In OpenAI Chat Completions streaming format, the tool call id and type
should only appear on the first chunk (tool declaration). Subsequent
argument delta chunks should only include index and function.arguments.
Including id on every delta chunk caused openai-to-claude.ts to emit
a new content_block_start for each chunk, breaking Claude Code ACP
sessions with malformed Claude-format streams.
Fixes#682
Co-authored-by: Chris Staley <christopher.staley@protonmail.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
When proxying Claude responses through OmniRoute, thinking blocks were being
emitted as regular content (delta.content) with <think>...</think> XML tags.
Clients like Claude Code, Cursor, and Windsurf look for delta.reasoning_content
to render the thinking panel — not <think> tags inside content.
Root cause (claude-to-openai.ts):
- content_block_start type:thinking → emitted { content: '<think>' }
- content_block_delta thinking_delta → emitted { content: delta.thinking }
- content_block_stop thinking block → emitted { content: '</think>' }
Fix:
- content_block_start → emits { reasoning_content: '' } (signals block start)
- thinking_delta → emits { reasoning_content: delta.thinking }
- content_block_stop → no extra chunk needed (thinking streamed via reasoning_content)
This fix applies when sourceFormat=CLAUDE targetFormat=OPENAI (Antigravity OAuth,
direct Claude API providers). The user reported 'Thinking Budget: passthrough'
was enabled but thinking was invisible — this is the root cause.
Fixes#289
Before this fix, geminiToClaudeResponse() called content_block_start +
content_block_stop on EVERY streaming chunk. Claude Code interprets each
content_block as a separate text element and renders them on separate
lines, making responses unreadable.
Fix: introduce state.openTextBlockIdx to track an open text block across
chunks. The block is opened on the first text chunk and only closed when:
- switching to a different block type (thinking, tool_use)
- the stream finishes (at finishReason)
All other block types (thinking, tool_use) are still opened+closed per
chunk as before since they are naturally atomic events.
Also: add 'Download opencode.json' button to the Agents dashboard.
The button appears when opencode is detected as installed and generates
a ready-to-use opencode.json config with the OmniRoute provider,
base URL and all models fetched from /v1/models.
- Issue #199: proxy_ prefix on tool names is now automatically disabled
when routing to non-Claude backends (Gemini, Compatible, etc.).
Prevents tool name mismatches in OpenCode, Cursor, and other clients.
- Issue #198: Added customRpm/customTpm support per provider connection.
Users can configure custom rate limits via connection settings,
overriding the default auto-learned limits from response headers.
- Add type annotations to all providerModels helper functions
- Introduce LegacyProvider interface in providerRegistry
- Refactor translator system to use self-registering module pattern
with bootstrapTranslatorRegistry() and per-file imports
- Simplify translator/index.ts by delegating to modular translators
- Remove hardcoded Gemini OAuth client secret for security
- Added tool_calls[].id and type:'function' to argument delta chunks
so OpenAI-compatible clients can associate argument fragments with
the correct tool call
- Changed finish_reason from hardcoded 'stop' to 'tool_calls' when
tool calls were emitted (in flush, response.completed, and final chunk)
- Fixes state desync in agentic clients (OpenCode, Claude Code, etc.)
where the assistant thinks tools already ran
OmniRoute is an intelligent API gateway that unifies 20+ AI providers behind a single
OpenAI-compatible endpoint. Features include intelligent routing with 6 strategies,
multi-format translation (OpenAI/Claude/Gemini/Responses API), circuit breakers,
semantic caching, combo fallback chains, real-time health monitoring, and a full
dashboard with provider management, analytics, and CLI tool integration.
Key highlights:
- 20+ providers (Claude Code, Codex, Gemini CLI, GitHub Copilot, iFlow, Qwen, Kiro, etc.)
- 6 routing strategies (Fill First, Round Robin, P2C, Random, Least Used, Cost Optimized)
- Export/Import database backup with full archive support
- Translator Playground with 4 modes (Playground, Chat Tester, Test Bench, Live Monitor)
- 100% TypeScript across src/ and open-sse/
- Docker support with multi-stage builds
- Comprehensive documentation and 9 dashboard screenshots
