Repository navigation
Releases
Releases · neuralnexustech/freeroute
v0.7.18 — Live 24h Usage Summary, Live Auto-Sync Modal & Topbar Spinner
Highlights in v0.7.18
- Zero CSS Changes: src/app/globals.css remains 100% untouched to avoid any styling collisions.
- Dashboard 24h Default: Usage Summary default filter is set to Last 24 Hours with an active pulsating LIVE indicator.
- Live Auto-Sync Modal: A non-intrusive floating popup displaying live stages (discovering, pulling, pinging), model counts, progress bar, real-time log feed, and clean UI/UX control buttons (Minimize, Close).
- Topbar Live Spinner: Appears smoothly right next to \Health: 99.98%\ while auto-sync is actively running in the background. Clicking re-opens the live popup; once sync completes, shows \✓ Synced\ and disappears cleanly.
v0.7.17 — Startup Auto-Sync, Sync All & Test, and Live Model Pinging
New Features & Improvements in v0.7.17
1. Server Startup Auto-Sync (\instrumentation.ts)
- Automatically registers a background worker on server boot.
- 3.5 seconds after startup, it automatically identifies all connected providers and silently syncs their latest model catalogs into the database.
- Runs non-blocking connectivity health checks on active models, measuring live response latency and marking verified models as live (\status: 'ok').
2. '⚡ Sync All & Test' in Dashboards
- Added a prominent *\⚡ Sync All & Test* button in the Models dashboard (/dashboard/models) and *\🔄 Sync All Models* in the Providers dashboard (/dashboard/providers).
- One-click discovers all latest models across every connected provider, enriches specs (context windows, pricing, modalities), and tests live status.
3. Row-Level Live Model Testing & Status Badges
- Live Status Badges:
- 🟢 OK (Latency ms): Verified live with actual ping response time (e.g. \OK 215ms).
- 🔴 Fail: Shows connection failure with hoverable tooltip explaining the exact error (e.g. \HTTP 401: Invalid API Key).
- ⚪ Untested: Indicates models awaiting initial check.
- Single-Click ⚡ Test Button on every row: Test any individual model directly from the table without running a full suite test.
- Status Filter: Easily filter the table by \🟢 OK / Verified, \🔴 Fail / Error, \⚪ Untested, or \All statuses.
4. Online-Only Ollama
- Configured Ollama strictly as an online/cloud provider (\ reeTier\ with bearer auth) and removed local \localhost:11434\ probing.
v0.7.16 — Fix Stream Unexpected EOF & Clean Stream Closure
Fixes in v0.7.16
Fixed 'stream reading error: unexpected EOF'
- Connection Timeout Scoped to Headers Only:
- Previously, \AbortSignal.timeout(30_000)\ was placed directly on upstream \ etch(). In Node.js, that signal stayed active across the entire streaming lifecycle and forcefully aborted active streams at 30 seconds.
- Now, a 30s connection timeout is used strictly for receiving the initial response headers. Once headers arrive and streaming begins, the timer is cleared via \clearTimeout()\ in \ inally, allowing AI models to stream uninterrupted for minutes.
- Clean Anthropic SSE Stream Termination:
- In \claude-translator.ts, if any network hiccup or provider interruption occurs, the stream pump now cleanly emits \content_block_stop, \message_delta\ (with stop_reason \end_turn), and \message_stop\ before closing the stream controller.
- Anthropic clients (e.g. Claude Code CLI, Cursor, Cline) will never experience an unclosed block or truncated stream triggering \unexpected EOF.
v0.7.15 — Faster API Response, 30s Timeout & Cleaner Combos UI
Fixes in v0.7.15
UI Fix
- Removed top healing banner — the 🔄 Auto-Healing In Progress bar at the top of the Combos page is gone. Healing status is still shown inline on each individual tier card (colored glow + countdown + Heal button), so you can still see which model is recovering and why.
Performance — Faster API Responses
- 30-second upstream timeout added to all provider fetch calls. Previously if a provider (e.g. OpenRouter under load) stalled, your request would hang indefinitely. Now it auto-fails over to the next tier within 30 seconds.
- Timed-out providers enter 30s cooldown — after a timeout, that provider is automatically skipped on subsequent requests until it recovers (same as 5xx errors).
- Faster first-token delivery (TTFT): peekStreamForFailover previously buffered up to 4 chunks before deciding if a stream is healthy. Now it commits on the first chunk as soon as healthy content (delta/content) is detected — eliminating 50–300ms of unnecessary buffering on every streaming request.
- 10-second stall detection in stream peek — if a provider sends an HTTP 200 but then stalls without sending any bytes, it's now detected and failed over within 10s.
v0.7.14 — Smart Auto-Healing & Combos Health Dashboard
🚀 What's New in v0.7.14
- Smart Auto-Healing Across All Strategies:
- Cooldowns and healing now apply universally across Failover, Round-Robin, Weighted, LKGP, Latency-Based, and Cost-Optimized combos.
- Automatically skips failing/exhausted models until they recover:
- RPM / Short Rate Limits (429): Cooldown for 60 seconds, then auto-re-enters routing pool.
- Daily Quotas / RPD Exceeded (403): Automatically waits until 00:00 UTC (tomorrow) before re-enabling.
- Server Outages (500 / 502 / 503 / 504): Cooldown for 30 seconds for transient recovery.
- Auth / Key Errors (401): 5-minute cooldown.
- Live Combos Health UI:
- Top auto-healing status banner showing count of recovering models.
- Per-model live status cards with glowing amber indicators, remaining countdown timers, and reason badges.
- Manual ⚡ Heal button to immediately reactivate models on demand.
- Streaming EOF Resilience:
- Normalized stream end-of-file handling (unexpected EOF / client disconnect) as clean stream completion rather than an upstream failure.
- Prevents false-positive failover loops when clients disconnect.
- Provider & Model Accuracy Improvements:
- Verified model capability queries and provider pulling.
v0.7.13
- Improved model spec accuracy and provider documentation extraction
- Enhanced combo routing and failover visibility
- Bug fixes and performance improvements