Audio endpoints (/v1/audio/speech and /v1/audio/transcriptions) only
supported hardcoded providers from audioRegistry.ts. Local inference
backends configured as provider_nodes (e.g., MLX-Audio, oMLX) could
not serve audio through OmniRoute.
This adds a Phase 3 fallback in the audio model parser that consults
provider_nodes from the database. Local providers with api_type=openai
are automatically available for audio routing via their prefix
(e.g., mlx-audio/tts-model, omlx/whisper-large-v3-turbo).
Design: injection pattern — Next.js route handlers load provider_nodes
(async DB query) and pass them to the sync parser as a parameter.
No cross-workspace imports, no breaking changes to existing parsers.
Changes:
- Add buildDynamicAudioProvider() in audioRegistry.ts
- Add Phase 3 (provider_nodes prefix match) to parseAudioModel()
- Extend parseSpeechModel/parseTranscriptionModel with optional
dynamicProviders parameter (backward compatible)
- Load and inject provider_nodes in speech/transcription route handlers
- Dynamic providers use authType=none (local, no credentials needed)
Claude OAuth tokens are short-lived and require refresh. The runtime
HealthCheck (open-sse) already refreshes them successfully, but the
Dashboard test endpoint was missing `refreshable: true` in its config.
This caused the Dashboard to show "auth failed / Token expired" for
Claude providers even though the tokens were being refreshed correctly
at runtime. The codex provider already had this flag set.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Anthropic API rejects requests containing {"type":"text","text":""} with
400 "text content blocks must be non-empty". Some clients like LiteLLM
passthrough and @ai-sdk/anthropic may forward empty text blocks as-is.
Filter out empty text content blocks from messages before calling
translateRequest, similar to how empty-name tools are already stripped.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- gemini/gemini-cli: removed gemini-3.1-pro/flash/preview (don't exist in Google API v1beta),
replaced with real models: gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash, gemini-1.5-*
- antigravity: removed gemini-3.1-pro-high/low and gemini-3-flash (internal aliases invalid),
replaced with gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash
- github: removed gemini-3-flash-preview and gemini-3-pro-preview, replaced with gemini-2.5-flash
- nvidia: corrected 'nvidia/llama-3.3-70b-instruct' to 'meta/llama-3.3-70b-instruct'
(NVIDIA NIM uses meta/ namespace, not nvidia/ namespace for Meta models)
- nvidia: added meta/llama-3.1-70b-instruct and nvidia/llama-3.1-405b-instruct
Also fixed free-stack combo on .15 DB:
- removed qw/qwen3-coder-plus (qwen provider has expired refresh token)
- corrected nvidia/llama-3.3-70b-instruct → nvidia/meta/llama-3.3-70b-instruct
- corrected gemini/gemini-3.1-flash → gemini/gemini-2.5-flash
- added if/deepseek-v3.2 as replacement for qw/qwen3-coder-plus
Local inference backends (oMLX, Ollama, LM Studio) configured as
provider_nodes have no health monitoring. When a local provider is
down, OmniRoute waits the full timeout before failing.
This adds a background health check that polls local provider_nodes:
- GET /models with 5s timeout for each local node (localhost only)
- In-memory health cache (no DB migration needed)
- Promise.allSettled for parallel checks (one slow node doesn't block)
- Exponential backoff on failures: 30s → 60s → 120s → 300s max
- Reset to 30s on first success after failure
- State transition logging (healthy ↔ unhealthy)
- Expose health status via GET /api/monitoring/health (localProviders)
- Auto-init on first import (same pattern as tokenHealthCheck)
- 401 treated as healthy (server up, auth required)
- isNodeHealthy() returns true if never checked (optimistic default)
Embedding endpoint (/v1/embeddings) only supports 6 hardcoded cloud
providers. Local inference backends (oMLX, Ollama) serving embeddings
via provider_nodes are inaccessible through OmniRoute.
This adds dynamic provider_node support for embeddings:
- Add EmbeddingProvider interface and buildDynamicEmbeddingProvider()
- Add Phase 2 (provider_nodes prefix match) in parseEmbeddingModel()
- Handler accepts resolvedProvider/resolvedModel from route (injection pattern)
- Handler supports authType=none for local providers (was missing — critical gap)
- Route loads local provider_nodes (localhost only — prevents auth bypass/SSRF)
- Route filters by apiType=chat|responses and localhost hostname
- buildDynamicEmbeddingProvider validates inputs (prefix + baseUrl required)
- Per-node try/catch in map — one bad row doesn't block all providers
- DB errors logged and fall back to hardcoded providers
Use shallow copy ({ ...body }) instead of direct reference assignment
so that later translatedBody.model = model does not mutate the
caller's original body object.
