The openai-to-claude translator was prefixing tool names with 'proxy_'
(e.g. Bash → proxy_Bash) even when routing Claude-format requests to
native Claude/Anthropic providers. Claude rejects unknown tool names,
causing 'No such tool available: proxy_Bash' errors.
Root cause: the _disableToolPrefix condition only disabled the prefix
for non-Claude providers, but it should be disabled for ALL providers
in the Claude passthrough path since tools are already in Claude format.
Fixes#618
- Fix Ollama Cloud base URL from api.ollama.com to ollama.com/v1/chat/completions
- Fix Ollama Cloud models URL to ollama.com/api/tags
- Add gemini-3.1-pro-preview and gemini-3.1-flash-lite-preview to Antigravity provider
Closes#643, closes#645
Thanks @brendandebeasi for this excellent contribution! 🎉 The bounded retry with exponential backoff is exactly the right approach for expired connections. Merged and will be included in v3.1.1.
Connections marked as 'expired' were permanently skipped by the health check scheduler (line 176: if testStatus === "expired" return). A single transient refresh failure could permanently disable auto-refresh, requiring manual re-authentication.
Replace the hard skip with a bounded retry mechanism: up to 3 attempts with exponential backoff (5min, 10min, 20min). On success, the connection is fully restored to active. On exhaustion, it remains expired (same as before). The existing circuit breaker (5 failures → 30min pause) provides additional protection.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
OpenAI-compatible clients (OpenCode, etc.) check capabilities/input_modalities fields on the /v1/models response to determine if a model supports image input. Omniroute was not emitting these fields, causing clients to assume text-only for all models routed through the proxy.
Add keyword-based vision detection (matching the existing playground heuristic) that annotates model entries with capabilities:{vision:true}, input_modalities:["text","image"], and output_modalities:["text"] for known multimodal models (GPT-4o/4-turbo, Claude 3+, Gemini, Pixtral, Qwen-VL, etc.).
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
- Rename in.json → hi.json: 'in' is Indonesian (ISO 639-1), Hindi is 'hi'.
Fixes Weblate locale conflict where id.json and in.json both claimed Indonesian.
- Move empty tool name filter before Codex passthrough: nativeCodexPassthrough
skipped all input sanitization, causing 400 'empty tool name' from upstream.
- Collapse 3+ consecutive newlines to \n\n in response sanitizer: thinking
models accumulate excessive line breaks between tool call blocks.
- OpenAI-to-Claude translator now maps reasoning_effort (low/medium/high/max)
to Claude's thinking.budget_tokens. Fixes clients like OpenCode sending
reasoning_effort via @ai-sdk/openai-compatible losing thinking configuration.
- Ensures max_tokens > budget_tokens for all thinking configs.
- Token health check now proactively refreshes tokens within 5 min of expiry,
regardless of the configured health check interval — addresses Qwen OAuth
token refresh failures between scheduled checks.
Adds structured YAML-based issue templates to improve issue quality.
Bug reports require version, install method, OS, repro steps, and
expected/actual behavior. Feature requests require use case and
proposed solution. Blank issues are still allowed for edge cases.
Adds a button next to the Auto-Sync toggle to clear all custom models
for a provider. Extends DELETE /api/provider-models to support ?all=true
parameter for bulk deletion.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Handle reasoning_details[] array (StepFun/OpenRouter format) in sanitizer and translator
- Handle 'reasoning' field alias → reasoning_content in streaming and non-streaming paths
- Cross-map input_tokens/output_tokens ↔ prompt_tokens/completion_tokens in filterUsageForFormat
- Fix extractUsage to accept input_tokens/output_tokens as alternative field names
- All 936 tests pass
- Add Claude Sonnet 4.5, Claude Sonnet 4, GPT 5, GPT 5 Mini
- Enable passthroughModels: true so users can access any model
Antigravity supports without waiting for registry updates
- Set clientSecretDefault for Antigravity provider (was empty, causing
'client_secret is missing' on token refresh for npm users)
- Add modelsUrl to opencode-zen registry for 'Import from /models'
Rewrote the account selector with a simpler, reliable approach:
- Fetch ALL connections once at startup (not per-provider)
- Filter by selectedProvider using ALIAS_TO_ID mapping
- Account/Key dropdown always visible when provider selected
- Shows 'Auto (N accounts)' default or individual account names
- Works for both OAuth accounts and API key providers
Import ALIAS_TO_ID mapping and resolve provider aliases (cx→codex,
kr→kiro, etc.) in loadConnections before filtering connections from
the API. The /v1/models endpoint returns alias-prefixed model IDs
but /api/providers/client returns provider IDs.
Playground:
- loadConnections() was parsing wrong API response shape (expected
providers[].connections[] but API returns flat connections[])
- Account selector now shows for any provider with ≥1 connection
- Uses conn.email as name fallback for OAuth providers
CLI Tools:
- getAllAvailableModels() now also fetches from /v1/models API
- Dynamic models supplement static PROVIDER_MODELS definitions
- Fixes providers like Kiro, OpenCode Zen showing 0 models
After stripping <antThinking>/<thinking> tags from streaming responses, the
surrounding newlines were left as artifacts (e.g. \n\n\n\n). Now collapses 3+
consecutive newlines to double-newline after any tag removal.
Also fixes PR #625 merge (Provider Limits light mode background).
The proxy test button in Settings was always failing with 'Socks5 Authentication
failed' because the frontend sent redacted credentials (***) from listProxies().
The backend received '***' as the password and tried to authenticate with it.
Fix: Frontend now sends proxyId in the test request body. The test endpoint
looks up the proxy from the DB with includeSecrets: true and uses the real
stored credentials for the SOCKS5 handshake.
Also: removed username/password from the frontend test payload since they
are always redacted and useless for testing.
Root cause: SOCKS5 proxies accept TCP connections (pass health check) but
can't relay HTTPS traffic. getCodexUsage() catches fetch errors internally
and returns {message: 'Failed to fetch...'} instead of throwing, so the
previous catch-based fallback never triggered.
Fix: After the initial proxied fetch, check the returned usage object for
network error indicators. If a proxy was active and the result contains
'fetch failed' / 'ECONNREFUSED' / etc., retry the entire operation
(credential refresh + usage fetch) without proxy context.
This is safe because usage fetching is read-only — showing limits data
without proxy is better than showing nothing.
bg-bg-subtle (#f0f0f5) appears gray against the page background in
light mode. Changed to bg-surface (#ffffff) for consistency with other
Card-based UI sections.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix: Limits usage fetch wraps BOTH token refresh and usage call inside proxy context (fixes SOCKS5 Codex accounts)
- Fix: CI integration test v1/models gracefully handles empty models list
- Fix: Settings proxy test button results now render with priority over health data
- Feat: Playground account selector dropdown for testing specific connections
- Merge: PR #623 LongCat API base URL path correction
