RichardAtCT · gustavokch · Mar 16, 2026 · Mar 16, 2026 · Mar 16, 2026 · Mar 16, 2026
diff --git a/.env.example b/.env.example
@@ -1,6 +1,11 @@
 # Claude CLI Configuration
 CLAUDE_CLI_PATH=claude
 
+# Gemini CLI Configuration
+# GEMINI_API_KEY=your-gemini-api-key-here
+# GOOGLE_API_KEY=your-google-api-key-here
+GEMINI_CLI_PATH=gemini
+
 # Authentication Method (optional - explicit selection)
 # Set this to override auto-detection. Values: cli, api_key, bedrock, vertex
 # If not set, auto-detects based on available env vars (ANTHROPIC_API_KEY, etc.)
@@ -13,6 +18,8 @@ CLAUDE_CLI_PATH=claude
 
 # Server Configuration
 PORT=8000
+# Maximum number of concurrent CLI processes allowed (default: 3)
+# MAX_CONCURRENT_PROCESSES=3
 # Host binding address - use 127.0.0.1 for local-only access, 0.0.0.0 for all interfaces
 # CLAUDE_WRAPPER_HOST=0.0.0.0
 # Maximum request body size in bytes (default: 10MB)
@@ -21,6 +28,10 @@ PORT=8000
 # Timeout Configuration (milliseconds)
 MAX_TIMEOUT=600000
 
+# Prewarming Configuration
+# Prompt to use during startup for prewarming the CLI backends (default: Hello)
+# PREWARM_PROMPT=Hello
+
 # CORS Configuration
 CORS_ORIGINS=["*"]
 

diff --git a/.gitignore b/.gitignore
@@ -40,6 +40,7 @@ logs/
 # Testing
 .coverage
 .pytest_cache/
+.hypothesis/
 htmlcov/
 
 # Claude Code
@@ -57,4 +58,5 @@ test_debug_*.py
 test_performance_*.py
 test_user_*.py
 test_new_*.py
-test_roocode_compatibility.py
+test_roocode_compatibility.py
+.worktrees/
diff --git a/PR.md b/PR.md
@@ -0,0 +1,24 @@
+# Release v2.3.0: Concurrency improvements, SDK options wiring, and critical bug fixes
+
+This PR introduces version 2.3.0, focusing on significant reliability improvements, full support for concurrent SDK calls, wiring of new Claude API options, and resolutions for several critical proxy bugs.
+
+## Features & Enhancements
+* **SDK Options Wiring:** Full support for `reasoning_effort`, `response_format`, `thinking`, `max_budget_usd`, and `user` fields passed directly to the Claude SDK.
+* **Concurrency:** Removed `os.environ` mutex (`_env_lock`) by passing auth via `options.env`, allowing fully concurrent SDK calls. `SessionManager` has been refactored to use `asyncio.Lock` with all session methods converted to async.
+* **Token & Reason Mapping:** Extracts real token counts directly from the SDK's `ResultMessage` and properly maps `stop_reason` to `finish_reason` (e.g., `max_tokens` → `length`).
+* **Tool Handling:** Changed `AnthropicMessagesRequest.enable_tools` default to `False` so simple message requests do not trigger unintended 10-turn loops. 
+
+## Bug Fixes
+* **Session Continuity:** Fixed session continuation by correcting `continue_session` to `continue_conversation` and replaced list appending with replacement to prevent exponential duplication.
+* **Timeouts & Hangs:** Wrapped async `query()` iterations with `asyncio.timeout` to prevent indefinite hangs when the SDK subprocess stalls.
+* **Proxy Reliability:** 
+  * Removed `filter_content()` from user input which was silently stripping XML-like tags.
+  * Secured `/v1/auth/status` endpoint with the `verify_api_key()` auth guard.
+  * Marked the Bash tool as `is_safe=False`.
+  * Replaced bare `except:` clauses with `except Exception:`.
+
+## Maintenance & Chores
+* Updated `poetry.lock` and the test suite for compatibility with `pydantic 2.13` and `poetry 2.3`.
+* Replaced deprecated `datetime.utcnow()` with `datetime.now(timezone.utc)`.
+* Ignored `.worktrees` directories in `.gitignore`.
+* Added diagnostic print statements for `/v1/messages` and improved the `test_message.py` script.
diff --git a/PR_GEMINI.md b/PR_GEMINI.md
@@ -0,0 +1,27 @@
+# Gemini CLI Proxy Support and Interactive Chat Client
+
+This PR introduces support for the Gemini CLI as an alternative backend, allowing users to use Gemini models (like Gemini 3 and 2.5) through the OpenAI-compatible proxy. It also includes a new interactive chat client with Markdown rendering.
+
+## New Features
+* **Gemini CLI Proxy:**
+  * New `GeminiCodeCLI` wrapper for the `@google/gemini-cli` tool.
+  * Real-time NDJSON stream parsing for low-latency responses.
+  * Full session continuity support using the CLI's `--resume` flag.
+  * Integrated model routing: models starting with `gemini-` or using aliases like `pro`, `flash`, `auto` are automatically routed to Gemini.
+* **Interactive Chat Client:**
+  * Added `examples/interactive_chat.py` which manages the background server, provides a rich TUI with `rich` for Markdown rendering, and supports live streaming.
+* **Unified Model Listing:**
+  * Updated `/v1/models` to return both Claude and Gemini models with correct metadata.
+
+## Enhancements
+* **Authentication:** Added support for `GEMINI_API_KEY` and `GOOGLE_API_KEY` in the `ClaudeCodeAuthManager`.
+* **Constants:** Defined the latest Gemini model IDs and aliases.
+* **Configuration:** Updated `.env.example` with Gemini-specific settings.
+
+## Bug Fixes & Refactoring
+* **Unified Interface:** Refactored `main.py` endpoints to use a common `get_cli_for_model` helper, making it easier to add more backends in the future.
+* **Metadata Extraction:** Improved metadata and usage parsing to handle both Anthropic and Gemini formats consistently.
+
+## Testing
+* Added `tests/test_gemini_cli_unit.py` with 100% coverage for the new wrapper.
+* Verified both streaming and non-streaming responses for both backends.
diff --git a/README.md b/README.md
@@ -1,6 +1,6 @@
 # Claude Code OpenAI API Wrapper
 
