+| **💰 BARATO**     | GLM-4.7           | $0.6/1M                      | Diário 10h        | Backup econômico               |
+|                   | MiniMax M2.1      | $0.2/1M                      | Rotativo 5h       | Opção mais barata              |
+|                   | Kimi K2           | $9/mês fixo                  | 10M tokens/mês    | Custo previsível               |
+| **🆓 GRATUITO**   | iFlow             | $0                           | Ilimitado         | 8 modelos gratuitos            |
+|                   | Qwen              | $0                           | Ilimitado         | 3 modelos gratuitos            |
+|                   | Kiro              | $0                           | Ilimitado         | Claude gratuito                |
+|                   | LongCat 🆕        | **$0** (50M tok/dia 🔥)      | 1 req/s           | Maior cota grátis do mundo     |
+|                   | Pollinations 🆕   | **$0** (sem chave API)       | 1 req/15s         | GPT-5, Claude, DeepSeek, Llama |
+|                   | Cloudflare AI 🆕  | **$0** (10K Neurons/dia)     | ~150 resp/dia     | 50+ modelos, edge global       |
+|                   | Scaleway AI 🆕    | **$0** (1M tokens total)     | Limitado por taxa | EU/GDPR, Qwen3 235B, Llama 70B |
 
 **💡 Dica Pro:** Comece com Gemini CLI (180K grátis/mês) + iFlow (ilimitado grátis) = $0 de custo!
 
@@ -879,16 +883,16 @@ Por que isso é relevante:
 
 ### 🎵 APIs Multi-Modal
 
-| Funcionalidade              | O que Faz                                                                                                   |
-| --------------------------- | ----------------------------------------------------------------------------------------------------------- |
-| 🖼️ **Geração de Imagem**    | `/v1/images/generations` — 10 provedores, 20+ modelos (cloud + local)                                       |
-| 📐 **Embeddings**           | `/v1/embeddings` — 6 provedores, 9+ modelos                                                                 |
-| 🎤 **Transcrição de Áudio** | `/v1/audio/transcriptions` — Whisper + Nvidia NIM, HuggingFace, Qwen3                                       |
-| 🔊 **Texto para Fala**      | `/v1/audio/speech` — ElevenLabs, Nvidia NIM, HuggingFace, Coqui, Tortoise, Qwen3, Inworld, Cartesia, PlayHT |
-| 🎬 **Geração de Vídeo**     | `/v1/videos/generations` — ComfyUI (AnimateDiff, SVD), SD WebUI                                             |
-| 🎵 **Geração de Música**    | `/v1/music/generations` — ComfyUI (Stable Audio Open, MusicGen)                                             |
-| 🛡️ **Moderações**           | `/v1/moderations` — Verificações de segurança                                                               |
-| 🔀 **Reranking**            | `/v1/rerank` — Reranking de relevância de documentos                                                        |
+| Funcionalidade              | O que Faz                                                                                                                                                                  |
+| --------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Geração de Imagem**    | `/v1/images/generations` — 10 provedores, 20+ modelos (cloud + local)                                                                                                      |
+| 📐 **Embeddings**           | `/v1/embeddings` — 6 provedores, 9+ modelos                                                                                                                                |
+| 🎤 **Transcrição de Áudio** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Texto para Fala**      | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🎬 **Geração de Vídeo**     | `/v1/videos/generations` — ComfyUI (AnimateDiff, SVD), SD WebUI                                                                                                            |
+| 🎵 **Geração de Música**    | `/v1/music/generations` — ComfyUI (Stable Audio Open, MusicGen)                                                                                                            |
+| 🛡️ **Moderações**           | `/v1/moderations` — Verificações de segurança                                                                                                                              |
+| 🔀 **Reranking**            | `/v1/rerank` — Reranking de relevância de documentos                                                                                                                       |
 
 ### 🛡️ Resiliência e Segurança
 
@@ -1223,6 +1227,53 @@ Modelos:
   kr/claude-haiku-4.5
 ```
 
+### LongCat AI (GRATUITO 50M tokens/dia!) 🆕
+
+1. Cadastre-se: [longcat.chat](https://longcat.chat) com e-mail ou telefone
+2. Gere uma chave de API gratuita
+3. Dashboard → Adicionar Provedor → LongCat
+
+**Modelos:**
+
+- `lc/LongCat-Flash-Lite` — **50M tokens/dia** 💥 (maior cota gratuita do mundo!)
+- `lc/LongCat-Flash-Chat` — 500K tokens/dia
+- `lc/LongCat-Flash-Thinking` — 500K tokens/dia (raciocínio)
+
+> 100% gratuito durante o beta público. Reset diário à meia-noite UTC.
+
+### Pollinations AI (SEM CHAVE NECESSÁRIA!) 🆕
+
+1. Adicione o provedor Pollinations no Dashboard
+2. Deixe o campo de chave API vazio (ou coloque qualquer string)
+3. Comece a usar imediatamente!
+
+**Modelos via `pol/`:** `openai` (GPT-5), `claude`, `gemini`, `deepseek`, `llama` (Llama 4)
+
+> Sem cadastro, sem chave, sem cartão de crédito. 1 req/15s ilimitado.
+
+### Cloudflare Workers AI (GRATUITO 10K Neurons/dia!) 🆕
+
+1. Cadastre-se: [dash.cloudflare.com](https://dash.cloudflare.com)
+2. Gere um API Token em Profile → API Tokens
+3. Copie seu Account ID (coluna direita do dashboard)
+4. Dashboard → Adicionar Provedor → Cloudflare AI
+   - API Key: seu token
+   - Account ID: seu account ID
+
+**Modelos via `cf/`:** `@cf/meta/llama-3.3-70b-instruct`, `@cf/google/gemma-3-12b-it`, 50+ mais
+
+> 10K Neurons/dia ≈ 150 respostas de LLM ou 500s de transcrição Whisper gratuita!
+
+### Scaleway AI (1M tokens gratuitos!) 🆕
+
+1. Cadastre-se: [console.scaleway.com](https://console.scaleway.com)
+2. Gere uma chave de API IAM
+3. Dashboard → Adicionar Provedor → Scaleway
+
+**Modelos via `scw/`:** `qwen3-235b-a22b-instruct-2507` (Qwen3 235B!), `llama-3.1-70b-instruct`
+
+> 1M tokens gratuitos para novas contas. Dados processados na 🇫🇷 França (EU/GDPR).
+
 </details>
 
 <details>

docs/i18n/pt/README.md

@@ -874,14 +874,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 APIs multimodais
 
-| Recurso                           | O que faz                                                   |
-| --------------------------------- | ----------------------------------------------------------- |
-| 🖼️ **Geração de imagens**         | `/v1/images/generations` — 4 provedores, mais de 9 modelos  |
-| 📐 **Incorporações**              | `/v1/embeddings` — 6 provedores, mais de 9 modelos          |
-| 🎤 **Transcrição de áudio**       | `/v1/audio/transcriptions` — Compatível com sussurro        |
-| 🔊 **Conversão de texto em fala** | `/v1/audio/speech` — Síntese de áudio multiprovedor         |
-| 🛡️ **Moderações**                 | `/v1/moderations` — Verificações de segurança de conteúdo   |
-| 🔀 **Reclassificação**            | `/v1/rerank` — Reclassificação da relevância dos documentos |
+| Recurso                           | O que faz                                                                                                                                                                  |
+| --------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Geração de imagens**         | `/v1/images/generations` — 4 provedores, mais de 9 modelos                                                                                                                 |
+| 📐 **Incorporações**              | `/v1/embeddings` — 6 provedores, mais de 9 modelos                                                                                                                         |
+| 🎤 **Transcrição de áudio**       | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Conversão de texto em fala** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderações**                 | `/v1/moderations` — Verificações de segurança de conteúdo                                                                                                                  |
+| 🔀 **Reclassificação**            | `/v1/rerank` — Reclassificação da relevância dos documentos                                                                                                                |
 
 ### 🛡️ Resiliência e segurança
 

docs/i18n/ro/README.md

@@ -875,14 +875,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 API-uri multimodale
 
-| Caracteristica            | Ce face                                                    |
-| ------------------------- | ---------------------------------------------------------- |
-| 🖼️ **Generarea imaginii** | `/v1/images/generations` — 4 furnizori, peste 9 modele     |
-| 📐 **Inglobări**          | `/v1/embeddings` — 6 furnizori, peste 9 modele             |
-| 🎤 **Transcriere audio**  | `/v1/audio/transcriptions` — Compatibil cu Whisper         |
-| 🔊 **Text-to-speech**     | `/v1/audio/speech` — Sinteză audio cu mai mulți furnizori  |
-| 🛡️ **Moderații**          | `/v1/moderations` — Verificări de siguranță a conținutului |
-| 🔀 **Reclasificare**      | `/v1/rerank` — Reclasificarea relevanței documentului      |
+| Caracteristica            | Ce face                                                                                                                                                                    |
+| ------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Generarea imaginii** | `/v1/images/generations` — 4 furnizori, peste 9 modele                                                                                                                     |
+| 📐 **Inglobări**          | `/v1/embeddings` — 6 furnizori, peste 9 modele                                                                                                                             |
+| 🎤 **Transcriere audio**  | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Text-to-speech**     | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderații**          | `/v1/moderations` — Verificări de siguranță a conținutului                                                                                                                 |
+| 🔀 **Reclasificare**      | `/v1/rerank` — Reclasificarea relevanței documentului                                                                                                                      |
 
 ### 🛡️ Reziliență și securitate
 

docs/i18n/ru/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Мультимодальные API
 
-| Функция                      | Что делает                                          |
-| ---------------------------- | --------------------------------------------------- |
-| 🖼️ **Генерация изображений** | `/v1/images/generations` — 4 провайдера, 9+ моделей |
-| 📐 **Embeddings**            | `/v1/embeddings` — 6 провайдеров, 9+ моделей        |
-| 🎤 **Транскрипция аудио**    | `/v1/audio/transcriptions` — Совместимо с Whisper   |
-| 🔊 **Текст в речь**          | `/v1/audio/speech` — Мульти-провайдерный синтез     |
-| 🛡️ **Модерация**             | `/v1/moderations` — Проверки безопасности контента  |
-| 🔀 **Reranking**             | `/v1/rerank` — Переранжирование релевантности       |
+| Функция                      | Что делает                                                                                                                                                                 |
+| ---------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Генерация изображений** | `/v1/images/generations` — 4 провайдера, 9+ моделей                                                                                                                        |
+| 📐 **Embeddings**            | `/v1/embeddings` — 6 провайдеров, 9+ моделей                                                                                                                               |
+| 🎤 **Транскрипция аудио**    | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Текст в речь**          | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Модерация**             | `/v1/moderations` — Проверки безопасности контента                                                                                                                         |
+| 🔀 **Reranking**             | `/v1/rerank` — Переранжирование релевантности                                                                                                                              |
 
 ### 🛡️ Устойчивость и безопасность
 

docs/i18n/sk/README.md

@@ -877,14 +877,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Multimodálne API
 
-| Funkcia                     | Čo to robí                                                       |
-| --------------------------- | ---------------------------------------------------------------- |
-| 🖼️ **Generovanie obrázkov** | `/v1/images/generations` — 4 poskytovatelia, 9+ modelov          |
-| 📐 **Vloženie**             | `/v1/embeddings` — 6 poskytovateľov, 9+ modelov                  |
-| 🎤 **Prepis zvuku**         | `/v1/audio/transcriptions` — Kompatibilné so šepotom             |
-| 🔊 **Prevod textu na reč**  | `/v1/audio/speech` — Zvuková syntéza od viacerých poskytovateľov |
-| 🛡️ **Moderovania**          | `/v1/moderations` — Kontroly bezpečnosti obsahu                  |
-| 🔀 **Reranking**            | `/v1/rerank` — Zmena poradia relevantnosti dokumentu             |
+| Funkcia                     | Čo to robí                                                                                                                                                                 |
+| --------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Generovanie obrázkov** | `/v1/images/generations` — 4 poskytovatelia, 9+ modelov                                                                                                                    |
+| 📐 **Vloženie**             | `/v1/embeddings` — 6 poskytovateľov, 9+ modelov                                                                                                                            |
+| 🎤 **Prepis zvuku**         | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Prevod textu na reč**  | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderovania**          | `/v1/moderations` — Kontroly bezpečnosti obsahu                                                                                                                            |
+| 🔀 **Reranking**            | `/v1/rerank` — Zmena poradia relevantnosti dokumentu                                                                                                                       |
 
 ### 🛡️ Odolnosť a bezpečnosť
 

docs/i18n/sv/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Multimodala API:er
 
-| Funktion                 | Vad det gör                                            |
-| ------------------------ | ------------------------------------------------------ |
-| 🖼️ **Bildgenerering**    | `/v1/images/generations` — 4 leverantörer, 9+ modeller |
-| 📐 **Inbäddningar**      | `/v1/embeddings` — 6 leverantörer, 9+ modeller         |
-| 🎤 **Ljudtranskription** | `/v1/audio/transcriptions` — Whisper-kompatibel        |
-| 🔊 **Text-till-tal**     | `/v1/audio/speech` — Ljudsyntes med flera leverantörer |
-| 🛡️ **Moderationer**      | `/v1/moderations` — Innehållssäkerhetskontroller       |
-| 🔀 **Omrankning**        | `/v1/rerank` — Omrankning av dokumentrelevans          |
+| Funktion                 | Vad det gör                                                                                                                                                                |
+| ------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Bildgenerering**    | `/v1/images/generations` — 4 leverantörer, 9+ modeller                                                                                                                     |
+| 📐 **Inbäddningar**      | `/v1/embeddings` — 6 leverantörer, 9+ modeller                                                                                                                             |
+| 🎤 **Ljudtranskription** | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Text-till-tal**     | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Moderationer**      | `/v1/moderations` — Innehållssäkerhetskontroller                                                                                                                           |
+| 🔀 **Omrankning**        | `/v1/rerank` — Omrankning av dokumentrelevans                                                                                                                              |
 
 ### 🛡️ Motståndskraft och säkerhet
 

docs/i18n/th/README.md

@@ -874,14 +874,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Multi-Modal API
 
-| คุณสมบัติ               | มันทำอะไร                                                     |
-| ----------------------- | ------------------------------------------------------------- |
-| 🖼️ **การสร้างภาพ**      | `/v1/images/generations` — ผู้ให้บริการ 4 ราย รุ่น 9+         |
-| 📐 **การฝัง**           | `/v1/embeddings` — ผู้ให้บริการ 6 ราย รุ่น 9+                 |
-| 🎶 **การถอดเสียง**      | `/v1/audio/transcriptions` — รองรับการกระซิบ                  |
-| 🔊 **ข้อความเป็นคำพูด** | `/v1/audio/speech` — การสังเคราะห์เสียงจากผู้ให้บริการหลายราย |
-| 🛡️ **การกลั่นกรอง**     | `/v1/moderations` — การตรวจสอบความปลอดภัยของเนื้อหา           |
-| 🔀 **จัดอันดับ**        | `/v1/rerank` — การจัดอันดับความเกี่ยวข้องของเอกสาร            |
+| คุณสมบัติ               | มันทำอะไร                                                                                                                             |
+| ----------------------- | ------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **การสร้างภาพ**      | `/v1/images/generations` — ผู้ให้บริการ 4 ราย รุ่น 9+                                                                                 |
+| 📐 **การฝัง**           | `/v1/embeddings` — ผู้ให้บริการ 6 ราย รุ่น 9+                                                                                         |
+| 🎶 **การถอดเสียง**      | `/v1/audio/transcriptions` — รองรับการกระซิบ                                                                                          |
+| 🔊 **ข้อความเป็นคำพูด** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise) |
+| 🛡️ **การกลั่นกรอง**     | `/v1/moderations` — การตรวจสอบความปลอดภัยของเนื้อหา                                                                                   |
+| 🔀 **จัดอันดับ**        | `/v1/rerank` — การจัดอันดับความเกี่ยวข้องของเอกสาร                                                                                    |
 
 ### 🛡️ ความยืดหยุ่นและความปลอดภัย
 

docs/i18n/uk-UA/README.md

@@ -878,14 +878,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 Мультимодальні API
 
-| Особливість                        | Що він робить                                         |
-| ---------------------------------- | ----------------------------------------------------- |
-| 🖼️ **Створення зображень**         | `/v1/images/generations` — 4 провайдери, 9+ моделей   |
-| 📐 **Вбудовування**                | `/v1/embeddings` — 6 провайдерів, 9+ моделей          |
-| 🎤 **Транскрипція аудіо**          | `/v1/audio/transcriptions` — сумісний із Whisper      |
-| 🔊 **Створення тексту в мовлення** | `/v1/audio/speech` — Багатопровайдерний аудіосинтез   |
-| 🛡️ **Модерації**                   | `/v1/moderations` — Перевірка безпеки вмісту          |
-| 🔀 **Переранжування**              | `/v1/rerank` — Переранжування релевантності документа |
+| Особливість                        | Що він робить                                                                                                                                                              |
+| ---------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Створення зображень**         | `/v1/images/generations` — 4 провайдери, 9+ моделей                                                                                                                        |
+| 📐 **Вбудовування**                | `/v1/embeddings` — 6 провайдерів, 9+ моделей                                                                                                                               |
+| 🎤 **Транскрипція аудіо**          | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Створення тексту в мовлення** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Модерації**                   | `/v1/moderations` — Перевірка безпеки вмісту                                                                                                                               |
+| 🔀 **Переранжування**              | `/v1/rerank` — Переранжування релевантності документа                                                                                                                      |
 
 ### 🛡️ Стійкість і безпека
 

docs/i18n/vi/README.md

@@ -874,14 +874,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 API đa phương thức
 
-| Tính năng                             | Nó làm gì                                                    |
-| ------------------------------------- | ------------------------------------------------------------ |
-| 🖼️ **Tạo hình ảnh**                   | `/v1/images/generations` — 4 nhà cung cấp, hơn 9 mô hình     |
-| 📐 **Nhúng**                          | `/v1/embeddings` — 6 nhà cung cấp, hơn 9 mô hình             |
-| 🎤 **Phiên âm âm thanh**              | `/v1/audio/transcriptions` — Tương thích với lời thì thầm    |
-| 🔊 **Chuyển văn bản thành giọng nói** | `/v1/audio/speech` — Tổng hợp âm thanh từ nhiều nhà cung cấp |
-| 🛡️ **Kiểm duyệt**                     | `/v1/moderations` — Kiểm tra an toàn nội dung                |
-| 🔀 **Sắp xếp lại**                    | `/v1/rerank` — Sắp xếp lại mức độ liên quan của tài liệu     |
+| Tính năng                             | Nó làm gì                                                                                                                                                                  |
+| ------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **Tạo hình ảnh**                   | `/v1/images/generations` — 4 nhà cung cấp, hơn 9 mô hình                                                                                                                   |
+| 📐 **Nhúng**                          | `/v1/embeddings` — 6 nhà cung cấp, hơn 9 mô hình                                                                                                                           |
+| 🎤 **Phiên âm âm thanh**              | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **Chuyển văn bản thành giọng nói** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **Kiểm duyệt**                     | `/v1/moderations` — Kiểm tra an toàn nội dung                                                                                                                              |
+| 🔀 **Sắp xếp lại**                    | `/v1/rerank` — Sắp xếp lại mức độ liên quan của tài liệu                                                                                                                   |
 
 ### 🛡️ Khả năng phục hồi và bảo mật
 

docs/i18n/zh-CN/README.md

@@ -873,14 +873,14 @@ npm run electron:build:linux   # Linux (.AppImage)
 
 ### 🎵 多模态 API
 
-| 功能              | 功能描述                                       |
-| ----------------- | ---------------------------------------------- |
-| 🖼️ **图像生成**   | `/v1/images/generations` — 4 个提供商，9+ 模型 |
-| 📐 **Embeddings** | `/v1/embeddings` — 6 个提供商，9+ 模型         |
-| 🎤 **音频转录**   | `/v1/audio/transcriptions` — Whisper 兼容      |
-| 🔊 **文字转语音** | `/v1/audio/speech` — 多提供商音频合成          |
-| 🛡️ **内容审核**   | `/v1/moderations` — 内容安全检查               |
-| 🔀 **重排序**     | `/v1/rerank` — 文档相关性重排序                |
+| 功能              | 功能描述                                                                                                                                                                   |
+| ----------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 🖼️ **图像生成**   | `/v1/images/generations` — 4 个提供商，9+ 模型                                                                                                                             |
+| 📐 **Embeddings** | `/v1/embeddings` — 6 个提供商，9+ 模型                                                                                                                                     |
+| 🎤 **音频转录**   | `/v1/audio/transcriptions` — 7 providers (Deepgram Nova 3, AssemblyAI, Groq Whisper, HuggingFace, ElevenLabs, OpenAI, Azure), auto-language detection, MP4/MP3/WAV support |
+| 🔊 **文字转语音** | `/v1/audio/speech` — 10 providers (ElevenLabs, OpenAI, Deepgram, Cartesia, PlayHT, HuggingFace, Nvidia NIM, Inworld, Coqui, Tortoise)                                      |
+| 🛡️ **内容审核**   | `/v1/moderations` — 内容安全检查                                                                                                                                           |
+| 🔀 **重排序**     | `/v1/rerank` — 文档相关性重排序                                                                                                                                            |
 
 ### 🛡️ 弹性与安全
 

docs/openapi.yaml

@@ -1,7 +1,7 @@
 openapi: 3.1.0
 info:
   title: OmniRoute API
-  version: 2.8.4
+  version: 3.0.0-rc.10
   description: |
     OmniRoute is a local-first AI API proxy router. It provides an OpenAI-compatible
     endpoint that routes requests to multiple AI providers with load balancing,

next.config.mjs

@@ -13,6 +13,7 @@ const nextConfig = {
   },
   output: "standalone",
   serverExternalPackages: [
+    "thread-stream",
     "better-sqlite3",
     "zod",
     "child_process",
@@ -37,8 +38,16 @@ const nextConfig = {
   images: {
     unoptimized: true,
   },
-  webpack: (config, { isServer }) => {
+  webpack: (config, { isServer, webpack }) => {
     if (isServer) {
+      // Webpack IgnorePlugin: skip thread-stream test files that contain
+      // intentionally broken syntax/imports (they cause Turbopack build errors)
+      config.plugins.push(
+        new webpack.IgnorePlugin({
+          resourceRegExp: /\/test\//,
+          contextRegExp: /thread-stream/,
+        })
+      );
       // ── Turbopack / Next.js 16 module-hash patch (#394, #396, #398) ────────
       //
       // Next.js 16 (with or without Turbopack) compiles the instrumentation hook

open-sse/config/constants.ts

@@ -4,9 +4,9 @@ import { loadProviderCredentials } from "./credentialLoader.ts";
 export const FETCH_TIMEOUT_MS = parseInt(process.env.FETCH_TIMEOUT_MS || "120000", 10);
 
 // Idle timeout for SSE streams (ms). Closes stream if no data for this duration.
-// Default: 300s to support extended-thinking models (claude-opus-4-6, o3, etc.)
-// that may pause for >60s during deep reasoning phases. Override with STREAM_IDLE_TIMEOUT_MS env var.
-export const STREAM_IDLE_TIMEOUT_MS = parseInt(process.env.STREAM_IDLE_TIMEOUT_MS || "300000", 10);
+// Default: 120s balances deep-reasoning pauses with fast zombie stream detection (#473).
+// Extended-thinking models rarely pause >90s between chunks. Override with STREAM_IDLE_TIMEOUT_MS env var.
+export const STREAM_IDLE_TIMEOUT_MS = parseInt(process.env.STREAM_IDLE_TIMEOUT_MS || "120000", 10);
 
 // Provider configurations
 // OAuth credentials read from env vars with hardcoded fallbacks for backward compatibility.

open-sse/config/embeddingRegistry.ts

@@ -17,6 +17,7 @@ export interface EmbeddingProvider {
 }
 
 export interface EmbeddingProviderNodeRow {
+  id?: string;
   prefix: string;
   name: string;
   baseUrl: string;

open-sse/config/providerRegistry.ts

@@ -6,6 +6,8 @@
  * is auto-generated from this registry.
