v0.9.9
cheaper-inference/OmniRoutev0.9.9Sep 10, 2026by tashfeenahmed
Upgrading?
AI Summary
Adds first-class adapters for Router9 and Septor Labs, optimizes routing for keyless platforms, enables response caching for streaming requests, and improves observability with `execution_id` and `caller` columns.
Key Highlights
- New adapters for Router9 and Septor Labs with key validation and platform colors.
- Keyless platforms no longer consume the attempt budget.
- Streaming requests can now hit the response cache.
- Added `execution_id` and `caller` column to request logs for better observability.
- Dashboard sign-in gate accessibility improvements.
New Features
- Router9 and Septor Labs adapters
- Keyless platform routing logic
- Streaming response caching
- execution_id in chat bodies
- caller column in requests table
- Dashboard accessibility labels
Full Release Notes
## Providers - **Router9 and Septor Labs** join the router with first-class adapters: key import aliases, key-picker entries, platform colours and account-wide quota observations. Router9 keys are validated with a no-inference probe; Septor responses that come back under a different model identity are rejected as retryable before any substituted content is streamed. (#1194) ## Routing - **Keyless platforms no longer eat the attempt budget.** Models whose platform has no enabled, healthy key are dropped before the routing walk, so the "all attempts failed" summary and the retry budget only count candidates that could actually run. Pinned models are still tried and still report against their own label. (#1105, thanks @chongjiazhen) ## Cache - **Streaming requests can hit the response cache.** `stream: true` completions are stored frame by frame in their own store and replayed as byte-identical SSE on a hit, with `X-FreeLLM-Cache: HIT`. Streams are only buffered when a cache key exists, buffering stops at 2 MB, truncated or broken streams are never stored, stream lookups feed the hit-rate counters, and both stores share the one `RESPONSE_CACHE_MAX_ENTRIES` cap. Off by default, as before. (#1099, #1197, thanks @suantea) ## Observability - **`execution_id` in every chat body.** Both completion routes now return the request's execution id on success, cache hits, idempotency replays and every error body, so a client can quote it against the request log. The `requests` table gains a `caller` column, written by the OpenAI, Responses, Messages, Ollama/Gemini and fusion paths. (#1103, #1196, thanks @suantea) ## Dashboard - **Sign-in gate accessibility.** Fields are labelled, validation errors are tied to their inputs with `aria-describedby` and announced with `role="alert"`, the desktop session-repair error is announced the same way, and the gate uses `min-h-dvh` so it centres correctly on mobile viewports. Covered by tests. (#1078, #1195, thanks @Kaban15) ## Docs - Streaming-pipeline, REST API, environment and idempotency pages in `docs/en` and `docs/zh-cn` now describe streaming replay; the observability page documents `execution_id` and `caller`. README credits @Kaban15. Full changelog: https://github.com/tashfeenahmed/freellmapi/compare/v0.9.8...v0.9.9 --- ā Like the free router? [Go Premium](https://freellmapi.co/?utm_source=github&utm_medium=release&utm_campaign=premium&utm_content=v0.9.9#pricing) ā the live signed catalog, $19/yr, cancel anytime.
Related Repositories
openclaw/openclawā
392,696
The AI that really does things. Any OS. Any Platform. The lobster way. š¦
nilbuild/developer-roadmapā
370,068
Interactive roadmaps, guides and other educational content to help developers grow in their careers.
morluto/reaā
357,715
Reverse engineer anything with agents, from app behavior down to native binaries.
anomalyco/opencodeā
219,674
The open source coding agent.
n8n-io/n8nā
208,251
Fair-code workflow automation platform with native AI capabilities. Combine visual building with custom code, self-host or cloud, 400+ integrations.