-An OpenAI API-compatible wrapper for Claude Code, allowing you to use Claude Code with any OpenAI client library. **Now powered by the official Claude Agent SDK v0.1.18** with enhanced authentication and features.
+An OpenAI API-compatible wrapper for Claude Code, allowing you to use Claude Code with any OpenAI client library. **Now powered by the official Claude Agent SDK v0.2.134+** with enhanced authentication and features.
 
 ## Version
 
@@ -9,11 +9,16 @@ An OpenAI API-compatible wrapper for Claude Code, allowing you to use Claude Cod
 - **Dynamic default Sonnet:** `DEFAULT_MODEL` resolves to the latest Sonnet at startup when `ANTHROPIC_API_KEY` is configured; falls back to `claude-sonnet-4-6` otherwise
 - **Operator overrides:** New `CLAUDE_MODELS_OVERRIDE`, `FAST_MODEL`, and `MODEL_LIST_*` env vars
 - **Updated catalog:** Claude 4.6 family added to the static fallback list
+- **Bug fixes:** `continue_conversation` SDK field corrected; `max_thinking_tokens` now wired through; real token counts from SDK; `finish_reason` mapped from actual `stop_reason`
+- **Concurrent requests:** Auth env vars passed via `options.env` — no more serialising lock
+- **New parameters:** `reasoning_effort`, `response_format`, `max_budget_usd`, `thinking` added to request models
+- **Async session manager:** `threading.Lock` replaced with `asyncio.Lock` for proper async safety
+- **SDK options refactor:** `run_completion` simplified to accept a `claude_options` dict, enabling generic passthrough of any SDK field
 
-**Upgrading from v1.x?**
+**Upgrading from v2.2.0:**
 1. Pull latest code: `git pull origin main`
 2. Update dependencies: `poetry install`
-3. Restart server - that's it!
+3. Restart server — no breaking changes to the OpenAI/Anthropic API surface
 