  */
 
+import { platform, arch } from "os";
+
 // ── Types ─────────────────────────────────────────────────────────────────
 
 export interface RegistryModel {
@@ -47,6 +49,8 @@ export interface RegistryEntry {
   executor: string;
   baseUrl?: string;
   baseUrls?: string[];
+  /** Override base URL used only for API key validation (e.g., opencode-go validates on zen/v1) */
+  testKeyBaseUrl?: string;
   responsesBaseUrl?: string;
   urlSuffix?: string;
   urlBuilder?: (base: string, model: string, stream: boolean) => string;
@@ -94,6 +98,32 @@ const KIMI_CODING_SHARED = {
   ] as RegistryModel[],
 } as const;
 
+function mapStainlessOs() {
+  switch (platform()) {
+    case "darwin":
+      return "MacOS";
+    case "win32":
+      return "Windows";
+    case "linux":
+      return "Linux";
+    default:
+      return `Other::${platform()}`;
+  }
+}
+
+function mapStainlessArch() {
+  switch (arch()) {
+    case "x64":
+      return "x64";
+    case "arm64":
+      return "arm64";
+    case "ia32":
+      return "x86";
+    default:
+      return `other::${arch()}`;
+  }
+}
+
 // ── Registry ──────────────────────────────────────────────────────────────
 
 export const REGISTRY: Record<string, RegistryEntry> = {
@@ -110,19 +140,19 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     headers: {
       "Anthropic-Version": "2023-06-01",
       "Anthropic-Beta":
-        "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14,context-management-2025-06-27",
+        "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14,context-management-2025-06-27,prompt-caching-scope-2026-01-05",
       "Anthropic-Dangerous-Direct-Browser-Access": "true",
-      "User-Agent": "claude-cli/1.0.83 (external, cli)",
+      "User-Agent": "claude-cli/2.1.63 (external, cli)",
       "X-App": "cli",
       "X-Stainless-Helper-Method": "stream",
       "X-Stainless-Retry-Count": "0",
       "X-Stainless-Runtime-Version": "v24.3.0",
-      "X-Stainless-Package-Version": "0.55.1",
+      "X-Stainless-Package-Version": "0.74.0",
       "X-Stainless-Runtime": "node",
       "X-Stainless-Lang": "js",
-      "X-Stainless-Arch": "arm64",
-      "X-Stainless-Os": "MacOS",
-      "X-Stainless-Timeout": "60",
+      "X-Stainless-Arch": mapStainlessArch(),
+      "X-Stainless-Os": mapStainlessOs(),
+      "X-Stainless-Timeout": "600",
     },
     oauth: {
       clientIdEnv: "CLAUDE_OAUTH_CLIENT_ID",
@@ -157,9 +187,13 @@ export const REGISTRY: Record<string, RegistryEntry> = {
       clientSecretDefault: "",
     },
     models: [
+      { id: "gemini-3.1-pro-high", name: "Gemini 3.1 Pro High" },
+      { id: "gemini-3.1-pro-low", name: "Gemini 3.1 Pro Low" },
       { id: "gemini-3.1-pro", name: "Gemini 3.1 Pro" },
       { id: "gemini-3-1-pro", name: "Gemini 3.1 Pro (Alt ID)" },
       { id: "gemini-3.1-pro-preview", name: "Gemini 3.1 Pro Preview" },
+      { id: "gemini-3.1-flash-lite-preview", name: "Gemini 3.1 Flash Lite Preview" },
+      { id: "gemini-3-flash-preview", name: "Gemini 3 Flash Preview" },
       { id: "gemini-2.5-pro", name: "Gemini 2.5 Pro" },
       { id: "gemini-2.5-flash", name: "Gemini 2.5 Flash" },
       { id: "gemini-2.5-flash-lite", name: "Gemini 2.5 Flash Lite" },
@@ -189,9 +223,13 @@ export const REGISTRY: Record<string, RegistryEntry> = {
       clientSecretDefault: "",
     },
     models: [
+      { id: "gemini-3.1-pro-high", name: "Gemini 3.1 Pro High" },
+      { id: "gemini-3.1-pro-low", name: "Gemini 3.1 Pro Low" },
       { id: "gemini-3.1-pro", name: "Gemini 3.1 Pro" },
       { id: "gemini-3-1-pro", name: "Gemini 3.1 Pro (Alt ID)" },
       { id: "gemini-3.1-pro-preview", name: "Gemini 3.1 Pro Preview" },
+      { id: "gemini-3.1-flash-lite-preview", name: "Gemini 3.1 Flash Lite Preview" },
+      { id: "gemini-3-flash-preview", name: "Gemini 3 Flash Preview" },
       { id: "gemini-2.5-pro", name: "Gemini 2.5 Pro" },
       { id: "gemini-2.5-flash", name: "Gemini 2.5 Flash" },
       { id: "gemini-2.5-flash-lite", name: "Gemini 2.5 Flash Lite" },
@@ -320,7 +358,11 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     alias: "ag",
     format: "antigravity",
     executor: "antigravity",
-    baseUrls: ["https://daily-cloudcode-pa.googleapis.com", "https://cloudcode-pa.googleapis.com"],
+    baseUrls: [
+      "https://daily-cloudcode-pa.googleapis.com",
+      "https://daily-cloudcode-pa.sandbox.googleapis.com",
+      "https://cloudcode-pa.googleapis.com",
+    ],
     urlBuilder: (base, model, stream) => {
       const path = stream
         ? "/v1internal:streamGenerateContent?alt=sse"
@@ -330,7 +372,7 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     authType: "oauth",
     authHeader: "bearer",
     headers: {
-      "User-Agent": "antigravity/1.104.0 darwin/arm64",
+      "User-Agent": `antigravity/1.107.0 ${platform()}/${arch()}`,
     },
     oauth: {
       clientIdEnv: "ANTIGRAVITY_OAUTH_CLIENT_ID",
@@ -359,9 +401,9 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     authHeader: "bearer",
     headers: {
       "copilot-integration-id": "vscode-chat",
-      "editor-version": "vscode/1.107.1",
-      "editor-plugin-version": "copilot-chat/0.26.7",
-      "user-agent": "GitHubCopilotChat/0.26.7",
+      "editor-version": "vscode/1.110.0",
+      "editor-plugin-version": "copilot-chat/0.38.0",
+      "user-agent": "GitHubCopilotChat/0.38.0",
       "openai-intent": "conversation-panel",
       "x-github-api-version": "2025-04-01",
       "x-vscode-user-agent-library-version": "electron-fetch",
@@ -495,6 +537,41 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     ],
   },
 
+  "opencode-go": {
+    id: "opencode-go",
+    alias: "opencode-go",
+    format: "openai",
+    executor: "opencode",
+    baseUrl: "https://opencode.ai/zen/go/v1",
+    // (#532) Key validation must hit the main zen endpoint (same key works for both tiers)
+    testKeyBaseUrl: "https://opencode.ai/zen/v1",
+    authType: "apikey",
+    authHeader: "Authorization",
+    authPrefix: "Bearer",
+    models: [
+      { id: "glm-5", name: "GLM-5" },
+      { id: "kimi-k2.5", name: "Kimi K2.5" },
+      { id: "minimax-m2.7", name: "MiniMax M2.7", targetFormat: "claude" },
+      { id: "minimax-m2.5", name: "MiniMax M2.5", targetFormat: "claude" },
+    ],
+  },
+
+  "opencode-zen": {
+    id: "opencode-zen",
+    alias: "opencode-zen",
+    format: "openai",
+    executor: "opencode",
+    baseUrl: "https://opencode.ai/zen/v1",
+    authType: "apikey",
+    authHeader: "Authorization",
+    authPrefix: "Bearer",
+    models: [
+      { id: "minimax-m2.5-free", name: "MiniMax M2.5 Free" },
+      { id: "big-pickle", name: "Big Pickle" },
+      { id: "gpt-5-nano", name: "GPT 5 Nano" },
+    ],
+  },
+
   openrouter: {
     id: "openrouter",
     alias: "openrouter",
@@ -709,6 +786,10 @@ export const REGISTRY: Record<string, RegistryEntry> = {
       "Anthropic-Beta": "claude-code-20250219,interleaved-thinking-2025-05-14",
     },
     models: [
+      // T12/T28: MiniMax default upgraded from M2.5 to M2.7
+      { id: "minimax-m2.7", name: "MiniMax M2.7" },
+      { id: "MiniMax-M2.7", name: "MiniMax M2.7 (Legacy Alias)" },
+      { id: "minimax-m2.7-highspeed", name: "MiniMax M2.7 Highspeed" },
       { id: "minimax-m2.5", name: "MiniMax M2.5" },
       { id: "MiniMax-M2.5", name: "MiniMax M2.5 (Legacy Alias)" },
       { id: "MiniMax-M2.1", name: "MiniMax M2.1" },
@@ -730,6 +811,9 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     },
     models: [
       // Keep parity with minimax to ensure model discovery works for minimax-cn connections.
+      { id: "minimax-m2.7", name: "MiniMax M2.7" },
+      { id: "MiniMax-M2.7", name: "MiniMax M2.7 (Legacy Alias)" },
+      { id: "minimax-m2.7-highspeed", name: "MiniMax M2.7 Highspeed" },
       { id: "minimax-m2.5", name: "MiniMax M2.5" },
       { id: "MiniMax-M2.5", name: "MiniMax M2.5 (Legacy Alias)" },
       { id: "MiniMax-M2.1", name: "MiniMax M2.1" },
@@ -883,6 +967,12 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     authType: "apikey",
     authHeader: "bearer",
     models: [
+      { id: "meta-llama/Llama-3.3-70B-Instruct-Turbo-Free", name: "Llama 3.3 70B Turbo (🆓 Free)" },
+      { id: "meta-llama/Llama-Vision-Free", name: "Llama Vision (🆓 Free)" },
+      {
+        id: "deepseek-ai/DeepSeek-R1-Distill-Llama-70B-Free",
+        name: "DeepSeek R1 Distill 70B (🆓 Free)",
+      },
       { id: "meta-llama/Llama-3.3-70B-Instruct-Turbo", name: "Llama 3.3 70B Turbo" },
       { id: "deepseek-ai/DeepSeek-R1", name: "DeepSeek R1" },
       { id: "Qwen/Qwen3-235B-A22B", name: "Qwen3 235B" },
@@ -1103,7 +1193,7 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     alias: "vertex",
     // Vertex AI uses Google's generateContent format (same as Gemini)
     format: "gemini",
-    executor: "default",
+    executor: "vertex",
     // URL uses {project_id} and {region} from providerSpecificData — handled by custom executor or fallback
     // Default to us-central1 / generic endpoint; users configure project via providerSpecificData
     baseUrl: "https://us-central1-aiplatform.googleapis.com/v1/projects",
@@ -1117,14 +1207,213 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     authType: "apikey",
     authHeader: "bearer",
     models: [
+      { id: "gemini-3.1-pro-preview", name: "Gemini 3.1 Pro Preview (Vertex)" },
+      { id: "gemini-3.1-flash-lite-preview", name: "Gemini 3.1 Flash Lite Preview (Vertex)" },
+      { id: "gemini-3-flash-preview", name: "Gemini 3 Flash Preview (Vertex)" },
       { id: "gemini-2.5-pro", name: "Gemini 2.5 Pro (Vertex)" },
       { id: "gemini-2.5-flash", name: "Gemini 2.5 Flash (Vertex)" },
       { id: "gemini-2.0-flash-thinking-exp", name: "Gemini 2.0 Flash Thinking Exp (Vertex)" },
       { id: "gemma-2-27b-it", name: "Gemma 2 27B (Vertex)" },
+      { id: "deepseek-v3.2", name: "DeepSeek V3.2 (Vertex Partner)" },
+      { id: "qwen3-next-80b", name: "Qwen3 Next 80B (Vertex Partner)" },
+      { id: "glm-5", name: "GLM-5 (Vertex Partner)" },
       { id: "claude-opus-4-5@20251101", name: "Claude Opus 4.5 (Vertex)" },
       { id: "claude-sonnet-4-5@20251101", name: "Claude Sonnet 4.5 (Vertex)" },
     ],
   },
+
+  alibaba: {
+    id: "alibaba",
+    alias: "ali",
+    format: "openai",
+    executor: "default",
+    // DashScope international OpenAI-compatible endpoint.
+    // China users should set providerSpecificData.baseUrl to:
+    //   https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions
+    baseUrl: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions",
+    modelsUrl: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/models",
+    authType: "apikey",
+    authHeader: "bearer",
+    models: [
+      { id: "qwen-max", name: "Qwen Max" },
+      { id: "qwen-max-2025-01-25", name: "Qwen Max (2025-01-25)" },
+      { id: "qwen-plus", name: "Qwen Plus" },
+      { id: "qwen-plus-2025-07-14", name: "Qwen Plus (2025-07-14)" },
+      { id: "qwen-turbo", name: "Qwen Turbo" },
+      { id: "qwen-turbo-2025-11-01", name: "Qwen Turbo (2025-11-01)" },
+      { id: "qwen3-coder-plus", name: "Qwen3 Coder Plus" },
+      { id: "qwen3-coder-flash", name: "Qwen3 Coder Flash" },
+      { id: "qwq-plus", name: "QwQ Plus (Reasoning)" },
+      { id: "qwq-32b", name: "QwQ 32B" },
+      { id: "qwen3-32b", name: "Qwen3 32B" },
+      { id: "qwen3-235b-a22b", name: "Qwen3 235B A22B" },
+    ],
+    passthroughModels: true,
+  },
+
+  // ── New Free Providers (2026) ─────────────────────────────────────────────
+
+  longcat: {
+    id: "longcat",
+    alias: "lc",
+    format: "openai",
+    executor: "default",
+    // (#536) Correct OpenAI-compatible base URL — was longcat.chat/api/v1/chat/completions
+    // which is the chat endpoint directly, not the base. Key validation and routing must
+    // use https://api.longcat.chat/openai which resolves /v1/models and /v1/chat/completions
+    baseUrl: "https://api.longcat.chat/openai",
+    authType: "apikey",
+    authHeader: "Authorization",
+    authPrefix: "Bearer",
+    // Free tier: 50M tokens/day (Flash-Lite) + 500K/day (Chat/Thinking) — 100% free while public beta
+    models: [
+      { id: "LongCat-Flash-Lite", name: "LongCat Flash-Lite (50M tok/day 🆓)" },
+      { id: "LongCat-Flash-Chat", name: "LongCat Flash-Chat (500K tok/day 🆓)" },
+      { id: "LongCat-Flash-Thinking", name: "LongCat Flash-Thinking (500K tok/day 🆓)" },
+      { id: "LongCat-Flash-Thinking-2601", name: "LongCat Flash-Thinking-2601 (🆓)" },
+      { id: "LongCat-Flash-Omni-2603", name: "LongCat Flash-Omni-2603 (🆓)" },
+    ],
+  },
+
+  pollinations: {
+    id: "pollinations",
+    alias: "pol",
+    format: "openai",
+    executor: "pollinations",
+    // No API key required for basic use. Proxy to GPT-5, Claude, Gemini, DeepSeek, Llama 4.
+    baseUrl: "https://text.pollinations.ai/openai/chat/completions",
+    authType: "apikey", // Optional — works without one too
+    authHeader: "bearer",
+    models: [
+      { id: "openai", name: "GPT-5 via Pollinations (🆓)" },
+      { id: "claude", name: "Claude via Pollinations (🆓)" },
+      { id: "gemini", name: "Gemini via Pollinations (🆓)" },
+      { id: "deepseek", name: "DeepSeek V3 via Pollinations (🆓)" },
+      { id: "llama", name: "Llama 4 via Pollinations (🆓)" },
+      { id: "mistral", name: "Mistral via Pollinations (🆓)" },
+    ],
+  },
+
+  puter: {
+    id: "puter",
+    alias: "pu",
+    format: "openai",
+    executor: "puter",
+    // OpenAI-compatible gateway with 500+ models (GPT, Claude, Gemini, Grok, DeepSeek, Qwen…)
+    // Auth: Bearer <puter_auth_token> from puter.com/dashboard → Copy Auth Token
+    // Model IDs use provider/model-name format for non-OpenAI models.
+    // Only chat completions (incl. streaming) are available via REST.
+    // Image gen, TTS, STT, video are puter.js SDK-only (browser).
+    baseUrl: "https://api.puter.com/puterai/openai/v1/chat/completions",
+    authType: "apikey",
+    authHeader: "bearer",
+    models: [
+      // OpenAI — use bare IDs
+      { id: "gpt-4o-mini", name: "GPT-4o Mini (🆓 Puter)" },
+      { id: "gpt-4o", name: "GPT-4o (Puter)" },
+      { id: "gpt-4.1", name: "GPT-4.1 (Puter)" },
+      { id: "gpt-4.1-mini", name: "GPT-4.1 Mini (Puter)" },
+      { id: "gpt-5-nano", name: "GPT-5 Nano (Puter)" },
+      { id: "gpt-5-mini", name: "GPT-5 Mini (Puter)" },
+      { id: "gpt-5", name: "GPT-5 (Puter)" },
+      { id: "o3-mini", name: "OpenAI o3-mini (Puter)" },
+      { id: "o3", name: "OpenAI o3 (Puter)" },
+      { id: "o4-mini", name: "OpenAI o4-mini (Puter)" },
+      // Anthropic Claude — use bare IDs (confirmed working)
+      { id: "claude-haiku-4-5", name: "Claude Haiku 4.5 (Puter)" },
+      { id: "claude-sonnet-4-5", name: "Claude Sonnet 4.5 (Puter)" },
+      { id: "claude-opus-4-5", name: "Claude Opus 4.5 (Puter)" },
+      { id: "claude-sonnet-4", name: "Claude Sonnet 4 (Puter)" },
+      { id: "claude-opus-4", name: "Claude Opus 4 (Puter)" },
+      // Google Gemini — use google/ prefix (confirmed working)
+      { id: "google/gemini-2.0-flash", name: "Gemini 2.0 Flash (Puter)" },
+      { id: "google/gemini-2.5-flash", name: "Gemini 2.5 Flash (Puter)" },
+      { id: "google/gemini-2.5-pro", name: "Gemini 2.5 Pro (Puter)" },
+      { id: "google/gemini-3-flash", name: "Gemini 3 Flash (Puter)" },
+      { id: "google/gemini-3-pro", name: "Gemini 3 Pro (Puter)" },
+      // DeepSeek — use deepseek/ prefix (confirmed working)
+      { id: "deepseek/deepseek-chat", name: "DeepSeek Chat (Puter)" },
+      { id: "deepseek/deepseek-r1", name: "DeepSeek R1 (Puter)" },
+      { id: "deepseek/deepseek-v3.2", name: "DeepSeek V3.2 (Puter)" },
+      // xAI Grok — use x-ai/ prefix
+      { id: "x-ai/grok-3", name: "Grok 3 (Puter)" },
+      { id: "x-ai/grok-3-mini", name: "Grok 3 Mini (Puter)" },
+      { id: "x-ai/grok-4", name: "Grok 4 (Puter)" },
+      { id: "x-ai/grok-4-fast", name: "Grok 4 Fast (Puter)" },
+      // Meta Llama — bare IDs (confirmed ✅)
+      { id: "llama-4-scout", name: "Llama 4 Scout (Puter)" },
+      { id: "llama-4-maverick", name: "Llama 4 Maverick (Puter)" },
+      { id: "llama-3.3-70b-instruct", name: "Llama 3.3 70B (Puter)" },
+      // Mistral — bare IDs (confirmed ✅)
+      { id: "mistral-small-latest", name: "Mistral Small (Puter)" },
+      { id: "mistral-medium-latest", name: "Mistral Medium (Puter)" },
+      { id: "open-mistral-nemo", name: "Mistral Nemo (Puter)" },
+      // Qwen — use qwen/ prefix (confirmed ✅)
+      { id: "qwen/qwen3-235b-a22b", name: "Qwen3 235B (Puter)" },
+      { id: "qwen/qwen3-32b", name: "Qwen3 32B (Puter)" },
+      { id: "qwen/qwen3-coder", name: "Qwen3 Coder 480B (Puter)" },
+    ],
+    passthroughModels: true, // 500+ models available — users can type any Puter model ID
+  },
+
+  "cloudflare-ai": {
+    id: "cloudflare-ai",
+    alias: "cf",
+    format: "openai",
+    executor: "cloudflare-ai",
+    // URL is dynamic: uses accountId from credentials. The executor builds it.
+    baseUrl: "https://api.cloudflare.com/client/v4/accounts",
+    authType: "apikey",
+    authHeader: "bearer",
+    // 10K Neurons/day free: ~150 LLM responses or 500s Whisper audio — global edge
+    models: [
+      { id: "@cf/meta/llama-3.3-70b-instruct", name: "Llama 3.3 70B (🆓 ~150 resp/day)" },
+      { id: "@cf/meta/llama-3.1-8b-instruct", name: "Llama 3.1 8B (🆓)" },
+      { id: "@cf/google/gemma-3-12b-it", name: "Gemma 3 12B (🆓)" },
+      { id: "@cf/mistral/mistral-7b-instruct-v0.2-lora", name: "Mistral 7B (🆓)" },
+      { id: "@cf/qwen/qwen2.5-coder-15b-instruct", name: "Qwen 2.5 Coder 15B (🆓)" },
+      { id: "@cf/deepseek-ai/deepseek-r1-distill-qwen-32b", name: "DeepSeek R1 Distill 32B (🆓)" },
+    ],
+  },
+
+  scaleway: {
+    id: "scaleway",
+    alias: "scw",
+    format: "openai",
+    executor: "default",
+    baseUrl: "https://api.scaleway.ai/v1/chat/completions",
+    authType: "apikey",
+    authHeader: "bearer",
+    // 1M tokens free for new accounts — EU/GDPR (Paris), no credit card needed under limit
+    models: [
+      { id: "qwen3-235b-a22b-instruct-2507", name: "Qwen3 235B A22B (1M free tok 🆓)" },
+      { id: "llama-3.1-70b-instruct", name: "Llama 3.1 70B (🆓 EU)" },
+      { id: "llama-3.1-8b-instruct", name: "Llama 3.1 8B (🆓 EU)" },
+      { id: "mistral-small-3.2-24b-instruct-2506", name: "Mistral Small 3.2 (🆓 EU)" },
+      { id: "deepseek-v3-0324", name: "DeepSeek V3 (🆓 EU)" },
+      { id: "gpt-oss-120b", name: "GPT-OSS 120B (🆓 EU)" },
+    ],
+  },
+
+  aimlapi: {
+    id: "aimlapi",
+    alias: "aiml",
+    format: "openai",
+    executor: "default",
+    baseUrl: "https://api.aimlapi.com/v1/chat/completions",
+    authType: "apikey",
+    authHeader: "bearer",
+    // $0.025/day free credits — 200+ models via single aggregator endpoint
+    models: [
+      { id: "gpt-4o", name: "GPT-4o (via AI/ML API)" },
+      { id: "claude-3-5-sonnet-20241022", name: "Claude 3.5 Sonnet (via AI/ML API)" },
+      { id: "gemini-1.5-pro", name: "Gemini 1.5 Pro (via AI/ML API)" },
+      { id: "meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo", name: "Llama 3.1 70B (via AI/ML API)" },
+      { id: "deepseek-chat", name: "DeepSeek Chat (via AI/ML API)" },
+      { id: "mistral-large-latest", name: "Mistral Large (via AI/ML API)" },
+    ],
+    passthroughModels: true,
+  },
 };
 
 // ── Generator Functions ───────────────────────────────────────────────────

open-sse/executors/antigravity.ts

@@ -44,12 +44,28 @@ export class AntigravityExecutor extends BaseExecutor {
     // stale/wrong client-side values causing 404/403 from Cloud Code endpoints.
     // Opt-in escape hatch: set OMNIROUTE_ALLOW_BODY_PROJECT_OVERRIDE=1.
     const projectId =
-      allowBodyProjectOverride && bodyProjectId ? bodyProjectId : credentialsProjectId || bodyProjectId;
+      allowBodyProjectOverride && bodyProjectId
+        ? bodyProjectId
+        : credentialsProjectId || bodyProjectId;
 
     if (!projectId) {
-      throw new Error(
-        "Missing Google projectId for Antigravity account. Please reconnect OAuth so OmniRoute can fetch your real Cloud Code project (loadCodeAssist)."
