fix(linux): add missing yaze_gfx_render dependency to yaze_gfx_debug

Fixes linker error on Linux where yaze_gfx_debug.a (performance_dashboard.cc)
was calling AtlasRenderer::Get() and AtlasRenderer::GetStats() but wasn't
linking against yaze_gfx_render which contains atlas_renderer.cc.

Root cause: yaze_gfx_debug was only linking to yaze_gfx_types and
yaze_gfx_resource, missing the yaze_gfx_render dependency.

This also fixes the undefined reference errors for HttpServer methods
which were already properly included in the agent.cmake source list.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

docs/internal/handoff/ai-api-phase2-handoff.md

@@ -0,0 +1,85 @@
+# AI API & Agentic Workflow Enhancement - Phase 2 Handoff
+
+**Date**: 2025-11-20
+**Status**: Phase 2 Implementation Complete
+**Previous Plan**: `docs/internal/AI_API_ENHANCEMENT_HANDOFF.md`
+
+## Overview
+This handoff covers the completion of Phase 2, which focused on unifying the UI for model selection and implementing the initial HTTP API server foundation. The codebase is now ready for building and verifying the API endpoints.
+
+## Completed Work
+
+### 1. UI Unification (`src/app/editor/agent/agent_chat_widget.cc`)
+- **Unified Model List**: Replaced the separate Ollama/Gemini list logic with a single, unified list derived from `ModelRegistry`.
+- **Provider Badges**: Models in the list now display their provider (e.g., `[ollama]`, `[gemini]`).
+- **Contextual Configuration**:
+  - If an **Ollama** model is selected, the "Ollama Host" input is displayed.
+  - If a **Gemini** model is selected, the "API Key" input is displayed.
+- **Favorites & Presets**: Updated to work with the unified `ModelInfo` structure.
+
+### 2. HTTP Server Implementation (`src/cli/service/api/`)
+- **`HttpServer` Class**:
+  - Wraps `httplib::Server` running in a background `std::thread`.
+  - Exposed via `Start(port)` and `Stop()` methods.
+  - Graceful shutdown handling.
+- **API Handlers**:
+  - `GET /api/v1/health`: Returns server status (JSON).
+  - `GET /api/v1/models`: Returns list of available models from `ModelRegistry`.
+- **Integration**:
+  - Updated `src/cli/agent.cmake` to include `http_server.cc`, `api_handlers.cc`, and `model_registry.cc`.
+  - Updated `src/app/main.cc` to accept `--enable_api` and `--api_port` flags.
+
+## Build & Test Instructions
+
+### 1. Building
+The project uses CMake. The new files are automatically included in the `yaze_agent` library via `src/cli/agent.cmake`.
+
+```bash
+# Generate build files (if not already done)
+cmake -B build -G Ninja
+
+# Build the main application
+cmake --build build --target yaze_app
+```
+
+### 2. Testing the UI
+1. Launch the editor:
+   ```bash
+   ./build/yaze_app --editor=Agent
+   ```
+2. Verify the **Model Configuration** panel:
+   - You should see a single list of models.
+   - Try searching for a model.
+   - Select an Ollama model -> Verify "Host" input appears.
+   - Select a Gemini model -> Verify "API Key" input appears.
+
+### 3. Testing the API
+1. Launch the editor with API enabled:
+   ```bash
+   ./build/yaze_app --enable_api --api_port=8080
+   ```
+   *(Check logs for "Starting API server on port 8080")*
+
+2. Test Health Endpoint:
+   ```bash
+   curl -v http://localhost:8080/api/v1/health
+   # Expected: {"status":"ok", "version":"1.0", ...}
+   ```
+
+3. Test Models Endpoint:
+   ```bash
+   curl -v http://localhost:8080/api/v1/models
+   # Expected: {"models": [{"name": "...", "provider": "..."}], "count": ...}
+   ```
+
+## Next Steps (Phase 3 & 4)
+
+### Phase 3: Tool Expansion
+- **FileSystemTool**: Implement safe file read/write operations (`src/cli/handlers/tools/filesystem_commands.h`).
+- **BuildTool**: Implement cmake/ninja triggers.
+- **Editor Integration**: Inject editor state (open files, errors) into the agent context.
+
+### Phase 4: Structured Output
+- Refactor `ToolDispatcher` to return JSON objects instead of capturing stdout strings.
+- Update API to expose a `/api/v1/chat` endpoint that returns these structured responses.
+