 **Migration Resources:**
 - [MIGRATION_STATUS.md](./MIGRATION_STATUS.md) - Detailed v2.0.0 migration status
@@ -22,7 +27,7 @@ An OpenAI API-compatible wrapper for Claude Code, allowing you to use Claude Cod
 ## Status
 
 🎉 **Production Ready** - All core features working and tested:
-- ✅ Chat completions endpoint with **official Claude Agent SDK v0.1.18**
+- ✅ Chat completions endpoint with **official Claude Agent SDK v0.2.134+**
 - ✅ **Anthropic Messages API** (`/v1/messages`) for native compatibility
 - ✅ Streaming and non-streaming responses
 - ✅ Full OpenAI SDK compatibility
@@ -32,7 +37,9 @@ An OpenAI API-compatible wrapper for Claude Code, allowing you to use Claude Cod
 - ✅ Model selection support with validation
 - ✅ **Fast by default** - Tools disabled for OpenAI compatibility (5-10x faster)
 - ✅ Optional tool usage (Read, Write, Bash, etc.) when explicitly enabled
-- ✅ **Real-time cost and token tracking** from SDK
+- ✅ **Real token counts** from SDK metadata (no more estimates)
+- ✅ **Accurate `finish_reason`** mapped from SDK `stop_reason`
+- ✅ **Fully concurrent requests** — no serialising lock for auth env vars
 - ✅ **Session continuity** with conversation history across requests
 - ✅ **Session management endpoints** for full session control
 - ✅ Health, auth status, and models endpoints
@@ -48,11 +55,13 @@ An OpenAI API-compatible wrapper for Claude Code, allowing you to use Claude Cod
 - Automatic model validation and selection
 
 ### 🛠 **Claude Agent SDK Integration**
-- **Official Claude Agent SDK** integration (v0.1.18) 🆕
+- **Official Claude Agent SDK** integration (v0.2.134+) 🆕
 - **Real-time cost tracking** - actual costs from SDK metadata
-- **Accurate token counting** - input/output tokens from SDK
+- **Real token counting** - input/output tokens directly from SDK (no estimation)
+- **Accurate finish_reason** - mapped from SDK `stop_reason` (`end_turn` → `stop`, `max_tokens` → `length`)
 - **Session management** - proper session IDs and continuity
 - **Enhanced error handling** with detailed authentication diagnostics
+- **Fully concurrent** - auth env vars passed via SDK options, no serialising mutex
 - **Modern SDK features** - Latest capabilities and improvements
 
 ### 🔐 **Multi-Provider Authentication**
@@ -66,6 +75,10 @@ An OpenAI API-compatible wrapper for Claude Code, allowing you to use Claude Cod
 - **System prompt support** via SDK options
 - **Optional tool usage** - Enable Claude Code tools (Read, Write, Bash, etc.) when needed
 - **Fast default mode** - Tools disabled by default for OpenAI API compatibility
+- **`reasoning_effort`** - Map OpenAI `reasoning_effort: "low"|"medium"|"high"` to SDK `effort`
+- **`response_format`** - Pass `{"type": "json_object"}` or JSON Schema for structured outputs
+- **`thinking`** - Explicit thinking config `{"type": "enabled", "budget_tokens": N}` (overrides `max_tokens` mapping)
+- **`max_budget_usd`** - Per-request cost cap in USD
 - **Development mode** with auto-reload (`uvicorn --reload`)
 - **Interactive API key protection** - Optional security with auto-generated tokens
 - **Comprehensive logging** and debugging capabilities
@@ -83,7 +96,7 @@ Get started in under 2 minutes:
 
 ```bash
 # 1. Clone and setup the wrapper
-git clone https://github.com/RichardAtCT/claude-code-openai-wrapper
+git clone https://github.com/gustavokch/claude-code-openai-wrapper
 cd claude-code-openai-wrapper
 poetry install  # Installs SDK with bundled Claude Code CLI
 
@@ -121,13 +134,13 @@ poetry run python test_endpoints.py
      ```
    - **Option C**: Use AWS Bedrock or Google Vertex AI (see Configuration section)
 
-> **Note:** The Claude Code CLI is bundled with the SDK (v0.1.18+). No separate Node.js or npm installation required!
+> **Note:** The Claude Code CLI is bundled with the SDK (v0.2.134+). No separate Node.js or npm installation required!
 
 ## Installation
 
 1. Clone the repository:
    ```bash
-   git clone https://github.com/RichardAtCT/claude-code-openai-wrapper
+   git clone https://github.com/gustavokch/claude-code-openai-wrapper
    cd claude-code-openai-wrapper
    ```
 
@@ -473,6 +486,71 @@ for chunk in stream:
         print(chunk.choices[0].delta.content, end="")
 ```
 