-      );
+      // (#489) Return a structured error instead of throwing — gives the client a clear signal
+      // to show a "Reconnect OAuth" prompt rather than an opaque "Internal Server Error".
+      const errorMsg =
+        "Missing Google projectId for Antigravity account. Please reconnect OAuth in Providers → Antigravity so OmniRoute can fetch your Cloud Code project.";
+      const errorBody = {
+        error: {
+          message: errorMsg,
+          type: "oauth_missing_project_id",
+          code: "missing_project_id",
+        },
+      };
+      const resp = new Response(JSON.stringify(errorBody), {
+        status: 422,
+        headers: { "Content-Type": "application/json" },
+      });
+      // Returning a Response object signals the executor to stop and forward it
+      return resp as unknown as never;
     }
 
     // Fix contents for Claude models via Antigravity

open-sse/executors/base.ts

@@ -2,6 +2,20 @@ import { HTTP_STATUS, FETCH_TIMEOUT_MS } from "../config/constants.ts";
 import { applyFingerprint, isCliCompatEnabled } from "../config/cliFingerprints.ts";
 import { getRotatingApiKey } from "../services/apiKeyRotator.ts";
 
+/**
+ * Sanitizes a custom API path to prevent path traversal attacks.
+ * Valid paths must start with '/', contain no '..' segments,
+ * no null bytes, and be reasonable in length.
+ */
+function sanitizePath(path: string): boolean {
+  if (typeof path !== "string") return false;
+  if (!path.startsWith("/")) return false;
+  if (path.includes("\0")) return false; // null byte
+  if (path.includes("..")) return false; // path traversal
+  if (path.length > 512) return false; // sanity limit
+  return true;
+}
+
 type JsonRecord = Record<string, unknown>;
 
 export type ProviderConfig = {
@@ -103,7 +117,9 @@ export class BaseExecutor {
       const psd = credentials?.providerSpecificData;
       const baseUrl = typeof psd?.baseUrl === "string" ? psd.baseUrl : "https://api.openai.com/v1";
       const normalized = baseUrl.replace(/\/$/, "");
-      const customPath = typeof psd?.chatPath === "string" && psd.chatPath ? psd.chatPath : null;
+      // Sanitize custom path: must start with '/', no path traversal, no null bytes
+      const rawPath = typeof psd?.chatPath === "string" && psd.chatPath ? psd.chatPath : null;
+      const customPath = rawPath && sanitizePath(rawPath) ? rawPath : null;
       if (customPath) return `${normalized}${customPath}`;
       const path = this.provider.includes("responses") ? "/responses" : "/chat/completions";
       return `${normalized}${path}`;

open-sse/executors/cloudflare-ai.ts

@@ -0,0 +1,59 @@
+import { BaseExecutor } from "./base.ts";
+import { PROVIDERS } from "../config/constants.ts";
+
+/**
+ * CloudflareAIExecutor — handles dynamic URL construction with accountId.
+ * Cloudflare Workers AI uses the authenticated user's account ID in the URL.
+ *
+ * URL pattern: https://api.cloudflare.com/client/v4/accounts/{accountId}/ai/v1/chat/completions
+ * Auth: Bearer <API Token>
+ * Docs: https://developers.cloudflare.com/workers-ai/
+ *
+ * Free tier: 10,000 Neurons/day = ~150 LLM responses or 500s Whisper audio
+ * API Token: dash.cloudflare.com/profile/api-tokens
+ * Account ID: right sidebar of dash.cloudflare.com
+ */
+export class CloudflareAIExecutor extends BaseExecutor {
+  constructor() {
+    super("cloudflare-ai", PROVIDERS["cloudflare-ai"] || { format: "openai" });
+  }
+
+  buildUrl(_model: string, _stream: boolean, _urlIndex = 0, credentials: any = null): string {
+    // Account ID can be stored in providerSpecificData or at top level credentials
+    const accountId =
+      credentials?.providerSpecificData?.accountId ||
+      credentials?.accountId ||
+      process.env.CLOUDFLARE_ACCOUNT_ID;
+
+    if (!accountId) {
+      throw new Error(
+        "Cloudflare Workers AI requires an Account ID. " +
+          "Add it in provider settings under 'Account ID'. " +
+          "Find it at: https://dash.cloudflare.com (right sidebar)."
+      );
+    }
+
+    return `https://api.cloudflare.com/client/v4/accounts/${accountId}/ai/v1/chat/completions`;
+  }
+
+  buildHeaders(credentials: any, stream = true): Record<string, string> {
+    const headers: Record<string, string> = {
+      "Content-Type": "application/json",
+      Authorization: `Bearer ${credentials.apiKey || credentials.accessToken}`,
+    };
+
+    if (stream) {
+      headers["Accept"] = "text/event-stream";
+    }
+
+    return headers;
+  }
+
+  transformRequest(_model: string, body: any, _stream: boolean, _credentials: any): any {
+    // Cloudflare uses full model paths like @cf/meta/llama-3.3-70b-instruct
+    // No transformation needed — user sends the full Cloudflare model path.
+    return body;
+  }
+}
+
+export default CloudflareAIExecutor;

open-sse/executors/codex.ts

@@ -3,6 +3,112 @@ import { CODEX_DEFAULT_INSTRUCTIONS } from "../config/codexInstructions.ts";
 import { PROVIDERS } from "../config/constants.ts";
 import { refreshCodexToken } from "../services/tokenRefresh.ts";
 
+// ─── T09: Codex vs Spark Scope-Aware Rate Limiting ────────────────────────
+// Codex has two independent quota pools: "codex" (standard) and "spark" (premium).
+// Exhausting one should NOT block requests to the other.
+// Ref: sub2api PR #1129 (feat(openai): split codex spark rate limiting from codex)
+
+/**
+ * Maps model name substrings to their rate-limit scope.
+ * Checked in order — first match wins.
+ */
+const CODEX_SCOPE_PATTERNS: Array<{ pattern: string; scope: "codex" | "spark" }> = [
+  { pattern: "codex-spark", scope: "spark" },
+  { pattern: "spark", scope: "spark" },
+  { pattern: "codex", scope: "codex" },
+  { pattern: "gpt-5", scope: "codex" }, // gpt-5.2-codex, gpt-5.3-codex, etc.
+];
+
+/**
+ * T09: Determine the rate-limit scope for a Codex model.
+ * Use this key as the suffix for per-scope rate limit state:
+ *   `${accountId}:${getModelScope(model)}`
+ *
+ * @param model - The Codex model ID (e.g. "gpt-5.3-codex", "codex-spark-mini")
+ * @returns "codex" | "spark"
+ */
+export function getCodexModelScope(model: string): "codex" | "spark" {
+  const lower = model.toLowerCase();
+  for (const { pattern, scope } of CODEX_SCOPE_PATTERNS) {
+    if (lower.includes(pattern)) return scope;
+  }
+  return "codex"; // default scope
+}
+
+/**
+ * T09: Get the scope-keyed rate limit identifier for an account+model combination.
+ * Use this as the key for rateLimitState maps to ensure scope isolation.
+ */
+export function getCodexRateLimitKey(accountId: string, model: string): string {
+  return `${accountId}:${getCodexModelScope(model)}`;
+}
+
+/**
+ * T03: Parsed quota snapshot from Codex response headers.
+ * Codex includes per-account usage windows that allow precise reset scheduling.
+ * Ref: sub2api PR #357 (feat(oauth): persist usage snapshots and window cooldown)
+ */
+export interface CodexQuotaSnapshot {
+  usage5h: number; // tokens used in 5h window
+  limit5h: number; // token limit for 5h window
+  resetAt5h: string | null; // ISO timestamp when 5h window resets
+  usage7d: number; // tokens used in 7d window
+  limit7d: number; // token limit for 7d window
+  resetAt7d: string | null; // ISO timestamp when 7d window resets
+}
+
+/**
+ * T03: Parse Codex-specific quota headers from a provider response.
+ * Returns null if none of the relevant headers are present.
+ *
+ * Extracts:
+ *   x-codex-5h-usage / x-codex-5h-limit / x-codex-5h-reset-at
+ *   x-codex-7d-usage / x-codex-7d-limit / x-codex-7d-reset-at
+ */
+export function parseCodexQuotaHeaders(headers: Headers): CodexQuotaSnapshot | null {
+  const usage5h = headers.get("x-codex-5h-usage");
+  const limit5h = headers.get("x-codex-5h-limit");
+  const resetAt5h = headers.get("x-codex-5h-reset-at");
+  const usage7d = headers.get("x-codex-7d-usage");
+  const limit7d = headers.get("x-codex-7d-limit");
+  const resetAt7d = headers.get("x-codex-7d-reset-at");
+
+  // Return null if none of the quota headers are present (not a quota-aware response)
+  if (!usage5h && !limit5h && !resetAt5h && !usage7d && !limit7d && !resetAt7d) {
+    return null;
+  }
+
+  return {
+    usage5h: usage5h ? parseFloat(usage5h) : 0,
+    limit5h: limit5h ? parseFloat(limit5h) : Infinity,
+    resetAt5h: resetAt5h ?? null,
+    usage7d: usage7d ? parseFloat(usage7d) : 0,
+    limit7d: limit7d ? parseFloat(limit7d) : Infinity,
+    resetAt7d: resetAt7d ?? null,
+  };
+}
+
+/**
+ * T03: Get the soonest quota reset time from a CodexQuotaSnapshot.
+ * 7d window takes priority (wider window, harder limit) but we use whichever
+ * is further in the future to avoid releasing the block too early.
+ *
+ * @returns Unix timestamp (ms) of the soonest effective reset, or null
+ */
+export function getCodexResetTime(quota: CodexQuotaSnapshot): number | null {
+  const times: number[] = [];
+  if (quota.resetAt7d) {
+    const t = new Date(quota.resetAt7d).getTime();
+    if (!isNaN(t) && t > Date.now()) times.push(t);
+  }
+  if (quota.resetAt5h) {
+    const t = new Date(quota.resetAt5h).getTime();
+    if (!isNaN(t) && t > Date.now()) times.push(t);
+  }
+  if (times.length === 0) return null;
+  return Math.max(...times); // Use furthest-out reset to avoid premature unblock
+}
+
 // Ordered list of effort levels from lowest to highest
 const EFFORT_ORDER = ["none", "low", "medium", "high", "xhigh"] as const;
 type EffortLevel = (typeof EFFORT_ORDER)[number];

open-sse/executors/default.ts

@@ -80,18 +80,14 @@ export class DefaultExecutor extends BaseExecutor {
   }
 
   /**
-   * For compatible providers, ensure the model name sent upstream
-   * is the clean model name without internal routing prefixes.
-   * e.g. "openapi-chat-anti/claude-opus-4-6-thinking" → "claude-opus-4-6-thinking"
+   * For compatible providers, the model name is already clean by the time
+   * it reaches the executor (chatCore sets body.model = modelInfo.model,
+   * which is the parsed model ID without internal routing prefixes).
+   *
+   * Models may legitimately contain "/" as part of their ID (e.g. "zai-org/GLM-5-FP8",
+   * "org/model-name") — we must NOT strip path segments. (Fix #493)
    */
   transformRequest(model, body, stream, credentials) {
-    if (
-      this.provider?.startsWith?.("openai-compatible-") ||
-      this.provider?.startsWith?.("anthropic-compatible-")
-    ) {
-      const cleanModel = model.includes("/") ? model.split("/").slice(1).join("/") : model;
-      return { ...body, model: cleanModel };
-    }
     return body;
   }
 

open-sse/executors/github.ts

@@ -1,4 +1,4 @@
-import { BaseExecutor } from "./base.ts";
+import { BaseExecutor, ExecuteInput } from "./base.ts";
 import { PROVIDERS, OAUTH_ENDPOINTS } from "../config/constants.ts";
 import { getModelTargetFormat } from "../config/providerModels.ts";
 
@@ -19,15 +19,82 @@ export class GithubExecutor extends BaseExecutor {
     return this.config.baseUrl;
   }
 
+  injectResponseFormat(messages: any[], responseFormat: any) {
+    if (!responseFormat) return messages;
+
+    let formatInstruction = "";
+    if (responseFormat.type === "json_object") {
+      formatInstruction =
+        "Respond only with valid JSON. Do not include any text before or after the JSON object.";
+    } else if (responseFormat.type === "json_schema" && responseFormat.json_schema) {
+      formatInstruction = `Respond only with valid JSON matching this schema:\n${JSON.stringify(
+        responseFormat.json_schema.schema,
+        null,
+        2
+      )}\nDo not include any text before or after the JSON.`;
+    }
+
+    if (!formatInstruction) return messages;
+
+    const systemIdx = messages.findIndex((m: any) => m.role === "system");
+    if (systemIdx >= 0) {
+      return messages.map((m: any, i: number) =>
+        i === systemIdx ? { ...m, content: `${m.content}\n\n${formatInstruction}` } : m
+      );
+    }
+
+    return [{ role: "system", content: formatInstruction }, ...messages];
+  }
+
+  transformRequest(model: string, body: any, stream: boolean, credentials: any): any {
+    const modifiedBody = JSON.parse(JSON.stringify(body));
+    if (modifiedBody.response_format && model.toLowerCase().includes("claude")) {
+      modifiedBody.messages = this.injectResponseFormat(
+        modifiedBody.messages,
+        modifiedBody.response_format
+      );
+      delete modifiedBody.response_format;
+    }
+    return modifiedBody;
+  }
+
+  async execute(input: ExecuteInput) {
+    const result = await super.execute(input);
+    if (!result || !result.response?.body) return result;
+
+    const isStreaming = input.stream === true;
+    if (isStreaming) {
+      const decoder = new TextDecoder();
+      const transformStream = new TransformStream({
+        transform(chunk, controller) {
+          const text = decoder.decode(chunk, { stream: true });
+          if (text.includes("data: [DONE]")) {
+            return;
+          }
+          controller.enqueue(chunk);
+        },
+      });
+
+      const newResponse = new Response(result.response.body.pipeThrough(transformStream), {
+        status: result.response.status,
+        statusText: result.response.statusText,
+        headers: result.response.headers, // Headers class carries over correctly
+      });
+      result.response = newResponse;
+    }
+
+    return result;
+  }
+
   buildHeaders(credentials, stream = true) {
     const token = credentials.copilotToken || credentials.accessToken;
     return {
       Authorization: `Bearer ${token}`,
       "Content-Type": "application/json",
       "copilot-integration-id": "vscode-chat",
-      "editor-version": "vscode/1.107.1",
-      "editor-plugin-version": "copilot-chat/0.26.7",
-      "user-agent": "GitHubCopilotChat/0.26.7",
+      "editor-version": "vscode/1.110.0",
+      "editor-plugin-version": "copilot-chat/0.38.0",
+      "user-agent": "GitHubCopilotChat/0.38.0",
       "openai-intent": "conversation-panel",
       "x-github-api-version": "2025-04-01",
       "x-request-id":
@@ -44,7 +111,7 @@ export class GithubExecutor extends BaseExecutor {
         headers: {
           Authorization: `token ${githubAccessToken}`,
           "User-Agent": "GithubCopilot/1.0",
-          "Editor-Version": "vscode/1.100.0",
+          "Editor-Version": "vscode/1.110.0",
           "Editor-Plugin-Version": "copilot/1.300.0",
           Accept: "application/json",
         },

open-sse/executors/index.ts

@@ -6,6 +6,11 @@ import { KiroExecutor } from "./kiro.ts";
 import { CodexExecutor } from "./codex.ts";
 import { CursorExecutor } from "./cursor.ts";
 import { DefaultExecutor } from "./default.ts";
+import { PollinationsExecutor } from "./pollinations.ts";
+import { CloudflareAIExecutor } from "./cloudflare-ai.ts";
+import { OpencodeExecutor } from "./opencode.ts";
+import { PuterExecutor } from "./puter.ts";
+import { VertexExecutor } from "./vertex.ts";
 
 const executors = {
   antigravity: new AntigravityExecutor(),
@@ -16,6 +21,15 @@ const executors = {
   codex: new CodexExecutor(),
   cursor: new CursorExecutor(),
   cu: new CursorExecutor(), // Alias for cursor
+  pollinations: new PollinationsExecutor(),
+  pol: new PollinationsExecutor(), // Alias
+  "cloudflare-ai": new CloudflareAIExecutor(),
+  cf: new CloudflareAIExecutor(), // Alias
+  "opencode-zen": new OpencodeExecutor("opencode-zen"),
+  "opencode-go": new OpencodeExecutor("opencode-go"),
+  puter: new PuterExecutor(),
+  pu: new PuterExecutor(), // Alias
+  vertex: new VertexExecutor(),
 };
 
 const defaultCache = new Map();
@@ -39,3 +53,8 @@ export { KiroExecutor } from "./kiro.ts";
 export { CodexExecutor } from "./codex.ts";
 export { CursorExecutor } from "./cursor.ts";
 export { DefaultExecutor } from "./default.ts";
+export { PollinationsExecutor } from "./pollinations.ts";
+export { CloudflareAIExecutor } from "./cloudflare-ai.ts";
+export { OpencodeExecutor } from "./opencode.ts";
+export { PuterExecutor } from "./puter.ts";
+export { VertexExecutor } from "./vertex.ts";

open-sse/executors/opencode.ts

@@ -0,0 +1,61 @@
+import { BaseExecutor, type ExecuteInput, type ProviderCredentials } from "./base.ts";
+import { PROVIDERS } from "../config/constants.ts";
+import { getModelTargetFormat } from "../config/providerModels.ts";
+
+export class OpencodeExecutor extends BaseExecutor {
+  _requestFormat: string | null = null;
+
+  constructor(provider: string) {
+    super(provider, PROVIDERS[provider] || PROVIDERS.openai);
+  }
+
+  async execute(input: ExecuteInput) {
+    this._requestFormat = getModelTargetFormat(this.provider, input.model) || "openai";
+    try {
+      return await super.execute(input);
+    } finally {
+      this._requestFormat = null;
+    }
+  }
+
+  buildUrl(
+    model: string,
+    stream: boolean,
+    urlIndex = 0,
+    credentials: ProviderCredentials | null = null
+  ) {
+    void urlIndex;
+    void credentials;
+
+    const base = this.config.baseUrl;
+    switch (this._requestFormat) {
+      case "claude":
+        return `${base}/messages`;
+      case "openai-responses":
+        return `${base}/responses`;
+      case "gemini":
+        return `${base}/models/${model}:${stream ? "streamGenerateContent?alt=sse" : "generateContent"}`;
+      default:
+        return `${base}/chat/completions`;
+    }
+  }
+
+  buildHeaders(credentials: ProviderCredentials | null, stream = true) {
+    const headers: Record<string, string> = { "Content-Type": "application/json" };
+    const key = credentials?.apiKey || credentials?.accessToken;
+
+    if (key) {
+      headers["Authorization"] = `Bearer ${key}`;
+    }
+
+    if (this._requestFormat === "claude") {
+      headers["anthropic-version"] = "2023-06-01";
+    }
+
+    if (stream) {
+      headers["Accept"] = "text/event-stream";
+    }
+
+    return headers;
+  }
+}

open-sse/executors/pollinations.ts

@@ -0,0 +1,46 @@
+import { BaseExecutor } from "./base.ts";
+import { PROVIDERS } from "../config/constants.ts";
+
+/**
+ * PollinationsExecutor — handles optional API key auth.
+ * Pollinations AI works WITHOUT any API key for basic use (1 req/15s).
+ * If an API key is provided, higher rate limits apply.
+ *
+ * Endpoint: https://text.pollinations.ai/openai/chat/completions
+ * Docs: https://pollinations.ai/docs
+ */
+export class PollinationsExecutor extends BaseExecutor {
+  constructor() {
+    super("pollinations", PROVIDERS["pollinations"] || { format: "openai" });
+  }
+
+  buildUrl(_model: string, _stream: boolean, _urlIndex = 0, _credentials = null): string {
+    return "https://text.pollinations.ai/openai/chat/completions";
+  }
+
+  buildHeaders(credentials: any, stream = true): Record<string, string> {
+    const headers: Record<string, string> = {
+      "Content-Type": "application/json",
+    };
+
+    // API key is OPTIONAL — skip Authorization header if no key provided
+    const key = credentials?.apiKey || credentials?.accessToken;
+    if (key) {
+      headers["Authorization"] = `Bearer ${key}`;
+    }
+
+    if (stream) {
+      headers["Accept"] = "text/event-stream";
+    }
+
+    return headers;
+  }
+
+  transformRequest(model: string, body: any, _stream: boolean, _credentials: any): any {
+    // Pollinations uses model names directly like "openai", "claude", "deepseek", etc.
+    // No transformation needed — the model name is already the Pollinations alias.
+    return body;
+  }
+}
+
+export default PollinationsExecutor;

open-sse/executors/puter.ts

@@ -0,0 +1,59 @@
+import { BaseExecutor } from "./base.ts";
+import { PROVIDERS } from "../config/constants.ts";
+
+/**
+ * PuterExecutor — OpenAI-compatible proxy for Puter AI.
+ *
+ * Puter exposes 500+ models (GPT, Claude, Gemini, Grok, DeepSeek, Qwen, Mistral...)
+ * through a single OpenAI-compatible REST endpoint.
+ *
+ * Endpoint: https://api.puter.com/puterai/openai/v1/chat/completions
+ * Auth:     Bearer <puter_auth_token>  (from puter.com/dashboard → Copy Auth Token)
+ * Docs:     https://docs.puter.com/AI/
+ *
+ * Model ID examples:
+ *   OpenAI:   "gpt-4o-mini", "gpt-4o", "gpt-4.1"
+ *   Claude:   "claude-sonnet-4-5", "claude-opus-4", "claude-haiku-4-5"
+ *   Gemini:   "google/gemini-2.0-flash", "google/gemini-2.5-pro"
+ *   DeepSeek: "deepseek/deepseek-chat", "deepseek/deepseek-r1"
+ *   Grok:     "x-ai/grok-3", "x-ai/grok-4"
+ *   Mistral:  "mistralai/mistral-small-3.2"
+ *   Meta:     "meta-llama/llama-3.3-70b-instruct"
+ *
+ * Note: Image generation, TTS, STT, and video are puter.js SDK-only features.
+ * Only text chat completions (with streaming SSE) are available via REST.
+ */
+export class PuterExecutor extends BaseExecutor {
+  constructor() {
+    super("puter", PROVIDERS["puter"] || { format: "openai" });
+  }
+
+  buildUrl(_model: string, _stream: boolean, _urlIndex = 0, _credentials = null): string {
+    return "https://api.puter.com/puterai/openai/v1/chat/completions";
+  }
+
+  buildHeaders(credentials: any, stream = true): Record<string, string> {
+    const headers: Record<string, string> = {
+      "Content-Type": "application/json",
+    };
+
+    const key = credentials?.apiKey || credentials?.accessToken;
+    if (key) {
+      headers["Authorization"] = `Bearer ${key}`;
+    }
+
+    if (stream) {
+      headers["Accept"] = "text/event-stream";
+    }
+
+    return headers;
+  }
+
+  transformRequest(model: string, body: any, _stream: boolean, _credentials: any): any {
+    // Puter accepts model IDs directly from its catalog.
+    // No transformation required — model string is passed as-is.
