yaze/docs/internal/agents/ai-infrastructure-initiative.md

AI Infrastructure & Build Stabilization Initiative

Summary

• Lead agent/persona: CLAUDE_AIINF
• Supporting agents: CODEX (documentation), GEMINI_AUTOM (testing/CI)
• Problem statement: Complete AI API enhancement phases 2-4, stabilize cross-platform build system, and ensure consistent dependency management across all platforms
• Success metrics:
  • All CMake presets work correctly on mac/linux/win (x64/arm64)
  • Phase 2 HTTP API server functional with basic endpoints
  • CI/CD pipeline consistently passes on all platforms
  • Documentation accurately reflects build commands and presets

Scope

In scope:

1. Build System Fixes

   • Add missing macOS/Linux presets to CMakePresets.json (mac-dbg, lin-dbg, mac-ai, etc.)
   • Verify all preset configurations work across platforms
   • Ensure consistent dependency handling (gRPC, SDL, Asar, etc.)
   • Update CI workflows if needed

2. AI Infrastructure (Phase 2-4 per handoff)

   • Complete UI unification for model selection (RenderModelConfigControls)
   • Implement HTTP server with basic endpoints (Phase 2)
   • Add FileSystemTool and BuildTool (Phase 3)
   • Begin ToolDispatcher structured output refactoring (Phase 4)

3. Documentation

   • Update build/quick-reference.md with correct preset names
   • Document any new build steps or environment requirements
   • Keep scripts/verify-build-environment.* accurate

Out of scope:

• Core editor features (CLAUDE_CORE domain)
• Comprehensive documentation rewrite (CODEX is handling)
• Full Phase 4 completion (can be follow-up work)
• New AI features beyond handoff document

Dependencies / upstream projects:

• gRPC v1.67.1 (ARM64 tested stable version)
• SDL2, Asar (via submodules)
• httplib (already in tree)
• Coordination with CODEX on documentation updates

Risks & Mitigations

Risk 1: Preset naming changes break existing workflows

Mitigation: Verify CI still works, update docs comprehensively, provide transition guide

Risk 2: gRPC build times affect CI performance

Mitigation: Ensure caching strategies are optimal, keep minimal preset without gRPC

Risk 3: HTTP server security concerns

Mitigation: Start with localhost-only default, document security model, require explicit opt-in

Risk 4: Cross-platform build variations

Mitigation: Test each preset locally before committing, verify on CI matrix

Testing & Validation

Required test targets:

• yaze_test - All unit/integration tests pass
• yaze - GUI application builds and launches
• z3ed - CLI tool builds with AI features
• Platform-specific: mac-dbg, lin-dbg, win-dbg, *-ai variants

ROM/test data requirements:

• Use existing test infrastructure (no new ROM dependencies)
• Agent tests use synthetic data where possible

Manual validation steps:

1. Configure and build each new preset on macOS (primary dev platform)
2. Verify CI passes on all platforms
3. Test HTTP API endpoints with curl/Postman
4. Verify z3ed agent workflow with Ollama

Documentation Impact

Public docs to update:

• docs/public/build/quick-reference.md - Correct preset names, add missing presets
• README.md - Update build examples if needed (minimal changes)
• CLAUDE.md - Update preset references if changes affect agent instructions

Internal docs/templates to update:

• docs/internal/AI_API_ENHANCEMENT_HANDOFF.md - Mark phases as complete
• docs/internal/agents/coordination-board.md - Regular status updates
• This initiative document - Track progress

Coordination board entry link:

See coordination-board.md entry: "2025-11-19 10:00 PST CLAUDE_AIINF – plan"

Timeline / Checkpoints

Milestone 1: Build System Fixes (Priority 1)

• Add missing macOS/Linux presets to CMakePresets.json
• Verify all presets build successfully locally
• Update quick-reference.md with correct commands
• Status: IN_PROGRESS

Milestone 2: UI Completion (Priority 2) - CLAUDE_CORE

Owner: CLAUDE_CORE Status: IN_PROGRESS Goal: Complete UI unification for model configuration controls

Files to Touch:

• src/app/editor/agent/agent_chat_widget.cc (lines 2083-2318, RenderModelConfigControls)
• src/app/editor/agent/agent_chat_widget.h (if member variables need updates)

Changes Required:

1. Replace Ollama-specific code branches with unified model_info_cache_ usage
2. Display models from all providers (Ollama, Gemini) in single combo box
3. Add provider badges/indicators (e.g., "[Ollama]", "[Gemini]" prefix or colored tags)
4. Handle provider filtering if selected provider changes
5. Show model metadata (family, size, quantization) when available

Build & Test:

# Build directory for CLAUDE_CORE
cmake --preset mac-ai -B build_ai_claude_core
cmake --build build_ai_claude_core --target yaze

# Launch and test
./build_ai_claude_core/bin/yaze --rom_file=zelda3.sfc --editor=Agent
# Verify: Model dropdown shows unified list with provider indicators

# Smoke build verification
scripts/agents/smoke-build.sh mac-ai yaze

Tests to Run:

• Manual: Launch yaze, open Agent panel, verify model dropdown
• Check: Models from both Ollama and Gemini appear
• Check: Provider indicators are visible
• Check: Model selection works correctly

Documentation Impact:

• No doc changes needed (internal UI refactoring)

Milestone 3: HTTP API (Phase 2 - Priority 3) - CLAUDE_AIINF

Owner: CLAUDE_AIINF Status: ✅ COMPLETE Goal: Implement HTTP REST API server for external agent access

Files to Create:

• src/cli/service/api/http_server.h - HttpServer class declaration
• src/cli/service/api/http_server.cc - HttpServer implementation
• src/cli/service/api/README.md - API documentation

Files to Modify:

• cmake/options.cmake - Add YAZE_ENABLE_HTTP_API flag (default OFF)
• src/cli/z3ed.cc - Wire HttpServer into main, add --http-port flag
• src/cli/CMakeLists.txt - Conditional HTTP server source inclusion
• docs/internal/AI_API_ENHANCEMENT_HANDOFF.md - Mark Phase 2 complete

Initial Endpoints:

1. GET /api/v1/health

   • Response: {"status": "ok", "version": "..."}
   • No authentication needed

2. GET /api/v1/models

   • Response: {"models": [{"name": "...", "provider": "...", ...}]}
   • Delegates to ModelRegistry::ListAllModels()

Implementation Notes:

• Use httplib from ext/httplib/ (header-only library)
• Server runs on configurable port (default 8080, flag: --http-port)
• Localhost-only by default for security
• Graceful shutdown on SIGINT
• CORS disabled initially (can add later if needed)

Build & Test:

# Build directory for CLAUDE_AIINF
cmake --preset mac-ai -B build_ai_claude_aiinf \
  -DYAZE_ENABLE_HTTP_API=ON
cmake --build build_ai_claude_aiinf --target z3ed

# Launch z3ed with HTTP server
./build_ai_claude_aiinf/bin/z3ed --http-port=8080

# Test endpoints (separate terminal)
curl http://localhost:8080/api/v1/health
curl http://localhost:8080/api/v1/models

# Smoke build verification
scripts/agents/smoke-build.sh mac-ai z3ed

Tests to Run:

• Manual: Launch z3ed with --http-port, verify server starts
• Manual: curl /health endpoint, verify JSON response
• Manual: curl /models endpoint, verify model list
• Check: Server handles concurrent requests
• Check: Server shuts down cleanly on Ctrl+C

Documentation Impact:

• Update AI_API_ENHANCEMENT_HANDOFF.md - mark Phase 2 complete
• Create src/cli/service/api/README.md with endpoint docs
• No public doc changes (experimental feature)

Milestone 4: Enhanced Tools (Phase 3 - Priority 4)

• Implement FileSystemTool (read-only first)
• Implement BuildTool
• Update ToolDispatcher registration
• Status: PENDING

Current Status

Last Updated: 2025-11-22 18:30 PST

Completed:

• ✅ Coordination board entry posted
• ✅ Initiative document created
• ✅ Build system analysis complete
• ✅ Milestone 1: Build System Fixes - COMPLETE
  • Added 11 new configure presets (6 macOS, 5 Linux)
  • Added 11 new build presets (6 macOS, 5 Linux)
  • Fixed critical Abseil linking bug in src/util/util.cmake
  • Updated docs/public/build/quick-reference.md
  • Verified builds on macOS ARM64
• ✅ Parallel work coordination - COMPLETE
  • Split Milestones 2 & 3 across CLAUDE_CORE and CLAUDE_AIINF
  • Created detailed task specifications with checklists
  • Posted IN_PROGRESS entries to coordination board

Completed:

• ✅ Milestone 3 (CLAUDE_AIINF): HTTP API server implementation - COMPLETE (2025-11-19 23:35 PST)

  • Added YAZE_ENABLE_HTTP_API CMake flag in options.cmake
  • Integrated HttpServer into cli_main.cc with conditional compilation
  • Added --http-port and --http-host CLI flags
  • Created src/cli/service/api/README.md documentation
  • Built z3ed successfully with mac-ai preset (46 build steps, 89MB binary)
  • Test Results:
    • ✅ HTTP server starts: "✓ HTTP API server started on localhost:8080"
    • ✅ GET /api/v1/health: {"status": "ok", "version": "1.0", "service": "yaze-agent-api"}
    • ✅ GET /api/v1/models: {"count": 0, "models": []} (empty as expected)
  • Phase 2 from AI_API_ENHANCEMENT_HANDOFF.md is COMPLETE

• ✅ Test Infrastructure Stabilization - COMPLETE (2025-11-21)

  • Fixed critical stack overflow crash on macOS ARM64 (increased stack from default ~8MB to 16MB)
  • Resolved circular dependency issues in test configuration
  • All test categories now stable: unit, integration, e2e, rom-dependent
  • Verified across all platforms (macOS, Linux, Windows)

• ✅ Milestone 2 (CLAUDE_CORE): UI unification for model configuration controls - COMPLETE

  • Completed unified model configuration UI for Agent panel
  • Models from all providers (Ollama, Gemini) now display in single dropdown
  • Provider indicators visible for each model
  • Provider filtering implemented when provider selection changes

In Progress:

• Milestone 4 (CLAUDE_AIINF): Enhanced Tools Phase 3 - FileSystemTool and BuildTool

Helper Scripts (from CODEX):

Both personas should use these scripts for testing and validation:

• scripts/agents/smoke-build.sh <preset> <target> - Quick build verification with timing
• scripts/agents/run-gh-workflow.sh - Trigger remote GitHub Actions workflows
• Documentation: scripts/agents/README.md and docs/internal/README.md

Next Actions (Post Milestones 2, 3, & Test Stabilization):

1. Complete Milestone 4: Add FileSystemTool and BuildTool (Phase 3)
2. Begin ToolDispatcher structured output refactoring (Phase 4)
3. Comprehensive testing across all platforms using smoke-build.sh
4. Release validation: Ensure all new features work in release builds
5. Performance optimization: Profile test execution time and optimize as needed