+## Advanced Parameters
+
+These extra fields extend the standard OpenAI request body and are passed through to the Claude Agent SDK.
+
+### `reasoning_effort`
+
+Controls the depth of Claude's thinking. Maps to the SDK `effort` field.
+
+```python
+response = client.chat.completions.create(
+    model="claude-sonnet-4-5-20250929",
+    messages=[{"role": "user", "content": "Solve this math problem..."}],
+    extra_body={"reasoning_effort": "high"}  # "low" | "medium" | "high"
+)
+```
+
+### `response_format`
+
+Request structured output. Passed through as SDK `output_format`.
+
+```python
+response = client.chat.completions.create(
+    model="claude-sonnet-4-5-20250929",
+    messages=[{"role": "user", "content": "Return JSON with name and age fields."}],
+    extra_body={"response_format": {"type": "json_object"}}
+)
+```
+
+### `thinking`
+
+Explicit thinking configuration — takes precedence over the `max_tokens → max_thinking_tokens` mapping.
+
+```python
+response = client.chat.completions.create(
+    model="claude-sonnet-4-6",
+    messages=[{"role": "user", "content": "Hard reasoning task"}],
+    extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
+)
+# Also: {"type": "adaptive"} or {"type": "disabled"}
+```
+
+### `max_budget_usd`
+
+Cap per-request cost in USD. The SDK will stop generation when the budget is reached.
+
+```python
+response = client.chat.completions.create(
+    model="claude-sonnet-4-5-20250929",
+    messages=[{"role": "user", "content": "Long task..."}],
+    extra_body={"max_budget_usd": 0.05}  # stop at $0.05
+)
+```
+
+### `max_tokens` / `max_completion_tokens`
+
+Maps to the SDK's `max_thinking_tokens` (best-effort). For precise control use `thinking` above.
+
+```python
+response = client.chat.completions.create(
+    model="claude-sonnet-4-5-20250929",
+    messages=[{"role": "user", "content": "Brief answer please"}],
+    max_tokens=512
+)
+```
+
 ## Supported Models
 
 The wrapper exposes Claude's full model catalog. When `ANTHROPIC_API_KEY` is set, `/v1/models` returns Anthropic's live list (cached for 1 hour) and the wrapper picks the latest Sonnet as `DEFAULT_MODEL` at startup. When the key is absent — for example, when running with Bedrock, Vertex, or Claude CLI subscription auth — the static list below is served and `claude-sonnet-4-6` is used as the fallback default. Operators who want a curated list regardless of auth can set `CLAUDE_MODELS_OVERRIDE`.
@@ -493,6 +571,30 @@ The wrapper exposes Claude's full model catalog. When `ANTHROPIC_API_KEY` is set
 
 **Note:** Claude 3.x models are not supported by the Claude Agent SDK. The model parameter is passed to Claude Code via the SDK's model selection.
 
+## Using non-Claude models via passthrough (e.g. GLM-5.2)
+
+The wrapper can serve any model that your Claude Code installation can reach,
+including non-Anthropic models such as **GLM-5.2**. The wrapper does not call
+the model provider directly — it forwards the model name to Claude Code, which
+must already be configured to reach the provider.
+
+**Prerequisite — point Claude Code at the provider.** Set these on the Claude
+Code process (your environment, not wrapper code):
+- `ANTHROPIC_BASE_URL` — your proxy that speaks the Anthropic API format and
+  forwards to the provider (e.g. a GLM endpoint).
+- `ANTHROPIC_AUTH_TOKEN` or `ANTHROPIC_API_KEY` — the credential your proxy
+  requires, if any.
+
+**Use it through the wrapper.** Send the model name in the request:
+```bash
+curl http://localhost:8000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{"model":"glm-5.2","messages":[{"role":"user","content":"hello"}]}'
+```
+
+To make GLM the default, set `DEFAULT_MODEL=glm-5.2` for the wrapper. `glm-5.2`
+is advertised in `GET /v1/models`.
+
 ## Session Continuity 🆕
 