+    return body;
+  }
+}
+
+export default PuterExecutor;

open-sse/executors/vertex.ts

@@ -0,0 +1,150 @@
+import { SignJWT, importPKCS8 } from "jose";
+import { BaseExecutor, ExecuteInput } from "./base.ts";
+import { PROVIDERS } from "../config/constants.ts";
+
+interface ServiceAccount {
+  type: string;
+  project_id: string;
+  private_key_id: string;
+  private_key: string;
+  client_email: string;
+  [key: string]: unknown;
+}
+
+const TOKEN_CACHE = new Map<string, { token: string; expiresAt: number }>();
+
+function parseSAFromApiKey(apiKey: string): ServiceAccount {
+  try {
+    return JSON.parse(apiKey);
+  } catch {
+    throw new Error("Vertex AI requires a valid Service Account JSON as the API key");
+  }
+}
+
+async function getAccessToken(sa: ServiceAccount): Promise<string> {
+  if (!sa.client_email || !sa.private_key) {
+    throw new Error(
+      "Service Account JSON is missing required fields (client_email or private_key)"
+    );
+  }
+
+  const cacheKey = sa.client_email;
+  const cached = TOKEN_CACHE.get(cacheKey);
+
+  // Buffer of 60 seconds
+  if (cached && Date.now() < cached.expiresAt - 60_000) {
+    return cached.token;
+  }
+
+  const privateKey = await importPKCS8(sa.private_key, "RS256");
+  const now = Math.floor(Date.now() / 1000);
+
+  const jwt = await new SignJWT({
+    iss: sa.client_email,
+    sub: sa.client_email,
+    aud: "https://oauth2.googleapis.com/token",
+    iat: now,
+    exp: now + 3600,
+    scope: "https://www.googleapis.com/auth/cloud-platform",
+  })
+    .setProtectedHeader({ alg: "RS256", kid: sa.private_key_id })
+    .sign(privateKey);
+
+  const tokenRes = await fetch("https://oauth2.googleapis.com/token", {
+    method: "POST",
+    headers: { "Content-Type": "application/x-www-form-urlencoded" },
+    body: new URLSearchParams({
+      grant_type: "urn:ietf:params:oauth:grant-type:jwt-bearer",
+      assertion: jwt,
+    }),
+  });
+
+  if (!tokenRes.ok) {
+    const errorText = await tokenRes.text();
+    throw new Error(
+      `Failed to exchange JWT for Vertex access token: ${tokenRes.status} ${errorText}`
+    );
+  }
+
+  const tokenData = await tokenRes.json();
+  const accessToken = tokenData.access_token;
+
+  if (!accessToken) {
+    throw new Error("Vertex AI token exchange succeeded but no access_token found");
+  }
+
+  TOKEN_CACHE.set(cacheKey, {
+    token: accessToken,
+    expiresAt: (now + 3600) * 1000,
+  });
+
+  return accessToken;
+}
+
+const PARTNER_MODELS = new Set([
+  "claude-3-5-sonnet",
+  "claude-3-opus",
+  "claude-3-haiku",
+  "deepseek-v3",
+  "deepseek-v3.2",
+  "deepseek-deepseek-r1",
+  "qwen3-next-80b",
+  "llama-3.1",
+  "mistral-",
+  "glm-5",
+  "meta/llama",
+]);
+
+function isPartnerModel(model: string) {
+  return [...PARTNER_MODELS].some((prefix) => model.startsWith(prefix));
+}
+
+export class VertexExecutor extends BaseExecutor {
+  constructor() {
+    super("vertex", PROVIDERS.vertex);
+  }
+
+  async execute(input: ExecuteInput) {
+    const { credentials, log } = input;
+    if (credentials.apiKey && !credentials.accessToken) {
+      try {
+        const sa = parseSAFromApiKey(credentials.apiKey);
+        credentials.accessToken = await getAccessToken(sa);
+      } catch (err: any) {
+        log?.error?.("VERTEX", `Failed to generate JWT token: ${err.message}`);
+        throw err;
+      }
+    }
+    return super.execute(input);
+  }
+
+  buildUrl(model: string, stream: boolean, urlIndex = 0, credentials: any = null) {
+    const region = credentials?.providerSpecificData?.region || "us-central1";
+    let project = "unknown-project";
+
+    if (credentials?.apiKey) {
+      try {
+        const sa = parseSAFromApiKey(credentials.apiKey);
+        if (sa.project_id) project = sa.project_id;
+      } catch {
+        // Ignored, handled in execute
+      }
+    }
+
+    if (isPartnerModel(model)) {
+      return `https://aiplatform.googleapis.com/v1/projects/${project}/locations/global/endpoints/openapi/chat/completions`;
+    }
+    return `https://aiplatform.googleapis.com/v1/projects/${project}/locations/${region}/publishers/google/models/${model}:${stream ? "streamGenerateContent?alt=sse" : "generateContent"}`;
+  }
+
+  buildHeaders(credentials: any, stream = true) {
+    const headers: Record<string, string> = { "Content-Type": "application/json" };
+    if (credentials.accessToken) {
+      headers["Authorization"] = `Bearer ${credentials.accessToken}`;
+    }
+    if (stream) {
+      headers["Accept"] = "text/event-stream";
+    }
+    return headers;
+  }
+}

open-sse/handlers/audioSpeech.ts

@@ -28,13 +28,17 @@ function upstreamErrorResponse(res, errText) {
   let errorMessage: string;
   try {
     const parsed = JSON.parse(errText);
-    errorMessage =
+    // Extract a human-readable message from various error response shapes.
+    // Guard against `parsed.error` being an object (e.g. ElevenLabs returns
+    // { error: { message: "...", status_code: 401 } } or { detail: { ... } })
+    const raw =
       parsed?.err_msg ||
       parsed?.error?.message ||
-      parsed?.error ||
+      (typeof parsed?.error === "string" ? parsed.error : null) ||
       parsed?.message ||
-      parsed?.detail ||
-      errText;
+      (typeof parsed?.detail === "string" ? parsed.detail : parsed?.detail?.message) ||
+      null;
+    errorMessage = raw ? String(raw) : errText || `Upstream error (${res.status})`;
   } catch {
     errorMessage = errText || `Upstream error (${res.status})`;
   }

open-sse/handlers/audioTranscription.ts

@@ -34,13 +34,15 @@ function upstreamErrorResponse(res, errText) {
   let errorMessage: string;
   try {
     const parsed = JSON.parse(errText);
-    errorMessage =
+    // Guard against `parsed.error` or `parsed.detail` being objects
+    const raw =
       parsed?.err_msg ||
       parsed?.error?.message ||
-      parsed?.error ||
+      (typeof parsed?.error === "string" ? parsed.error : null) ||
       parsed?.message ||
-      parsed?.detail ||
-      errText;
+      (typeof parsed?.detail === "string" ? parsed.detail : parsed?.detail?.message) ||
+      null;
+    errorMessage = raw ? String(raw) : errText || `Upstream error (${res.status})`;
   } catch {
     errorMessage = errText || `Upstream error (${res.status})`;
   }
@@ -65,13 +67,67 @@ function getUploadedFileName(file: Blob & { name?: unknown }): string {
   return typeof file.name === "string" && file.name.length > 0 ? file.name : "audio.wav";
 }
 
+/**
+ * Infer a suitable Content-Type for Deepgram from the browser-provided MIME
+ * type and the original filename.  Deepgram accepts `audio/*` and many raw
+ * formats, but `video/*` causes it to silently fail with "no speech detected".
+ *
+ * Strategy:
+ * 1. If the browser says `audio/*`, keep it as-is.
+ * 2. If it's `video/*` (e.g. `.mp4`), remap to the audio equivalent so
+ *    Deepgram extracts the audio track.  `.mp4` → `audio/mp4`, etc.
+ * 3. Fall back to `application/octet-stream` which tells Deepgram to
+ *    auto-detect from the raw bytes (most reliable for unknown formats).
+ */
+function resolveAudioContentType(file: Blob & { name?: unknown }): string {
+  const browserType = (file.type || "").toLowerCase();
+  const fileName = typeof file.name === "string" ? file.name.toLowerCase() : "";
+
+  // 1) Browser already says it's audio — trust it
+  if (browserType.startsWith("audio/")) return browserType;
+
+  // 2) Derive from file extension (covers video/* and empty MIME)
+  const ext = fileName.includes(".") ? fileName.split(".").pop() : "";
+  const EXT_TO_MIME: Record<string, string> = {
+    mp3: "audio/mpeg",
+    mp4: "audio/mp4",
+    m4a: "audio/mp4",
+    wav: "audio/wav",
+    ogg: "audio/ogg",
+    flac: "audio/flac",
+    webm: "audio/webm",
+    aac: "audio/aac",
+    wma: "audio/x-ms-wma",
+    opus: "audio/opus",
+  };
+  if (ext && EXT_TO_MIME[ext]) return EXT_TO_MIME[ext];
+
+  // 3) Fallback — let Deepgram auto-detect from raw bytes
+  return "application/octet-stream";
+}
+
 /**
  * Handle Deepgram transcription (raw binary audio, model via query param)
  */
-async function handleDeepgramTranscription(providerConfig, file, modelId, token) {
+async function handleDeepgramTranscription(
+  providerConfig,
+  file,
+  modelId,
+  token,
+  formData?: FormData
+) {
   const url = new URL(providerConfig.baseUrl);
   url.searchParams.set("model", modelId);
   url.searchParams.set("smart_format", "true");
+  url.searchParams.set("punctuate", "true");
+
+  // Language: if caller specified one, use it; otherwise let Deepgram auto-detect
+  const langParam = formData?.get("language");
+  if (typeof langParam === "string" && langParam.trim()) {
+    url.searchParams.set("language", langParam.trim());
+  } else {
+    url.searchParams.set("detect_language", "true");
+  }
 
   const arrayBuffer = await file.arrayBuffer();
 
@@ -79,7 +135,7 @@ async function handleDeepgramTranscription(providerConfig, file, modelId, token)
     method: "POST",
     headers: {
       ...buildAuthHeaders(providerConfig, token),
-      "Content-Type": file.type || "audio/wav",
+      "Content-Type": resolveAudioContentType(file),
     },
     body: arrayBuffer,
   });
@@ -212,7 +268,7 @@ async function handleHuggingFaceTranscription(providerConfig, file, modelId, tok
     method: "POST",
     headers: {
       ...buildAuthHeaders(providerConfig, token),
-      "Content-Type": file.type || "audio/wav",
+      "Content-Type": resolveAudioContentType(file),
     },
     body: arrayBuffer,
   });
@@ -283,7 +339,7 @@ export async function handleAudioTranscription({
 
   // Route to provider-specific handler
   if (providerConfig.format === "deepgram") {
-    return handleDeepgramTranscription(providerConfig, file, modelId, token);
+    return handleDeepgramTranscription(providerConfig, file, modelId, token, formData);
   }
 
   if (providerConfig.format === "assemblyai") {

open-sse/handlers/chatCore.ts

@@ -16,6 +16,9 @@ import { resolveModelAlias } from "../services/modelDeprecation.ts";
 import { getUnsupportedParams } from "../config/providerRegistry.ts";
 import { createErrorResult, parseUpstreamError, formatProviderError } from "../utils/error.ts";
 import { HTTP_STATUS } from "../config/constants.ts";
+import { classifyProviderError, PROVIDER_ERROR_TYPES } from "../services/errorClassifier.ts";
+import { updateProviderConnection } from "@/lib/db/providers";
+import { logAuditEvent } from "@/lib/compliance";
 import { handleBypassRequest } from "../utils/bypassHandler.ts";
 import {
   saveRequestUsage,
@@ -23,8 +26,13 @@ import {
   appendRequestLog,
   saveCallLog,
 } from "@/lib/usageDb";
-import { getModelNormalizeToolCallId } from "@/lib/db/models";
+import { getModelNormalizeToolCallId, getModelPreserveOpenAIDeveloperRole } from "@/lib/localDb";
 import { getExecutor } from "../executors/index.ts";
+import {
+  parseCodexQuotaHeaders,
+  getCodexResetTime,
+  getCodexModelScope,
+} from "../executors/codex.ts";
 import { translateNonStreamingResponse } from "./responseTranslator.ts";
 import { extractUsageFromResponse } from "./usageExtractor.ts";
 import { parseSSEToOpenAIResponse, parseSSEToResponsesOutput } from "./sseParser.ts";
@@ -44,11 +52,17 @@ import { getIdempotencyKey, checkIdempotency, saveIdempotency } from "@/lib/idem
 import { createProgressTransform, wantsProgress } from "../utils/progressTracker.ts";
 import { isModelUnavailableError, getNextFamilyFallback } from "../services/modelFamilyFallback.ts";
 import { computeRequestHash, deduplicate, shouldDeduplicate } from "../services/requestDedup.ts";
+import {
+  getBackgroundTaskReason,
+  getDegradedModel,
+  getBackgroundDegradationConfig,
+} from "../services/backgroundTaskDetector.ts";
 import {
   shouldUseFallback,
   isFallbackDecision,
   EMERGENCY_FALLBACK_CONFIG,
 } from "../services/emergencyFallback.ts";
+import { resolveStreamFlag, stripMarkdownCodeFence } from "../utils/aiSdkCompat.ts";
 
 export function shouldUseNativeCodexPassthrough({
   provider,
@@ -93,7 +107,9 @@ export async function handleChatCore({
   userAgent,
   comboName,
 }) {
-  const { provider, model, extendedContext } = modelInfo;
+  let { provider, model, extendedContext } = modelInfo;
+  const requestedModel =
+    typeof body?.model === "string" && body.model.trim().length > 0 ? body.model : model;
   const startTime = Date.now();
   const persistFailureUsage = (statusCode: number, errorCode?: string | null) => {
     saveRequestUsage({
@@ -112,6 +128,67 @@ export async function handleChatCore({
     }).catch(() => {});
   };
 
+  const persistCodexQuotaState = async (
+    headers: Headers | Record<string, string> | null,
+    status = 0
+  ) => {
+    if (provider !== "codex" || !connectionId || !headers) return;
+
+    try {
+      const quota = parseCodexQuotaHeaders(headers as Headers);
+      if (!quota) return;
+
+      const existingProviderData =
+        credentials?.providerSpecificData && typeof credentials.providerSpecificData === "object"
+          ? credentials.providerSpecificData
+          : {};
+      const scope = getCodexModelScope(model || requestedModel || "");
+      const quotaState = {
+        usage5h: quota.usage5h,
+        limit5h: quota.limit5h,
+        resetAt5h: quota.resetAt5h,
+        usage7d: quota.usage7d,
+        limit7d: quota.limit7d,
+        resetAt7d: quota.resetAt7d,
+        scope,
+        updatedAt: new Date().toISOString(),
+      };
+
+      const nextProviderData: Record<string, unknown> = {
+        ...existingProviderData,
+        codexQuotaState: quotaState,
+      };
+
+      // T03/T09: on 429, persist exact reset time per scope to avoid global over-blocking.
+      if (status === 429) {
+        const resetTimeMs = getCodexResetTime(quota);
+        if (resetTimeMs && resetTimeMs > Date.now()) {
+          const scopeUntil = new Date(resetTimeMs).toISOString();
+          const scopeMapRaw =
+            existingProviderData &&
+            typeof existingProviderData === "object" &&
+            existingProviderData.codexScopeRateLimitedUntil &&
+            typeof existingProviderData.codexScopeRateLimitedUntil === "object"
+              ? existingProviderData.codexScopeRateLimitedUntil
+              : {};
+
+          nextProviderData.codexScopeRateLimitedUntil = {
+            ...(scopeMapRaw as Record<string, unknown>),
+            [scope]: scopeUntil,
+          };
+        }
+      }
+
+      await updateProviderConnection(connectionId, {
+        providerSpecificData: nextProviderData,
+      });
+
+      credentials.providerSpecificData = nextProviderData;
+    } catch (err) {
+      log?.debug?.("CODEX", `Failed to persist codex quota state: ${err?.message || err}`);
+    }
+  };
+
   // ── Phase 9.2: Idempotency check ──
   const idempotencyKey = getIdempotencyKey(clientRawRequest?.headers);
   const cachedIdemp = checkIdempotency(idempotencyKey);
@@ -157,6 +234,37 @@ export async function handleChatCore({
   // Detect source format and get target format
   // Model-specific targetFormat takes priority over provider default
 
+  // ── Background Task Redirection (T41) ──
+  const bgConfig = getBackgroundDegradationConfig();
+  const backgroundReason = bgConfig.enabled
+    ? getBackgroundTaskReason(body, clientRawRequest?.headers)
+    : null;
+  if (backgroundReason) {
+    const degradedModel = getDegradedModel(model);
+    if (degradedModel !== model) {
+      const originalModel = model;
+      log?.info?.(
+        "BACKGROUND",
+        `Background task redirect (${backgroundReason}): ${originalModel} → ${degradedModel}`
+      );
+      model = degradedModel;
+      if (body && typeof body === "object") {
+        body.model = model;
+      }
+
+      logAuditEvent({
+        action: "routing.background_task_redirect",
+        actor: apiKeyInfo?.name || "system",
+        target: connectionId || provider || "chat",
+        details: {
+          original_model: originalModel,
+          redirected_to: degradedModel,
+          reason: backgroundReason,
+        },
+      });
+    }
+  }
+
   // Apply custom model aliases (Settings → Model Aliases → Pattern→Target) before routing (#315, #472)
   // Custom aliases take priority over built-in and must be resolved here so the
   // downstream getModelTargetFormat() lookup AND the actual provider request use
@@ -173,7 +281,12 @@ export async function handleChatCore({
   const targetFormat = modelTargetFormat || getTargetFormat(provider);
 
   // Default to false unless client explicitly sets stream: true (OpenAI spec compliant)
-  const stream = body.stream === true;
+  const acceptHeader =
+    clientRawRequest?.headers && typeof clientRawRequest.headers.get === "function"
+      ? clientRawRequest.headers.get("accept") || clientRawRequest.headers.get("Accept")
+      : (clientRawRequest?.headers || {})["accept"] || (clientRawRequest?.headers || {})["Accept"];
+
+  const stream = resolveStreamFlag(body?.stream, acceptHeader);
 
   // ── Phase 9.1: Semantic cache check (non-streaming, temp=0 only) ──
   if (isCacheable(body, clientRawRequest?.headers)) {
@@ -308,6 +421,27 @@ export async function handleChatCore({
                   }
                   return [];
                 }
+                // (#527) tool_result → convert to text instead of dropping.
+                // When Claude Code + superpowers routes through Codex, it sends tool_result
+                // blocks in user messages. Silently dropping them causes Codex to loop
+                // because it never receives the tool response and keeps re-requesting it.
+                if (block.type === "tool_result") {
+                  const toolId = block.tool_use_id ?? block.id ?? "unknown";
+                  const resultContent = block.content ?? block.text ?? block.output ?? "";
+                  const resultText =
+                    typeof resultContent === "string"
+                      ? resultContent
+                      : Array.isArray(resultContent)
+                        ? resultContent
+                            .filter((c: Record<string, unknown>) => c.type === "text")
+                            .map((c: Record<string, unknown>) => c.text)
+                            .join("\n")
+                        : JSON.stringify(resultContent);
+                  if (resultText.length > 0) {
+                    return [{ type: "text", text: `[Tool Result: ${toolId}]\n${resultText}` }];
+                  }
+                  return [];
+                }
                 // Unknown types: drop silently
                 log?.debug?.("CONTENT", `Dropped unsupported content part type="${block.type}"`);
                 return [];
@@ -317,7 +451,16 @@ export async function handleChatCore({
         }
       }
 
-      const normalizeToolCallId = getModelNormalizeToolCallId(provider || "", model || "");
+      const normalizeToolCallId = getModelNormalizeToolCallId(
+        provider || "",
+        model || "",
+        sourceFormat
+      );
+      const preserveDeveloperRole = getModelPreserveOpenAIDeveloperRole(
+        provider || "",
+        model || "",
+        sourceFormat
+      );
       translatedBody = translateRequest(
         sourceFormat,
         targetFormat,
@@ -327,7 +470,7 @@ export async function handleChatCore({
         credentials,
         provider,
         reqLogger,
-        { normalizeToolCallId }
+        { normalizeToolCallId, preserveDeveloperRole }
       );
     }
   } catch (error) {
@@ -427,7 +570,7 @@ export async function handleChatCore({
       // Non-stream responses need cloning for shared dedup consumers.
       const status = rawResult.response.status;
       const statusText = rawResult.response.statusText;
-      const headers = Array.from(rawResult.response.headers.entries());
+      const headers = Array.from(rawResult.response.headers.entries()) as [string, string][];
       const payload = await rawResult.response.text();
 
       return {
@@ -502,6 +645,7 @@ export async function handleChatCore({
       path: clientRawRequest?.endpoint || "/v1/chat/completions",
       status: error.name === "AbortError" ? 499 : HTTP_STATUS.BAD_GATEWAY,
       model,
+      requestedModel,
       provider,
       connectionId,
       duration: Date.now() - startTime,
@@ -573,6 +717,8 @@ export async function handleChatCore({
     }
   }
 
+  await persistCodexQuotaState(providerResponse.headers, providerResponse.status);
+
   // Check provider response - return error info for fallback handling
   if (!providerResponse.ok) {
     trackPendingRequest(model, provider, connectionId, false);
@@ -580,6 +726,54 @@ export async function handleChatCore({
       providerResponse,
       provider
     );
+
+    // T06/T10/T36: classify provider errors and persist terminal account states.
+    const errorType = classifyProviderError(statusCode, message);
+    if (connectionId && errorType) {
+      try {
+        if (errorType === PROVIDER_ERROR_TYPES.FORBIDDEN) {
+          await updateProviderConnection(connectionId, {
+            isActive: false,
+            testStatus: "banned",
+            lastErrorType: errorType,
+            lastError: message,
+            errorCode: statusCode,
+          });
+          console.warn(
+            `[provider] Node ${connectionId} banned (${statusCode}) — disabling permanently`
+          );
+        } else if (errorType === PROVIDER_ERROR_TYPES.QUOTA_EXHAUSTED) {
+          await updateProviderConnection(connectionId, {
+            testStatus: "credits_exhausted",
+            lastErrorType: errorType,
+            lastError: message,
+            errorCode: statusCode,
+          });
+          console.warn(`[provider] Node ${connectionId} exhausted quota (${statusCode})`);
+        } else if (errorType === PROVIDER_ERROR_TYPES.ACCOUNT_DEACTIVATED) {
+          await updateProviderConnection(connectionId, {
+            isActive: false,
+            testStatus: "expired",
+            lastErrorType: errorType,
+            lastError: message,
+            errorCode: statusCode,
+          });
+          console.warn(
+            `[provider] Node ${connectionId} account deactivated (${statusCode}) — marked expired`
+          );
+        } else if (errorType === PROVIDER_ERROR_TYPES.UNAUTHORIZED) {
+          // Normal 401 (token/session auth issue): keep account active for refresh/re-auth.
+          await updateProviderConnection(connectionId, {
+            lastErrorType: errorType,
+            lastError: message,
+            errorCode: statusCode,
+          });
+        }
+      } catch {
+        // Best-effort state update; request flow should continue with fallback handling.
+      }
+    }
+
     appendRequestLog({ model, provider, connectionId, status: `FAILED ${statusCode}` }).catch(
       () => {}
     );
@@ -588,6 +782,7 @@ export async function handleChatCore({
       path: clientRawRequest?.endpoint || "/v1/chat/completions",
       status: statusCode,
       model,
+      requestedModel,
       provider,
       connectionId,
       duration: Date.now() - startTime,
@@ -778,6 +973,7 @@ export async function handleChatCore({
       path: clientRawRequest?.endpoint || "/v1/chat/completions",
       status: 200,
       model,
+      requestedModel,
       provider,
       connectionId,
       duration: Date.now() - startTime,
@@ -818,6 +1014,28 @@ export async function handleChatCore({
       ? translateNonStreamingResponse(responseBody, targetFormat, sourceFormat)
       : responseBody;
 
+    // T26: Strip markdown code blocks if provider format is Claude
+    if (sourceFormat === "claude" && !stream) {
+      if (typeof translatedResponse?.choices?.[0]?.message?.content === "string") {
+        translatedResponse.choices[0].message.content = stripMarkdownCodeFence(
+          translatedResponse.choices[0].message.content
+        ) as string;
+      }
+    }
+
+    // T18: Normalize finish_reason to 'tool_calls' if tool calls are present
+    if (translatedResponse?.choices) {
+      for (const choice of translatedResponse.choices) {
+        if (
+          choice.message?.tool_calls &&
+          choice.message.tool_calls.length > 0 &&
+          choice.finish_reason !== "tool_calls"
+        ) {
+          choice.finish_reason = "tool_calls";
+        }
+      }
+    }
+
     // Sanitize response for OpenAI SDK compatibility
     // Strips non-standard fields (x_groq, usage_breakdown, service_tier, etc.)