When both source and target formats are Claude, skip all request
modification and forward the body untouched. This prevents
prepareClaudeRequest from corrupting valid Claude-native requests
destined for anthropic-compatible provider nodes.
When Claude Code compacts conversation context to fit within token
limits, it may remove assistant messages containing tool_use/tool_calls
while leaving the corresponding tool_result/function_call_output
messages intact. This creates orphaned tool results that cause
providers to reject requests with errors like "tool result's tool id
not found" or "No tool call found for function call output".
Prevents infinite retry loops when models generate tool calls with
empty function names. The normalizeToolName function converted these
to "placeholder_tool" which does not exist in any client's tool
registry, causing repeated error-retry cycles.
Reasoning models (o1, o1-pro, o3, o3-mini) reject standard parameters
like temperature and top_p with 400 Bad Request. OmniRoute's default
executor forwards all parameters without filtering.
This fix adds declarative parameter filtering:
- Add unsupportedParams[] field to RegistryModel interface
- Add REASONING_UNSUPPORTED frozen constant shared across entries
- Add o1-pro, o3, o3-mini to OpenAI registry (were missing)
- Add getUnsupportedParams() helper with:
- O(1) precomputed map lookup (not O(N×M) scan)
- Cross-provider routing support via precomputed map
- Prefixed model ID support (e.g., "openai/o3" → "o3")
- Strip unsupported params in chatCore.ts before executor call
- Use Object.hasOwn() for safe property check (no prototype chain)
- Log stripped params at WARN level for visibility
- Remove apiKey===null heuristic (too broad — could match cloud providers
with non-standard auth). Use URL-based detection only.
- Guard local 404 branch with provider && model check — if either is null,
fall through to standard connection lockout (safer behavior).
- Document LOCAL_HOSTNAMES as module-load-time constant (restart required).
- Document PROVIDER_PROFILES.local as intentionally not yet wired.
When a local inference backend (oMLX, Ollama, LM Studio) returns 404
for an unknown model, OmniRoute previously locked the entire connection
for 2 minutes — blocking all valid models on that connection.
This fix introduces local provider detection and changes the 404
behavior for local providers:
- Model-only lockout (5s) instead of connection-level lockout (2min)
- Connection stays active — other models continue working immediately
- Detection via URL heuristic (localhost/127.0.0.1) + apiKey===null fallback
- Configurable via LOCAL_HOSTNAMES env var for Docker setups
Also fixes a pre-existing bug where the model parameter was not passed
to markAccountUnavailable() from chat.ts, preventing per-model lockouts
from working at all.
Changes:
- Add isLocalProvider(baseUrl) helper in providerRegistry.ts
- Add COOLDOWN_MS.notFoundLocal (5s) and PROVIDER_PROFILES.local
- Add local 404 branch in markAccountUnavailable() in auth.ts
- Pass model param to markAccountUnavailable() in chat.ts (bug fix)
Kilo Gateway (api.kilo.ai/api/gateway) is an OpenAI-compatible API
offering 335+ models via a single API key, including 6 free models
and 3 auto-routing models (frontier/balanced/free).
This is distinct from the existing KiloCode provider which uses
OAuth + /api/openrouter/ endpoint.
- Register kilo-gateway in providerRegistry.ts (alias: kg)
- Add to APIKEY_PROVIDERS in providers.ts
- Add models endpoint config in route.ts
- Add official Kilo AI icon (favicon)
Even with EXPERIMENTAL_TURBOPACK=0 and NEXT_PRIVATE_BUILD_WORKER=0, Next.js 16
instrumentation chunks still emit require('better-sqlite3-<16hexchars>') and
require('zod-<16hexchars>') into the compiled .js files inside .next/server/.
The webpack externals function in next.config.mjs patches the runtime bundler
but does NOT rewrite already-compiled chunks. Added step 5.6 to prepublish.mjs:
walks all .js files in app/.next/server/ and strips the 16-char hex suffix from
any require() string that matches the Turbopack hash pattern.
Also updated deploy-vps workflow: npm registry rejects 299MB packages, so
deployment now uses npm pack + scp + npm install -g /tmp/omniroute-*.tgz.
PM2 entry point is app/server.js inside the npm global package.