 The wrapper now supports **session continuity**, allowing you to maintain conversation context across multiple requests. This is a powerful feature that goes beyond the standard OpenAI API.
@@ -607,26 +709,32 @@ See `examples/session_continuity.py` for comprehensive Python examples and `exam
 ### 🚫 **Current Limitations**
 - **Images in messages** are converted to text placeholders
 - **Function calling** not supported (tools work automatically based on prompts)
-- **OpenAI parameters** not yet mapped: `temperature`, `top_p`, `max_tokens`, `logit_bias`, `presence_penalty`, `frequency_penalty`
+- **OpenAI parameters** not mapped: `temperature`, `top_p`, `logit_bias`, `presence_penalty`, `frequency_penalty` (ignored with a warning)
 - **Multiple responses** (`n > 1`) not supported
 
-### 🛣 **Planned Enhancements** 
-- [ ] **Tool configuration** - allowed/disallowed tools endpoints  
-- [ ] **OpenAI parameter mapping** - temperature, top_p, max_tokens support
-- [ ] **Enhanced streaming** - better chunk handling
+### 🛣 **Planned Enhancements**
+- [ ] **Token-level streaming** - `include_partial_messages` for finer chunks
 - [ ] **MCP integration** - Model Context Protocol server support
-
-### ✅ **Recent Improvements (v2.2.0)**
-- **Interactive Landing Page**: API explorer with live endpoint testing
-- **Anthropic Messages API**: Native `/v1/messages` endpoint
-- **Explicit Auth Selection**: `CLAUDE_AUTH_METHOD` env var
-- **Tool Execution Fix**: `enable_tools: true` now works correctly
+- [ ] **Temperature/top_p** - native SDK mapping when available
+
+### ✅ **Recent Improvements (v2.3.0)**
+- **Bug fixes:** `continue_conversation` field corrected; `max_thinking_tokens` now wired through to SDK
+- **Real token counts**: response `usage` comes from SDK metadata, not character estimation
+- **Accurate `finish_reason`**: mapped from SDK `stop_reason` (`max_tokens` → `length`, etc.)
+- **Concurrent requests**: auth env vars via `options.env` — no serialising mutex
+- **New parameters**: `reasoning_effort`, `response_format`, `max_budget_usd`, `thinking`
+- **Async session manager**: `threading.Lock` → `asyncio.Lock` for proper async safety
+
+### ✅ **v2.2.0 Features**
+- Interactive Landing Page: API explorer with live endpoint testing
+- Anthropic Messages API: Native `/v1/messages` endpoint
+- Explicit Auth Selection: `CLAUDE_AUTH_METHOD` env var
+- Tool Execution Fix: `enable_tools: true` now works correctly
 
 ### ✅ **v2.0.0 - v2.1.0 Features**
-- Claude Agent SDK v0.1.18 with bundled CLI
+- Claude Agent SDK v0.2.134+ with bundled CLI
 - Multi-provider auth (CLI, API key, Bedrock, Vertex AI)
 - Session continuity and management
-- Real-time cost and token tracking
 - System prompt support
 
 ## Troubleshooting
@@ -679,13 +787,16 @@ curl http://localhost:8000/v1/auth/status | python -m json.tool
 ### ⚙️ **Development Tools**
 ```bash
 # Install development dependencies
-poetry install --with dev
+poetry install
 
 # Format code
 poetry run black .
 
-# Run full tests (when implemented)
-poetry run pytest tests/
+# Run unit tests (no server required)
+PYTHONPATH=$(pwd) poetry run pytest tests/test_claude_cli_unit.py tests/test_session_manager_unit.py tests/test_models_unit.py -v
+
+# Run full test suite (unit + integration, server must be running for integration tests)
+PYTHONPATH=$(pwd) poetry run pytest tests/ -v
 ```
 
 ### ✅ **Expected Results**