     // Extracts <think> tags into reasoning_content
@@ -891,6 +1109,7 @@ export async function handleChatCore({
       path: clientRawRequest?.endpoint || "/v1/chat/completions",
       status: streamStatus || 200,
       model,
+      requestedModel,
       provider,
       connectionId,
       duration: Date.now() - startTime,

open-sse/handlers/imageGeneration.ts

@@ -16,6 +16,7 @@
  */
 
 import { getImageProvider, parseImageModel } from "../config/imageRegistry.ts";
+import { mapImageSize } from "../translator/image/sizeMapper.ts";
 import { saveCallLog } from "@/lib/usageDb";
 import {
   submitComfyWorkflow,
@@ -95,11 +96,21 @@ export async function handleImageGeneration({ body, credentials, log, resolvedPr
     });
   }
 
-  // Route to format-specific handler
   if (providerConfig.format === "gemini-image") {
     return handleGeminiImageGeneration({ model, providerConfig, body, credentials, log });
   }
 
+  if (providerConfig.format === "imagen3") {
+    return handleImagen3ImageGeneration({
+      model,
+      provider,
+      providerConfig,
+      body,
+      credentials,
+      log,
+    });
+  }
+
   if (providerConfig.format === "hyperbolic") {
     return handleHyperbolicImageGeneration({
       model,
@@ -539,7 +550,7 @@ async function handleNanoBananaImageGeneration({
       ? body.aspectRatio
       : typeof body.aspect_ratio === "string"
         ? body.aspect_ratio
-        : inferAspectRatioFromSize(body.size) || "1:1";
+        : mapImageSize(body.size);
 
   let resolution =
     typeof body.resolution === "string"
@@ -856,18 +867,6 @@ async function normalizeNanoBananaTaskResult(taskData, body, log) {
   return [];
 }
 
-function inferAspectRatioFromSize(size) {
-  if (typeof size !== "string") return null;
-  const [wRaw, hRaw] = size.split("x");
-  const width = Number(wRaw);
-  const height = Number(hRaw);
-  if (!Number.isFinite(width) || !Number.isFinite(height) || width <= 0 || height <= 0) return null;
-
-  const gcd = (a, b) => (b === 0 ? a : gcd(b, a % b));
-  const div = gcd(Math.round(width), Math.round(height));
-  return `${Math.round(width / div)}:${Math.round(height / div)}`;
-}
-
 function inferResolutionFromSize(size) {
   if (typeof size !== "string") return null;
   const [wRaw, hRaw] = size.split("x");
@@ -1081,3 +1080,113 @@ async function handleComfyUIImageGeneration({ model, provider, providerConfig, b
     return { success: false, status: 502, error: `Image provider error: ${err.message}` };
   }
 }
+
+/**
+ * Handle Imagen 3 image generation
+ */
+async function handleImagen3ImageGeneration({
+  model,
+  provider,
+  providerConfig,
+  body,
+  credentials,
+  log,
+}: any) {
+  const startTime = Date.now();
+  const token = credentials.apiKey || credentials.accessToken;
+  const aspectRatio = mapImageSize(body.size);
+
+  const upstreamBody = {
+    prompt: body.prompt,
+    aspect_ratio: aspectRatio,
+    number_of_images: body.n ?? 1,
+  };
+
+  if (log) {
+    const promptPreview = String(body.prompt ?? "").slice(0, 60);
+    log.info(
+      "IMAGE",
+      `${provider}/${model} (imagen3) | prompt: "${promptPreview}..." | aspect_ratio: ${aspectRatio}`
+    );
+  }
+
+  try {
+    const response = await fetch(providerConfig.baseUrl, {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+        Authorization: `Bearer ${token}`,
+      },
+      body: JSON.stringify(upstreamBody),
+    });
+
+    if (!response.ok) {
+      const errorText = await response.text();
+      if (log)
+        log.error("IMAGE", `${provider} error ${response.status}: ${errorText.slice(0, 200)}`);
+
+      saveCallLog({
+        method: "POST",
+        path: "/v1/images/generations",
+        status: response.status,
+        model: `${provider}/${model}`,
+        provider,
+        duration: Date.now() - startTime,
+        error: errorText.slice(0, 500),
+        requestBody: upstreamBody,
+      }).catch(() => {});
+
+      return { success: false, status: response.status, error: errorText };
+    }
+
+    const data = await response.json();
+
+    // Normalize response to OpenAI format
+    const images: any[] = [];
+    if (Array.isArray(data.images)) {
+      images.push(
+        ...data.images.map((img: any) => ({
+          b64_json: img.image || img.b64_json || img.url || img,
+          revised_prompt: body.prompt,
+        }))
+      );
+    } else if (Array.isArray(data.data)) {
+      images.push(...data.data);
+    } else if (data.url || data.b64_json || data.image) {
+      images.push({
+        b64_json: data.image || data.b64_json || data.url,
+        url: data.url,
+        revised_prompt: body.prompt,
+      });
+    }
+
+    saveCallLog({
+      method: "POST",
+      path: "/v1/images/generations",
+      status: 200,
+      model: `${provider}/${model}`,
+      provider,
+      duration: Date.now() - startTime,
+      responseBody: { images_count: images.length },
+    }).catch(() => {});
+
+    return {
+      success: true,
+      data: { created: data.created || Math.floor(Date.now() / 1000), data: images },
+    };
+  } catch (err: any) {
+    if (log) log.error("IMAGE", `${provider} fetch error: ${err.message}`);
+
+    saveCallLog({
+      method: "POST",
+      path: "/v1/images/generations",
+      status: 502,
+      model: `${provider}/${model}`,
+      provider,
+      duration: Date.now() - startTime,
+      error: err.message,
+    }).catch(() => {});
+
+    return { success: false, status: 502, error: `Image provider error: ${err.message}` };
+  }
+}

open-sse/handlers/responseTranslator.ts

@@ -20,6 +20,51 @@ function toNumber(value: unknown, fallback = 0): number {
   return Number.isFinite(parsed) ? parsed : fallback;
 }
 
+function extractMessageOutputText(item: JsonRecord): string {
+  if (!Array.isArray(item.content)) return "";
+  let text = "";
+  for (const part of item.content) {
+    if (!part || typeof part !== "object") continue;
+    const partObj = toRecord(part);
+    if (partObj.type === "output_text" && typeof partObj.text === "string") {
+      text += partObj.text;
+    }
+  }
+  return text;
+}
+
+/**
+ * T19: Pick the last non-empty message output text from Responses API output.
+ * Falls back to the last message item even when all message texts are empty.
+ */
+function findBestMessageText(output: unknown[]): {
+  text: string;
+  selectedMessageIndex: number;
+  messageItems: JsonRecord[];
+} {
+  const messageItems = output
+    .map((item) => toRecord(item))
+    .filter((item) => item.type === "message" && Array.isArray(item.content));
+
+  for (let i = messageItems.length - 1; i >= 0; i -= 1) {
+    const text = extractMessageOutputText(messageItems[i]);
+    if (text.trim().length > 0) {
+      return { text, selectedMessageIndex: i, messageItems };
+    }
+  }
+
+  if (messageItems.length > 0) {
+    const lastIndex = messageItems.length - 1;
+    return {
+      text: extractMessageOutputText(messageItems[lastIndex]),
+      selectedMessageIndex: lastIndex,
+      messageItems,
+    };
+  }
+
+  return { text: "", selectedMessageIndex: -1, messageItems: [] };
+}
+
 /**
  * Translate non-streaming response to OpenAI format
  * Handles different provider response formats (Gemini, Claude, etc.)
@@ -44,7 +89,8 @@ export function translateNonStreamingResponse(
     const output = Array.isArray(response.output) ? response.output : [];
     const usage = toRecord(response.usage ?? responseRoot.usage);
 
-    let textContent = "";
+    const messageSelection = findBestMessageText(output);
+    let textContent = messageSelection.text;
     let reasoningContent = "";
     const toolCalls: JsonRecord[] = [];
 
@@ -56,9 +102,7 @@ export function translateNonStreamingResponse(
         for (const part of itemObj.content) {
           if (!part || typeof part !== "object") continue;
           const partObj = toRecord(part);
-          if (partObj.type === "output_text" && typeof partObj.text === "string") {
-            textContent += partObj.text;
-          } else if (partObj.type === "summary_text" && typeof partObj.text === "string") {
+          if (partObj.type === "summary_text" && typeof partObj.text === "string") {
             reasoningContent += partObj.text;
           }
         }
@@ -103,6 +147,18 @@ export function translateNonStreamingResponse(
       message.content = "";
     }
 
+    if (process.env.DEBUG_RESPONSES_SSE_TO_JSON === "true") {
+      console.log(
+        `[ResponsesSSE] ${output.length} output items, ${messageSelection.messageItems.length} message items`
+      );
+      messageSelection.messageItems.forEach((item, idx) => {
+        const textLen = extractMessageOutputText(item).length;
+        console.log(`  [${idx}] text length: ${textLen}`);
+      });
+      console.log(`  → Selected message index: ${messageSelection.selectedMessageIndex}`);
+      console.log(`  → Final text content length: ${textContent.length}`);
+    }
+
     const createdAt = toNumber(response.created_at, Math.floor(Date.now() / 1000));
     const model = toString(response.model || responseRoot.model, "openai-responses");
     const finishReason = toolCalls.length > 0 ? "tool_calls" : "stop";

open-sse/handlers/sseParser.ts

@@ -23,9 +23,18 @@ export function parseSSEToOpenAIResponse(rawSSE, fallbackModel) {
   const first = chunks[0];
   const contentParts = [];
   const reasoningParts = [];
+  const accumulatedToolCalls = new Map<string, any>();
+  let unknownToolCallSeq = 0;
   let finishReason = "stop";
   let usage = null;
 
+  const getToolCallKey = (toolCall: any) => {
+    if (toolCall?.id) return `id:${toolCall.id}`;
+    if (Number.isInteger(toolCall?.index)) return `idx:${toolCall.index}`;
+    unknownToolCallSeq += 1;
+    return `seq:${unknownToolCallSeq}`;
+  };
+
   for (const chunk of chunks) {
     const choice = chunk?.choices?.[0];
     const delta = choice?.delta || {};
@@ -36,6 +45,40 @@ export function parseSSEToOpenAIResponse(rawSSE, fallbackModel) {
     if (typeof delta.reasoning_content === "string" && delta.reasoning_content.length > 0) {
       reasoningParts.push(delta.reasoning_content);
     }
+
+    // T18: Accumulate tool calls correctly across streamed chunks
+    if (delta.tool_calls) {
+      for (const tc of delta.tool_calls) {
+        const key = getToolCallKey(tc);
+        const existing = accumulatedToolCalls.get(key);
+        const deltaArgs = typeof tc?.function?.arguments === "string" ? tc.function.arguments : "";
+
+        if (!existing) {
+          accumulatedToolCalls.set(key, {
+            id: tc?.id ?? null,
+            index: Number.isInteger(tc?.index) ? tc.index : accumulatedToolCalls.size,
+            type: tc?.type || "function",
+            function: {
+              name: tc?.function?.name || "unknown",
+              arguments: deltaArgs,
+            },
+          });
+        } else {
+          existing.id = existing.id || tc?.id || null;
+          if (!Number.isInteger(existing.index) && Number.isInteger(tc?.index)) {
+            existing.index = tc.index;
+          }
+          if (tc?.function?.name && !existing.function?.name) {
+            existing.function = existing.function || {};
+            existing.function.name = tc.function.name;
+          }
+          existing.function = existing.function || {};
+          existing.function.arguments = `${existing.function.arguments || ""}${deltaArgs}`;
+          accumulatedToolCalls.set(key, existing);
+        }
+      }
+    }
+
     if (choice?.finish_reason) {
       finishReason = choice.finish_reason;
     }
@@ -46,12 +89,22 @@ export function parseSSEToOpenAIResponse(rawSSE, fallbackModel) {
 
   const message: Record<string, unknown> = {
     role: "assistant",
-    content: contentParts.join(""),
+    content: contentParts.length > 0 ? contentParts.join("") : null,
   };
   if (reasoningParts.length > 0) {
     message.reasoning_content = reasoningParts.join("");
   }
 
+  const finalToolCalls = [...accumulatedToolCalls.values()].filter(Boolean).sort((a, b) => {
+    const ai = Number.isInteger(a?.index) ? a.index : 0;
+    const bi = Number.isInteger(b?.index) ? b.index : 0;
+    return ai - bi;
+  });
+  if (finalToolCalls.length > 0) {
+    finishReason = "tool_calls"; // T18 normalization
+    message.tool_calls = finalToolCalls;
+  }
+
   const result: Record<string, unknown> = {
     id: first.id || `chatcmpl-${Date.now()}`,
     object: "chat.completion",

open-sse/mcp-server/server.ts

@@ -433,23 +433,48 @@ async function handleListModelsCatalog(args: { provider?: string; capability?: s
   const start = Date.now();
   try {
     let path = "/v1/models";
-    const params = new URLSearchParams();
-    if (args.provider) params.set("provider", args.provider);
-    if (args.capability) params.set("capability", args.capability);
-    if (params.toString()) path += `?${params.toString()}`;
+    let isProviderSpecific = false;
+    let source = "local_catalog";
+    let warning = undefined;
+
+    if (args.provider && !args.capability) {
+      // Use direct provider fetch to get real-time API status
+      path = `/api/providers/${encodeURIComponent(args.provider)}/models`;
+      isProviderSpecific = true;
+    } else {
+      const params = new URLSearchParams();
+      if (args.provider) params.set("provider", args.provider);
+      if (args.capability) params.set("capability", args.capability);
+      if (params.toString()) path += `?${params.toString()}`;
+    }
 
     const raw = toRecord(await omniRouteFetch(path));
+
+    // If we used the direct provider endpoint
+    let rawModels = [];
+    if (isProviderSpecific) {
+      rawModels = Array.isArray(raw.models) ? raw.models : [];
+      source = typeof raw.source === "string" ? raw.source : "api";
+      if (raw.warning) warning = String(raw.warning);
+    } else {
+      rawModels = Array.isArray(raw.data) ? raw.data : [];
+      source = "local_catalog";
+      // OmniRoute's global /v1/models is always a cached/local catalog
+    }
+
     const result = {
-      models: toArray(raw.data).map((rawModel) => {
+      models: rawModels.map((rawModel) => {
         const model = toRecord(rawModel);
         return {
           id: toString(model.id, ""),
-          provider: toString(model.owned_by, toString(model.provider, "unknown")),
+          provider: toString(model.owned_by, toString(model.provider, args.provider || "unknown")),
           capabilities: toStringArray(model.capabilities, ["chat"]),
           status: toString(model.status, "available"),
           pricing: model.pricing,
         };
       }),
+      source,
+      ...(warning ? { warning } : {}),
     };
 
     await logToolCall(

open-sse/services/accountFallback.ts

@@ -8,6 +8,46 @@ import {
 } from "../config/constants.ts";
 import { getProviderCategory } from "../config/providerRegistry.ts";
 
+// T06 (sub2api PR #1037): Signals that indicate permanent account deactivation.
+// When a 401 body contains these strings, the account is permanently dead
+// and should NOT be retried after token refresh.
+export const ACCOUNT_DEACTIVATED_SIGNALS = [
+  "account_deactivated",
+  "account has been deactivated",
+  "account has been disabled",
+  "your account has been suspended",
+  "this account is deactivated",
+];
+
+// T10 (sub2api PR #1169): Signals that indicate billing credits are exhausted.
+// Distinct from rate-limit 429 — the account won't recover until credits are added.
+export const CREDITS_EXHAUSTED_SIGNALS = [
+  "insufficient_quota",
+  "billing_hard_limit_reached",
+  "exceeded your current quota",
+  "credit_balance_too_low",
+  "your credit balance is too low",
+  "credits exhausted",
+  "out of credits",
+  "payment required",
+];
+
+/**
+ * T06: Returns true if response body indicates the account is permanently deactivated.
+ */
+export function isAccountDeactivated(errorText: string): boolean {
+  const lower = String(errorText || "").toLowerCase();
+  return ACCOUNT_DEACTIVATED_SIGNALS.some((sig) => lower.includes(sig));
+}
+
+/**
+ * T10: Returns true if response body indicates credits/quota are permanently exhausted.
+ */
+export function isCreditsExhausted(errorText: string): boolean {
+  const lower = String(errorText || "").toLowerCase();
+  return CREDITS_EXHAUSTED_SIGNALS.some((sig) => lower.includes(sig));
+}
+
 // ─── Provider Profile Helper ────────────────────────────────────────────────
 
 /**
@@ -201,6 +241,14 @@ export function classifyErrorText(errorText) {
   ) {
     return RateLimitReason.QUOTA_EXHAUSTED;
   }
+  // T10: credits_exhausted signals
+  if (isCreditsExhausted(errorText)) {
+    return RateLimitReason.QUOTA_EXHAUSTED;
+  }
+  // T06: account_deactivated signals
+  if (isAccountDeactivated(errorText)) {
+    return RateLimitReason.AUTH_ERROR;
+  }
   if (
     lower.includes("rate limit") ||
     lower.includes("too many requests") ||
@@ -294,13 +342,67 @@ export function checkFallbackError(
   errorText,
   backoffLevel = 0,
   model = null,
-  provider = null
+  provider = null,
+  headers = null
 ) {
+  const errorStr = (errorText || "").toString();
+
+  function parseResetFromHeaders(headers, errorStr = "") {
+    if (!headers) return null;
+
+    // Retry-After header
+    const retryAfter =
+      typeof headers.get === "function"
+        ? headers.get("retry-after")
+        : headers["retry-after"] || headers["Retry-After"];
+
+    if (retryAfter) {
+      const seconds = parseInt(retryAfter, 10);
+      if (!isNaN(seconds) && String(seconds) === String(retryAfter).trim()) {
+        return Date.now() + seconds * 1000;
+      }
+      const date = new Date(retryAfter);
+      if (!isNaN(date.getTime())) return date.getTime();
+    }
+
+    // X-RateLimit-Reset
+    const rlReset =
+      typeof headers.get === "function"
+        ? headers.get("x-ratelimit-reset")
+        : headers["x-ratelimit-reset"] || headers["X-RateLimit-Reset"];
+
+    if (rlReset) {
+      const ts = parseInt(rlReset, 10);
+      if (!isNaN(ts)) {
+        return ts > 10000000000 ? ts : ts * 1000;
+      }
+    }
+    return null;
+  }
   // Check error message FIRST - specific patterns take priority over status codes
   if (errorText) {
-    const errorStr = typeof errorText === "string" ? errorText : JSON.stringify(errorText);
     const lowerError = errorStr.toLowerCase();
 
+    // T06 (sub2api #1037): Permanent account deactivation — do NOT retry, mark as permanent failure
+    if (isAccountDeactivated(errorStr)) {
+      return {
+        shouldFallback: true,
+        cooldownMs: 365 * 24 * 60 * 60 * 1000, // 1 year = effectively permanent
+        reason: RateLimitReason.AUTH_ERROR,
+        permanent: true,
+      };
+    }
+
+    // T10 (sub2api #1169): Credits/quota exhausted — long cooldown, distinct from rate limit
+    if (isCreditsExhausted(errorStr)) {
+      return {
+        shouldFallback: true,
+        cooldownMs: COOLDOWN_MS.paymentRequired ?? 3600 * 1000, // 1h cooldown
+        reason: RateLimitReason.QUOTA_EXHAUSTED,
+        creditsExhausted: true,
+      };
+    }
+
     if (lowerError.includes("no credentials")) {
       return {
         shouldFallback: true,
@@ -325,6 +427,18 @@ export function checkFallbackError(
       lowerError.includes("capacity") ||
       lowerError.includes("overloaded")
     ) {
+      const resetTime = parseResetFromHeaders(headers);
+      if (resetTime) {
+        const waitMs = resetTime - Date.now();
+        if (waitMs > 60_000) {
+          return {
+            shouldFallback: true,
+            cooldownMs: waitMs,
+            newBackoffLevel: 0,
+            reason: RateLimitReason.RATE_LIMIT_EXCEEDED,
+          };
+        }
+      }
       const newLevel = Math.min(backoffLevel + 1, BACKOFF_CONFIG.maxLevel);
       const reason = classifyErrorText(errorStr);
       return {
@@ -362,6 +476,19 @@ export function checkFallbackError(
 
   // 429 - Rate limit with exponential backoff
   if (status === HTTP_STATUS.RATE_LIMITED) {
+    const resetTime = parseResetFromHeaders(headers);
+    if (resetTime) {
+      const waitMs = resetTime - Date.now();
+      if (waitMs > 60_000) {
+        return {
+          shouldFallback: true,
+          cooldownMs: waitMs,
+          newBackoffLevel: 0,
+          reason: RateLimitReason.RATE_LIMIT_EXCEEDED,
+        };
+      }
+    }
+
     const newLevel = Math.min(backoffLevel + 1, BACKOFF_CONFIG.maxLevel);
     return {
       shouldFallback: true,
@@ -381,6 +508,19 @@ export function checkFallbackError(
     HTTP_STATUS.GATEWAY_TIMEOUT,
   ];
   if (transientStatuses.includes(status)) {
+    const resetTime = parseResetFromHeaders(headers, errorStr);
+    if (resetTime) {
+      const waitMs = resetTime - Date.now();
+      if (waitMs > 60_000) {
+        return {
+          shouldFallback: true,
+          cooldownMs: waitMs,
+          newBackoffLevel: 0,
+          reason: RateLimitReason.SERVER_ERROR,
+        };
+      }
+    }
+
     const profile = provider ? getProviderProfile(provider) : null;
     const baseCooldown = profile?.transientCooldown ?? COOLDOWN_MS.transientInitial;
     const maxLevel = profile?.maxBackoffLevel ?? BACKOFF_CONFIG.maxLevel;

open-sse/services/backgroundTaskDetector.ts

@@ -47,16 +47,16 @@ const DEFAULT_DETECTION_PATTERNS = [
 
 const DEFAULT_DEGRADATION_MAP: Record<string, string> = {
   // Premium → Cheap alternatives
-  "claude-opus-4-6": "gemini-2.5-flash",
-  "claude-opus-4-6-thinking": "gemini-2.5-flash",
-  "claude-opus-4-5-20251101": "gemini-2.5-flash",
-  "claude-sonnet-4-5-20250929": "gemini-2.5-flash",
-  "claude-sonnet-4-20250514": "gemini-2.5-flash",
-  "claude-sonnet-4": "gemini-2.5-flash",
-  "gemini-3.1-pro": "gemini-3.1-flash",
-  "gemini-3.1-pro-high": "gemini-3.1-flash",
+  "claude-opus-4-6": "gemini-3-flash",
+  "claude-opus-4-6-thinking": "gemini-3-flash",
+  "claude-opus-4-5-20251101": "gemini-3-flash",
+  "claude-sonnet-4-5-20250929": "gemini-3-flash",
+  "claude-sonnet-4-20250514": "gemini-3-flash",
+  "claude-sonnet-4": "gemini-3-flash",
+  "gemini-3.1-pro": "gemini-3-flash",
+  "gemini-3.1-pro-high": "gemini-3-flash",
   "gemini-3-pro-preview": "gemini-3-flash-preview",
-  "gemini-2.5-pro": "gemini-2.5-flash",
+  "gemini-2.5-pro": "gemini-3-flash",
   "gpt-4o": "gpt-4o-mini",
   "gpt-5": "gpt-5-mini",
   "gpt-5.1": "gpt-5-mini",
@@ -114,12 +114,93 @@ interface BackgroundMessage {
 interface BackgroundTaskBody {
   messages?: BackgroundMessage[];
   input?: BackgroundMessage[];
+  max_tokens?: unknown;
+  max_completion_tokens?: unknown;
+  max_output_tokens?: unknown;
 }
 
 function toMessageArray(value: unknown): BackgroundMessage[] {
   return Array.isArray(value) ? (value as BackgroundMessage[]) : [];
 }
 
+function toFiniteNumber(value: unknown): number | null {
+  if (typeof value === "number" && Number.isFinite(value)) return value;
+  if (typeof value === "string" && value.trim().length > 0) {
+    const parsed = Number(value);
+    return Number.isFinite(parsed) ? parsed : null;
+  }
+  return null;
+}
+
+function headerValue(headers: Record<string, string> | null, key: string): string {
+  if (!headers) return "";
+  const value = headers[key] ?? headers[key.toLowerCase()] ?? headers[key.toUpperCase()];
+  return typeof value === "string" ? value.trim() : "";
+}
+
+/**
+ * Get reason label when request is a background/utility task.
