Peekaboo MCP

--- summary: 'OpenAI provider architecture and migration status in Peekaboo' read_when: - 'debugging OpenAI model integration or tool calling' - 'planning changes to the OpenAI provider layer' - 'explaining the Assistants→Chat Completions migration' --- # OpenAI in Peekaboo ## Status - Migration from Assistants API to Chat Completions is **complete** (as of 2025-11). Legacy Assistants code was removed; protocol-based message architecture is the sole implementation. - Streaming, tool calling, and session persistence are wired through the shared model interface used by other providers. ## Key migration outcomes 1. Protocol-based message and tool abstractions for strong typing. 2. Native streaming via `AsyncThrowingStream` with event handling. 3. Type-safe tool system with generic context. 4. Integrated with PeekabooCore services and session resume. 5. Removed feature flags and legacy Assistants artifacts. ## Implementation notes - Follows the same architecture used for Anthropic/Grok/Ollama: provider-conforming model types with shared error handling. - Tool results and errors are normalized so agents and CLI renderers stay provider-agnostic. - When adding new OpenAI models, update the provider registry and regression tests for model capabilities (vision, tools, JSON). ## References and snippets - Docs/examples for OpenAI live under `examples/docs/` (see `agentCloning.ts`, `chatLoop.ts`, `basicStreaming.ts`, etc.)—useful when cross-checking CLI/MCP behaviors. - If new event types appear, mirror the Anthropic streaming verification playbook: add golden tests for partial deltas and tool-call payloads.