+ *
+ * @param {object} body - Request body
+ * @param {object} [headers] - Request headers (optional)
+ * @returns {string | null} Reason label or null when not detected
+ */
+export function getBackgroundTaskReason(
+  body: BackgroundTaskBody | unknown,
+  headers: Record<string, string> | null = null
+): string | null {
+  if (!body || typeof body !== "object") return null;
+  const typedBody = body as BackgroundTaskBody;
+
+  // 1. Check explicit header
+  if (headers) {
+    const taskType = headerValue(headers, "x-task-type");
+    const priority = headerValue(headers, "x-request-priority");
+    const initiator = headerValue(headers, "x-initiator");
+    const explicitValue = [taskType, priority, initiator].find(Boolean);
+    if (explicitValue && explicitValue.toLowerCase() === "background") {
+      return "header_background";
+    }
+  }
+
+  // 2. Very low max tokens usually indicates utility/background tasks
+  const maxTokens = toFiniteNumber(
+    typedBody.max_tokens ?? typedBody.max_completion_tokens ?? typedBody.max_output_tokens
+  );
+  if (maxTokens !== null && maxTokens > 0 && maxTokens < 50) {
+    return "low_max_tokens";
+  }
+
+  // 3. Check system prompt for background task patterns
+  const messages = toMessageArray(typedBody.messages ?? typedBody.input ?? []);
+  if (!Array.isArray(messages) || messages.length === 0) return null;
+
+  // Find system message
+  const systemMsg = messages.find(
+    (message: BackgroundMessage) => message.role === "system" || message.role === "developer"
+  );
+  if (!systemMsg) return null;
+
+  const systemContent =
+    typeof systemMsg.content === "string" ? systemMsg.content.toLowerCase() : "";
+
+  if (!systemContent) return null;
+
+  // Check against detection patterns
+  const matched = _config.detectionPatterns.some((pattern) =>
+    systemContent.includes(pattern.toLowerCase())
+  );
+
+  if (!matched) return null;
+
+  // 4. Additional heuristic: background tasks typically have very few messages
+  // (system + 1-2 user messages)
+  const userMessages = messages.filter((message: BackgroundMessage) => message.role === "user");
+  if (userMessages.length > 3) return null; // Too many turns for a background task
+
+  return "system_prompt_pattern";
+}
+
 /**
  * Check if a request is a background/utility task.
  *
@@ -131,44 +212,7 @@ export function isBackgroundTask(
   body: BackgroundTaskBody | unknown,
   headers: Record<string, string> | null = null
 ): boolean {
-  if (!body || typeof body !== "object") return false;
-  const typedBody = body as BackgroundTaskBody;
-
-  // 1. Check explicit header
-  if (headers) {
-    const priority =
-      headers["x-request-priority"] || headers["X-Request-Priority"] || headers["x-initiator"];
-    if (priority === "background" || priority === "Background") return true;
-  }
-
-  // 2. Check system prompt for background task patterns
-  const messages = toMessageArray(typedBody.messages ?? typedBody.input ?? []);
-  if (!Array.isArray(messages) || messages.length === 0) return false;
-
-  // Find system message
-  const systemMsg = messages.find(
-    (message: BackgroundMessage) => message.role === "system" || message.role === "developer"
-  );
-  if (!systemMsg) return false;
-
-  const systemContent =
-    typeof systemMsg.content === "string" ? systemMsg.content.toLowerCase() : "";
-
-  if (!systemContent) return false;
-
-  // Check against detection patterns
-  const matched = _config.detectionPatterns.some((pattern) =>
-    systemContent.includes(pattern.toLowerCase())
-  );
-
-  if (!matched) return false;
-
-  // 3. Additional heuristic: background tasks typically have very few messages
-  // (system + 1-2 user messages)
-  const userMessages = messages.filter((message: BackgroundMessage) => message.role === "user");
-  if (userMessages.length > 3) return false; // Too many turns for a background task
-
-  return true;
+  return getBackgroundTaskReason(body, headers) !== null;
 }
 
 /**

open-sse/services/combo.ts

@@ -447,8 +447,10 @@ export async function handleComboChat({
   const handleSingleModelWrapped = combo.context_cache_protection
     ? async (b, modelStr) => {
         const res = await handleSingleModel(b, modelStr);
-        // Inject tag only on success and only for non-streaming non-binary responses
-        if (res.ok && !b.stream) {
+        if (!res.ok) return res;
+
+        // Non-streaming: inject tag into JSON response (existing logic)
+        if (!b.stream) {
           try {
             const json = await res.clone().json();
             const msgs = Array.isArray(json?.messages) ? json.messages : [];
@@ -460,10 +462,71 @@ export async function handleComboChat({
               });
             }
           } catch {
-            /* non-JSON or stream — skip tagging */
+            /* non-JSON — skip tagging */
           }
+          return res;
         }
-        return res;
+
+        // Streaming (Fix #490 + #511): prepend omniModel tag into the first
+        // non-empty content chunk so it arrives BEFORE finish_reason:stop.
+        // SDKs close the connection on finish_reason, so anything sent after
+        // that marker is silently dropped.
+        if (!res.body) return res;
+        const tagContent = `\\n<omniModel>${modelStr}</omniModel>\\n`;
+        const encoder = new TextEncoder();
+        const decoder = new TextDecoder();
+        let tagInjected = false;
+
+        const transform = new TransformStream({
+          transform(chunk, controller) {
+            if (tagInjected) {
+              // Already injected — passthrough
+              controller.enqueue(chunk);
+              return;
+            }
+
+            const text = decoder.decode(chunk, { stream: true });
+
+            // Look for the first SSE data line with non-empty content
+            // Pattern: "content":"<non-empty>" — we inject tag at the start
+            const contentMatch = text.match(/"content":"([^"]+)/);
+            if (contentMatch) {
+              // Inject tag at the beginning of the first content value
+              const injected = text.replace(
+                /"content":"([^"]+)/,
+                `"content":"${tagContent.replace(/"/g, '\\"')}$1`
+              );
+              tagInjected = true;
+              controller.enqueue(encoder.encode(injected));
+              return;
+            }
+
+            // No content yet — passthrough
+            controller.enqueue(chunk);
+          },
+          flush(controller) {
+            // If stream ends without ever finding content (edge case),
+            // inject tag as a standalone chunk before the stream closes
+            if (!tagInjected) {
+              const tagChunk = `data: ${JSON.stringify({
+                choices: [
+                  {
+                    delta: { content: tagContent },
+                    index: 0,
+                    finish_reason: null,
+                  },
+                ],
+              })}\n\n`;
+              controller.enqueue(encoder.encode(tagChunk));
+            }
+          },
+        });
+
+        const transformedStream = res.body.pipeThrough(transform);
+        return new Response(transformedStream, {
+          status: res.status,
+          headers: res.headers,
+        });
       }
     : handleSingleModel;
   // ─────────────────────────────────────────────────────────────────────────
@@ -778,7 +841,8 @@ export async function handleComboChat({
         errorText,
         0,
         null,
-        provider
+        provider,
+        result.headers
       );
 
       // Record failure in circuit breaker for transient errors
@@ -802,6 +866,12 @@ export async function handleComboChat({
       if (!lastStatus) lastStatus = result.status;
       if (i > 0) fallbackCount++;
       log.warn("COMBO", `Model ${modelStr} failed, trying next`, { status: result.status });
+
+      if ([502, 503, 504].includes(result.status) && cooldownMs > 0 && cooldownMs <= 5000) {
+        log.info("COMBO", `Waiting ${cooldownMs}ms before fallback to next model`);
+        await new Promise((r) => setTimeout(r, cooldownMs));
+      }
+
       break; // Move to next model
     }
   }
@@ -823,7 +893,20 @@ export async function handleComboChat({
     );
   }
 
-  const status = lastStatus || 406;
+  if (!lastStatus) {
+    return new Response(
+      JSON.stringify({
+        error: {
+          message: "Service temporarily unavailable: all upstream accounts are inactive",
+          type: "service_unavailable",
+          code: "ALL_ACCOUNTS_INACTIVE",
+        },
+      }),
+      { status: 503, headers: { "Content-Type": "application/json" } }
+    );
+  }
+
+  const status = lastStatus;
   const msg = lastError || "All combo models unavailable";
 
   if (earliestRetryAfter) {
@@ -878,7 +961,7 @@ async function handleRoundRobinCombo({
 
   const modelCount = orderedModels.length;
   if (modelCount === 0) {
-    return unavailableResponse(406, "Round-robin combo has no models");
+    return unavailableResponse(503, "Round-robin combo has no models");
   }
 
   // Get and increment atomic counter
@@ -1014,7 +1097,8 @@ async function handleRoundRobinCombo({
           errorText,
           0,
           null,
-          provider
+          provider,
+          result.headers
         );
 
         // Transient errors → mark in semaphore AND record circuit breaker failure
@@ -1043,6 +1127,12 @@ async function handleRoundRobinCombo({
         if (!lastStatus) lastStatus = result.status;
         if (offset > 0) fallbackCount++;
         log.warn("COMBO-RR", `${modelStr} failed, trying next model`, { status: result.status });
+
+        if ([502, 503, 504].includes(result.status) && cooldownMs > 0 && cooldownMs <= 5000) {
+          log.info("COMBO-RR", `Waiting ${cooldownMs}ms before fallback to next model`);
+          await new Promise((r) => setTimeout(r, cooldownMs));
+        }
+
         break;
       }
     } finally {
@@ -1073,7 +1163,20 @@ async function handleRoundRobinCombo({
     );
   }
 
-  const status = lastStatus || 406;
+  if (!lastStatus) {
+    return new Response(
+      JSON.stringify({
+        error: {
+          message: "Service temporarily unavailable: all upstream accounts are inactive",
+          type: "service_unavailable",
+          code: "ALL_ACCOUNTS_INACTIVE",
+        },
+      }),
+      { status: 503, headers: { "Content-Type": "application/json" } }
+    );
+  }
+
+  const status = lastStatus;
   const msg = lastError || "All round-robin combo models unavailable";
 
   if (earliestRetryAfter) {

open-sse/services/comboAgentMiddleware.ts

@@ -34,7 +34,11 @@ interface Message {
 
 // ── Context Caching Tag ─────────────────────────────────────────────────────
 
-const CACHE_TAG_PATTERN = /<omniModel>([^<]+)<\/omniModel>/;
+// Handles both actual newlines (U+000A) and literal \n sequences injected
+// by combo.ts streaming around the <omniModel> tag (#531). Non-global so that
+// .exec() and .test() stay stateless; callers that need full replacement use
+// String.prototype.replace() which replaces all non-overlapping matches.
+const CACHE_TAG_PATTERN = /(?:\\n|\n)?<omniModel>([^<]+)<\/omniModel>(?:\\n|\n)?/;
 
 /**
  * Inject the model tag into the last assistant message (or append a new one).
@@ -52,7 +56,15 @@ export function injectModelTag(messages: Message[], providerModel: string): Mess
 
   // Find last assistant message with string content
   const lastAssistantIdx = cleaned.map((m) => m.role).lastIndexOf("assistant");
-  if (lastAssistantIdx === -1) return cleaned;
+
+  // #474: If no assistant message exists yet (first turn), append a synthetic one
+  // so the tag is present when the client sends the next request with the response.
+  if (lastAssistantIdx === -1) {
+    return [
+      ...cleaned,
+      { role: "assistant", content: `\n<omniModel>${providerModel}</omniModel>` },
+    ];
+  }
 
   const msg = cleaned[lastAssistantIdx];
   if (typeof msg.content !== "string") return cleaned;
@@ -157,7 +169,11 @@ export function applyComboAgentMiddleware(
   if (comboConfig.context_cache_protection) {
     pinnedModel = extractPinnedModel(messages);
     if (pinnedModel) {
-      // Model is pinned — caller should override model selection
+      // (#535) Model is pinned via <omniModel> tag — override body.model so the combo
+      // router uses exactly this model instead of picking a different one. Without this,
+      // the extracted pinnedModel is returned but body.model is unchanged, breaking
+      // context cache sessions by sending subsequent turns to a different model.
+      body = { ...body, model: pinnedModel };
     }
   }
 

open-sse/services/errorClassifier.ts

@@ -0,0 +1,53 @@
+import { isAccountDeactivated, isCreditsExhausted } from "./accountFallback.ts";
+
+export const PROVIDER_ERROR_TYPES = {
+  RATE_LIMITED: "rate_limited", // 429 — transient, retry with backoff
+  UNAUTHORIZED: "unauthorized", // 401 — token expired, refresh
+  ACCOUNT_DEACTIVATED: "account_deactivated", // 401 + deactivation signal
+  FORBIDDEN: "forbidden", // 403 — account banned/revoked, disable node
+  SERVER_ERROR: "server_error", // 500/502/503 — retry limited
+  QUOTA_EXHAUSTED: "quota_exhausted", // 402/429/400 + billing signals
+};
+
+function responseBodyToString(responseBody: unknown): string {
+  if (typeof responseBody === "string") return responseBody;
+  if (responseBody !== null && typeof responseBody === "object") {
+    try {
+      return JSON.stringify(responseBody);
+    } catch {
+      return "";
+    }
+  }
+  return "";
+}
+
+export function classifyProviderError(statusCode: number, responseBody: unknown): string | null {
+  const bodyStr = responseBodyToString(responseBody);
+  const creditsExhausted = isCreditsExhausted(bodyStr);
+  const accountDeactivated = isAccountDeactivated(bodyStr);
+
+  // T10: credits exhausted is terminal and can appear as 400/402/429 depending on provider.
+  if (
+    creditsExhausted &&
+    (statusCode === 400 || statusCode === 402 || statusCode === 429 || statusCode === 403)
+  ) {
+    return PROVIDER_ERROR_TYPES.QUOTA_EXHAUSTED;
+  }
+
+  if (statusCode === 429) {
+    return PROVIDER_ERROR_TYPES.RATE_LIMITED;
+  }
+
+  // T06: only deactivation-like 401s should be treated as permanent account expiry.
+  if (statusCode === 401) {
+    return accountDeactivated
+      ? PROVIDER_ERROR_TYPES.ACCOUNT_DEACTIVATED
+      : PROVIDER_ERROR_TYPES.UNAUTHORIZED;
+  }
+
+  if (statusCode === 402) return PROVIDER_ERROR_TYPES.QUOTA_EXHAUSTED;
+  if (statusCode === 403) return PROVIDER_ERROR_TYPES.FORBIDDEN;
+  if (statusCode >= 500) return PROVIDER_ERROR_TYPES.SERVER_ERROR;
+
+  return null;
+}

open-sse/services/ipFilter.ts

@@ -4,6 +4,8 @@
  * IP-based access control with blacklist, whitelist, priority modes, and temporary bans.
  */
 
+import { isIP } from "node:net";
+
 // In-memory IP lists
 let _config = {
   enabled: false,
@@ -161,10 +163,10 @@ export function createIPFilterMiddleware() {
  */
 export function checkRequestIP(request) {
   const ip =
-    request.headers?.get?.("x-forwarded-for")?.split(",")[0].trim() ||
-    request.headers?.get?.("x-real-ip") ||
-    request.headers?.get?.("cf-connecting-ip") ||
-    request.ip ||
+    pickFirstValidIp(request.headers?.get?.("cf-connecting-ip")) ||
+    pickFirstValidIp(request.headers?.get?.("x-forwarded-for")) ||
+    pickFirstValidIp(request.headers?.get?.("x-real-ip")) ||
+    normalizeIP(request.ip || "") ||
     "unknown";
   return checkIP(ip);
 }
@@ -177,6 +179,18 @@ function normalizeIP(ip) {
   return ip.replace(/^::ffff:/, "").trim();
 }
 
+function pickFirstValidIp(rawValue) {
+  if (typeof rawValue !== "string" || rawValue.trim().length === 0) return null;
+  const candidates = rawValue.split(",");
+  for (const candidate of candidates) {
+    const normalized = normalizeIP(candidate);
+    if (normalized && isIP(normalized) !== 0) {
+      return normalized;
+    }
+  }
+  return null;
+}
+
 function matchesAny(ip, ipSet) {
   // Direct match
   if (ipSet.has(ip)) return true;
@@ -225,12 +239,13 @@ function matchesWildcard(ip, pattern) {
 }
 
 function extractClientIP(req) {
+  const headers = req.headers || {};
   return (
-    req.headers?.["x-forwarded-for"]?.split(",")[0].trim() ||
-    req.headers?.["x-real-ip"] ||
-    req.headers?.["cf-connecting-ip"] ||
-    req.socket?.remoteAddress ||
-    req.ip ||
+    pickFirstValidIp(headers["cf-connecting-ip"]) ||
+    pickFirstValidIp(headers["x-forwarded-for"]) ||
+    pickFirstValidIp(headers["x-real-ip"]) ||
+    pickFirstValidIp(req.socket?.remoteAddress) ||
+    pickFirstValidIp(req.ip) ||
     "unknown"
   );
 }

open-sse/services/modelDeprecation.ts

@@ -18,6 +18,8 @@ const BUILT_IN_ALIASES: Record<string, string> = {
   "gemini-1.5-flash": "gemini-2.5-flash",
   "gemini-1.0-pro": "gemini-2.5-pro",
   "gemini-2.0-flash": "gemini-2.5-flash",
+  "gemini-3-pro-high": "gemini-3.1-pro-high",
+  "gemini-3-pro-low": "gemini-3.1-pro-low",
 
   // Claude legacy → current
   "claude-3-opus-20240229": "claude-opus-4-20250514",

open-sse/services/modelFamilyFallback.ts

@@ -101,6 +101,7 @@ const MODEL_UNAVAILABLE_FRAGMENTS = [
   "does not support",
   "not enabled for",
   "access to model",
+  "improperly formed request", // Kiro 400 (model unavailable)
 ];
 
 /**

open-sse/services/rateLimitManager.ts

@@ -12,6 +12,7 @@ import Bottleneck from "bottleneck";
 import { parseRetryAfterFromBody, lockModel } from "./accountFallback.ts";
 import { getProviderCategory } from "../config/providerRegistry.ts";
 import { DEFAULT_API_LIMITS } from "../config/constants.ts";
+import { getCodexRateLimitKey } from "../executors/codex.ts";
 
 interface LearnedLimitEntry {
   provider: string;
@@ -195,8 +196,15 @@ export function isRateLimitEnabled(connectionId) {
 /**
  * Get or create a limiter for a given provider+connection combination
  */
+function getLimiterKey(provider, connectionId, model = null) {
+  if (provider === "codex" && model) {
+    return `${provider}:${getCodexRateLimitKey(connectionId, model)}`;
+  }
+  return `${provider}:${connectionId}`;
+}
+
 function getLimiter(provider, connectionId, model = null) {
-  const key = model ? `${provider}:${connectionId}:${model}` : `${provider}:${connectionId}`;
+  const key = getLimiterKey(provider, connectionId, model);
 
   if (!limiters.has(key)) {
     const limiter = new Bottleneck({
@@ -235,7 +243,7 @@ export async function withRateLimit(provider, connectionId, model, fn) {
     return fn();
   }
 
-  const limiter = getLimiter(provider, connectionId, null);
+  const limiter = getLimiter(provider, connectionId, model);
   return limiter.schedule(fn);
 }
 
@@ -320,7 +328,7 @@ export function updateFromHeaders(provider, connectionId, headers, status, model
   if (!enabledConnections.has(connectionId)) return;
   if (!headers) return;
 
-  const limiter = getLimiter(provider, connectionId, null);
+  const limiter = getLimiter(provider, connectionId, model);
   const headerMap =
     provider === "claude" || provider === "anthropic" ? ANTHROPIC_HEADERS : STANDARD_HEADERS;
 
@@ -339,14 +347,19 @@ export function updateFromHeaders(provider, connectionId, headers, status, model
   // Handle 429 — rate limited
   if (status === 429) {
     const retryAfterMs = parseResetTime(retryAfterStr) || 60000; // Default 60s
+    const counts = limiter.counts();
+    const limiterKey = getLimiterKey(provider, connectionId, model);
     console.log(
-      `🚫 [RATE-LIMIT] ${provider}:${connectionId.slice(0, 8)} — 429 received, pausing for ${Math.ceil(retryAfterMs / 1000)}s`
+      `🚫 [RATE-LIMIT] ${provider}:${connectionId.slice(0, 8)} — 429 received, pausing for ${Math.ceil(retryAfterMs / 1000)}s, dropping ${counts.QUEUED} queued request(s)`
     );
 
-    limiter.updateSettings({
-      reservoir: 0,
-      reservoirRefreshAmount: limit || 60,
-      reservoirRefreshInterval: retryAfterMs,
+    // Stop the limiter and drop all waiting jobs so they fail immediately
+    // instead of hanging in the queue until reservoir refreshes (which can
+    // be hours for providers like Codex with long rate limit windows).
+    // This lets upstream callers (e.g. LiteLLM) trigger fallback to other providers.
+    // After stop, delete from Map so getLimiter() creates a fresh instance.
+    limiter.stop({ dropWaitingJobs: true }).finally(() => {
+      limiters.delete(limiterKey);
     });
     return;
   }
@@ -392,7 +405,12 @@ export function updateFromHeaders(provider, connectionId, headers, status, model
     limiter.updateSettings(updates);
 
     // Persist learned limits (debounced)
-    recordLearnedLimit(provider, connectionId, { limit, remaining, minTime: updates.minTime });
+    recordLearnedLimit(
+      provider,
+      connectionId,
+      { limit, remaining, minTime: updates.minTime },
+      model
+    );
   }
 }
 
@@ -454,9 +472,10 @@ export function getLearnedLimits() {
 function recordLearnedLimit(
   provider: string,
   connectionId: string,
-  limits: Partial<Omit<LearnedLimitEntry, "provider" | "connectionId" | "lastUpdated">>
+  limits: Partial<Omit<LearnedLimitEntry, "provider" | "connectionId" | "lastUpdated">>,
+  model: string | null = null
 ) {
-  const key = `${provider}:${connectionId}`;
+  const key = getLimiterKey(provider, connectionId, model);
   learnedLimits[key] = {
     ...limits,
     provider,

open-sse/services/roleNormalizer.ts

@@ -76,26 +76,35 @@ function supportsSystemRole(provider: string, model: string): boolean {
 }
 
 /**
- * Normalize the `developer` role to `system` for non-OpenAI providers.
- * OpenAI introduced `developer` as a replacement for `system` in newer models,
- * but most other providers still expect `system`.
+ * Normalize the `developer` role to `system` when the upstream does not support it.
+ * OpenAI Responses API sends `developer`; MiniMax and most OpenAI-compatible gateways
+ * only accept system/user/assistant/tool and return "role param error" otherwise.
+ *
+ * Logic:
+ * - When targetFormat !== "openai": always convert developer → system (Claude, Gemini, etc.).
+ * - When targetFormat === "openai": convert only when preserveDeveloperRole === false.
+ *   This covers OpenAI-compatible providers (MiniMax, etc.) that use targetFormat "openai"
+ *   but do not accept the developer role; the per-model preserveDeveloperRole flag is set
+ *   via the dashboard "Compatibility" toggle ("Do not preserve developer role").
+ * - When targetFormat === "openai" && preserveDeveloperRole !== false: keep developer (e.g. official OpenAI).
  *
  * @param messages - Array of messages
  * @param targetFormat - The target format (e.g., "openai", "claude", "gemini")
- * @returns Modified messages array
+ * @param preserveDeveloperRole - For targetFormat openai: undefined/true = keep developer (legacy default); false = map to system (MiniMax and other OpenAI-compatible gateways that reject developer)
  */
 export function normalizeDeveloperRole(
   messages: NormalizedMessage[] | unknown,
-  targetFormat: string
+  targetFormat: string,
+  preserveDeveloperRole?: boolean
 ): NormalizedMessage[] | unknown {
   if (!Array.isArray(messages)) return messages;
 
-  // For OpenAI format, keep developer role as-is (it's valid)
-  // For all other formats, convert developer → system
-  if (targetFormat === "openai") return messages;
+  if (targetFormat === "openai" && preserveDeveloperRole !== false) return messages;
 
   return messages.map((msg: NormalizedMessage) => {
-    if (msg.role === "developer") {
+    if (!msg || typeof msg !== "object") return msg;
+    const role = typeof msg.role === "string" ? msg.role : "";
+    if (role.toLowerCase() === "developer") {
       return { ...msg, role: "system" };
     }
     return msg;
@@ -169,25 +178,25 @@ export function normalizeSystemRole(
 /**
  * Full role normalization pipeline.
  * Call this before sending the request to the provider.
+ * Applies developer→system (when needed) then system→user for providers/models that do not support system role.
  *
- * @param messages - Array of messages
- * @param provider - Provider name/id
- * @param model - Model name
- * @param targetFormat - Target API format
- * @returns Normalized messages array
+ * @param messages - Array of messages to normalize (or non-array, returned as-is)
+ * @param provider - Provider id for capability lookup (e.g. system role support)
+ * @param model - Model id for capability lookup
+ * @param targetFormat - Target request format (e.g. "openai", "claude", "gemini"); see {@link normalizeDeveloperRole}
+ * @param preserveDeveloperRole - Optional; see {@link normalizeDeveloperRole}. When false, developer role is mapped to system.
+ * @returns Normalized messages array, or the original value if messages is not an array
  */
 export function normalizeRoles(
   messages: NormalizedMessage[] | unknown,
   provider: string,
   model: string,
-  targetFormat: string
+  targetFormat: string,
+  preserveDeveloperRole?: boolean
 ): NormalizedMessage[] | unknown {
   if (!Array.isArray(messages)) return messages;
 
-  // Step 1: Normalize developer → system (for non-OpenAI formats)
-  let result = normalizeDeveloperRole(messages, targetFormat);
-
-  // Step 2: Normalize system → user (for providers that don't support system role)
+  let result = normalizeDeveloperRole(messages, targetFormat, preserveDeveloperRole);
   result = normalizeSystemRole(result, provider, model);
 
   return result;

open-sse/services/sessionManager.ts

@@ -41,7 +41,13 @@ const SESSION_TTL_MS = 30 * 60 * 1000;
 const _cleanupTimer = setInterval(() => {
   const now = Date.now();
   for (const [key, entry] of sessions) {
-    if (now - entry.lastActive > SESSION_TTL_MS) sessions.delete(key);
+    if (now - entry.lastActive > SESSION_TTL_MS) {
+      sessions.delete(key);
+      for (const [apiKeyId, sessionSet] of activeSessionsByKey) {
+        sessionSet.delete(key);
+        if (sessionSet.size === 0) activeSessionsByKey.delete(apiKeyId);
+      }
+    }
   }
 }, 60_000);
 _cleanupTimer.unref();
@@ -173,6 +179,114 @@ export function getActiveSessions(): Array<SessionEntry & { sessionId: string; a
  */
 export function clearSessions(): void {
   sessions.clear();
+  activeSessionsByKey.clear();
+}
+
+// ─── T08: Per-API-Key Session Limit ─────────────────────────────────────────
+// Tracks concurrent sticky sessions per API key and enforces max_sessions limits.
+// Ref: sub2api PR #634 (fix: stabilize session hash + add user-level session limit)
+
+// Map: apiKeyId → Set<sessionId>
+const activeSessionsByKey = new Map<string, Set<string>>();
+
+/**
+ * T08: Get the number of currently active sessions for an API key.
+ * @param apiKeyId - The API key's UUID from the database
+ */
+export function getActiveSessionCountForKey(apiKeyId: string): number {
+  return activeSessionsByKey.get(apiKeyId)?.size ?? 0;
+}
+
+/**
+ * Snapshot of active session counts per API key.
+ */
+export function getAllActiveSessionCountsByKey(): Record<string, number> {
+  const out: Record<string, number> = {};
+  for (const [apiKeyId, sessionIds] of activeSessionsByKey) {
+    out[apiKeyId] = sessionIds.size;
+  }
+  return out;
+}
+
+/**
+ * T08: Register a session as belonging to an API key.
+ * Call this after session creation is allowed (i.e., limit check passed).
+ */
+export function registerKeySession(apiKeyId: string, sessionId: string): void {
+  if (!activeSessionsByKey.has(apiKeyId)) {
+    activeSessionsByKey.set(apiKeyId, new Set());
+  }
+  activeSessionsByKey.get(apiKeyId)!.add(sessionId);
+}
+
+/**
+ * Check whether a given session is already registered for an API key.
+ */
+export function isSessionRegisteredForKey(apiKeyId: string, sessionId: string): boolean {
+  return activeSessionsByKey.get(apiKeyId)?.has(sessionId) === true;
+}
+
+/**
+ * T08: Unregister a session from an API key's active set.
+ * Call this when the request closes or the session TTL expires.
+ */
+export function unregisterKeySession(apiKeyId: string, sessionId: string): void {
+  activeSessionsByKey.get(apiKeyId)?.delete(sessionId);
+  // Clean up empty sets to avoid memory leaks
+  if (activeSessionsByKey.get(apiKeyId)?.size === 0) {
+    activeSessionsByKey.delete(apiKeyId);
+  }
+}
+
+/**
+ * T08: Check whether adding a new session would exceed the key's max_sessions limit.
+ * Returns null if allowed, or an error object to return as a 429 response.
+ *
+ * @param apiKeyId - The API key's UUID
+ * @param maxSessions - The limit from the DB (0 = unlimited)
+ */
+export function checkSessionLimit(
+  apiKeyId: string,
+  maxSessions: number
+): { code: "SESSION_LIMIT_EXCEEDED"; message: string; limit: number; current: number } | null {
+  if (!maxSessions || maxSessions <= 0) return null; // unlimited
+  const current = getActiveSessionCountForKey(apiKeyId);
+  if (current < maxSessions) return null;
+  return {
+    code: "SESSION_LIMIT_EXCEEDED",
+    message:
+      `You have reached the maximum number of active sessions (${maxSessions}). ` +
+      `Please close unused sessions or wait for them to expire.`,
+    limit: maxSessions,
+    current,
+  };
+}
+
+/**
+ * T04: Extract an external session ID from request headers.
+ * Accepts both hyphenated and underscore forms for Nginx compatibility.
+ * Nginx drops headers with underscores by default — use `underscores_in_headers on`
+ * in nginx.conf, or use X-Session-Id (hyphenated) which passes cleanly.
+ *
+ * Ref: sub2api README + PR #634
+ *
+ * @param headers - Request headers (Headers object or plain object with .get())
+ * @returns External session ID with "ext:" prefix, or null
+ */
+export function extractExternalSessionId(
+  headers: Headers | { get?: (n: string) => string | null } | null | undefined
+): string | null {
+  if (!headers || typeof (headers as Headers).get !== "function") return null;
+  const h = headers as Headers;
+  const raw =
+    h.get("x-session-id") ?? // Preferred: hyphenated (passes through Nginx)
+    h.get("x_session_id") ?? // Underscore variant (direct HTTP / custom clients)
+    h.get("x-omniroute-session") ?? // OmniRoute-specific form
+    h.get("session-id") ?? // Bare session-id
+    null;
+  if (!raw || !raw.trim()) return null;
+  // Prefix "ext:" to ensure no collision with internal SHA-256 hash IDs
+  return `ext:${raw.trim().slice(0, 64)}`; // max 64 chars to avoid abuse
 }
 
 // ─── Internal Helpers ───────────────────────────────────────────────────────

open-sse/services/thinkingBudget.ts

@@ -13,21 +13,27 @@ export const ThinkingMode = {
   ADAPTIVE: "adaptive", // Scale based on request complexity
 };
 
+import { capThinkingBudget, getDefaultThinkingBudget } from "@/shared/constants/modelSpecs";
+
 // Effort → budget token mapping
 export const EFFORT_BUDGETS = {
   none: 0,
   low: 1024,
   medium: 10240,
-  high: 131072,
+  high: 131072, // Handled globally by capThinkingBudget later
+  max: 131072, // T11: Claude "max" / "xhigh" — full budget
+  xhigh: 131072, // T11: explicit alias used internally
 };
 
 // thinkingLevel string → budget token mapping
 // Used when clients send string-based thinking levels (e.g., VS Code Copilot)
 export const THINKING_LEVEL_MAP = {
   none: 0,
-  low: 1024,
-  medium: 10240,
-  high: 131072,
+  low: 4096,
+  medium: 8192,
+  high: 24576,
+  max: 131072, // T11: max = full Claude budget (sub2api: xhigh)
+  xhigh: 131072, // T11: explicit xhigh alias
 };
 
 // Default config (passthrough = backward compatible)
@@ -68,8 +74,9 @@ export function normalizeThinkingLevel(body) {
 
   // Handle top-level thinkingLevel or thinking_level string fields
   const levelStr = result.thinkingLevel || result.thinking_level;
-  if (typeof levelStr === "string" && THINKING_LEVEL_MAP[levelStr] !== undefined) {
-    const budget = THINKING_LEVEL_MAP[levelStr];
+  if (typeof levelStr === "string" && THINKING_LEVEL_MAP[levelStr.toLowerCase()] !== undefined) {
+    const rawBudget = THINKING_LEVEL_MAP[levelStr.toLowerCase()];
+    const budget = capThinkingBudget(result.model || "", rawBudget);
     // Convert to Claude thinking format as canonical representation
     result.thinking = {
       type: budget > 0 ? "enabled" : "disabled",
@@ -83,15 +90,22 @@ export function normalizeThinkingLevel(body) {
   const geminiLevel =
     result.generationConfig?.thinkingConfig?.thinkingLevel ||
     result.generationConfig?.thinking_config?.thinkingLevel;
-  if (typeof geminiLevel === "string" && THINKING_LEVEL_MAP[geminiLevel] !== undefined) {
-    const budget = THINKING_LEVEL_MAP[geminiLevel];
+  if (
+    typeof geminiLevel === "string" &&
+    THINKING_LEVEL_MAP[geminiLevel.toLowerCase()] !== undefined
+  ) {
+    const rawBudget = THINKING_LEVEL_MAP[geminiLevel.toLowerCase()];
+    const budget = capThinkingBudget(result.model || "", rawBudget);
     result.generationConfig = {
       ...result.generationConfig,
-      thinking_config: { thinking_budget: budget },
+      thinkingConfig: { ...result.generationConfig.thinkingConfig, thinkingBudget: budget },
     };
-    // Clean up camelCase variant if it was the source
+    // Clean up string variants
     if (result.generationConfig.thinkingConfig) {
-      delete result.generationConfig.thinkingConfig;
+      delete result.generationConfig.thinkingConfig.thinkingLevel;
+    }
+    if (result.generationConfig.thinking_config) {
+      delete result.generationConfig.thinking_config;
     }
   }
 
@@ -118,7 +132,7 @@ export function ensureThinkingConfig(body) {
   const result = { ...body };
   result.thinking = {
     type: "enabled",
-    budget_tokens: EFFORT_BUDGETS.medium, // 10240 default
+    budget_tokens: getDefaultThinkingBudget(model) || EFFORT_BUDGETS.medium,
   };
   return result;
 }
@@ -198,7 +212,7 @@ function setCustomBudget(body, budget) {
     };
   }
 
-  // OpenAI reasoning_effort mapping
+  // OpenAI reasoning_effort mapping (T11: add 'max' tier for full budget)
   if (result.reasoning_effort !== undefined || result.reasoning !== undefined) {
     if (budget <= 0) {
       delete result.reasoning_effort;
@@ -207,8 +221,10 @@ function setCustomBudget(body, budget) {
       result.reasoning_effort = "low";
     } else if (budget <= 10240) {
       result.reasoning_effort = "medium";
-    } else {
+    } else if (budget < 131072) {
       result.reasoning_effort = "high";
+    } else {
+      result.reasoning_effort = "max"; // T11: full budget → "max"
     }
   }
 
@@ -251,8 +267,11 @@ function applyAdaptiveBudget(body, cfg) {
   if (toolCount > 3) multiplier += 0.5;
   if (lastMsgLength > 2000) multiplier += 0.3;
 
-  const baseBudget = EFFORT_BUDGETS[cfg.effortLevel] || EFFORT_BUDGETS.medium;
-  const budget = Math.min(Math.ceil(baseBudget * multiplier), 131072);
+  const baseBudget =
+    EFFORT_BUDGETS[cfg.effortLevel] ||
+    getDefaultThinkingBudget(body.model || "") ||
+    EFFORT_BUDGETS.medium;
+  const budget = capThinkingBudget(body.model || "", Math.ceil(baseBudget * multiplier));
 
   return setCustomBudget(body, budget);
 }

open-sse/transformer/responsesTransformer.ts

@@ -1,5 +1,5 @@
-import * as fs from 'fs';
-import * as path from 'path';
+import * as fs from "fs";
+import * as path from "path";
 /**
  * Responses API Transformer
  * Converts OpenAI Chat Completions SSE to Codex Responses API SSE format
@@ -402,6 +402,16 @@ export function createResponsesApiTransformStream(logger = null) {
             const newCallId = tc.id;
             const funcName = tc.function?.name;
 
+            // T37: Prevent merging if a new tool_call uses the same index
+            if (state.funcCallIds[tcIdx] && newCallId && state.funcCallIds[tcIdx] !== newCallId) {
+              closeToolCall(controller, tcIdx);
+              delete state.funcCallIds[tcIdx];
+              delete state.funcNames[tcIdx];
+              delete state.funcArgsBuf[tcIdx];
+              delete state.funcArgsDone[tcIdx];
+              delete state.funcItemDone[tcIdx];
+            }
+
             if (funcName) state.funcNames[tcIdx] = funcName;
 
             if (!state.funcCallIds[tcIdx] && newCallId) {

open-sse/translator/helpers/geminiHelper.ts

@@ -172,6 +172,9 @@ function convertEnumValuesToStrings(obj) {
 
   if (obj.enum && Array.isArray(obj.enum)) {
     obj.enum = obj.enum.map((v) => String(v));
+    if (!obj.type) {
+      obj.type = "string";
+    }
   }
 
   for (const value of Object.values(obj)) {

open-sse/translator/helpers/responsesApiHelper.ts

@@ -93,10 +93,11 @@ export function convertResponsesApiFormat(body) {
   }
 
   // Cleanup Responses API specific fields
+  // Note: prompt_cache_key is intentionally preserved — it is used by Codex and other
+  // providers as a cache-affinity signal. Stripping it breaks prompt caching (#517).
   delete result.input;
   delete result.instructions;
   delete result.include;
-  delete result.prompt_cache_key;
   delete result.store;
   delete result.reasoning;
 

open-sse/translator/image/sizeMapper.ts

@@ -0,0 +1,22 @@
+const OPENAI_SIZE_TO_ASPECT_RATIO: Record<string, string> = {
+  "256x256": "1:1",
+  "512x512": "1:1",
+  "1024x1024": "1:1",
+  "1792x1024": "16:9",
+  "1024x1792": "9:16",
+  "1536x1024": "3:2",
+  "1024x1536": "2:3",
+};
+
+// Supports direct aspect ratios (e.g. "16:9")
+const ASPECT_RATIO_PASSTHROUGH = /^\d+:\d+$/;
+
+export function mapImageSize(sizeParam?: string | null): string {
+  if (!sizeParam) return "1:1"; // default
+
+  // Native aspect ratio (e.g. "16:9") — pass-through
+  if (ASPECT_RATIO_PASSTHROUGH.test(sizeParam)) return sizeParam;
+
+  // Map OpenAI sizes to aspect ratios
+  return OPENAI_SIZE_TO_ASPECT_RATIO[sizeParam] ?? "1:1";
+}

open-sse/translator/index.ts

@@ -67,6 +67,7 @@ function normalizeOpenAIResponsesRequest(body) {
 }
 
 /** @param options.normalizeToolCallId - When true, use 9-char tool call ids (e.g. Mistral); when false, leave ids as-is */
+/** @param options.preserveDeveloperRole - undefined/true: keep developer for OpenAI format (default); false: map to system */
 // Translate request: source -> openai -> target
 export function translateRequest(
   sourceFormat,
@@ -77,10 +78,11 @@ export function translateRequest(
   credentials = null,
   provider = null,
   reqLogger = null,
-  options?: { normalizeToolCallId?: boolean }
+  options?: { normalizeToolCallId?: boolean; preserveDeveloperRole?: boolean }
 ) {
   let result = body;
   const use9CharId = options?.normalizeToolCallId === true;
+  const preserveDeveloperRole = options?.preserveDeveloperRole;
 
   // Phase 2: Apply thinking budget control before normalization
   result = applyThinkingBudget(result);
@@ -94,9 +96,17 @@ export function translateRequest(
   // Fix missing tool responses (insert empty tool_result if needed)
   fixMissingToolResponses(result);
 
-  // Normalize roles: developer→system for non-OpenAI, system→user for incompatible models
+  // Normalize roles: developer→system unless preserved, system→user for incompatible models.
+  // This handles (1) sourceFormat openai with messages containing developer → non-openai target
+  // or preserveDeveloperRole=false, and (2) all other paths where result.messages already exists.
   if (result.messages && Array.isArray(result.messages)) {
-    result.messages = normalizeRoles(result.messages, provider || "", model || "", targetFormat);
+    result.messages = normalizeRoles(
+      result.messages,
+      provider || "",
+      model || "",
+      targetFormat,
+      preserveDeveloperRole
+    );
   }
 
   // If same format, skip translation steps
@@ -143,6 +153,24 @@ export function translateRequest(
     result = normalizeOpenAIResponsesRequest(result);
   }
 
+  // Second role normalization: only for OPENAI_RESPONSES. Here messages are built from input
+  // after the translation step, so the first normalizeRoles (above) did not see them. For
+  // sourceFormat openai with messages already on the body, the first block handles developer
+  // → system (non-openai target or preserveDeveloperRole=false); no second pass needed.
+  if (
+    sourceFormat === FORMATS.OPENAI_RESPONSES &&
+    result.messages &&
+    Array.isArray(result.messages)
+  ) {
+    result.messages = normalizeRoles(
+      result.messages,
+      provider || "",
+      model || "",
+      targetFormat,
+      preserveDeveloperRole
+    );
+  }
+
   // Ensure unique tool_call ids on final payload (translators may have introduced duplicates)
   ensureToolCallIds(result, { use9CharId });
   fixMissingToolResponses(result);

open-sse/translator/request/claude-to-gemini.ts

@@ -1,6 +1,10 @@
 import { register } from "../registry.ts";
 import { FORMATS } from "../formats.ts";
-import { DEFAULT_SAFETY_SETTINGS, tryParseJSON } from "../helpers/geminiHelper.ts";
+import {
+  DEFAULT_SAFETY_SETTINGS,
+  tryParseJSON,
+  cleanJSONSchemaForAntigravity,
+} from "../helpers/geminiHelper.ts";
 import { DEFAULT_THINKING_GEMINI_SIGNATURE } from "../../config/defaultThinkingSignature.ts";
 
 /**
@@ -154,7 +158,9 @@ export function claudeToGeminiRequest(model, body, stream) {
         functionDeclarations.push({
           name: tool.name,
           description: tool.description || "",
-          parameters: tool.input_schema || { type: "object", properties: {} },
+          parameters: cleanJSONSchemaForAntigravity(
+            tool.input_schema || { type: "object", properties: {} }
+          ),
         });
       }
     }

open-sse/translator/request/openai-responses.ts

@@ -227,10 +227,11 @@ export function openaiResponsesToOpenAIRequest(
   });
 
   // Cleanup Responses API specific fields
+  // Note: prompt_cache_key is intentionally preserved — it is used by Codex and other
+  // providers as a cache-affinity signal. Stripping it breaks prompt caching (#517).
   delete result.input;
   delete result.instructions;
   delete result.include;
-  delete result.prompt_cache_key;
   delete result.store;
   delete result.reasoning;
 

open-sse/translator/request/openai-to-claude.ts

@@ -27,6 +27,60 @@ type ClaudeTool = {
   defer_loading?: boolean;
 };
 
+/**
+ * T02: Recursively strips empty text blocks from content arrays.
+ * Anthropic returns 400 "text content blocks must be non-empty" if any
+ * text block has text: "". Must also recurse into nested tool_result.content.
+ * Ref: sub2api PR #1212
+ */
+export function stripEmptyTextBlocks(content: unknown[] | undefined): unknown[] {
+  if (!Array.isArray(content)) return content ?? [];
+  return content
+    .filter((block: unknown) => {
+      if (
+        block &&
+        typeof block === "object" &&
+        (block as Record<string, unknown>).type === "text"
+      ) {
+        const text = (block as Record<string, unknown>).text;
+        if (text === "" || text == null) return false;
+      }
+      return true;
+    })
+    .map((block: unknown) => {
+      if (
+        block &&
+        typeof block === "object" &&
+        (block as Record<string, unknown>).type === "tool_result" &&
+        Array.isArray((block as Record<string, unknown>).content)
+      ) {
+        // Recurse into nested tool_result.content
+        return {
+          ...(block as Record<string, unknown>),
+          content: stripEmptyTextBlocks((block as Record<string, unknown>).content as unknown[]),
+        };
+      }
+      return block;
+    });
+}
+
+/**
+ * T15: Normalize content to string form.
+ * Handles both string and array-of-blocks forms (Cursor, Codex 2.x, etc.).
+ * Ref: sub2api PR #1197
+ */
+export function normalizeContentToString(content: string | unknown[] | null | undefined): string {
+  if (!content) return "";
+  if (typeof content === "string") return content;
+  if (Array.isArray(content)) {
+    return (content as Array<Record<string, unknown>>)
+      .filter((b) => b.type === "text")
+      .map((b) => String(b.text ?? ""))
+      .join("\n");
+  }
+  return "";
+}
+
 // Convert OpenAI request to Claude format
 export function openaiToClaudeRequest(model, body, stream) {
   // Check if tool prefix should be disabled (configured per-provider or global)
@@ -61,11 +115,11 @@ export function openaiToClaudeRequest(model, body, stream) {
   const systemParts = [];
 
   if (body.messages && Array.isArray(body.messages)) {
-    // Extract system messages
+    // Extract system messages (T15: handle both string and array content)
     for (const msg of body.messages) {
       if (msg.role === "system") {
         systemParts.push(
-          typeof msg.content === "string" ? msg.content : extractTextContent(msg.content)
+          typeof msg.content === "string" ? msg.content : normalizeContentToString(msg.content)
         );
       }
     }
@@ -270,10 +324,14 @@ function getContentBlocksFromMessage(msg, toolNameMap = new Map(), disableToolPr
   const blocks = [];
 
   if (msg.role === "tool") {
+    // T02: Strip empty text blocks from nested tool_result content to avoid Anthropic 400
+    const toolContent = Array.isArray(msg.content)
+      ? stripEmptyTextBlocks(msg.content)
+      : msg.content;
     blocks.push({
       type: "tool_result",
       tool_use_id: msg.tool_call_id,
-      content: msg.content,
+      content: toolContent,
     });
   } else if (msg.role === "user") {
     if (typeof msg.content === "string") {
@@ -287,10 +345,14 @@ function getContentBlocksFromMessage(msg, toolNameMap = new Map(), disableToolPr
         } else if (part.type === "tool_result") {
           // Skip tool_result with no tool_use_id (would be useless and may cause errors)
           if (!part.tool_use_id) continue;
+          // T02: strip empty text blocks from nested content before passing to Anthropic
+          const resultContent = Array.isArray(part.content)
+            ? stripEmptyTextBlocks(part.content)
+            : part.content;
           blocks.push({
             type: "tool_result",
             tool_use_id: part.tool_use_id,
-            content: part.content,
+            content: resultContent,
             ...(part.is_error && { is_error: part.is_error }),
           });
         } else if (part.type === "image_url") {

open-sse/translator/request/openai-to-gemini.ts

@@ -3,6 +3,11 @@ import { FORMATS } from "../formats.ts";
 import { DEFAULT_THINKING_GEMINI_SIGNATURE } from "../../config/defaultThinkingSignature.ts";
 import { ANTIGRAVITY_DEFAULT_SYSTEM } from "../../config/constants.ts";
 import { openaiToClaudeRequestForAntigravity } from "./openai-to-claude.ts";
+import {
+  capMaxOutputTokens,
+  capThinkingBudget,
+  getDefaultThinkingBudget,
+} from "../../../src/shared/constants/modelSpecs.ts";
 
 function generateUUID() {
   return crypto.randomUUID();
@@ -88,7 +93,9 @@ function openaiToGeminiBase(model, body, stream) {
     result.generationConfig.topK = body.top_k;
   }
   if (body.max_tokens !== undefined) {
-    result.generationConfig.maxOutputTokens = body.max_tokens;
+    result.generationConfig.maxOutputTokens = capMaxOutputTokens(model, body.max_tokens);
+  } else {
+    result.generationConfig.maxOutputTokens = capMaxOutputTokens(model);
   }
 
   // Build tool_call_id -> name map
@@ -283,8 +290,12 @@ export function openaiToGeminiCLIRequest(model, body, stream) {
 
   // Add thinking config for CLI
   if (body.reasoning_effort) {
-    const budgetMap = { low: 1024, medium: 8192, high: 32768 };
-    const budget = budgetMap[body.reasoning_effort] || 8192;
+    const budgetMap = {
+      low: 1024,
+      medium: getDefaultThinkingBudget(model) || 8192,
+      high: capThinkingBudget(model, 32768),
+    };
+    const budget = budgetMap[body.reasoning_effort] || getDefaultThinkingBudget(model) || 8192;
     gemini.generationConfig.thinkingConfig = {
       thinkingBudget: budget,
       include_thoughts: true,

open-sse/translator/response/openai-responses.ts

@@ -257,6 +257,17 @@ function emitToolCall(state, emit, tc) {
   const newCallId = tc.id;
   const funcName = tc.function?.name;
 
+  // T37: If we already have a tool call at this index but the ID changed,
+  // we must close the current one and start a new one to prevent merging.
+  if (state.funcCallIds[tcIdx] && newCallId && state.funcCallIds[tcIdx] !== newCallId) {
+    closeToolCall(state, emit, tcIdx);
+    delete state.funcCallIds[tcIdx];
+    delete state.funcNames[tcIdx];
+    delete state.funcArgsBuf[tcIdx];
+    delete state.funcArgsDone[tcIdx];
+    delete state.funcItemDone[tcIdx];
+  }
+
   if (funcName) state.funcNames[tcIdx] = funcName;
 
   if (!state.funcCallIds[tcIdx] && newCallId) {

open-sse/utils/aiSdkCompat.ts

@@ -0,0 +1,31 @@
+/**
+ * AI SDK compatibility helpers (T26).
+ */
+
+/**
+ * Detects when a client explicitly prefers JSON (non-SSE) responses.
+ */
+export function clientWantsJsonResponse(acceptHeader: unknown): boolean {
+  if (typeof acceptHeader !== "string") return false;
+  const normalized = acceptHeader.toLowerCase();
+  return normalized.includes("application/json") && !normalized.includes("text/event-stream");
+}
+
+/**
+ * Resolves stream behavior from request body + Accept header.
+ * OpenAI-compatible behavior: stream only when `stream: true` and client did not force JSON.
+ */
+export function resolveStreamFlag(bodyStream: unknown, acceptHeader: unknown): boolean {
+  return bodyStream === true && !clientWantsJsonResponse(acceptHeader);
+}
+
+/**
+ * Removes surrounding markdown code fences when Claude wraps JSON payloads.
+ * Example: ```json\n{"ok":true}\n``` -> {"ok":true}
+ */
+export function stripMarkdownCodeFence(text: unknown): unknown {
+  if (typeof text !== "string") return text;
+  const codeBlockRegex = /^```(?:json|javascript|typescript|js|ts)?\s*\n?([\s\S]*?)\n?```\s*$/i;
+  const match = text.trim().match(codeBlockRegex);
+  return match ? match[1].trim() : text;
+}

open-sse/utils/ollamaTransform.ts

@@ -12,6 +12,7 @@ type PendingToolCall = {
 export function transformToOllama(response, model) {
   let buffer = "";
   let pendingToolCalls: Record<number, PendingToolCall> = {};
+  const completedToolCalls: PendingToolCall[] = [];
 
   const transform = new TransformStream({
     transform(chunk, controller) {
@@ -41,6 +42,13 @@ export function transformToOllama(response, model) {
           if (toolCalls) {
             for (const tc of toolCalls) {
               const idx = tc.index;
+
+              // T37: Prevent merging tool_calls on same index if ID changes
+              if (pendingToolCalls[idx] && tc.id && pendingToolCalls[idx].id !== tc.id) {
+                completedToolCalls.push(pendingToolCalls[idx]);
+                delete pendingToolCalls[idx];
+              }
+
               if (!pendingToolCalls[idx]) {
                 pendingToolCalls[idx] = { id: tc.id, function: { name: "", arguments: "" } };
               }
@@ -59,7 +67,7 @@ export function transformToOllama(response, model) {
 
           const finishReason = parsed.choices?.[0]?.finish_reason;
           if (finishReason === "tool_calls" || finishReason === "stop") {
-            const toolCallsArr = Object.values(pendingToolCalls);
+            const toolCallsArr = [...completedToolCalls, ...Object.values(pendingToolCalls)];
             if (toolCallsArr.length > 0) {
               const formattedCalls = toolCallsArr.map((tc) => ({
                 function: {

open-sse/utils/proxyFetch.ts

@@ -6,6 +6,7 @@ import {
   proxyUrlForLogs,
 } from "./proxyDispatcher.ts";
 import tlsClient from "./tlsClient.ts";
+import { isProxyReachable } from "@/lib/proxyHealth";
 
 function isTlsFingerprintEnabled() {
   return process.env.ENABLE_TLS_FINGERPRINT === "true";
@@ -134,6 +135,22 @@ export async function runWithProxyContext(proxyConfig, fn) {
 
   const resolvedProxyUrl = proxyConfig ? proxyConfigToUrl(proxyConfig) : null;
 
+  // T14: Proxy Fast-Fail
+  // Perform a short TCP reachability check before issuing upstream requests.
+  if (resolvedProxyUrl) {
+    const reachable = await isProxyReachable(resolvedProxyUrl);
+    if (!reachable) {
+      const proxyLabel = proxyUrlForLogs(resolvedProxyUrl);
+      const err = new Error(`[Proxy Fast-Fail] Proxy unreachable: ${proxyLabel}`) as Error & {
+        code?: string;
+        statusCode?: number;
+      };
+      err.code = "PROXY_UNREACHABLE";
+      err.statusCode = 503;
+      throw err;
+    }
+  }
+
   return proxyContext.run(proxyConfig || null, async () => {
     if (resolvedProxyUrl) {
       console.log(

open-sse/utils/stream.ts

@@ -222,16 +222,17 @@ export function createSSEStream(options: StreamOptions = {}) {
                   const extracted = extractUsage(parsed);
                   if (extracted) {
                     // Non-destructive merge: never overwrite a positive value with 0
-                    // message_start carries input_tokens, message_delta carries output_tokens
-                    if (!usage) usage = {};
-                    if (extracted.prompt_tokens > 0) usage.prompt_tokens = extracted.prompt_tokens;
-                    if (extracted.completion_tokens > 0)
-                      usage.completion_tokens = extracted.completion_tokens;
-                    if (extracted.total_tokens > 0) usage.total_tokens = extracted.total_tokens;
-                    if (extracted.cache_read_input_tokens)
-                      usage.cache_read_input_tokens = extracted.cache_read_input_tokens;
-                    if (extracted.cache_creation_input_tokens)
-                      usage.cache_creation_input_tokens = extracted.cache_creation_input_tokens;
+                    // message_start carries input_tokens, message_delta carries output_tokens;
+                    if (!usage) usage = {} as any;
+                    const u = usage as Record<string, number>;
+                    const eu = extracted as Record<string, number>;
+                    if (eu.prompt_tokens > 0) u.prompt_tokens = eu.prompt_tokens;
+                    if (eu.completion_tokens > 0) u.completion_tokens = eu.completion_tokens;
+                    if (eu.total_tokens > 0) u.total_tokens = eu.total_tokens;
+                    if (eu.cache_read_input_tokens)
+                      u.cache_read_input_tokens = eu.cache_read_input_tokens;
+                    if (eu.cache_creation_input_tokens)
+                      u.cache_creation_input_tokens = eu.cache_creation_input_tokens;
                   }
                   // Track content length and accumulate from Claude format
                   if (parsed.delta?.text) {
@@ -263,6 +264,11 @@ export function createSSEStream(options: StreamOptions = {}) {
                     }
                   }
 
+                  // T18: Track if we saw tool calls
+                  if (delta?.tool_calls && delta.tool_calls.length > 0) {
+                    (state as any).passthroughHasToolCalls = true;
+                  }
+
                   const content = delta?.content || delta?.reasoning_content;
                   if (content && typeof content === "string") {
                     totalContentLength += content.length;
@@ -278,6 +284,20 @@ export function createSSEStream(options: StreamOptions = {}) {
                   }
 
                   const isFinishChunk = parsed.choices?.[0]?.finish_reason;
+
+                  // T18: Normalize finish_reason to 'tool_calls' if tool calls were used
+                  if (
+                    isFinishChunk &&
+                    (state as any).passthroughHasToolCalls &&
+                    parsed.choices[0].finish_reason !== "tool_calls"
+                  ) {
+                    parsed.choices[0].finish_reason = "tool_calls";
+                    // If we modify it, we must output the modified object
+                    if (!injectedUsage && hasValidUsage(parsed.usage)) {
+                      output = `data: ${JSON.stringify(parsed)}\n`;
+                      injectedUsage = true;
+                    }
+                  }
                   if (isFinishChunk && !hasValidUsage(parsed.usage)) {
                     const estimated = estimateUsage(body, totalContentLength, FORMATS.OPENAI);
                     parsed.usage = filterUsageForFormat(estimated, FORMATS.OPENAI);
@@ -529,19 +549,17 @@ export function createSSEStream(options: StreamOptions = {}) {
                 if (!state.usage) {
                   state.usage = extracted;
                 } else {
-                  if (extracted.prompt_tokens > 0)
-                    state.usage.prompt_tokens = extracted.prompt_tokens;
-                  if (extracted.completion_tokens > 0)
-                    state.usage.completion_tokens = extracted.completion_tokens;
-                  if (extracted.total_tokens > 0) state.usage.total_tokens = extracted.total_tokens;
-                  if (extracted.cache_read_input_tokens > 0)
-                    state.usage.cache_read_input_tokens = extracted.cache_read_input_tokens;
-                  if (extracted.cache_creation_input_tokens > 0)
-                    state.usage.cache_creation_input_tokens = extracted.cache_creation_input_tokens;
-                  if (extracted.cached_tokens > 0)
-                    state.usage.cached_tokens = extracted.cached_tokens;
-                  if (extracted.reasoning_tokens > 0)
-                    state.usage.reasoning_tokens = extracted.reasoning_tokens;
+                  const su = state.usage as Record<string, number>;
+                  const eu = extracted as Record<string, number>;
+                  if (eu.prompt_tokens > 0) su.prompt_tokens = eu.prompt_tokens;
+                  if (eu.completion_tokens > 0) su.completion_tokens = eu.completion_tokens;
+                  if (eu.total_tokens > 0) su.total_tokens = eu.total_tokens;
+                  if (eu.cache_read_input_tokens > 0)
+                    su.cache_read_input_tokens = eu.cache_read_input_tokens;
+                  if (eu.cache_creation_input_tokens > 0)
+                    su.cache_creation_input_tokens = eu.cache_creation_input_tokens;
+                  if (eu.cached_tokens > 0) su.cached_tokens = eu.cached_tokens;
+                  if (eu.reasoning_tokens > 0) su.reasoning_tokens = eu.reasoning_tokens;
                 }
               }
 

open-sse/utils/streamHandler.ts

@@ -134,7 +134,36 @@ export function createDisconnectAwareStream(transformStream, streamController) {
         controller.enqueue(value);
       } catch (error) {
         streamController.handleError(error);
-        controller.error(error);
+
+        // T35: Encapsulate mid-stream errors as SSE events instead of abruptly aborting
+        // This prevents TransferEncodingError on the client side
+        const errorMsg = error instanceof Error ? error.message : "Upstream stream error";
+        const statusCode =
+          typeof error === "object" && error !== null && "statusCode" in error
+            ? (error as any).statusCode
+            : 500;
+
+        const errorEvent = {
+          object: "chat.completion.chunk",
+          choices: [
+            {
+              index: 0,
+              delta: {},
+              finish_reason: "error",
+            },
+          ],
+          error: {
+            message: errorMsg,
+            type: "upstream_error",
+            code: statusCode,
+          },
+        };
+
+        const encoder = new TextEncoder();
+        controller.enqueue(encoder.encode(`data: ${JSON.stringify(errorEvent)}\n\n`));
+        controller.enqueue(encoder.encode(`data: [DONE]\n\n`));
+
+        controller.close();
       }
     },
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "omniroute",
-  "version": "2.8.4",
+  "version": "3.0.0-rc.10",
   "description": "Smart AI Router with auto fallback — route to FREE & cheap models, zero downtime. Works with Cursor, Cline, Claude Desktop, Codex, and any OpenAI-compatible tool.",
   "type": "module",
   "bin": {
@@ -81,8 +81,10 @@
     "system-info": "node scripts/system-info.mjs"
   },
   "dependencies": {
+    "@lobehub/icons": "^5.0.1",
     "@modelcontextprotocol/sdk": "^1.27.1",
     "@monaco-editor/react": "^4.7.0",
+    "@swc/helpers": "0.5.19",
     "bcryptjs": "^3.0.3",
     "better-sqlite3": "^12.6.2",
     "bottleneck": "^2.19.5",
@@ -94,7 +96,7 @@
     "jose": "^6.1.3",
     "lowdb": "^7.0.1",
     "monaco-editor": "^0.55.1",
-    "next": "^16.1.6",
+    "next": "^16.0.10",
     "next-intl": "^4.8.3",
     "node-machine-id": "^1.1.12",
     "open": "^11.0.0",
@@ -110,8 +112,7 @@
     "uuid": "^13.0.0",
     "wreq-js": "^2.0.1",
     "zod": "^4.3.6",
-    "zustand": "^5.0.10",
-    "@swc/helpers": "0.5.19"
+    "zustand": "^5.0.10"
   },
   "devDependencies": {
     "@playwright/test": "^1.58.2",
@@ -124,7 +125,7 @@
     "concurrently": "^9.2.1",
     "cross-env": "^10.1.0",
     "eslint": "^9.39.2",
-    "eslint-config-next": "16.1.6",
+    "eslint-config-next": "^16.0.10",
     "husky": "^9.1.7",
     "lint-staged": "^16.2.7",
     "prettier": "^3.8.1",

public/providers/cloudflare-ai.svg

@@ -0,0 +1 @@
+<svg role="img" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><title>Cloudflare</title><path d="M16.5088 16.8447c.1475-.5068.0908-.9707-.1553-1.3154-.2246-.3164-.6045-.499-1.0615-.5205l-8.6592-.1123a.1559.1559 0 0 1-.1333-.0713c-.0283-.042-.0351-.0986-.021-.1553.0278-.084.1123-.1484.2036-.1562l8.7359-.1123c1.0351-.0489 2.1601-.8868 2.5537-1.9136l.499-1.3013c.0215-.0561.0293-.1128.0147-.168-.5625-2.5463-2.835-4.4453-5.5499-4.4453-2.5039 0-4.6284 1.6177-5.3876 3.8614-.4927-.3658-1.1187-.5625-1.794-.499-1.2026.119-2.1665 1.083-2.2861 2.2856-.0283.31-.0069.6128.0635.894C1.5683 13.171 0 14.7754 0 16.752c0 .1748.0142.3515.0352.5273.0141.083.0844.1475.1689.1475h15.9814c.0909 0 .1758-.0645.2032-.1553l.12-.4268zm2.7568-5.5634c-.0771 0-.1611 0-.2383.0112-.0566 0-.1054.0415-.127.0976l-.3378 1.1744c-.1475.5068-.0918.9707.1543 1.3164.2256.3164.6055.498 1.0625.5195l1.8437.1133c.0557 0 .1055.0263.1329.0703.0283.043.0351.1074.0214.1562-.0283.084-.1132.1485-.204.1553l-1.921.1123c-1.041.0488-2.1582.8867-2.5527 1.914l-.1406.3585c-.0283.0713.0215.1416.0986.1416h6.5977c.0771 0 .1474-.0489.169-.126.1122-.4082.1757-.837.1757-1.2803 0-2.6025-2.125-4.727-4.7344-4.727"/></svg>

public/providers/scaleway.svg

@@ -0,0 +1 @@
+<svg role="img" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><title>Scaleway</title><path d="M16.605 11.11v5.72a1.77 1.77 0 01-1.54 1.69h-4a1.43 1.43 0 01-1.31-1.22 1.09 1.09 0 010-.18 1.37 1.37 0 011.37-1.36h1.74a1 1 0 001-1v-3.62a1.4 1.4 0 011.18-1.39h.17a1.37 1.37 0 011.39 1.36zm-6.46 1.74V9.26a1 1 0 011-1h1.85a1.37 1.37 0 001.37-1.37 1 1 0 000-.17 1.45 1.45 0 00-1.41-1.2h-3.96a1.81 1.81 0 00-1.58 1.66v5.7a1.37 1.37 0 001.37 1.37h.21a1.4 1.4 0 001.15-1.4zm12-4.29V20a4.53 4.53 0 01-4.15 4h-7.58a8.57 8.57 0 01-8.56-8.57V4.54A4.54 4.54 0 016.395 0h7.18a8.56 8.56 0 018.56 8.56zm-2.74 0a5.83 5.83 0 00-5.82-5.82h-7.19a1.79 1.79 0 00-1.8 1.8v10.89a5.83 5.83 0 005.82 5.8h7.44a1.79 1.79 0 001.54-1.48z"/></svg>

scripts/check-docs-sync.mjs

@@ -48,7 +48,8 @@ function extractChangelogSections(content) {
 }
 
 function isSemver(value) {
-  return /^\d+\.\d+\.\d+$/.test(value);
+  // Accept X.Y.Z and X.Y.Z-prerelease.N (e.g. 3.0.0-rc.1, 3.0.0-beta.2)
+  return /^\d+\.\d+\.\d+(-[a-zA-Z0-9.]+)?$/.test(value);
 }
 
 let hasFailure = false;

scripts/run-next.mjs

@@ -16,7 +16,8 @@ const { dashboardPort } = runtimePorts;
 const env = bootstrapEnv();
 
 const args = ["./node_modules/next/dist/bin/next", mode, "--port", String(dashboardPort)];
-if (mode === "dev") {
+// Default: use webpack (stable). Set OMNIROUTE_USE_TURBOPACK=1 to use Turbopack (faster dev).
+if (mode === "dev" && process.env.OMNIROUTE_USE_TURBOPACK !== "1") {
   args.splice(2, 0, "--webpack");
 }
 

src/app/(dashboard)/dashboard/analytics/SearchAnalyticsTab.tsx

@@ -0,0 +1,196 @@
+/**
+ * Search Analytics Tab
+ *
+ * Shows search request stats from call_logs (request_type = 'search'),
+ * provider breakdown, cache hit rate, and cost summary.
+ */
+
+"use client";
+
+import { useEffect, useState } from "react";
+
+interface SearchStats {
+  total: number;
+  today: number;
+  cached: number;
+  errors: number;
+  totalCostUsd: number;
+  byProvider: Record<string, { count: number; costUsd: number }>;
+  last24h: Array<{ hour: string; count: number }>;
+  cacheHitRate: number;
+  avgDurationMs: number;
+}
+
+function StatCard({
+  icon,
+  label,
+  value,
+  sub,
+}: {
+  icon: string;
+  label: string;
+  value: string | number;
+  sub?: string;
+}) {
+  return (
+    <div className="card p-4 flex flex-col gap-1">
+      <div className="flex items-center gap-2 text-text-muted text-sm">
+        <span className="material-symbols-outlined text-[18px]">{icon}</span>
+        {label}
+      </div>
+      <div className="text-2xl font-bold text-text">{value}</div>
+      {sub && <div className="text-xs text-text-muted">{sub}</div>}
+    </div>
+  );
+}
+
+function ProviderBar({
+  provider,
+  count,
+  total,
+  costUsd,
+}: {
+  provider: string;
+  count: number;
+  total: number;
+  costUsd: number;
+}) {
+  const pct = total > 0 ? Math.round((count / total) * 100) : 0;
+  return (
+    <div className="flex flex-col gap-1">
+      <div className="flex justify-between text-sm">
+        <span className="font-medium text-text">{provider}</span>
+        <span className="text-text-muted">
+          {count} queries · ${costUsd.toFixed(4)}
+        </span>
+      </div>
+      <div className="h-2 rounded-full bg-bg-muted overflow-hidden">
+        <div
+          className="h-full rounded-full bg-primary transition-all"
+          style={{ width: `${pct}%` }}
+        />
+      </div>
+      <div className="text-xs text-text-muted text-right">{pct}%</div>
+    </div>
+  );
+}
+
+export default function SearchAnalyticsTab() {
+  const [stats, setStats] = useState<SearchStats | null>(null);
+  const [loading, setLoading] = useState(true);
+  const [error, setError] = useState<string | null>(null);
+
+  useEffect(() => {
+    fetch("/api/v1/search/analytics")
+      .then((r) => r.json())
+      .then((d) => {
+        setStats(d);
+        setLoading(false);
+      })
+      .catch((e) => {
+        setError(e.message);
+        setLoading(false);
+      });
+  }, []);
+
+  if (loading) {
+    return (
+      <div className="flex items-center justify-center py-16 text-text-muted">
+        <span className="material-symbols-outlined animate-spin mr-2">progress_activity</span>
+        Loading search analytics…
+      </div>
+    );
+  }
+
+  if (error || !stats) {
+    return (
+      <div className="card p-6 text-center text-text-muted">
+        <span className="material-symbols-outlined text-[32px] mb-2 block">search_off</span>
+        {error || "No search data available yet."}
+        <p className="text-xs mt-2">
+          Search requests will appear here after the first search via /v1/search.
+        </p>
+      </div>
+    );
+  }
+
+  const providers = Object.entries(stats.byProvider).sort(([, a], [, b]) => b.count - a.count);
+
+  return (
+    <div className="flex flex-col gap-6">
+      {/* KPI Cards */}
+      <div className="grid grid-cols-2 md:grid-cols-4 gap-4">
+        <StatCard
+          icon="manage_search"
+          label="Total Searches"
+          value={stats.total.toLocaleString()}
+          sub={`${stats.today} today`}
+        />
+        <StatCard
+          icon="cached"
+          label="Cache Hit Rate"
+          value={`${stats.cacheHitRate}%`}
+          sub={`${stats.cached} cached requests`}
+        />
+        <StatCard
+          icon="attach_money"
+          label="Total Cost"
+          value={`$${stats.totalCostUsd.toFixed(4)}`}
+          sub="search API costs"
+        />
+        <StatCard
+          icon="timer"
+          label="Avg Response"
+          value={`${stats.avgDurationMs}ms`}
+          sub={stats.errors > 0 ? `${stats.errors} errors` : "No errors"}
+        />
+      </div>
+
+      {/* Provider Breakdown */}
+      {providers.length > 0 && (
+        <div className="card p-5">
+          <h3 className="font-semibold text-text mb-4 flex items-center gap-2">
+            <span className="material-symbols-outlined text-primary text-[20px]">hub</span>
+            Provider Breakdown
+          </h3>
+          <div className="flex flex-col gap-4">
+            {providers.map(([prov, data]) => (
+              <ProviderBar
+                key={prov}
+                provider={prov}
+                count={data.count}
+                total={stats.total}
+                costUsd={data.costUsd}
+              />
+            ))}
+          </div>
+        </div>
+      )}
+
+      {/* Empty state */}
+      {stats.total === 0 && (
+        <div className="card p-8 text-center text-text-muted">
+          <span className="material-symbols-outlined text-[48px] mb-3 block text-primary opacity-50">
+            travel_explore
+          </span>
+          <p className="font-medium text-text">No searches yet</p>
+          <p className="text-sm mt-1">
+            Use <code className="bg-bg-muted px-1 rounded">POST /v1/search</code> to start routing
+            web searches.
+          </p>
+        </div>
+      )}
+
+      {/* Free tier note */}
+      <div className="text-xs text-text-muted border border-border rounded-lg p-3 flex items-start gap-2">
+        <span className="material-symbols-outlined text-[16px] text-green-500 mt-0.5">
+          check_circle
+        </span>
+        <span>
+          <strong>Free tier available:</strong> Serper (2,500/mo), Brave (2,000/mo), Exa (1,000/mo),
+          Tavily (1,000/mo) — total 6,500+ free searches/month with automatic failover.
+        </span>
+      </div>
+    </div>
+  );
+}